Economics and Information Systems, Volume 1

HANDBOOKS IN INFORMATION SYSTEMS
VOLUME 1
Handbooks in
Information Systems
Advisory Editors Editor

Ba, Sulin Andrew B. Whinston
University of Connecticut
Duan, Wenjing
The George Washington University
Geng, Xianjun
University of Washington
Gupta, Alok Volume 1

University of Minnesota
Hendershott, Terry
University of California at Berkeley
Rao, H.R.
SUNY at Buffalo
Santanam, Raghu T.
Arizona State University
Zhang, Han
Georgia Institute of Technology
Amsterdam – Boston – Heidelberg – London – New York – Oxford – Paris – San Diego
San Francisco – Singapore – Sydney – Tokyo
Economics and Information
Systems
Edited by
T. Hendershott
University of California
Amsterdam – Boston – Heidelberg – London – New York – Oxford – Paris – San Diego
San Francisco – Singapore – Sydney – Tokyo
Elsevier
Radarweg 29, PO Box 211, 1000 AE Amsterdam, The Netherlands
The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, UK
First edition 2006
Copyright r 2006 Elsevier B.V. All rights reserved
No part of this publication may be reproduced, stored in a retrieval system

or transmitted in any form or by any means electronic, mechanical, photocopying,
recording or otherwise without the prior written permission of the publisher
Permissions may be sought directly from Elsevier’s Science & Technology Rights
Department in Oxford, UK: phone (+44) (0) 1865 843830; fax (+44) (0) 1865 853333;
email: permissions@elsevier.com. Alternatively, you can submit your request online by
visiting the Elsevier web site at http://elsevier.com/locate/permissions, and selecting
Obtaining permission to use Elsevier material
Notice
No responsibility is assumed by the publisher for any injury and/or damage to persons
or property as a matter of products liability, negligence or otherwise, or from any use
or operation of any methods, products, instructions or ideas contained in the material
herein. Because of rapid advances in the medical sciences, in particular, independent
verification of diagnoses and drug dosages should be made
Library of Congress Cataloging-in-Publication Data

A catalog record for this book is available from the Library of Congress
British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library
ISBN-13: 978-0-444-51771-5 (this volume)

ISBN-10: 0-444-51771-5 (this volume)
ISSN: 1574-0145 (series)
For information on all Elsevier publications

visit our website at books.elsevier.com
Printed and bound in The Netherlands
06 07 08 09 10 10 9 8 7 6 5 4 3 2 1
Contents
Introduction xv
CHAPTER 1
Diffusion of Information and Communication Technologies to
Businesses
C. Forman and A. Goldfarb 1
1. Introduction 1
2. ICT diffusion and its impact 4
2.1. Diffusion modeling 4
2.2. The impact of ICT diffusion 8
3. ICT adoption and organizational characteristics 9
3.1. Adoption, internal firm organization, and organizational change 9
3.2. Adoption and firm boundaries 14
3.3. Adoption and size 17
3.4. Technical infrastructure 18
3.5. Adoption, assimilation, intra-firm diffusion, and usage 19
4. Geographic differences in adoption 21
4.1. Adoption of ICT across urban and rural areas 23
4.2. Evidence of how ICT use influences location patterns 24
4.3. Future research 24
5. Trade-offs between organization and environment 25
5.1. Co-invention 26
5.2. Strategic issues in technology adoption 28
6. Network effects 30
6.1. Theoretical literature on direct and indirect network externalities 32
6.2. Evidence of network effects of any kind in ICT 33
6.3. Evidence of positive network externalities in ICT 35
7. Internet diffusion across countries 37
8. Conclusion 42
Acknowledgments 43
References 43
CHAPTER 2
Economics of Data Communications
P. Afèche 53
1. Introduction 53
2. Problem definition and discussion framework 56
2.1. Service demand 57
2.2.. Service contracts: QoS guarantees and service designs 59
v
vi Contents
2.3. Service production 63

2.4. Distinctive features in comparison to telephony 67
2.5. Pricing decisions 67
2.6. Discussion framework 71
3. Pricing guaranteed services 73
3.1. Features, modeling and overview 73
3.2. Basic pricing and allocation principles 75
3.3. Incomplete information and adaptive pricing 78
3.4. More differentiation 78
3.5. Optimal dynamic pricing 80
3.6. Conclusions and directions 81
4. Pricing best effort services 83
5. Pricing flexible bandwidth-sharing services 106
6. Discussion 118
6.1. Auctions versus posted prices 118
6.2. Flat-rate versus usage-based pricing 124
6.3. Providing QoS: overprovisioning versus control 126
7. Conclusions 128
7.1. Research directions 128
7.2. Further topics and readings 129
References 129
CHAPTER 3
Firms and Networks in Two-Sided Markets
D. F. Spulber 137
1. Introduction 137
2. Firms in two-sided markets 140
2.1. Firms in two-sided markets: matchmaking 142
2.2. Firms in two-sided markets: market making 143
2.3. Information systems and platforms 147
3. Networks in two-sided markets 151
3.1. Transactions on networks in two-sided markets 151
3.2. Basics of network theory 154
4. Assignments of buyers and sellers in a network: costly communication 156
Contents vii
4.1. Assignments with homogenous products 157

4.2. Assignments with differentiated products 161
4.3. Second-best assignments 165
5. Networks and the core in a two-sided market 166
5.1. The core with homogenous products 166
5.2. The core with differentiated products 167
6. Stable assignments in a decentralized two-sided market: costly
computation 168
6.1. Stable assignments with homogenous products 169
6.2. Stable assignments with differentiated products 172
7. Firms and stable assignments in a centralized two-sided market 173
7.1. Firms and stable assignments with homogenous products 174
7.2. Firms and stable assignments with differentiated products 179
8. Matchmaking and market making by a firm using double auctions 181
8.1. Market making by a firm using double auctions for homogenous
products 182
8.2. Matchmaking and market making by a firm using double auctions
for differentiated products 183
9. Two-sided markets in random networks 188
9.1. Search and random assignments 188
9.2. Markets and random networks 193
10. Conclusion 196
Acknowledgments 197
References 197
CHAPTER 4
Organization Structure
T. Marschak 201
1. Introduction 201
2. Goals, mechanisms, and informational costs: the ‘‘incentive-free’’
case, where individuals obey the designer’s rules without inducement 204
2.1. Two general frameworks for judging the organization’s actions 204
2.2. How the organization finds its current action when incentives are not an
issue 207
2.3. Finite approximations of mechanisms whose message spaces are
continua 230
2.4. The dynamics of a mechanism 233
2.5. Constructing an informationally efficient mechanism 236
2.6. Finding a best action rule (outcome function) once a mechanism has
conveyed information about the environment to each person: the
methods of the Theory of Teams 238
2.7. Designing an organization "from scratch": choosing its members, what
each observes, and the speak-once-only mechanism that they use 240
3. Models in which the designer is concerned with incentives as well
as Informational costs 264
3.1. The message-space size required for implementation of a goal 265
viii Contents
3.2. Models in which the organization’s mechanism is partly designed by

its self-interested members, who bear some of the informational costs 266
3.3. Networks of self-interested decision-makers, who bear the network’s
informational costs 275
4. Organizational models in which the primitive is a ‘‘task’’,
‘‘problem’’, ‘‘project’’, or ‘‘item’’ 278
5. Concluding remarks 280
References 281
CHAPTER 5
Open Source Software: The New Intellectual Property Paradigm
S. M. Maurer and S. Scotchmer 285
1. Introduction 285
2. Incentives for R&D 287
2.1. Intellectual property and open source 288
2.2. Own use 290
2.3. Complementary goods and services 290
2.4. Signaling 293
2.5. Education 295
2.6. Achieving network externalities and denying them to others 295
2.7. Social psychology 296
3. Stability and organizational issues 300
3.1. Who contributes, and how much? 300
3.2. Who pays? 301
3.3. Why licenses? 302
3.4. Why leadership? 304
3.5. Network effects 305
4. Efficiency implications 306
4.1. Disclosure of code 306
4.2. Meeting users’ needs 307
4.3. Deadweight loss and pricing 308
4.4. Training and using programmers 308
4.5. Free riding 310
4.6. Modularity and the organization of the research effort 310
5. Open source and proprietary software 312
5.1. Competition between open source and proprietary software 312
5.2. Market segmentation 314
6. Limitations and extensions 315
6.1. Limits to open source software 315
6.2. Beyond software: drug discovery, geographic information systems,
and Wikipedia 315
7. Conclusion 318
Acknowledgments 319
References 319
Contents ix
CHAPTER 6
Information, Search, and Price Dispersion
M. R. Baye, J. Morgan, and P. Scholten 323
1. Introduction 323
2. Theoretical models of price dispersion 331
2.1. Search-theoretic models of price dispersion 332
2.1.1. The Stigler model 333
2.1.2. The Rothschild critique and Diamond’s paradox 336
2.1.3. The Reinganum model and optimal sequential search 338
2.1.4. Remarks on fixed versus sequential search 343
2.1.5. The MacMinn model 343
2.1.6. The Burdett and Judd model 346
2.2. Models with an ‘‘Information Clearinghouse’’ 348
2.2.1. The Rosenthal model 351
2.2.2. The Varian model 352
2.2.3. The Baye and Morgan model 354
2.2.4. Models with asymmetric consumers 357
2.5.5. Cost heterogeneities and the Spulber model 357
2.3. Bounded rationality models of price dispersion 358
2.4. Concluding remarks: theory 359
3. Empirical analysis of price dispersion 360
3.1. Measuring price dispersion 360
3.2. Price dispersion in the field 363
3.2.1. Dispersion and the ‘‘benefits’’ of search 363
3.2.2. Dispersion and the ‘‘cost’’ of search 365
3.2.3. Dispersion and the number of sellers 367
3.2.4. Dispersion and price persistence 369
3.3. Concluding remarks: empirics 370
Acknowledgments 371
References 371
CHAPTER 7
Behavior-Based Price Discrimination and Customer Recognition
D. Fudenberg and J. M. Villas-Boas 377
1. Introduction 377
2. Monopoly 379
2.1. Two-period model 380
Base model 380
No customer recognition 382
Customer recognition and behavior-based price discrimination 382
The role of commitment 383
2.2. Overlapping generations of consumers 384
No constant prices in equilibrium 386
Price cycles in equilibrium 387
2.3. Long-lived consumers 389
x Contents
Long-term contracts 391

Relationship to durable goods and bargaining 393
2.4. Two-good monopoly 396
3. Competition 398
3.1. Two periods, short-term contracts 399
Analysis of the two-period model under the MHR assumption 400
Discrete distributions 403
Welfare 404
3.2. Infinite lived firms, overlapping generations of consumers, and
short-term contracts 404
3.3. Long-term contracts 406
3.4. Switching costs 408
4. Behavior-based pricing with multiple products, and product design 413
4.1. Upgrades and buybacks with an anonymous second-handmarket 414
4.2. Upgrades and buybacks with non-anonymous consumers 417
4.3. Endogenous innovation 418
4.4. Endogenous location choice in duopoly 420
5. Related topics: privacy, credit markets, and customized pricing 422
5.1. Privacy 422
5.2. Credit markets 426
5.3. Customized pricing 429
6. Conclusion 431
Acknowledgements 433
References 433
CHAPTER 8
Information Technology and Switching Costs
Pei-yu Chen and L. M. Hitt 437
1. Introduction 438
2. Switching cost: definition and measurement issues 440
3. Switching costs, competition, and firm strategy 444
3.1. Switching costs and competition 444
3.2. Endogenous switching costs 444
3.3. Switching costs in information-intensive markets 445
3.4. Empirical evidence 448
Switching costs in software and other ‘‘high-tech’’ markets 449
Switching costs in online markets 450
4. Endogenous switching costs and firm strategy in information-intensive
markets 451
5. A framework for managing switching costs 455
5.1. Introduction 455
5.2. A model of customer retention 457
5.3. Measuring switching costs 461
6. Conclusion 464
Acknowledgments 466
References 467
Contents xi
CHAPTER 9
The Economics of Privacy
K.-L. Hui and I.P.L. Png 471
1. Introduction 471
2. ‘‘Free market’’ approach 475
3. Within-market consequential externalities 476
3.1. Non-productive information 477
3.2. Productive information 480
4. Cross-market consequential externalities 481
5. Direct externalities 483
6. Property rights 485
7. Regulation 487
8. Empirical evidence 489
9. Future directions 492
Acknowledgment 493
References 493
CHAPTER 10
Product Bundling
X. Geng, M. B. Stinchcombe, and A. B. Whinston 499
1. Introduction 499
2. Bundling for price discrimination: the case of two products 502
2.1. The base model 503
2.2. Issues to be considered in monopoly bundling 503
2.2.1. Pure bundling, mixed bundling, and the combinatorial issue 504
2.2.2. Bundling products and bundling buyers 504
2.2.3. Monitoring purchase 504
2.2.4. Posted price and auction 505
2.3. Pure bundling 505
2.4. Mixed bundling 508
2.5. Extension: bundling complements or substitutes 510
3. Bundling for price discrimination: the case of many products 510
3.1. The integer programming approach 512
3.2. Pure or simple mixed bundling of many products 513
4. Bundling as competition tools 515
4.1. A monopoly market plus a competitive market 517
4.2. A monopoly market plus a potentially duopoly market: the additive
valuations case 517
4.3. A monopoly market plus a potentially duopoly market: the
complements case 519
4.4. Duopoly bundling 520
5. Concluding remarks 522
Acknowledgment 523
xii Contents
CHAPTER 11
Dynamic Pricing in the Airline Industry
R. P. McAfee and Vera L. te Velde 527
1. Airline pricing 529
2. Existing literature 529
3. Dynamic price discrimination with price commitment 534
4. Continuous time theory 536
5. Efficiency in the Gallego and van Ryzin model 541
6. Efficiently allocating limited capacity under uncertainty 541
7. The log normal case 546
8. Options and interruptible sales 548
9. Actual airline pricing patterns 551
10. Research projects and mysteries 556
11. Conclusion 561
Appendix 561
References 567
CHAPTER 12
Online Auctions
A. Ockenfels, D. H. Reiley, and A. Sadrieh 571
1. Why do information systems make auctions (even) more popular? 571
2. Single-object auctions: theory and experiments 573
2.1. Standard auction mechanisms and models 573
2.2. Bidding behavior and auction outcomes in theory 575
2.3. Bidding behavior in controlled laboratory and field experiments 577
3. Reserve prices, minimum bids, and shill bids 582
3.1. Theoretical considerations 584
3.2. Empirical and experimental observations 587
3.2.1. Entry and revenue effects of public and secret reserve prices 587
3.2.2. Auction fever 590
3.2.3. Shill bids 591
4. Late and incremental bidding 594
5. The buy-now option 599
5.1. Explaining the buy-now option with risk-aversion 600
5.2. Explaining the buy-now option with impatience and other transaction
costs 601
5.3. Explaining the buy-now option with a sequence of transaction
opportunities 602
5.4. Empirical and experimental evidence 603
6. Parallel markets and other outside options 605
7. Multi-item auctions 608
7.1. Standard multi-unit auction mechanisms 608
7.2. Bid shading and demand reduction in multi-unit auctions 610
7.3. Complementarities and combinatorial auctions 612
Contents xiii
8. Design of online auctions 614

8.1. The advantages of long, open auctions 614
8.2. Controlling the pace of bidding 616
8.3. Design aspects in multi-unit auctions 620
Acknowledgements 621
References 622
CHAPTER 13
Reputation Mechanisms
C. Dellarocas 629
1. Introduction 629
2. Signaling and sanctioning role of reputation mechanisms 633
3. Reputation in game theory and economics 635
3.1. Basic concepts 635
3.2. Reputation dynamics 638
3.3. When is reputation bad? 640
3.4. Other extensions to the basic theory 642
4. New opportunities and challenges of online mechanisms 643
4.1. Eliciting sufficient and honest feedback 643
4.2. Exploring the design space of feedback mediators 645
4.3. Coping with cheap online identities 647
4.4. Understanding the consequences of strategic manipulation 648
4.5. Distributed reputation mechanisms 649
5. Empirical and experimental studies 651
5.1. Empirical studies and field experiments 651
5.2. Controlled experiments 654
6. Conclusions: opportunities for IS research 655
References 657
Subject Index 661

This page intentionally left blank
INTRODUCTION
Technological advances in processing and communicating information

have begun what may be the largest transformation of the economy since
the industrial revolution. The dramatic improvements in information tech-
nology (IT) and information systems (IS) decrease the cost of activities and
makes activities previously too expensive or cumbersome attractive.1 As IT
adoption proliferates both demand- and supply-side externalities increase,
making the technology cheaper and more valuable to additional users,
spurring further adoption and development of new applications. This vir-
tuous circle has enabled the dramatic growth of applications and busi-
nesses. The economic models upon which many of these businesses are
based on are old, e.g., auctions trace back to the Egyptians, but the ability
to easily connect vast parts of the population to businesses and other users
anywhere in the world makes raise new issues in new settings and for far
wider ranges of applications.
At the beginning of the Internet boom, the popular press promulgated the
view—often the view of firms either supplying the technology, using the
technology, or funding the firms using and supplying the technology—that
the Internet would cause everything to be different; with the primary
difference being that competition between firms would be enhanced by
creating frictionless markets and that information would be free.2 One of
the first and best attempts to deflate this hype was Information Rules by
Carl Shapiro and Hal Varian. This book succeeded in bringing to bear in
lucid text some of the academic literature on price discrimination and
damaged goods, customer switching costs and lock in, and network exter-
nalities and the associated compatibility issues and standard wars. It is my
hope that this volume will help further our scholarly understanding on these
and other topics, which will be discussed further.
1
The distinction between IS and IT is subtle and much debated, with IT often defined as actual
information technology and IS as its implementation in complex activities and systems of activities. This
volume focuses more on the economics of the IS uses, but these closely depend on the details of the
economics of the underlying IT. See Footnote 3 in Dan Spulber’s chapter for a more detailed industry
definition of information systems.
2
Table 1 in Dan Spulber’s chapter in this volume provides an excellent summary of how reducing
communication and computation costs can affect the way buyers and sellers interact.
xv
xvi Introduction
Information has risen to prominence in economics, for example, the 2001

Nobel Prize or Fig. 1 in Baye, Morgan, and Scholten in this volume.
However, economists often model technology in abstract ways. In contrast,
IS researchers often focus on the details of the technology without fully
exploring the implications of the systems’ capabilities for economic inter-
action. This volume attempts to bridge these approaches by highlighting
areas where IT is changing the importance of different economic forces and
illustrating the economics that lay behind the use of many types of infor-
mation systems.
While the areas of economics and other related disciplines impacted by
the growth and improvements in IS are too numerous to catalogue and
examine in a single volume, this book’s chapters survey many of the most
significant issues and our progress on them. The chapters in this volume
focus on various individual interrelated subjects regarding the economics of
information systems: the adoption and diffusion of information technol-
ogies; the pricing of data communications; the ways firms organize and
transform themselves as information is better captured and disseminated
within and across firms and customers; the means and tactics firms use to
compete with each other; the manner in which firms interact with and
distribute goods to customers; the methods and mechanisms for anony-
mous and infrequent interactions between users, firms, and customers in far
reaching locations; and the type and use of information on customers and
their behavior. These issues span areas of economics and disciplines within
business schools.
The chapters that follow flesh these areas out in detail. With the chapters
spanning such a broad topic from a variety of perspectives, there are many
ways to structure and order the chapters. I choose simply to start with two
chapters studying the aspects of the economics of the IT infrastructure:
diffusion, adoption, and pricing of information and communication tech-
nologies. The next three chapters examine the impact IT is having on mar-
kets (networks of buyers and sellers), organization of firms, and methods of
innovation. The remaining chapters focus on how IT is transforming firm
competition through the frictions in competition, firms’ knowledge of their
customers, and improvements in the mechanisms by which prices are
adjusted and information is collected.
In Diffusion of Information and Communication Technologies to Busi-
nesses, Chris Forman and Avi Goldfarb survey the literature on the adop-
tion and diffusion of information and communication technologies with a
focus on technologies that facilitate communication within and across firm
boundaries. These technologies improve the monitoring and coordination
capabilities of organizations. The chapter examines work on adoption costs
and benefits and how these are influenced by firms’ internal organization
and firms’ external environment. For the internal aspects, they examine
how organizational characteristics, firm boundaries, internal decision-mak-
ing rights, and individual incentives influence adoption. For the external
Introduction xvii
factors, they examine firms’ geographic location—both rural versus urban

and across countries, the role of network effects, and the interaction be-
tween the internal and external influences.
In Economics of Data Communications, Philipp Afèche surveys core as-
pects and roles of pricing data transport services with a focus on three
fundamentally different types of service contracts: guaranteed services,
congestion-prone best effort services, and adaptive bandwidth sharing. The
discussion is organized according to a unifying framework that compares
and contrasts the key features of service demand and delivery, develops the
basic pricing principles for socially optimal allocation, studies issues of
service differentiation and customer incentives, and considers iterative and
dynamic price mechanisms. The chapter then examines the value of auc-
tions for data transport services and closes with a discussion of the relative
merits and challenges of alternative quality of service approaches.
After the first two chapters on the economics of the adoption and pricing
of information communications technology, the book moves to three chap-
ters covering IS and markets, organizations, and innovation at high levels.
In Networks and Two-Sided Markets, Daniel Spulber represents decentral-
ized and centralized two-sided markets using network theory to shed light
on decentralized mechanisms in which consumers transact directly and
centralized mechanisms operated by one or more firms acting as interme-
diaries. After an excellent introduction—as with many of the chapters,
I recommend readers to study the introductory sections for less technical
background on the topics; this chapter’s introduction is particularly ex-
ceptional, the chapter explores the implications of costly communication
for the design and efficiency of market mechanisms for allocations of ho-
mogeneous and differentiated goods in buyer–seller networks. Firms’ em-
ploy IS to improve communication between buyers and sellers and to
improve computation through centralized market mechanisms. Firms
charging subscribers for admission to a communications network provide
centralized allocation mechanisms that can increase economic efficiency
and reduce transaction costs relative to decentralized exchange. The chapter
also presents how the network can be used via double auctions to link
buyers and sellers. Finally, the chapter considers the connection between
search models and random networks and compares the efficiency of buyer–
seller search with centralized assignments.
In Organizational Structure, Thomas Marschak examines formal models
of organizations that regularly acquire information about a changing en-
vironment in order to optimize their actions.
Each member of the organization privately learns about particular aspects
of the new environment. The organization operates a mechanism to process
this information and act on it. The mechanism has various informational and
agency costs and balances these costs against the benefits in performance. As
costs drop, due to improved IT, the properties of good mechanisms, and
hence the structure of the organizations that adopt them change.
xviii Introduction
In Open Source Software: The New Intellectual Property Paradigm, Step-

hen Maurer and Suzanne Scotchmer study how a new form of innovation
and intellectual property is developing in information systems: open source
methods for creating software which rely on voluntarily revelation of code.
Open source incentives differ from other intellectual property paradigms,
leading to different types of inefficiencies and different biases in R&D in-
vestment. Open source remedies a defect of intellectual property protec-
tion—the lack of requirements or encouragement to disclose source code.
The lack of disclosure in proprietary innovation hampers interoperability in
complex systems. The chapter analyzes developers’ incentive to participate
in open source collaborations, studies when open source is most likely to
predominate, and evaluates the extent to which open source may improve
welfare compared to proprietary development.
As noted above, the Internet was claimed to move us into a frictionless
utopia where the ‘‘law of one price’’ would truly apply. In Information,
Search, and Price Dispersion, Michael Baye, John Morgan, and Patrick
Scholten establish that remarkably little progress has been made toward
this idealistic state. The chapter provides a unified treatment of various
search models that have been proposed to rationalize price dispersion in
markets for homogeneous products: sequential search, fixed sample search,
and clearinghouses. These models reveal that reduction or elimination of
consumer search costs do not always reduce or eliminate price dispersion.
The links the authors draw between the models is a significant contribution.
The chapter further connects the search literature to mechanism design by
showing how auction tools can simplify and even generalize existing results.
The chapter concludes with an overview of the burgeoning empirical lit-
erature, which suggests that price dispersion in both online and offline
markets is sizeable, pervasive, and persistent.
Just as IT changes consumers’ ability to gather information about firms
and their prices, information technologies and the Internet allow firms to
keep, gather, and process more information about their past customers.
This increase in information has led to the proliferation of customer re-
lationship management practices in most industries. In Behavior-Based
Price Discrimination and Customer Recognition, Drew Fudenberg and Mi-
guel Villas-Boas examine how when firms have information about con-
sumers’ previous purchases, they can use it to offer different prices and/or
products to consumers with different purchase histories—practice ‘‘be-
havior-based price discrimination’’ (BBPD). Throughout the chapter firms’
commitment problem arises: although having more information helps ex-
tract more surplus with its current prices, consumers may anticipate this
possibility and alter their initial purchases. A second theme is that more
information may lead to more intense competition between firms, creating a
potential prisoner’s dilemma where each firm would gain from practicing
BBPD, but industry profits fall when all practice it. The chapter also
surveys the literature on firm competition, short- and long-term contracts
Introduction xix
between firms and customers, and firms’ optimal product lines under
numerous variations.
One of the most important aspects of multi-period interactions between
firms and customers is how consumers’ past purchases impact their future
ones, which implicitly produce one type of demand-side externality: switch-
ing costs. In Information Technology and Switching Costs, Pei-yu Chen and
Lorin Hitt investigate how firms can influence consumer-switching costs
with a focus on ‘‘information-intensive’’ markets which often have signifi-
cant standardization and compatibility issues. The chapter presents a for-
mal definition of switching costs, clarifies some general points about
switching costs, and reviews some theoretical and empirical studies of IT
and switching costs. The chapter highlights how switching costs arise en-
dogenously in high-tech and information-intensive markets while discussing
instruments for influencing these costs. Finally, the chapter provides a dis-
crete choice model for managing and estimating customer switching costs.
While IT could benefit consumers by more precisely identifying their
needs, it can also be used to price discriminate (as in BBPD) or to exclude
individuals with less attractive characteristics. Furthermore, organizations
sell customer information to third parties, subjecting their customers to
their information being used ‘‘against’’ them more broadly. In The Eco-
nomics of Privacy, Ivan Png and Kai-Lung Hui examine the issue of who
should control what information is tracked and stored through an eco-
nomic analysis of privacy. The chapter begins with the ‘‘free market’’ cri-
tique of privacy regulation. Because welfare may be non-monotone in the
quantity of information—due to the cost of information or some consumers
being priced out of the market when it is socially efficient for them to
consume—there may be excessive incentive to collect information. This
result applies to both non-productive and productive information and is
exacerbated when personal information is exploited across markets. Fur-
thermore, the ‘‘free market’’ critique does not apply to overt and covert
collection of information that directly causes harm, for example, a flood of
unsolicited promotions. The chapter then reviews research on property
rights and the challenges in determining their optimal allocation to examine
whether or not individuals might voluntarily or be paid to reveal their
personal information.
IT enabling more flexible pricing policies has increased price discrimina-
tion in many ways. In Product Bundling, Xianjun Geng, Maxwell Stinchco-
mbe, and Andrew Whinston study product bundling, especially for
information goods whose low marginal cost and flexibility facilitates bun-
dling. While bundling can reduce integration, transaction, and distribution
costs, the straightforwardness of this argument has limited its presence in
the literature. Therefore, this chapter focuses on bundling for price dis-
crimination and bundling as a competition tool. Price discrimination arises
most easily where buyer valuations over two products are negatively cor-
related. In general, bundling benefits a seller when it reduces valuation
xx Introduction
heterogeneity and if marginal cost is low. However, these are not necessary
conditions and the chapter explores when bundling emerges in a broader
range of cases. Research on using bundling as a competition tool falls into
two categories: entry deterrence (tying) and product differentiation. While
the literature on using bundling for entry deterrence focuses on how a seller
can fend off all rivals, the literature on using bundling for product differ-
entiation asks the question of when two or more ex-ante homogeneous
sellers can coexist and both reap positive profits using bundling.
Another aspect of pricing impacted by information technologies’ ability
to track and analyze information is dynamic pricing, also known as yield
management or revenue management, which is most useful when the prod-
ucts expire at a point in time and where the capacity is fixed. Airlines are a
natural instantiation of these characteristics and in the 1980s airline pio-
neered the use of complex IS to constantly monitor and vary their prices. In
Dynamic Pricing in the Airline Industry, Preston McAfee and Vera te Velde
go beyond surveying of yield management research to expand a common
existing model to a more standard case. Then, by examining the efficient
allocation, rather than the profit-maximizing allocation, the chapter shows
that many of the conclusions attributed to profit-maximization are actually
consequences of dynamic efficiency. The chapter proposes the perspective
of selling options and suggests that airlines should sell two kinds of tickets:
a guaranteed use ticket and a ticket that can be delayed at the airline’s
request. Finally, airline pricing data is used to generate stylized facts about
the determinants of pricing, facilitating the evaluation of different models.
The prior chapters on firms’ pricing decisions implicitly assume variants
on posted price schemes. Improvements in IT also enhance the attractive-
ness of more general selling mechanism such as auctions. Auctions have the
advantage of price discovery and the disadvantage of higher transaction
costs. IS improve both sides of the trade-off in auctions favor by reducing
the transaction costs of auctions and raising the value of price discovery by
increasing the number of potential participants. In Online Auctions, Axel
Ockenfels, David Reiley, and Abdolkarim Sadrieh provide an overview of
some of the theoretical, empirical, and experimental research on online
auctions. The chapter first compares theoretical single-object-auction
results to experimental findings and empirical observations online. The
chapter then focuses on auction design details: public and secret reserve
prices including minimum bids and shill bids, late and incremental bidding,
and the buy-now option. The buy-now option creates an outside option for
bidders, which is also provided by parallel auctions and other outside op-
tions by other sellers. The chapter also discusses multi-object auctions,
although these are both theoretically and empirically more difficult than the
single-object auctions. Finally, the chapter concludes with general remarks
on the design of online auctions.
When large numbers of buyer and sellers interact with one infrequently,
reputation and brand become weaker and the adverse selection and model
Introduction xxi
hazard problems become more acute. Auctions are the most extreme exam-
ple of this, but it is a widespread phenomenon. Online reputation mecha-
nisms using the Internet’s bi-directional communication capabilities mitigate
these concerns by allowing individuals to share opinions and experiences on
a wide range of topics. In Reputation Mechanisms, Chrysanthos Dellarocas
surveys our progress in understanding the limitations inherent in these
mechanisms. The chapter discusses how technology-based reputation mech-
anisms differ from traditional word-of-mouth networks, for example, the
global reach and anonymity of online interactions. The chapter combines an
overview of relevant work in game theory and economics as well as insights
from computer science, marketing, and psychology to evaluate the design,
evaluation, and use of reputation mechanisms in online environments.
I am very pleased with this volume. The authors deserve congratulations
for their outstanding work. I want to thank all of them for this and for their
forbearance in the delays in the publication process. I believe that the vol-
ume will provide a survey of our current state of knowledge in theses areas
and frame the most fruitful directions for future research.
Terrence Hendershott
Terrence Hendershott, Ed., Handbooks in Information Systems, Vol. 1
Copyright r 2006 by Elsevier B.V.
Chapter 1
Diffusion of Information and Communication

Technologies to Businesses
Chris Forman
Tepper School of Business, Carnegie Mellon University, Pittsburgh, PA, USA
Avi Goldfarb
Rotman School of Management, University of Toronto, Toronto, ON, Canada
Abstract
We survey the literature on the adoption and diffusion of information and

communication technologies (ICT) in businesses. We identify two key di-
mensions that have been the focus of most of the literature. First, research
can be categorized as focusing on ICT adoption costs or ICT adoption ben-
efits. Second, research can be categorized as focusing on how adoption is
influenced by either the internal organization of the firm or by the external
environment. Major themes are highlighted as opportunities for future re-
search.
1 Introduction
The invention of new technology is only the first step in economic

progress. To contribute to economic growth, new technology must first be
used in the course of productive economic activity. However, new technol-
ogy often diffuses unevenly across economic agents. Such variance in
adoption rates is important, as it influences the rate of technological
progress. It may also be a source of sustainable competitive advantage
(Mata et al., 1995).
Research on information and communication technology (ICT) diffusion
to businesses seeks to understand why firms adopt ICT at different rates.
Researchers attempt to identify and measure characteristics of the organ-
ization and its environment that influence the barriers to ICT investment
1
2 C. Forman and A. Goldfarb
and the benefit of ICT investment to the organization. By investigating how

business characteristics shape ICT investment decisions through their im-
pact on (unobservable) costs and benefits of adoption, this literature is
complementary to business-value research, which seeks to identify how
observable ICT investment decisions influence observable gross benefits.
This line of research is important to understanding the fundamental ques-
tion in the information systems literature of ‘‘Why do some organizations
succeed with their information technology investments while others do
not?’’ (Dhar and Sundararajan, 2004).
This review examines research on the diffusion of ICT to businesses.
Because of the breadth of this field, we limit the scope of our review in
several ways. First, we focus upon recent research on ICT that facilitates
communication within and across firm boundaries. Internet technology is
one common example of such technology, but it is not the only one.1
Information technologies with data processing functions but no commu-
nications functions are not considered. By lowering the costs of communi-
cations, ICTs improve the monitoring and coordination capabilities of
organizations. Moreover, they will be most susceptible to network effects: the
value of the technology will be increasing in the number of other users. A
large theory literature has studied how the communications capabilities of
ICT will alter the organization of firms (Marschak, 2006), their interrelation-
ships (Spulber, 2006), and their location (Gaspar and Glaeser, 1998). Owing
largely to data constraints, empirical testing of these hypotheses has only been
possible over the past 15 years. This review will examine recent advances
made in this empirical literature, and highlight areas for future research.
Moreover, we will focus attention upon ICT that involves significant
costs of adoption at the organizational level.2 Thus, our review generally
excludes technologies such as cell phones that have significant infrastruc-
ture costs but which involve little idiosyncratic adaptation costs at the
organizational level to be used successfully. As in any review article, the
content of our discussion will be shaped by the extant literature. Thus, our
review will include detailed discussion on EDI but less on newer technol-
ogies such as RFID for which the literature is presently small.
Last, we examine ICT diffusion through the lens of economic analysis.
Economic analysis offers a rich theoretical framework for understanding the
diffusion of ICT. Moreover, the increasing availability of micro data has led
to rapid growth in this field. We consider this review to be complementary to
other excellent review articles in the information systems (IS) literature
that have focused on the diffusion of IT through an organizational lens
(Fichman, 1992, 2000), or to research in economics that has reviewed the
1
Other examples include electronic data interchange (EDI) and automated clearing house (ACH)
technology in banking, both of which involve communication that does not occur over the Internet
backbone.
2
The main exception is Section 7, in which we examine country-level diffusion of Internet technology.
Ch. 1. Diffusion of Information and Communication Technologies 3
literature on the diffusion of other technologies (Hall and Khan, 2003). By

focusing on the economic analysis of ICT diffusion to businesses, we address
a growing area of the literature that has not recently been reviewed.
Table 1 provides an overview of some of the major questions in ICT
diffusion research and our framework for addressing them in this paper.
Research in this area can be classified as exploring either how characteristics
of the organization or of its external environment shape the (net) benefits of
ICT adoption. Further, as noted above, some characteristics influence pri-
marily the costs and some influence primarily the benefits of adoption (and
some influence both). In Table 1, we classify some of the major questions in
diffusion research along these two dimensions, and then indicate with sec-
tion numbers where research on these questions can be found in the paper.
Following the introduction, we begin Section 2 by describing the empir-
ical models that are commonly used in the economic analysis of ICT diffu-
sion, and provide a summary of recent work that examines the economic
outcomes from ICT investment.
Sections 3–7 constitute the main body of the paper. In Section 3, we
describe how organizational characteristics influence the rate with which
firms adopt ICT. Some research in this area focuses on the adaptations
needed by organizations to successfully implement ICT, the costs of these
adaptations, and how firms overcome them. Other research examines the
relationship between ICT adoption, organizational characteristics, and in-
dividual incentives. In particular, an evolving body of work seeks to un-
derstand how the location of firm boundaries and internal decision-making
rights influence the speed of ICT adoption.
Sections 4–7 review literature that examines how an organization’s external
environment influences the decision to adopt ICT. Section 4 describes how
the geographic location of a firm affects the speed of ICT adoption. In
particular, we emphasize recent research that examines whether the commu-
nications capabilities of ICT imply that the marginal benefit to ICT adoption
is highest in rural areas. Section 5 explores research into the extent to which
internal and external factors jointly influence an organization’s decision to
adopt ICT. Section 6 shows how network effects shape the pattern of ICT
diffusion. Though prior literature has begun to identify the role of network
effects on the diffusion of new technologies, new data sets offer opportunities
to draw a tighter link between theory and evidence than has previously been
possible. In Section 7, we review research on cross-country differences in the
speed of Internet technology diffusion. Here we change focus somewhat from
the micro-level determinants of firm adoption to macro-level factors that
shape differences in country-level ICT diffusion. Moreover, in this section we
focus primarily on the diffusion of Internet technology because of the im-
portant policy implications of this research.3 Section 8 concludes.
3
Because of the different unit of analysis (country versus organization), cross-country differences in
Internet use are not included in the framework in Table 1.
Table 1
Summary of research on ICT diffusion
How adopter characteristics influence the net benefits of

internet adoption
Source of variation in Adoption costs Adoption benefits

adopter characteristics
Internal organization How do firm boundaries How does ICT influence the
influence the speed of ICT optimal location of decision-
adoption? (see Section making rights within firms?
3.2) (see Section 3.1)
How does firm size influence Which organizational
the speed of ICT characteristics are
adoption? (see Section complementary with ICT?
3.3) (see Section 3.1)
How do prior investments How does ICT investment
influence the speed of new influence firm boundaries?
ICT adoption? (see (see Section 3.2)
Section 3.4)
How does co-invention How does firm size influence
shape the diffusion of the speed of ICT adoption?
ICT? (see Section 5.1) (see Section 3.3)
External environment Is ICT use a complement or Is ICT use a complement or
substitute for urban substitute for urban
agglomeration? (see agglomeration? (see Section
Section 4.1) 4.1)
How does co-invention How does ICT use influence the
shape the diffusion of location decisions of firms?
ICT? (see Section 5.1) (see Section 4.2)
Do network effects shape Do network effects shape the
the adoption of ICT? (see adoption of ICT? (see Section
Section 6.2) 6.2)
How do network How do network externalities
externalities shape the shape the adoption of ICT?
adoption of ICT? (see (see Section 6.3)
Section 6.3)
Note: Table presents major research questions in the literature on ICT adoption classified along two
dimensions: variation in adopter characteristics and how adopter characteristics influence the net ben-
efits to adoption
2 ICT diffusion and its impact
2.1 Diffusion modeling
Before examining the economics of ICT diffusion, it is necessary to give a

brief overview of diffusion research methodology. An excellent literature
review on diffusion modeling in economics is Stoneman (2002). In this

section, we provide an overview of diffusion modeling, highlighting some of
the common themes that will appear in our analysis of ICT diffusion.
Adoption is the individual-level decision to use a new technology. Diffu-
sion is the aggregation of a number of adoption decisions. Rogers (1995, p. 5)
defines it as ‘‘the process by which an innovation is communicated through
certain channels over time among the members of a social system.’’ Diffusion
research is then concerned with finding patterns across a large number of
adoption decisions.
The earliest economic models of diffusion were epidemic models. These
models assumed that the diffusion of new technology is like that of an
infectious disease. Non-adopters adopt a new technology when they come
into contact with adopters and learn about the new technology. Over time,
the number of users increases, leading to an increased probability of any
given non-adopter learning about the technology. This increases the rate of
diffusion. As more people adopt, the number of non-adopters declines,
which decreases the rate of diffusion. This pattern of diffusion leads to the
common S-shaped curve on the rate of technology diffusion with respect to
time (see Fig. 1).
The first modern technology diffusion study by Ryan and Gross (1943).
They used an epidemic model to study the diffusion of hybrid corn to Iowa
farmers and found that social networks matter. Methods for measuring
epidemic effects developed around the same time in sociology and eco-
nomics. Sociologists Coleman et al. (1957) were among the first to use
epidemic diffusion models outside of an agricultural setting. Their work
examined physician choices of new drug prescriptions.
Epidemic models are commonly used to help forecast the rate of aggregate
technology diffusion. Bass (1969) uses an epidemic model to help predict the
# of Adopters
Time
Fig. 1. The typical pattern of diffusion.

rate at which a product will diffuse.4 The central themes of these models—
communications and social networks—are also prominent in recent eco-
nomic research on technology diffusion. In Section 6.2, we discuss papers
that have examined how these themes may have influenced the diffusion
of personal computers (Goolsbee and Klenow, 2002), use of online grocery
services (Bell and Song, 2004), and early networks such as BITNET
(Gurbaxani, 1990).
As noted above, in epidemic models technology spreads through inter-
personal contact and information dissemination. These models do not ex-
plicitly model the adoption decisions of individual users, nor do they allow
for differences in the costs and benefits of adoption by different members of
the population. As a result, these models omit many important aspects of
economic behavior. Later models explicitly include these elements.
Probit (or rank) models emphasize population heterogeneity. Pioneered
by David (1969), the most basic model assumes that the entire population
has perfect information about the technology. Individuals (or firms) adopt
the technology when the net benefit of adopting is positive. Since the probit
model is the one most commonly used in economic diffusion modeling, it is
worthwhile to consider it further. In general under the probit model, an
establishment i will adopt a new ICT at time t if the following conditions
hold
NBðxi ; tÞ Bðxi ; tÞ Cðxi ; tÞ40 (1)
0
NBðxi ; tÞ4NBðxi ; t0 Þ=ð1 þ rÞt 1 8t0 at (2)
where NB is the net benefit of adoption, B the gross benefit of adoption, and
C the cost of adoption. All functions represent the present value of profits
discounted to time t. We let xi be a vector of firm characteristics that in-
fluence the value of adopting the new technology. These equations say that a
firm will adopt if two conditions hold—first, that the expected benefits less
expected costs (or net benefits) are positive and second, that the net benefits
of adopting at time t are greater than the net benefits of adopting any other
time t0 at: A technology diffuses throughout the population because either
the benefits of adopting are increasing over time (due, e.g., to improvements
in the technology as a result of technological change), @Bðxi ; tÞ=@t40; or
because the costs of adoption are declining, @Cðxi ; tÞ=@t40: Most diffusion
papers are unable to separately identify benefits and costs and instead iden-
tify @NBðxi ; tÞ=@xi ; the change in the net benefits to adoption that occur as
a result of the changes in firm characteristics. As Hall and Khan (2003)
note, due to high sunk costs of adoption, adoption is usually an absorbing
4
Fichman (2000) refers to epidemic models used to forecast the rate, pattern, and extent of technology
diffusion as ‘‘diffusion modeling studies.’’ For surveys of such studies, see Mahajan et al. (1990) and
Mahajan and Peterson (1985).
state. That is, we rarely observe organizations ‘‘unadopting’’ a new tech-

nology, and analysts rarely worry about this decision in their econometric
modeling.
The basic probit model underlies any diffusion modeling that explicitly
considers agents’ trade-offs between the costs and benefits of adopting, and
it is the workhorse for many of the models we discuss in Sections 3–6.5
However, in contrast to epidemic models, the probit model examines ex-
clusively how internal firm factors shape the benefits to adoption and as-
signs no role to the behavior of other users. Clearly this may be too limiting.
Recent economic models of ICT diffusion have extended the probit model
to allow a role for other users’ behavior.
In addition to epidemic and probit models, Karshenas and Stoneman
(1993) consider two ways in which other users’ behavior may influence
technology adoption, which they term ‘‘stock and order effects.’’ Stock
models argue that if new technologies are cost reducing, they will increase
the output that a firm produces. As a result, increasing adoption of new
technologies eventually decreases the profits of adopters and non-adopters
alike. Under certain conditions, the difference in profits between adopters
and non-adopters declines over time (Reinganum, 1981), leading to de-
creasing net benefits from new technology adoption. Stock models assume
that the profits among all adopters are identical—as are the profits among
all non-adopters. In contrast, order models assume that the benefits of
new adoption decrease monotonically with the number of prior adopters.
Despite this, early adopters continue to benefit disproportionately from
the technology (Fudenberg and Tirole, 1985). The intuition behind why
early adopters enjoy higher profits than later adopters is that there may
be first-mover advantages to adoption, due, for example, to the ability
of early movers to capture scarce inputs such as labor (Ireland and
Stoneman, 1985).
Order and stock effects are examples of negative network externalities.
That is, the benefits of adopting a new technology are declining in the
number of other users. However, the benefits of adopting a new technology
can also increase as others adopt, generating positive network externalities.6
Positive network externalities can further be categorized as direct or indirect.
The telephone provides an example of a direct positive network externality:
the value of adoption relies explicitly on the value of communicating with
other users. Alternatively, positive network externalities may be indirect,
as in the case of video game consoles such as the Sony Playstation: the
value of adoption increases in the adoption by other users because of the
increased availability of a complementary good (games). Recent research
5
In Section 7, the unit of analysis is frequently the country, making it more difficult to identify the
underlying economic model that is generating the observed empirical pattern.
6
The description at the beginning of Section 6 provides further details on positive network exter-
nalities.
has attempted to understand how direct and indirect network externalities

shape user adoption behavior; however, the econometric identification issues
in this research are daunting. A finding of a statistical correlation between
one user’s adoption and another’s may reflect network externalities. Alter-
natively, it may reflect unobserved characteristics that are common across
users and which increase the value to adoption. In Section 6.3, we discuss the
conditions under which these models are identified.
The main review of diffusion research in fields other than economics
is Rogers (1995). Emphasizing communications and sociology, Rogers
focuses on the role of communications networks in technology diffu-
sion. He details the process through which innovations move from one
population to another and discusses the role of five key factors in
the individual decision to adopt: relative advantage, complexity, compat-
ability, trialability, and observability. He emphasizes that these factors
are only relevant after informative contact with the innovation, and
much of this work focuses on the roles of different communications
networks in initiating this contact. This contact is achieved by a ‘‘change
agent.’’ The change agent brings an innovation into a new social net-
work and can effectively communicate its benefits. Managers aiming to
generate technology adoption should think of themselves as change
agents. In their emphasis on how characteristics of the adopter or technol-
ogy shape the value of adopting, the five key factors are similar to the
emphasis on heterogeneous adoption benefits and costs in the probit
model. Moreover, Rogers’ emphasis on communication and change agents
is analogous to the importance of information transmission in epidemic
models of diffusion.
2.2 The impact of ICT diffusion
There is considerable evidence that ICT investment has a positive impact

on firm performance. One line of research has demonstrated using micro-
economic production theory that ICT investment had a large impact on
the company behavior and productivity in the late 1990s. For example,
Jorgenson and Stiroh (2000), Oliner and Sichel (2000), and Baily and Law-
rence (2001) credit ICT with the rapid growth of the US economy in the late
1990s. Stiroh (2002a) argues that this productivity acceleration was broad
based and finds an increase in productivity related to ICT use in nearly two-
thirds of industries from 1995 to 2000. Baily and Lawrence (2001, p. 308)
claim about the 1990s, ‘‘In particular, there has been a substantial structural
acceleration of total factor productivity outside of the computer sector. And
there is clear supportive evidence of an acceleration of productivity in service
industries that are purchasing [ICT].’’ In another study of ICT and pro-
ductivity growth in the 1990s, Brynjolfsson and Hitt (2003) use firm-level
data to find that substantial long-term productivity gains result from ICT
use.7 Although these studies focus on the role of generic information tech-
nology investment, recent work has demonstrated a link between computer
networking and acceleration in establishment-level productivity (Atrostic
and Nguyen, 2002; Stiroh, 2002b).
Another line of work has examined how adoption of ICT improves
business processes. This ‘‘process-oriented’’ framework (Barua and
Mukhopadhyay, 2000) examines the relationship between ICT and interme-
diate process-level variables that are more closely related to ICT investment
than the measures of output, sales, or value added that are traditionally used
in production theory studies. For example, Mukhopadhyay and several
co-authors have examined the impact of ICT investment on supply
chain performance (Srinivasan et al., 1994; Mukhopadhyay et al., 1995;
Mukhopadhyay and Kekre, 2002). Hubbard (2003) shows that on-board
computers increase efficiency in trucking. Athey and Stern (2002) document
the role of ICT in improving emergency health care response outcomes.
Recent research has argued that the link between IT adoption and
firm performance will depend on usage. Using data on hospitals, Devaraj
and Kohli (2003) argue that ICT use is a much better predictor of per-
formance than ICT adoption at the firm level. Many firms adopt a tech-
nology on the surface, but, unless it is frequently and properly used, it will
not have a positive impact and may even have a negative one. We review
recent research that has examined IT usage separately from IT adoption in
Section 3.5.
In summary, ICT has had an important impact on productivity at both
the micro- and macro-levels. The rest of this paper details patterns in
ICT diffusion. We first look at how organizational characteristics influ-
ence technology adoption. We then examine the external environment in
Sections 4–7.
3 ICT adoption and organizational characteristics
The decision to adopt a technology can be influenced by factors relating

directly to the firm. In this section we consider how organizational char-
acteristics influence ICT adoption (see Table 2).
3.1 Adoption, internal firm organization, and organizational change
Bresnahan and Trajtenberg (1995, p. 84) define general purpose technol-

ogies (GPTs) as ‘‘enabling technologies, opening up new opportunities
rather than offering complete, final solutions.’’ Several authors have argued
that ICT is a GPT (e.g., David, 1990; Bresnahan and Greenstein, 1996;
7
For firm-level evidence on the productivity benefits of ICT investment from an earlier time period,
see Brynjolfsson and Hitt (1996).
Table 2
Summary of research on ICT adoption and organization characteristics (Section 3)
Open question Papers Results
How does ICT Gurbaxani and Whang (1991), ICT can lead both to
influence the optimal George and King (1991), centralization and
location of decision- Marschak (2004), Hubbard decentralization of decision-
making rights within (2000), Barua et al. (1995, making rights, depending on
firms? (see Section 1996), Barua and Whinston the technology and
3.1) (1998), Hitt and organization characteristics.
Brynjolfsson (1997), To date, cross-industry
Bresnahan et al. (2002), empirical work has
Dunlop and Weil (1996), suggested that
decentralization
predominates
Which organizational Mendelson and Pillai (1998), ICT adoption is more valuable
characteristics are Dewan et al. (1998), Banker for firms that operate in
complementary with et al. (2003), Forman (2005) dynamic business
ICT? (see Section environments, firms with
3.1) higher coordination costs,
and firms that encourage
participatory behavior by
value chain partners
How does ICT Gurbaxani and Whang (1991), Decreases in coordination
investment influence Malone et al. (1987), costs engendered by ICT
firm boundaries? Clemons et al. (1993), lead to less vertical
(see Section 3.2) Brynolfsson et al. (1994), integration. However,
Hitt (1999), Dewan et al. decreases in monitoring
(1998), Kraemer et al. costs may lead to more
(2002), Baker and Hubbard vertical integration
(2003, 2004)
How do firm Forman and Gron (2005), Decreases in vertical
boundaries influence Gertner and Stillman (2001) integration lead to increases
the speed of ICT in frictions that slow ICT
adoption? (Section adoption, other things equal
3.2)
How does firm size Kimberly and Evanisko Size is positively correlated
influence the speed (1981), Eveland and with adoption. However,
of ICT adoption? Tornatzky (1990), Attewell most studies are unable to
(see Section 3.3) (1992), Hannan and identify the theoretical
McDowell (1984), Charles et explanation for this
al. (2002), Forman et al. empirical result
(2002, 2006), Astebro (2002,
2004), Banker et al. (2003),
Hubbard (2000), Kauffman
et al. (2000)
How do prior Tornatzky and Fleischer Firms that have made more
investments (1990), Bresnahan and recent ICT investments or
Table 2. (Continued )
influence the speed Greenstein (1996), Forman that have more experience
of new ICT (2005), Zhu and Kraemer with ICT will adopt faster,
adoption? (see (2005) ceteris paribus. However, if
Section 3.4) these investments are
embedded they may lead to
slower adoption
How does intra-firm Fichman and Kemerer (1997, There is considerable evidence
diffusion differ from 1999), Attewell (1992), that the pattern of ICT
inter-firm diffusion? Astebro (1995, 2004), usage differs from that of
(see Section 3.5) Battisti and Stoneman adoption. However,
(2003), Battisti et al. (2004), theoretical development and
Copper and Zmud (1990), empirical testing of the
Goldfarb and Prince (2005), reasons behind these
Kraut et al. (1998) differences are at an early
stage
How do individuals Examples include Davis The TAM model has focused
use ICT once it has (1989), Davis et al. (1989), on how perceived usefulness
been adopted by an Szajna (1996), Venkatesh and ease of use predict long-
organization? (see and Davis (2000), others run usage
Section 3.5)
Note: First column describes major questions that explore how organizational characteristics influence
ICT adoption. Second column describes some of the major papers that have addressed these issues.
Third column describes a generalization of the results of these papers.
Harris, 1998; Forman et al., 2005).8 A GPT is an enabling technology

that can be deployed in a multitude of ways in many sectors of the
economy. However, adapting these general solutions to the needs of indivi-
dual industries and idiosyncratic users often poses great challenges for
organizations.
A large case study literature has detailed the challenges of implementing
general ICT systems to fulfill organizational needs. This case study liter-
ature has shown there to be significant technical development risks when
implementing new ICT systems in firms (e.g., Kemerer and Sosa, 1991).
Moreover, organizations face significant challenges in adapting business
processes and organizational routines to new ICT systems (e.g., Attewell
and Rule, 1984; Scott Morton, 1991; Davenport, 1998). Finally, new ICT
systems may require a different skill set than may have been previously
available, requiring either additional education or skills (Autor et al., 2002;
Levy and Murname, 2004), or they may also be ‘‘de-skilling,’’ requiring
workers to be less skilled than previously. Despite this rich case study
8
Examples of other historical GPTs include the steam engine (Rosenberg and Trajtenberg, 2004) and
the dynamo (David, 1990).
literature, there remains relatively little empirical adoption research in this

area that applies an economic perspective. The main work in this area has,
instead, explored theories of organizational change.9
New information technology often changes the organization of work
within firms. Beginning with Leavitt and Whisler (1958), a long-running
question within the IS field is how improvements in information processing
capabilities change the optimal location of decision rights within organi-
zations. Improvements in information processing capabilities can decrease
monitoring costs, leading to lower agency costs when delegating decision-
making rights. However, lower costs of information transmission can also
lower the costs of centralizing information within the organization by im-
proving the quality and speed of information processing (Gurbaxani and
Whang, 1991). A rich case study and small sample empirical literature has
found support for both hypotheses within the context of information
processing IT (George and King, 1991). Marschak (2004) formalizes many
of these ideas in an economic model. However, only recently have there
been large-scale empirical studies that examine the relationship between
ICT adoption and decision rights within firms.
Hubbard (2000) pursues this debate in the context of trucking carriers’
decisions to adopt electronic monitoring technologies. He notes that mon-
itoring ICT can create two benefits for organizations. First, it can lower
agency costs. For example, on-board computers in the trucking industry can
monitor drivers’ speeds. Thus, such technologies increase the value of de-
centralization, all else equal. Second, monitoring ICT can improve the re-
source-allocation decisions of managers, for example, by monitoring the
relative locations of trucks. By lowering information transmission costs,
these technologies can also increase the value of centralization. Hubbard
shows that both types of monitoring technologies are adopted by trucking
firms. Moreover, the likelihood that each is adopted will depend on the
relative values of centralization and decentralization across different carriers.
Other research has argued that ICT is more valuable when adopted by
firms that use innovative organizational and human resource practices, such
as self-managing teams, flatter organizational hierarchies, and broader job
descriptions that involve decentralization. Milgrom and Roberts (1990) ar-
gue that investments in new technology may have indirect effects that in-
crease the value of utilizing such innovative practices. They say that groups
of activities such as new technology investment and organizational vari-
ables are complements when an ‘‘increase in the levels of any subset of
activities implies that the marginal return to increases in any or all of the
remaining activities increases.’’ Researchers in the IS literature have argued
that there exist complementarities between IS design, worker incentives,
9
For recent exceptions, see the discussion on research on co-invention in Section 5.1. See also Doms et
al. (1997) for an econometric examination of the relationship between worker skills and adoption of
computer-automated design, numerically controlled machines, and programmable controllers.
and organizational characteristics (e.g., Barua et al., 1995, 1996; Barua and
Whinston, 1998).
A number of recent papers have empirically tested the assertions of
Milgrom and Roberts (1990). Hitt and Brynjolfsson (1997) and Bresnahan
et al. (2002) use a large cross-industry study to show that decentralized
decision rights, innovative human resource practices, and workplace in-
vestments in human capital are complementary with ICT investment.
Brynjolfsson et al. (1997) develop a tool for managers to understand the
indirect effects between ICT investments and organizational practices, and
how these interactions should shape investment in new ICT. Dunlop and
Weil (1996) study the use of ICT and modular assembly (team production)
in the apparel industry. They find that modular production is adopted in
conjunction with new communication systems like EDI to reduce lead times
and inventories.
Empirical research has tested a number of other assertions about which
organizational characteristics are complementary with ICT adoption. For
example, ICT adoption may be more valuable when coordination costs are
high or when real-time coordination is especially valuable. Mendelson and
Pillai (1998) show that firms operating in dynamic business environments
need to process information more rapidly. These firms are more likely to
adopt real-time communication technologies such as pagers and video-
conferencing, and they are more likely to have EDI connections. Dewan
et al. (1998) show that firms with higher coordination costs arising from
diversification or vertical integration have greater ICT investments. Banker
et al. (2003) find that manufacturing plants that employ customer and
supplier participation practices adopt EDI more rapidly. Forman (2005)
shows that firms that are geographically decentralized adopt Internet tech-
nology more rapidly.
As noted by Athey and Stern (2003), there are significant challenges in
testing theories about complementarities solely by looking at investment
behavior. As a result, several researchers have sought to test theories relating
ICT to human resource practices through an ICT-productivity or business-
value approach. Research in this literature generally examines how ICT,
organizational practices, and their interaction influence labor productivity
(e.g., Black and Lynch, 2001; Bresnahan et al., 2002; Hempell, 2003;
Bertschek and Kaiser, 2004) and stock market returns (Brynjolfsson et al.,
2002).10 These papers generally proceed by regressing outcome measures on
ICT, organizational practices, and interactions of these variables. Research
in this area has explored how ICT and labor practices influence outcomes,
and provided evidence that the marginal returns to each increase with the
presence of one of the others.
10
For recent reviews, see Barua and Mukhopadhyay (2000) and Brynjolfsson and Hitt (2000). Co-
rrado et al. (2005) provides an overview of recent work that attempts to measure the value of organ-
izational investments.
Empirical work on ICT adoption and organizational change remains a

fertile area of research. Despite a rich theoretical and case study literature in
this area, the empirical literature remains small. Hypotheses in this area
typically focus on long-run effects. However, data sets that allow for testing
these long-run hypotheses are rare. New research has used ICT adoption
decisions to examine the influence of medium- to long-run organization-
level decisions on short-run ICT adoption decisions. This allows for the
identification of complementarities in the short run. In the long run, these
complementarities may influence organizational decisions and outputs.
Research in this area has proceeded along two independent lines. Large
cross-industry studies such as Bresnahan et al. (2002) have attempted to
identify complementarities between ICT investment and organizational
characteristics that are common to multiple industries. Single-industry
studies such as Hubbard (2000) are able to make very precise statements
about the interactions between incentives and ICT in an industry, but at the
cost of generalizability. Research in this area should continue along these
separate paths to more fully identify the relationships among ICT, incen-
tives, organizational design, and outcomes.
3.2 Adoption and firm boundaries
In the section above, we described a set of papers that analyze the re-
lationship between ICT investment and the location of decision-making
authority within a firm. ICT investment can influence firm organization in
another way, by altering the costs and benefits of locating economic ac-
tivities outside the boundaries of the firm.
One can classify the costs of operations into internal coordination costs,
external coordination costs, and production costs (Gurbaxani and Whang,
1991). Production costs refer to the actual physical costs of producing a
good or service. Internal coordination costs refer to the costs of managing
activities within the boundaries of the firm. External coordination costs
represent the search costs of identifying suppliers, the costs of writing con-
tracts, and potential transaction costs arising from opportunistic behavior
by upstream and downstream partners (Williamson, 1975). All else equal,
market-based exchange should have lower production costs but higher co-
ordination costs (Malone et al., 1987).
Investments in ICT can lower the costs of internal coordination; this will
decrease the costs of managing large enterprises, enterprises that are ge-
ographically dispersed, enterprises that are diversified, and enterprises that
are vertically integrated. However, ICT investments lower the costs of
market transactions, as well, by lowering the communication and coordi-
nation costs required to undertake arm’s-length transactions with external
parties and by lowering the risks of opportunism on the part of trading
partners (Clemons et al., 1993). Malone et al. (1987) argue that the marginal
impact of ICT investment will be greater on external coordination
costs than on internal coordination costs. This leads to the conclusion

that ICT investments facilitate more market transactions and less vertical
integration.
Empirical work that has researched the relationship between ICT invest-
ment and vertical integration across multiple industries has generally sup-
ported the assertion that ICT investment will lead to more market-based
transactions. Research in this area usually proceeds by examining whether
changes in the dollar value of ICT spending are associated with an increase
or decrease in vertical integration. Brynjolfsson et al. (1994) use Bureau of
Economic Analysis investment data to show that greater ICT investments
are associated with a significant decline in average industry firm size. Hitt
(1999) examines these hypotheses empirically at the firm level using a panel
of 549 large firms. He explores how ICT investments influence vertical
integration and internal diversification. He argues that a negative relation-
ship between ICT investment and vertical integration implies that ICT de-
creases external coordination costs, while a positive relationship between
ICT and diversification implies that ICT decreases internal coordination
costs. He shows that ICT investment leads to a significant decline in vertical
integration and a smaller increase in diversification. Moreover, he shows
that increases in vertical integration lead to less ICT investment.11
Other papers have explored this relationship from the other direction: do
firms with greater internal and external coordination costs have greater ICT
investments? Dewan et al. (1998) show that vertical integration is negatively
related to the level of ICT investment, and that ICT therefore has a larger
impact on external coordination costs than on internal coordination costs.
They also demonstrate that diversified firms will have greater levels of ICT
investment. Thus, their results are consistent with the results of Malone et
al. (1987) and Hitt (1999). Kraemer et al. (2005) also find that firms that are
more global are more likely to adopt business-to-business electronic com-
merce. This result may reflect higher coordination costs among global firms,
but it may also reflect the effects of stronger competition or external pres-
sure from trading partners.
While cross-industry studies have consistently found that ICT investment
is associated with less vertical integration, a set of recent single-industry
studies have found less consistency in this relationship. These studies have
generally examined the adoption of industry-specific ICTs, as opposed to
the dollar value of ICT spending. By focusing on a single industry, these
studies are able to precisely examine how ICT investments influence incen-
tives and coordination costs among firms within a particular industry en-
vironment. However, as always, this precision comes at a potential loss of
generalizability.
11
Acemoglu et al. (2004) also examine the role of general technology investment and vertical inte-
gration. They find a negative relationship between intensity of R&D and investment and vertical in-
tegration.
Building on the work of Hubbard (2000), Baker and Hubbard (2003,

2004) examine the role of ICT adoption on the organization of firms in the
trucking industry. They argue that while some margins of ICT investment
primarily lower coordination costs, other margins of ICT investment also
lower monitoring costs, improving agents’ incentives. The effects of ICT on
monitoring and coordination create differing predictions on how ICT will
influence firm boundaries based on the margin of ICT investment and the
characteristics of the firm.
Baker and Hubbard (2003) examine the decision of shippers—purchasers
of trucking services—to own their own trucking fleets or to contract for
carrier services. They show that adoption of certain kinds of ICT leads to
more shipper ownership of trucks (vertical integration) by improving mon-
itoring capabilities (and thereby lowering agency costs associated with
complex job designs). In contrast, they also find that other margins of ICT
investment enhance coordination capabilities that improve the comparative
advantage of for-hire trucks, thereby leading to less shipper ownership of
trucks.
Baker and Hubbard (2004) investigate how ICT investment influences the
decision of truckers to own their own trucks. They show that driver own-
ership of trucks declines with adoption of on-board computers that improve
monitoring capabilities. These on-board computers decrease the agency
costs of trucker ownership by lowering monitoring costs.
Some recent single-industry studies have also examined how vertical in-
tegration influences the short-run decision to adopt new ICTs. In contrast
to studies that look at the impact of ICT investment on firm boundaries,
these papers have consistently found vertical integration and ICT adoption
to be complements. This literature has generally focused on the contracting
costs of making interorganizational ICT investments. For example, less
vertically integrated firms find interorganizational Internet investments less
valuable because downstream partners may appropriate some of the surplus
generated by the new investment. Moreover, ICT investments such as e-
commerce that create new distribution channels may be harder to imple-
ment when retailers and producers are different firms because of potential
channel conflict. In other words, while long-run studies argue that ICT
investments reduce the frictions associated with market transactions, short-
run studies demonstrate that such frictions can have significant effects on
the speed of ICT adoption.
Forman and Gron (2005) investigate how vertical integration influences
adoption of electronic commerce applications in the insurance industry. In
their context, electronic commerce influences the distribution relationship
between insurers and insurance agents. They argue that insurers that are
vertically integrated with their agents will have lower transaction costs of
adopting electronic commerce technologies. Their work demonstrates that
such insurers adopt e-commerce technologies faster than similar insurers
who are not vertically integrated.
Gertner and Stillman (2001) examine how vertical integration influences

the incentives of apparel firms to offer their products online. In their setting,
transaction costs and channel conflict each play a role in increasing the
costs of Internet adoption for apparel firms that do not own the distribution
channel. They provide evidence that vertically integrated apparel retailers
who also own retail outlets such as the Gap started to sell products online
sooner than non-integrated apparel companies like Nautica.12
In all, the current literature on ICT investment and firm boundaries
draws on a rich array of theoretical perspectives—transaction costs eco-
nomics, property rights theory of the firm, principal agent theory—and
arrives at a broad array of conclusions. Although cross-industry research
suggests that, on average, ICT investment may lead to smaller, less ver-
tically integrated firms, single-industry studies suggest there may be sub-
stantial heterogeneity in the effects of ICT investment on firm boundaries
across industries and technologies. More research is needed to identify the
link between ICT investment, agent incentives, and organizational out-
comes. Moreover, prior research has identified some differences in the
short- and long-run relationship between ICT and firm boundaries. To the
extent possible, future research should explore how this relationship
changes as organizations adopt, implement, extend, and maintain new ICT
innovations.
3.3 Adoption and size
Prior research in the diffusion of innovations literature has consistently

shown a positive relationship between organization size and innovativeness
(Rogers, 1995). The most common reasons offered for this relationship are
economies of scale (Kimberly and Evanisko, 1981), slack resources (Eveland
and Tornatzky, 1990), access to outside resources (Attewell, 1992), and
ability to bear adoption risks (Hannan and McDowell, 1984).
Recent work on ICT adoption has continued to show a positive relation-
ship between firm size and adoption. Large-scale descriptive studies on In-
ternet use have shown that even as late as 2000, adoption of basic Internet
technologies such as e-mail varies with establishment size, and that small
establishments rarely adopt complex technologies such as e-commerce (e.g.,
Census, 2002; Charles et al., 2002; Forman et al., 2002).13 Academic research
has also shown a positive relationship between size and ICT adoption (e.g.,
Hubbard, 2000; Kauffman et al., 2000; Astebro, 2002; Banker et al., 2003;
Forman et al., 2006).
12
Helper (1995) and Lane (1991) have also found a positive relationship between relationship ‘‘close-
ness’’ between suppliers and buyers and general (non-IT) technology adoption. These papers focus on
how tighter relationships between upstream and downstream firms increase the stability of demand and
thereby decrease the risks of new technology adoption.
13
A similar relationship has also been found in the adoption of knowledge management practices. See,
for example, Kremp and Mairesse (2004).
Measures of size are typically included as a control in recent research on

firm adoption of ICT. However, the theoretical reasons for why size in-
fluences adoption are not widely understood. Empirical research has been
unable to inform theory because of the difficulty in separately identifying
the various explanations for this phenomenon. This prevents researchers
from making strong statements about why size influences technology adop-
tion. One notable exception is Astebro (2002, 2004), who demonstrates that
faster adoption of computer-aided design (CAD) and computer numerically
controlled (CNC) machine tools among large manufacturing plants is due
to the large non-capital investment costs such as learning that are required
to use these technologies. Future work should seek to further understand
how and why size influences technology adoption.
3.4 Technical infrastructure
Some research has also emphasized how prior IT investments influence

the value of adopting new ICT. Tornatzky and Fleischer (1990) list tech-
nological context as one of the key aspects of a firm’s adoption decision in
their TOE framework.14 Because prior work has found that IT investments
can also capture organizational differences (e.g., Bresnahan and Greenstein,
1996; Forman, 2005), we include technical infrastructure under organiza-
tional characteristics in our framework.
Technical infrastructure can influence the value of new technology adop-
tion because compatibility with new innovations influences the costs of
adoption. Rogers (1995) lists compatibility as one of his five key factors
influencing an individual’s decision to adopt new technology. However,
prior investments in hardware and software can also proxy for an organ-
ization’s overall technology competence (Bharadwaj, 2000) or technological
sophistication (Raymond and Paré, 1992). Technological sophistication re-
flects the number and diversity of information technologies used by organ-
izations, and is a key component of IT sophistication (Raymond and Paré,
1992). Iacovou et al. (1995) note that organizations with high levels of IT
sophistication are less likely to feel intimidated by technology and are more
likely to have access to the technological and managerial resources necessary
to adopt new technologies. Other empirical research has also demonstrated
that IS departments with higher levels of technical competence or more
recent infrastructure investments are more likely to adopt new ICT (e.g.,
Zhu et al., 2003; Forman, 2005; Zhu and Kraemer, 2005). However, Forman
(2005) and Zhu et al. (2006) have shown that if such investments are specific
to prior generations of ICT, they may in fact slow adoption.15
14
The TOE framework identifies three key components that influence an organization’s technology
adoption decision: technological context, organizational context, and environmental context.
15
Prior investments in related technologies can also create switching costs that can influence IT vendor
choice. For more on empirical research in this area, see Chen and Forman (2006), Chen and Hitt (2006),
Forman and Chen (2005), and Greenstein (1993).
3.5 Adoption, assimilation, intra-firm diffusion, and usage
The adoption of new ICT involves four separate decisions that may occur
at separate time periods. One decision is simply whether to adopt ICT at all.
This is the decision most commonly studied in the literature on ICT diffu-
sion, and is often labeled inter-firm diffusion (Battisti and Stoneman, 2003,
2005). The second decision involves which capabilities of an innovation to
use—variously labeled infusion (Cooper and Zmud, 1990), assimilation
(Fichman and Kemerer, 1997, 1999), and depth of adoption (Astebro,
2004). The third decision refers to the rate at which new technology dis-
places old within the organization, and has been labeled intra-firm diffusion
(Battisti and Stoneman, 2003, 2005; Astebro, 2004). The fourth decision is
the individual-level long-term decision of how often to use the technology.16
Recent research has explored the relationship between the first two de-
cisions, i.e., the extent to which an organization may not fully assimilate or
deploy all of the features of an ICT innovation once the organization
adopts it. In most survey-based research, organizations report whether they
have a particular application installed. However, patterns of assimilation
may differ systematically from that of installation if there exist significant
post-investment costs of using new ICT and if these costs are unknown or
uncertain ex ante to potential adopters (Fichman and Kemerer, 1999).
Empirical work examining process innovations such as software develop-
ment tools or CAD/CNC tools has demonstrated that these differences
exist and can be significant (Fichman and Kemerer, 1999; Astebro, 2004).
As a result, some argue that researchers should focus on technology
assimilation rather than adoption (Fichman and Kemerer, 1997, 1999). For
example, Fichman and Kemerer (1997) show how related and unrelated
knowledge and learning economies influence the assimilation of software
process innovations, while Zhu and Kraemer (2005) demonstrate how
technology competence, financial commitment, competitive pressure, and
regulatory support influence e-business use. Because assimilation is a newer
concept, little is known about how the factors influencing assimilation differ
systematically from those influencing adoption. Astebro (2004) shows that
plant size and learning costs influence adoption of CAD/CNC tools more
than they do assimilation. Cooper and Zmud (1990) show that task tech-
nology fit plays an important role in understanding whether an organiza-
tion adopts manufacturing resource planning but is less successful in
explaining assimilation.
The third area of research on ICT diffusion examines the rate with which
new technology displaces old within an organization, termed intra-firm
diffusion. Research in technologies as diverse client/server computing
16
These distinctions are only relevant if assimilation costs, intra-firm diffusion costs, and usage costs
are not perfectly anticipated by the firm. If firms have perfect foresight, then the issues relating to these
factors can be mapped back into the first decision on firm-level adoption.
(Bresnahan and Greenstein, 1996; Ito, 1996), CNC tools (Battisti and
Stoneman, 2003, 2005), electronic mail (Astebro, 1995), and video-
conferencing (Kraut et al., 1998; Tucker, 2005) have shown a significant
lag between initial adoption and widespread diffusion within an organiza-
tion. Depending upon the innovation, the internal spread of new ICT in-
novations may be driven by individual user adoption decisions or by
organization-level capital stock adjustment decisions. This distinction has
important implications for the factors driving intra-firm diffusion, and for
the modeling technology used by the econometrician. User studies in this
area have examined how social networks and network externalities influ-
ence user decisions to adopt new communication technologies (Astebro,
1995; Kraut et al., 1998; Tucker, 2005).17,18 In contrast, studies that ex-
amine firm-level capital adjustment patterns have emphasized adjustment
costs (Bresnahan and Greenstein, 1996; Ito, 1996) as well as the importance
of available complementary technologies and managerial techniques
(Battisti and Stoneman, 2005).
The fourth area of research relates to individual-level long-term usage
within the firm. Behavioral and psychological approaches are particularly
important in this area. In particular, much of this research draws upon the
‘‘technology acceptance model (TAM),’’ based on the theory of reasoned
action from social psychology (Davis, 1989; Davis et al., 1989). The TAM
model predicts that perceived usefulness and perceived ease of use is key to
predicting long-run usage. The idea is that factors that influence behavior,
such as user characteristics and system design, do so indirectly through
attitudes and subjective norms.
The TAM has proven to be a robust model that is frequently employed to
study user acceptance of information technology—as of January 2005, the
Social Science Citation index reported 511 citations for Davis et al. (1989).
It has also inspired several extensions. Szajna (1996) improves on the details
of the model and provides further tests. Venkatesh and Davis (2000) extend
the TAM model to explain perceived usefulness and ease of use in terms of
social influence and cognitive instrumental processes, naming their model
TAM2. Kim and Malhotra (2005) show that belief updating, self-percep-
tion, and habit help explain usage behavior when added to the TAM.
Most research on assimilation, intra-firm diffusion, and usage has been
able to demonstrate that there are significant differences in the factors in-
fluencing adoption versus these various subsequent decisions. Despite the
challenging data requirements, recent gains have been made in this area.
Goldfarb and Prince (2005) show differences in Internet adoption and
usage patterns at the household level. Their work demonstrates that the
17
We discuss the role of social networks in technology adoption further in Section 6.
18
A related literature on media richness theory has explored user choice of different communication
media, and how this choice is shaped by communication needs and the characteristics of the media. See,
for example, Daft and Lengel (1984), Daft et al. (1987), Markus (1994), and Hinds and Kiesler (1995).
demographic characteristics of early adopters are very different from those

of heavy users. Battisti and Stoneman (2003) note, however, that our
knowledge of this process is much less developed than our knowledge
of inter-firm diffusion. As a result, this is an important area for future
research.
4 Geographic differences in adoption
In the previous section, we described how ICT can reshape the nature of
contractual relationships along the value chain. Until recently, a somewhat
less-explored notion is how ICT can alter the geographic dispersion of
economic activity.19 The open question is whether ICT leads to more or less
concentration in economic activity, i.e., whether ICT is a complement or
substitute to urban agglomeration. Research in this literature commonly
examines whether ICT adoption and use is more or less common in cities.
Less commonly, research in this literature has also examined whether ICT
use leads to clustering or dispersion in the location decisions of economic
agents (see Table 3).20
One school of thought argues that ICT reduces the costs of performing
isolated economic activities, particularly in rural settings, even when de-
ployment costs are high. In this view, ICT decreases the costs of coordi-
nating economic activity that occurs across long distances within and
between firms. These distance-related coordination costs may be in addition
to those arising from communication across firm boundaries (Section 3.2).
For example, these costs may arise due to time lags inherent in transporting
physical goods across long distances, or due to the costs of face-to-face
communication among geographically separated individuals.
This hypothesis argues that the gross benefits for ICT adoption will be
decreasing in the size or density of a firm’s location, other things equal
(Cairncross, 1997; Forman et al., 2005). There may be several potential
explanations for this hypothesis. First, while all business establishments
benefit from an increase in capabilities, establishments in rural or small
urban areas derive the most benefit from overcoming the disadvantages
associated with a remote location. Second, establishments in rural areas
lack substitute data communication technologies for lowering communica-
tion costs, such as fixed private lines. Third, advanced tools such as group-
ware, knowledge management, web meetings, and others also may
effectively facilitate collaboration over distances. These alternative expla-
nations all lead to the same empirical prediction: that ICT and urban ag-
glomeration will be substitutes.
19
This research builds upon the seminal work of Griliches (1957), who examined economic factors
shaped the geographic variance in hybrid seed adoption.
20
Kolko (2002) refers to increasing clustering of economic activity as concentration, while shifts in
economic activity away from places where it has traditionally been concentrated as convergence.
Table 3
Summary of research on geographic differences in adoption (Section 4)
Is ICT use a complement Forman et al. (2005), Empirical work has shown
or substitute for urban Gaspar and Glaeser that ICT can be either a
agglomeration? (see (1998), Duranton and complement or substitute.
Section 4.1) Puga (2004), Sinai and Depending on the
Waldfogel (2004), technology, the benefits of
Charlot and Duranton ICT use can either be
(2006), Kolko (2000) increasing or decreasing in
location size. Cities often
contain complementary
resources that can increase
the net benefits of ICT use.
How does ICT use Kolko (2002), Fitoussi ICT use leads industries and
influence the location (2004) firms to become more evenly
decisions of firms? (see distributed geographically.
Section 4.2)
Note: First column describes major questions that explore how geographic location influences ICT
adoption. Second column describes some of the major papers that have addressed these issues. Third
column describes a generalization of the results of these papers.
A second school of thought argues that ICT will lead to increasing con-
centration of economic activity. There are two reasons why ICT may lead to
increases in concentration. First, increases in the size or population density
of a location may increase the marginal benefit to electronic communication
(Gaspar and Glaeser, 1998). This view argues that improvements in elec-
tronic communications will increase the prevalence of face-to-face meetings,
thereby increasing the value of locating in cities.21 Moreover, increases in
location size will increase the availability of complementary products and
services that increase the net benefits of ICT investment. For example, urban
areas may offer (1) availability of complementary information technology
infrastructure, such as broadband services;22 (2) labor-market thickness for
complementary services or specialized skills; and (3) knowledge spillovers
and earlier access to new ideas (Duranton and Puga, 2004).23 Each of these
21
This view is consistent with that of IS researchers who study how different types of communication
media have different levels of information richness (Daft and Lengel, 1984; Daft et al., 1987). Media
such as face-to-face communication, e-mail, and telephone communication differ in terms of feedback
capability, communication channels, utilization, source, and language (Bodensteiner, 1970; Holland et
al., 1976). As a result of these differing capabilities, these media may be used to transmit different kinds
of information.
22
By 1998, almost all but the poorest and most remote geographic areas were serviced by dial-up
internet service providers (Downes and Greenstein, 2002). Yet, broadband access was disproportion-
ately an urban technology (U. S. Department of Agriculture, 2001; Crandall and Alleman, 2002).
23
These are closely related to the three major reasons given for industrial agglomeration (e.g., Mar-
shall, 1890; Krugman, 1991).
concepts leads to the same empirical prediction: ICT use and urban agglo-
meration will be complementary.
4.1 Adoption of ICT across urban and rural areas
Adoption papers in this literature generally proceed by examining differ-

ences in ICT use across urban and rural areas. However, as noted above,
differences in location may affect the value of ICT in several ways. A major
challenge for empirical papers in this literature is to identify how these
competing forces simultaneously shape geographic variation in ICT adoption
and use.
Using a large survey of Internet use among US firms, Forman et al. (2005)
show that, on average, firms in large cities adopted Internet technology
faster than those in small cities or rural areas. However, they demonstrate
that this pattern is due in part to the disproportionate presence of ICT-
intensive firms in large locations. Controlling for industry composition, they
find a very different relationship between location size and ICT adoption.
They show that use of Internet technology for basic purposes like e-mail or
web browsing is more likely in rural areas than in urban areas, other things
equal. This is particularly true for technologies that involve communication
between establishments, which are associated with ending economic isola-
tion. However, use of frontier Internet technologies is more common in
cities, even with industry controls. This is particularly true for Internet
technologies used for within-establishment communication. They argue that
this pattern is consistent with better complementary resources in cities to
help overcome the co-invention costs for complex Internet technology.
Sinai and Waldfogel (2004) examine how location shapes the use of In-
ternet technology among individuals. Like Forman et al. (2005), they find
evidence of both complementarity and substitutability between Internet use
and cities. In particular, they provide evidence of increasing availability of
local online content in large cities. Individuals use the Internet more when
there is more local content; however, controlling for content, individuals in
large markets use the Internet less. On balance, there is no systematic re-
lationship between Internet use and geographic location.
In contrast, examining individual use of electronic communication tech-
nologies among French firms, Charlot and Duranton (2006) find increasing
use of all electronic communication technologies within cities. This is con-
sistent with evidence found in Gaspar and Glaeser (1998).
Research that has examined investment in Internet infrastructure among
suppliers of Internet services has also found a complementary relationship
between city size and ICT investment. Kolko (2000) shows that domain
name registrations are especially prevalent in medium and large cities.
Greenstein (2000) and Downes and Greenstein (2002) show that, in the
mid-1990s, one important barrier to Internet adoption was local availabil-
ity. Internet access was simply not available in isolated areas. By 1998,
however, almost all areas had access, and by 2000 almost all had more than
one provider in their local market. Availability ceased to be a barrier to
adoption. Augereau and Greenstein (2001) find evidence of an urban bias in
the adoption of high-speed enabling Internet technology—56 K modem and
ISDN—among Internet service providers (ISPs) in the 1990s. This bias
likely reflects the impact of larger local market demand and greater local
competition on the incentives for suppliers of Internet services to invest in
new ICT.
4.2 Evidence of how ICT use influences location patterns
A related question asks how the adoption of ICT influences the location
decisions of firms. Given that widespread interorganizational communica-
tions are a relatively new phenomenon, it has been difficult to test this
hypothesis using current data because insufficient time has elapsed to un-
derstand how ICT use influences firm location decisions. However, a small
number of papers have begun to investigate this important question.
Kolko (2002) defines employment convergence as the tendency for an
industry to become more uniformly distributed geographically over time.
He shows that while there exists an overall trend toward employment con-
vergence, ICT-intensive industries exhibit slower convergence than others.
However, he also demonstrates that slower convergence is not due to ICT
usage per se, but rather is because ICT-intensive industries tend to value
more highly the complementary resources found in cities. In particular,
ICT-intensive industries hire more highly educated workers who are dis-
proportionately found in cities. Controlling for labor market effects, he
finds that the direct effect of ICT use is to speed convergence.
Fitoussi (2004) argues that ICT adoption might allow firms to relocate
employees to remote locations. Alternatively, it might make local assets
even more important if ICT and face-to-face communication are comple-
ments. Based on a sample of Fortune 1000 manufacturing firms, he shows
that there is unlikely to be massive relocation due to the advent of the
Internet. He also finds that Internet use does induce cost saving through
reduced communication costs.
4.3 Future research
The relationship between ICT adoption and firm location remains a

fruitful area of research. Although the studies above have provided some
useful findings on this subject, they have only scratched the surface. These
papers demonstrate a variety of different relationships between ICT use and
location size, depending on the use of the technology, its complexity, and
the importance of complementary resources that depend on market scale.
More research is needed on different technologies in different settings in
order to understand exactly what features of a location shape ICT use.
Future work should also investigate how ICT use affects the long-run
location decisions of firms and the agglomeration of economic activity. For
example, inexpensive communications may mean that establishments relo-
cate from high-cost, high-density areas to low-cost, low-density areas. These
remain open questions, however. Further work should compare the location
decisions in industries where interorganizational ICT use is prevalent with
those in other industries. This will help complete the picture of how the
Internet affects geographic variance in productivity and employment.
Furthermore, future work should continue with regard to how pooled
local resources influence ICT investment decisions. These resources have
the potential to significantly alter co-invention costs by providing access to
highly skilled resources that firms may not have internally. For example,
one open question is how the thickness of local labor markets and third-
party service firms influences the ability of firms to overcome co-invention
costs. A further question is when do firms rely on external channels to
overcome co-invention costs and when do they choose to use internal re-
sources. Forman et al. (2006) represent one step toward addressing this
question.
Finally, increasing use of ICT may eventually decrease the costs of using
externally sourced resources such as ICT services firms. This premise lies
behind much of the recent movement to ICT and business-process off-
shoring. Surprisingly, to date there has been no systematic empirical work
that has examined this issue. This represents an important area for future
research.
Research on urban/rural differences in technology usage has important
public policy implications. Rural areas are often the recipients of telecom-
munications subsidies. The argument for subsidizing rural areas relates to
the high fixed costs of building telecommunications infrastructure in low-
density areas. If there are positive externalities to having the entire nation
online, then subsidizing rural areas may make sense. However, the results of
Forman et al. (2003) suggest that firms in rural areas have already adopted
the Internet in large numbers. Further subsidies would simply provide
money to firms that would adopt anyway. Advocates of subsidies need to
provide more compelling evidence that adoption lags in rural areas.
5 Trade-offs between organization and environment
Many adoption decisions depend on the interaction of factors that are

external and internal to the firm. For example, adapting GPTs to the id-
iosyncratic needs of organizations often relies on complementary inputs
that can be obtained from inside the organization or from external sources
(Bresnahan and Greenstein, 1996). Moreover, strategic decisions relating to
rival actions and the optimal timing of investment relate to both internal
and external factors. This section explores all of these issues (see Table 4).
Table 4
Summary of research on trade-offs between organization and environment (Section 5)
Open Question Papers Results
How does co-invention Bresnahan and Greenstein High co-invention costs lead
shape the diffusion of (1996), Forman (2005), to slower adoption of new
ICT? (see Section 5.1) Forman et al. (2005, ICT
2006), Borzekowski
(2004)
Does real options analysis Dos Santos (1991), Kumar Real options analysis can
lead to better decisions for (1996), Benaroch and lead to better timing of
when to adopt ICT? (see Kauffman (1999, 2000), adoption decisions
Section 5.2) Taudes et al. (2000),
Fichman (2004)
Note: First column describes major questions that explore how organization and environment influence
ICT adoption. Second column describes some of the major papers that have addressed these issues.
Third column describes a generalization of the results of these papers.
5.1 Co-invention
Bresnahan and Greenstein (1996) argue that GPTs require complemen-

tary investments (‘‘co-invention’’) to adapt general technologies to the id-
iosyncratic needs of organizations. Such co-invention can involve great time
and expense and may involve both organizational change as well as tech-
nical adaptation of the organization. Because these changes arise within the
boundaries of the organization, co-invention theory draws upon some of
the same ideas on organizational characteristics that were discussed in
Section 3. However, co-invention theory emphasizes the role of third parties
in enabling the transformations necessary for new technology adoption. As
a result, research drawing upon co-invention theory spans both the organ-
izational and environmental perspectives on technology adoption.
In their study of the firm adoption of client/server networking technol-
ogy, Bresnahan and Greenstein (1996) demonstrate that co-invention costs
can be a significant barrier to technology adoption. They further demon-
strate that co-invention costs are highest among high-value users because
ICT is most embedded in the business processes and legacy investments in
such organizations. As a result, high-value users may be slow to adopt new
ICT innovations such as client/server. They use this result to explain the
slow conversion of many data-processing centers from mainframe to client/
server technology. Though Bresnahan and Greenstein (1996) is unique as
an econometric study that shows how co-invention can be used to explain
firm-level ICT adoption, other studies have used co-invention to explain
macrodiffusion patterns. For example, Forman (2005) argues that the rapid
diffusion of Internet technologies such as access to the World Wide Web
and e-mail can be explained by their low co-invention costs. In contrast,
technologies such as consumer electronic commerce or business-to-business

integration require substantial co-invention, and have consequently
diffused more slowly.
A technology that requires significant co-invention to be useful is likely to
face resistance. Therefore, for the technology to diffuse within a firm,
management must be an effective change agent. Rogers (1995, pp. 346–354)
argues that effective change agents have high social status and technological
competence but are otherwise similar to their target group. He maintains
that managerial attempts to drive technology adoption must adhere to these
principles. Co-invention, therefore, implies that ICT adoption and usage
will be particularly driven by effective management.
Co-invention can be accomplished either through innovative activity by
users or by third parties. For example, third parties such as ICT outs-
ourcing firms or ISPs may have economies of scale advantages because of
their ability to spread the fixed costs of innovation across multiple clients
(Ang and Straub, 1998; Greenstein, 2001). Use of third-party resources may
be less costly in cities since such locations may have thicker labor markets
for complementary services or specialized skills such as outsourcing. In such
cases, thicker markets lower the price of obtaining workers to perform
development activities in-house and/or lower the price of obtaining co-
invention services such as contract programming. Forman et al. (2006)
show that, if third-party resources are less costly in cities, adoption of
frontier technology will be faster, other things equal. Wheeler (2005) shows
that computer adoption is increasing in the employment of a firm’s local
industry even controlling for the population of a city, suggesting that these
third-party resources may be industry specific.
Firms may utilize internal or external channels or both when adapting
new ICT. In other words, these channels substitute for one another (For-
man et al., 2006). Forman et al. show that organizations with little ICT
capabilities but that are located in a city are at least as likely to adopt new
ICT as similar organizations with greater capabilities but that are located in
smaller areas. Thus, a variety of papers have demonstrated how co-inven-
tion costs shape the diffusion of new information technology across firms,
industries, and locations.
While these papers have demonstrated the importance of co-invention to
new technology adoption, relatively little work has explored how firms
overcome co-invention barriers. One reason is the stringent data require-
ments required for such analysis. For example, to identify how ICT outs-
ourcing influences co-invention costs, the econometrician faces the daunting
task of identifying two simultaneous discrete decisions: outsourcing and
adoption. In the absence of quality instruments, such analysis requires a
time series of data, preferably in a single-industry setting. One such paper is
Borzekowski (2004), which examines the joint decision to outsource ICT
and adopt Internet technology by credit unions. Borzekowski demonstrates
the importance of controlling for unobserved adopter heterogeneity, which
has a significant impact on the outsourcing decision. He demonstrates that,

controlling for buyer type, the decision to outsource has little effect on the
Internet adoption decision.
Despite the wealth of case study evidence, there is still much research
to be done in this area. With the exception of Bresnahan and Greenstein
(1996), little work has examined how co-invention shapes the diffusion
of other ICT. The co-invention necessary to adopt Internet technology is
likely to differ substantially from that in the transition from mainframe to
client/server. For example, a current challenge for many firms is the elec-
tronic integration of vertical supply chains (e.g., Reddy and Reddy, 2001).
Besides the obvious technical challenges of enabling communication be-
tween heterogeneous software systems, a major challenge to these integra-
tion efforts are incentive problems related to the misuse of this information
by supply-chain partners. Co-invention is also likely to differ across in-
dustries and, as noted above, may be based on the location of the firm and
its external environment. In other words, more work needs to explore how
co-invention shapes the diffusion of ICT across different industries and
locations.
In addition, more empirical work is needed to understand how firms
undertake co-invention. When do firms decide to develop new ICT projects
in-house, and when do they rely on third parties? Do spillovers play an
important role in obtaining new ideas necessary for co-invention? How do a
firm’s co-invention activities evolve after initial adoption?
5.2 Strategic issues in technology adoption
Firms often face strategic trade-offs in the decision of when to adopt a

new technology. A rival firm’s adoption may influence its own adoption.
New strategic opportunities may depend on technology adoption choices.
For example, Debruyne and Reibstein (2005) show that retail investment
brokers adopt e-commerce in response to similar rivals. In particular, all
else equal, a particular broker is much more likely to adopt e-commerce in
the quarter after its closest rival adopts.24 In another exploration of the
strategic value of adoption, Sadowski et al. (2002) examine the Internet
adoption decisions of 264 small and medium Dutch enterprises. They use a
rank (logit) framework to show that adoption is not based on current
strategic factors such as intensity of competition, but on potential strategic
opportunities relating to future communication requirements. Management
adopts the technology in anticipation of future needs.
One recent area of research in the ICT investment and adoption litera-
ture has used real options analysis to understand the optimal timing of
ICT investment decisions. These studies argue that traditional financial
24
In Section 6, we consider additional ways in which one user’s adoption may influence the adoption
decisions of others.
analysis of IT investment projects such as NPV is inadequate for IT

projects with inherently high uncertainty (McGrath, 1997). The value
of such options will depend upon internal organizational and external
factors.
Dos Santos (1991) makes an early case in the IS literature for why NPV
analysis undervalues ICT investment projects and provides a numerical
example of how real options analysis can improve ICT investment deci-
sions. Kumar (1996) shows that, unlike financial options, the value of real
options can go up or down with the variance of second-stage projects.
Benaroch and Kauffman (1999, 2000) provide a case for using real options
analysis to model the timing of ICT investment decisions and show how the
Black–Scholes and binomial pricing models can be used for this purpose.
Further, they apply an approximation of the Black–Scholes model to study
a real investment decision, the Yankee 24 banking network’s decision to
provide point-of-sale (POS) debit card network to member firms. They
further discuss the assumptions needed to use models designed for the
pricing of financial instruments within the setting of an ICT investment
decision, and demonstrate the robustness of real options analysis to changes
in these assumptions. Other research has used Margrabe’s (1978) formula
for valuing the exchange of one asset for another to evaluate IT investment
decisions. Kumar (1996, 1999) has used the Margrabe formula to value
decision support systems (DSS). Taudes et al. (2000) uses the Margrabe
formula to examine the decision to adopt SAP R/3 at a central European
manufacturing firm.
Schwartz and Zozaya-Gorostiza (2003) argue that application of the
Black–Scholes model involves strong assumptions that are sometimes not
consistent with reality. For example, the Black–Scholes model only applies
to European options, while most real options are American options since
managers have discretion over when to exercise the option. Further, Black–
Scholes assumes there is an underlying tradable asset, while most IT in-
vestment projects do not typically involve tradable assets. Schwartz and
Zozaya-Gorostiza develop new models for valuing IT investments. They
develop separate models for when managers wish to value IT projects in
which they develop an IT asset and for when they wish to acquire an IT
asset, depending upon the time it takes to start benefitting from an IT asset
after the initial investment.
Fichman (2004) develops a conceptual model that draws from both the
real options and ICT adoption perspectives. Drawing upon four streams of
research from the innovation literature—technology strategy, organiza-
tional learning, innovation bandwagons, and technology adaptation—he
discusses the factors that lead to increases and decreases in real options
value. One potential area for future research could be to test some of these
assumptions. This would require a broader cross-section analysis than has
generally been used, however, and would be challenging because of the
detailed data required to execute options-pricing analysis.
6 Network effects
The term ‘‘network effect’’ has been used to describe a number of differ-
ent phenomena in the literature. Researchers have used ‘‘network effects’’
to refer to three distinct concepts: direct network externalities, indirect
network externalities, and social network effects (see Table 5).
In the economics literature, a ‘‘network effect’’ is another name for a
positive externality. For this reason, we will refer to this type of network
Table 5
Summary of research on network effects (Section 6)
How do network Farrell and Saloner (1986a, When network externalities

externalities shape 1986b), Katz and are present, the private
technology adoption? Shapiro (1986), Riggins benefits of adoption differ
(theory work) (see et al. (1994), Wang and from the social benefits. This
Section 6.1) Seidmann (1995), Nault can lead to a variety of sub-
and Dexter (1994), Dai optimal equilibrium
and Kauffman (2006), outcomes, including ‘‘excess
others inertia’’ and ‘‘excess
momentum’’
Do network effects shape Goolsbee and Klenow The probability of adopting
the adoption of ICT? (2002), Goldfarb (2006), by a given date is positively
(see Section 6.2) Bell and Song (2004), related to the proportion of
Gurbaxani (1990), firms in the peer group that
Iacovou (1995), Chwelos have already adopted.
et al. (2001), Premkumar However, most research has
and Ramamurthy been unable to identify
(1995), Premkumar et al. whether this empirical
(1994), Hart and pattern is caused by network
Saunders (1997, 1998), externalities, knowledge
Bertschek and Fryges spillovers, or other
(2002), Forman (2005) ‘‘bandwagon effects’’
How do network Saloner and Shepard Larger network size increases
externalities shape the (1995), Kauffman et al. the speed of adoption when
adoption of ICT? (see (2000), Gowrisankaran there exist common
Section 6.3) and Stavins (2004), standards, other things
Augereau et al. (2004). equal. When standards are
not set and can be used to
differentiate, adoption may
exhibit negative network
externalities
Note: First column describes major questions that explore how network effects influence ICT adoption.
Second column describes some of the major papers that have addressed these issues. Third column
describes a generalization of the results of these papers.
effect as a ‘‘network externality.’’ There are two types of positive exter-

nalities: direct and indirect. The simplest example of a technology that
exhibits direct network externalities is the telephone. If only one person
owns a telephone, it has no value. The benefits of using a telephone are
increasing in the number of other users. An indirect network externality
exists when increasing consumption of a good leads to the provision of
complementary goods.25 For example, more people adopt a video game
system when more firms produce content for it, and similarly more firms
produce content for a system when more people adopt it. Another example
of this type of network externality is a local Yellow Pages directory. More
people use it if there are more advertisers, and more people advertise if
there are more users (Rysman, 2004). While software and operating sys-
tems often display such indirect network effects related to complementary
inputs such as compatible software and user skills (e.g., Gandal, 1994;
Shapiro and Varian, 1999), information-processing ICT does not exhibit
direct network effects. This is a key difference between information-
processing IT and ICT.
In sociology and communications, ‘‘network effects’’ usually refer to the
communication of ideas through social ties. The rapid diffusion of Hot-
mail e-mail is an example of social network effects. New customers learned
about the product from friends through e-mail. Rogers (1995) gives a
detailed literature review. These models are driven by the importance of
personal interaction in learning about a new technology. As mentioned in
Section 2.1, epidemic models of diffusion rely on communication of ideas
through social ties, though modern work generally attempts to measure
social network effects using probit models. Manski (1993) discusses the
difficulties of econometrically identifying social network effects separately
from underlying similarities among people in the same communication
network or from positive externalities.
Recent literature on ICT diffusion has focused on identifying some
type of network effect, but has been unable to separately identify social
network effects from network externalities. In this section, we first
give a brief review of the theoretical literature on network externalities. We
then discuss papers that examine network effects that may arise either
from social network effects or from network externalities. These papers do
not separately identify the source of network effects. Our focus in
this section will be on empirical work that identifies any type of network
effect using revealed preference data on adoption decisions; however, we
also mention some papers that use a survey-based approach. Finally,
we describe the small number of papers that have identified positive net-
work externalities in ICT diffusion using data on adoption decisions.
25
Rohlfs (2001) provides numerous case studies of direct and indirect network externalities.
6.1 Theoretical literature on direct and indirect network externalities
A large theoretical literature has developed showing the implications

of network externalities. This literature has shown how network external-
ities can lead to under-adoption of a new technology or adoption of a
technologically inferior product, and has examined at length how network
externalities can influence supplier strategies, including the decision of
whether to produce compatible or incompatible products. This literature is
far too extensive to survey here; for comprehensive reviews, see Farrell and
Klemperer (2004) and Spulber (2006). Instead, we briefly review early results
on how network externalities influence technology adoption, as these are the
results that have been empirically tested most frequently. We also note sev-
eral theory papers in the IS literature that have examined adoption of in-
terorganizational systems (IOS). These latter works provide an interesting
contrast with empirical work on the adoption of IOS, as we shall see below.
When network externalities are present, user adoption influences the
utilities of past and future users of a technology; however, these external-
ities are not internalized. In other words, the private benefits of adoption
differ from the social benefits. This diversion of private benefits from social
ones can engender a number of equilibrium outcomes that are not social
welfare maximizing. For example, it can create ‘‘excess inertia’’ when an
industry is trapped in an inferior standard, and can also create ‘‘excess
momentum’’ when users move to a new standard that ‘‘strands’’ users of the
existing standard (Farrell and Saloner, 1986a,b). Katz and Shapiro (1986)
show that sponsored technologies have a strategic advantage over techno-
logies that are not sponsored and may be adopted even when they are
inferior.26,27
A number of theory papers in the IS literature have looked at the role of
network externalities in adoption of EDI and other IOS. A common theme
is the presence of negative network externalities: increases in network size
decrease the value of adopting EDI. These negative externalities are similar
to what Karshenas and Stoneman (1993) refer to as stock effects.28 Riggins
et al. (1994) show that in buyer-driven networks, buyers may provide a
subsidy in the second period of a two-period game to encourage marginal
suppliers to adopt. These latter-stage subsidies may distort first-period
adoption incentives for suppliers, leading to slower network growth and
26
A sponsor is an entity with property rights to a technology and who may make investments to
promote it.
27
Choi and Thum (1998) extend Katz and Shapiro (1986) by examining how waiting alters their
conclusions. Further, they find that consumers do not appropriately value network effects and adopt too
early. Au and Kauffman (2001) extend Choi and Thum’s (1998) work to explain the adoption of
electronic billing and find that, due to network externalities, agents may adopt a new technology too
early even when the next technology may be superior.
28
Karshenas and Stoneman (1993) show that stock effects play little role in explaining the diffusion of
CNC machine tools in the UK.
lower buyer profits. Wang and Seidmann (1995) also examine the impact of
competitive externalities in the adoption of EDI systems. Like Riggins et al.
(1994), they argue that supplier benefits from adopting EDI will be de-
creasing in the number of other suppliers who adopt. Also, a buyer’s profits
will be increasing in the number of suppliers that adopt; however, the
marginal profit from supplier adoption decreases monotonically with the
number of adopters. They show that buyers will offer a price premium to
suppliers that adopt EDI, while prices offered to non-adopters will fall from
their ex ante levels. This leads to a concentration in the production of
upstream inputs, a result that is consistent with empirical evidence showing
that EDI use tends to decrease the number of suppliers that are used. They
further examine conditions under which buyers may require sellers to adopt
or provide a subsidy to encourage adoption.
Other research has examined how network externalities influence the
adoption of other kinds of IOS. Nault and Dexter (1994) examine how
franchise agreements influence electronic network size and franchise incen-
tives for investment. Parthasarathy and Bhattacherjee (1998) find that the
presence of indirect network externalities can reduce the likelihood that an
adopter of online services will eventually discontinue use of the service.
While traditional theoretical work that has examined the adoption of IOS
has focused on the adoption of proprietary EDI applications, new business-
to-business applications based on Internet protocols create new decisions
for buyers adopting IOS. Buyers can choose to adopt extranet systems that
are based on Internet protocols but which maintain the ‘‘one buyer to many
sellers’’ characteristics of traditional EDI systems. Alternatively, they can
opt to join electronic markets that lower the search costs of identifying low
prices but which may also provide lower incentives for suppliers to make
non-contractible investments. Dai and Kauffman (2006) examine this trade-
off, finding that a buyer’s decision on an e-procurement approach will
depend upon the importance of four factors: (1) lower search and operation
costs; (2) the importance of information sharing between suppliers; (3) the
extent of competition in the supplier market; and (4) the desired levels of
supplier relationship-specific investments.
6.2 Evidence of network effects of any kind in ICT
Empirical research in network effects faces two inherently challenging

identification problems. First, observed statistical correlations in user be-
havior in cross-sectional regressions may be the result of an underlying
relationship between one user’s adoption decision and another’s, or may
simply reflect common unobserved factors that increase the likelihood that
both users adopt. For example, if an econometrician observes that two
firms in the same location adopt ICT, this result may be due to network
effects or to unobserved lower adoption costs in that location. In the
absence of long panels of data with sufficient cross-sectional heterogeneity,

these alternative explanations are not usually separately identified.
Second, even when network effects are themselves identified, it is often
the case that the source of the network effects is not. In this case, the term
‘‘bandwagon effects’’ is often used. Bandwagon effects may be the result of
network externalities, social network effects, or even competitive effects.
Stated succinctly, the bandwagon hypothesis argues that ‘‘the probability of
adoption by a firm at a given date is positively related to the proportion of
firms in the industry that have already adopted’’ (Jensen, 1982).
Goolsbee and Klenow (2002), Bell and Song (2004), and Goldfarb (2006)
examine the impact of network effects on Internet adoption by consumers.
Goolsbee and Klenow (2002) use instrumental variables estimation to
examine the importance of local spillovers such as learning and network
externalities on consumer home PC adoption. They show that these spill-
overs are connected to Internet usage and argue that this provides evidence
of network effects in adoption. Though they are able to demonstrate that
network effects exist, they are unable to show whether they are the result of
direct externalities related to the use of -email or the Internet, or whether
they are related to learning spillovers. In other words, they are unable to
separately identify network externalities from social network effects.
Goldfarb (2006) is also able to identify network effects without separately
identifying these phenomena. He shows that the impact of prior university
attendance on Internet use is much higher for people who attended uni-
versity in the mid-1990s than for others. This is not true of other computing
technologies such as word processing. Universities may have taught stu-
dents to use the Internet, suggesting a social network effect. Alternatively,
network externalities may be the driving factor: students may have an extra
benefit from using the Internet because they know more people online. Bell
and Song (2004) find strong neighborhood effects in the adoption of online
grocery services, but they do not separately identify social network effects
from spurious correlation due to the fact that people with similar prefer-
ences often choose to live in the same neighborhoods.
A number of papers have examined network effects using a survey ap-
proach. Iacovou et al. (1995) develop a model using Rogers’ (1995) diffusion
theories to understand how organizational readiness, external pressure, and
perceived benefits influence EDI adoption and examine the applicability of
this framework using seven case studies. They show that external pressure
from trading partners plays an important role in determining whether an
organization adopts EDI. Chwelos et al. (2001) extend this model and use it
to examine EDI adoption by a group of over 300 Canadian purchasing
managers. They break the effects of external pressure into four constructs:
competitive pressure, dependency on trading partner, trading partner power,
and industry pressure. They find that external pressure plays a significant role
in explaining EDI adoption; in particular, they find competitive pressure and
trading partner power-influenced adopter intentions. This research supported
findings in earlier work on EDI adoption that found a role for competitive
pressure and trading partner power (e.g., Premkumar and Ramamurthy,
1995; Hart and Saunders, 1997, 1998; Premkumar et al., 1994).29
Surprisingly, despite the extensive body of work that has examined how
network effects and competitive pressure has influenced EDI adoption,
relatively little work has examined how such factors influence adoption of
Internet technology by firms. Currently, firms are migrating from tradi-
tional EDI to Internet-enabled supply-chain applications (Saloner and
Spence, 2002). Like EDI, these applications automate the transfer of in-
formation between trading partners, saving on labor costs and decreasing
error rates. However, by integrating with existing enterprise applications,
they also allow for the possibility of real-time knowledge of production and
inventory levels of trading partners. A small number of papers have begun
to investigate the role of network effects on Internet-enabled IOS.
Bertschek and Fryges (2002) find that bandwagon effects play a role in
the adoption of business-to-business electronic commerce among 3,000
German firms, while Forman (2005) finds they influence adoption of ad-
vanced business applications in a sample of over 6,000 US firms.30 Neither
is able to identify between competing explanations for bandwagon effects
nor to control for unobservable differences in preferences across industries
or locations. Zhu et al. (2006) use survey methods to show that network
effects significantly influence firm decisions to adopt Internet-based IOS.
There exists a disconnect between the theory and empirical literatures on
how competitor adoption influences the speed with which organizations
adopt IOS. While theory work emphasizes how negative network external-
ities reduce the incremental benefit of adopting IOS for later adopters,
empirical work focuses on the role of bandwagon effects, generally finding a
positive relationship between competitor adoption and the speed with
which an organization adopts IOS. Researchers that wish to reconcile
these findings will need to identify network effects from unobserved hetero-
geneity.
6.3 Evidence of positive network externalities in ICT
A handful of studies have separately identified network externalities from

social network effects. These studies have examined adoption of enabling
network infrastructure that is subject to strong network externalities but
which is unlikely to be subject to social network effects.
29
For a recent example of how external pressure may influence EDI adoption as viewed through a
sociological lens, see Teo et al. (2003). They show that mimetic pressures, coercive pressures, and
normative pressures all have a significant influence on firm decisions to adopt EDI.
30
Another exception is Lee and Cheung (2004), who find that environmental factors are one of the key
drivers of Internet retailing.
As noted above, network externalities arise when the value for partic-
ipating in the network increases in network size. Saloner and Shepard
(1995) examine how network size increases the speed with which commer-
cial banks adopt proprietary ATM technology during 1971–1979. Using the
number of bank branches as a proxy for network size, they find that banks
with many ATM branches adopt ATM technology earlier than banks with
fewer branches (controlling for the number of depositors), suggesting the
presence of network effects. Kauffman et al. (2000) examine banks’ deci-
sions to join the Yankee 24 electronic banking network. In contrast to
Saloner and Shepard (1995), they examine how the potential size of an
interorganizational banking network influences the decision to adopt a new
ATM technology. They demonstrate that, other things equal, larger po-
tential network size increases the speed of adoption; they also show that the
banks with a larger investment in proprietary network technology adopt
more slowly because of lower net benefits from an interorganizational net-
work.
While the papers above attempt to identify how potential network size
influences the adoption of new ICT through network externalities, a second
class of papers seeks to explicitly measure the externality that arises from
user adoption. Rigorously identifying whether such network externalities
exist is difficult for the reasons described above. As Gowrisankaran and
Stavins (2004) note, time series data are inadequate because price and costs
are decreasing over time, while quality is increasing. Use of cross-sectional
data also presents problems since local differences may be caused by un-
observable heterogeneity in supply or preferences. In general, identification
is only possible when the analyst has long panels with sufficient cross-
sectional heterogeneity. Gowrisankaran and Stavins (2004) utilize a panel
data set to demonstrate that network externalities influence commercial
banks’ adoption of automated clearing house (ACH) electronic payments
systems. In their application, direct network effects are likely the most
important. For ACH to work, bill payers and payees have to coordinate on
a method of payment, and their banks must support that method. They use
three separate reduced-form analyses to demonstrate the presence of net-
work effects: (1) fixed-effects estimation, (2) instrumental variables, and (3)
a quasi-natural experiment using the adoption decisions of small, remote
branches of banks.
Tucker (2005) examines how network externalities shape intra-firm adop-
tion of videoconferencing technology. Her very detailed data enable her to
measure communication patterns within a firm, and to show that network
effects arise from externalities related to communication with other firm
employees, rather than spillovers related to learning. Network externalities
arise only for adoption by employees that communicate with likely adopters
frequently. Moreover, the adoption decisions of employees who are ‘‘in-
formation brokers’’ that communicate directly with many other employees
are most important in shaping others’ adoption decisions. Moreover, she
finds that the influence of network externalities varies across individuals.

Tucker’s research is the first in the economics literature to empirically
identify such heterogeneity in the causes of network externalities and how
they influence adoption behavior.
Network externalities have also played a role in the adoption of enabling
Internet technologies. Augereau et al. (2004) show how network external-
ities can influence users’ decision of which standard to adopt. They examine
the adoption of competing 56 K modem technologies by ISPs in 1997. Prior
to standardization by the International Telecommunication Union (ITU),
these model standards were incompatible. Adoption and choice of 56 K
modem technology by ISPs was influenced by network effects: the value of
adopting a particular model technology was increasing in the number of
consumers with 56 K modems and the particular technology they had
adopted; however, the value of standard adoption was decreasing in the
number of competitors offering the same standard. Network effects in this
market operate at the local level because of flat-rate local pricing of tele-
phone service in the US. Augereau et al. demonstrate that network effects
have a significant impact on consumers’ choice of model technology in this
market. Further, they show that despite these network effects, local ISP
markets failed to standardize, as ISPs chose to differentiate their offerings
rather than conform to a common standard. In other words, while con-
sumer adoption of a 56 K modem standard exhibited positive externalities,
firm adoption exhibited negative externalities.
Research on network externalities has important implications for public
policy. The positive network externalities associated with ICT imply that
the private benefits to adoption may be lower than the overall welfare effect
of an individual adoption decision. Subsidies then can lead to improved
efficiency. Implementation of this idea, however, is difficult. If people who
receive the subsidy would have adopted anyway, then the tax distortions
from the subsidy may outweigh the benefits in overcoming the externality.
Therefore, while there is a theoretical argument for government policy to
subsidize ICT, it is difficult to implement in practice.
7 Internet diffusion across countries
A number of studies have sought to understand cross-country differences

in the rate of Internet diffusion. These studies argue that if Internet use has
a substantial positive impact on the rate of productivity growth, then cross-
country differences in Internet use can exacerbate existing income inequal-
ities between countries. However, some authors have argued that adoption
of ICT such as the Internet can offer the opportunity for late-industrializing
countries to catch up with richer countries (e.g., Kagami and Tsuji, 2001;
Steinmuller, 2001) (see Table 6).
38
Table 6
International diffusion of internet technology (Section 7)
What is the relationship between Comin and Hobjin (2003), Beilock and Dimitrova Per capita income is positively correlated with
per capita income and Internet (2003), Chinn and Fairlie (2004), Kiiski and Internet diffusion
diffusion? Pohjola (2002), Hargittai (1999), Wong (2002),
Dewan et al. (2005)
What is the relationship between Coming and Hobjin (2003), Beilock and Dimirova Internet diffusion has been faster among countries
government institutions and (2003), Oxley and Yeung (2001), Chinn and that are democratic, those with political
C. Forman and A. Goldfarb

Internet diffusion? Fairlie (2004), Hargittai (1999), Kenney (2003), freedom, and that have rule of law.
Wallsten (2003) Telecommunications and regulatory policy also
play a role
What is the relationship between Chinn and Fairlie (2004), Kiiski and Pohjola Internet use will diffuse more quickly to countries
country-level education and (2002), Dewan et al. (2005) with higher education
Internet diffusion?
What is the relationship between Beilock and Dimitrova (2003), Chinn and Fairlie Internet diffusion is faster in countries with better
existing telecommunications (2004), Kiiski and Pohjola (2002), Hargittai telecommunications infrastructure. The
infrastructure and Internet (1999), Dewan et al. (2005) relationship between Internet diffusion and
diffusion? telecommunications prices is less clear
What is the relationship between Hargittai (1999), Kiiski and Pohjola (2002), There is no clear consensus in the literature
English language use and Gandal (2006)
Internet diffusion?
(How) can developing countries Antonelli (2003), James (2003) There is no clear consensus in the literature.
use ICT to catch up to
developed countries?
Note: First column describes major questions that explore the international diffusion of Internet technology. Second column describes some of the major papers
that have addressed these issues. Third column describes a generalization of the results of these papers.
One of the most persistent findings in all of these studies is that Internet
use is correlated with per capita income: Internet technology diffuses
fastest in rich countries (e.g., Hargittai, 1999; Kiiski and Pohjola, 2002;
Wong, 2002; Beilock and Dimitrova, 2003; Comin and Hobijn, 2003;
Chinn and Fairlie, 2004; Dewan et al., 2005). This is true even when one
compares differences in adoption rates within groups of industrialized
countries such as the OECD (Kiiski and Pohjola, 2002) and geo-
graphic regions such as Asia (Wong, 2002). Antonelli (2003) provides
one explanation for this finding, arguing that new technological change
may be ‘‘biased’’ in that it may increase the productivity of inputs with
relatively low factor prices in the country of origin. Antonelli argues
that ICT fits this profile, as it is a skilled labor and capital-intensive tech-
nology, i.e., unskilled labor saving. Thus, he argues it is much better suited
for developed countries such as the US than for developing countries.
Antonelli argues that ICT is complementary with many of the managerial
and organizational changes listed in Section 3.1 and that have been un-
dertaken in the US As a result, he argues that new innovations in ICT
work to increase the competitive advantages of developed nations such as
the US. James (2003) argues that ICT produced with poor countries in
mind can overcome this difficulty. He argues that much of the current
framework for thinking of ICT adoption is inappropriate for developing
countries.
Several studies have also sought to understand how a country’s govern-
mental institutions can encourage or discourage the diffusion of new
ICT such as the Internet. Beilock and Dimitrova (2003) and Comin and
Hobijn (2003) argue that ICT diffuses more quickly to democratic coun-
tries and those with political freedom, while Oxley and Yeung (2001) argue
that rule-of-law plays a major role. Hargittai (1999), Kenney (2003),
Wallsten (2003), and Chinn and Fairlie (2004) show that telecommunica-
tions and regulatory policy are also important in explaining cross-
country diffusion rates. Some studies have argued that Internet use will
diffuse more quickly to countries with higher education (Kiiski and Po-
hjola, 2002; Chinn and Fairlie, 2004; Dewan et al., 2005). Telecommuni-
cations infrastructure also influences the rate of Internet diffusion (Beilock
and Dimitrova, 2003; Chinn and Fairlie, 2004). Kiiski and Pohjola
(2002) argue that Internet access prices are an important determinant of
the rate of Internet diffusion. In contrast, Hargittai (1999) finds they
play little role. Dewan et al. (2005) show that telephone costs influence the
rate of Internet diffusion; this is particularly true at low levels of pene-
tration.
Another factor that may influence Internet use is language. Since most
web sites are in English, English language countries may have an advantage
in Internet use. Several cross-country studies of the Internet have explored
the role of English language use in diffusion (e.g., Hargittai, 1999; Kiiski
and Pohjola, 2002); however, these studies have generally been unable to
uncover a systematic role for language. The role of English in Internet

adoption remains an open question.31
Overall, the empirical evidence supports the view that, in the short run at
least, Internet technology has diffused fastest to industrialized nations
with greater income, more open political institutions, and with more-
developed telecommunications infrastructures. At this aggregate level of
analysis, these studies suggest that there is an international digital divide
between rich and poor country use of the Internet. Some authors argue,
however, that this divide may be narrowing over time (Dewan et al.,
2005).
One challenge faced by researchers studying the cross-country digital
divide is the difficulty in controlling for unobserved cross-country differ-
ences. Countries differ in complex political, cultural, and economic ways
that are impossible to control for in any regression. Developed and deve-
loping countries are particularly different. Hypothesized relationships are
difficult to identify because covariates may pick up other unobserved
differences across countries.
Some researchers have attempted to address the problem of unobserved
country-level differences by comparing countries within similar global re-
gions or demographics. Kiiski and Pohjola (2002) examine OECD coun-
tries, while Kraemer and Dedrick (2002) and Wong (2002) show there exists
significant variation in Internet use among Asian countries.32 While such
studies must still face some residual inter-country differences, they do allow
for easier comparisons among countries than studies that are not so re-
stricted. Other studies have sought to understand how income and other
factors affect the use of Internet technology at different points along the
diffusion path (Dewan et al., 2005). By examining how Internet use is
shaped among Internet-intensive countries and non-Internet-intensive
countries, this research is able to compare Internet use among countries
that are more similar, removing some of the problems of unobserved
heterogeneity.
Another strategy that has been used to understand the factors influencing
country-level Internet investment is to pursue case studies on individual
countries. For example, Brousseau and Chaves (2004) compare the diffu-
sion of Internet-based e-commerce technology in France to that in Den-
mark and Germany. One finding in their paper is that the existence of a
prior e-commerce technology that preceded the Internet influenced the
speed with which French businesses and individuals adopted Internet tech-
nology and the manner in which they used Internet technology. Tachiki
31
Other studies have examined how national cultural differences impact ICT adoption. For example,
Png et al. (2001) demonstrate in a multinational study that high uncertainty avoidance lowered the speed
of frame relay adoption.
32
Kauffman and Techatassanasoontorn (2005) have found evidence of multinational regional con-
tagion in the diffusion of wireless technology.
et al. (2004) study the diffusion of e-commerce in Japan.33 They find that
keiretsu adopted business-to-business technologies quickly but were slower
than small- and medium-sized enterprises at adopting business-to-consumer
technologies.
Finally, many studies have addressed country-level differences by shifting
the unit of observation from the country to the organization or, in some
cases, to the individual. Some studies have examined how factors such as
telecommunications infrastructure, regulatory policy, and education influ-
ence an organization’s decision to adopt ICT in a single-country setting.
Gandal (2006) uses this strategy to examine the role of language on ICT
diffusion. He studies Internet use in Quebec, whose population speaks both
English and French. He finds that Internet use among French-speaking
20–45-year-olds in Canada is significantly lower than that of similar English
speakers. Several US studies have also examined how factors such as edu-
cation and government policy influence Internet adoption. For example,
Goldfarb (2006) examines the role of education on Internet use within the
US, while Goolsbee (2000) shows how high local sales taxes in the US made
online purchasing more appealing.
While single-country econometric studies represent a useful way of ex-
amining the factors influencing Internet penetration, thus far such studies
have used primarily data from developed countries. However, some re-
search has argued that the theoretical models created for industrialized
countries (and the associated empirical results) may not apply to developing
countries because of differences in culture, regulatory climate, and eco-
nomic environment (e.g., Austin, 1990; Rosenzweig, 1994; Xu et al., 2004).
To overcome this difficulty, several studies have examined organization-
level adoption of Internet technology using mixed samples of firms from
developed and developing countries (Zhu et al., 2002; Xu et al., 2004; Zhu
and Kraemer, 2005). These studies confirm the importance of the local
regulatory environment on e-business. Moreover, they also find a role for
organizational factors, such as firm size and scope, technology competence,
and competitive pressure. More studies that include developing country
data are needed.
In summary, there is considerable evidence of variance in Internet use
across countries, suggesting the presence of what some authors have titled
an international digital divide. Though differences in income explain most
of this variance, a variety of other factors have been found to explain the
residual not explained by income differences. Identifying the causes and
consequences of this divide is difficult because of unobserved country-level
heterogeneity and because the frameworks created for understanding diffu-
sion in rich countries may not apply in developing countries. A number of
approaches have been used to overcome these difficulties, including focused
33
For other examples of country-level case studies of e-commerce diffusion, see Kraemer et al. (2002),
Tan and Ouyang (2004), Tigre (2003), Wong and Ho (2004).
empirical approaches that look at similar countries or that look at Internet

use within a particular country, and case studies that are able to explain in
detail how the Internet is used within a country. Considerable progress has
been made, but much more work needs to be done to resolve the conflicting
evidence in the current literature.
8 Conclusion
In this review, we have examined how internal organizational factors and

external factors such as geography and network effects influence business
adoption of ICT. Although research on diffusion has a long and rich his-
tory, the communication capabilities of ICT have led to a number of new
research topics. In particular, the rapidly developing communication capa-
bilities of ICT over the past 20 years have given rise to new ways that ICT
can influence the economic and geographic relationships between firms.
Adoption research provides one lens through which to understand who
benefits most from new capabilities. It examines how factors such as or-
ganizational investments, economic relationships, and geographic location
shape the returns to new ICT adoption. Such studies form a useful com-
plement to other approaches, such as business-value studies, that also ex-
plore the economic effects of ICT.
We have explored the diffusion of ICT from a variety of perspectives;
however, it is apparent that research has only scratched the surface on
many of these topics. As noted above, one reason is that many of these
questions have been driven by new developments in ICT. Tests of a tech-
nology’s ability to reduce the costs of geographic isolation were only rel-
evant with the lower costs of communication engendered by ICT in general,
and the Internet in particular. Moreover, the micro data sets necessary to
test many of these hypotheses have only recently become available: by its
very nature, adoption research requires detailed data on a heterogeneous
pool of economic agents. Our understanding of ICT adoption will continue
to improve as more micro data sets become available.
This review has highlighted a number of issues that require additional
exploration. While there is an understanding that internal factors matter,
there is little work that provides a deep understanding of how co-invention
occurs in ICT adoption. Moreover, additional research is needed to un-
derstand the relationship between ICT investment and the economic or-
ganization of firms. Research in this area would complement existing
business-value work on the same topic. Moreover, additional research is
needed to develop an understanding of the intensive margin of ICT in-
vestment beyond adoption. The core remaining questions about the impact
of location on adoption relate to the effect of Internet use on isolated areas.
Does Internet access substitute for cities? Do we see an increase in tele-
commuting? How does new ICT influence the location decisions of firms?
The core problem with respect to network effects involves separately es-
timating network externalities from social network effects. This is an ex-
cellent topic for future research.
Further research is also needed to understand cross-country differences in
the diffusion of Internet technology: additional empirical and case study
research that examines Internet diffusion within countries other than the
US would be especially helpful in this regard. Finally, there are many op-
portunities to study how government policy influences ICT diffusion. From
India to the United States, numerous government policies have been im-
plemented over the last 10 years that aim to encourage ICT adoption. Solid
empirical work estimating the impact of these policies will provide us with
an important understanding of what works and why.
Acknowledgments
We thank Shane Greenstein and Terry Hendershott for comments and

advice, and Kristina Steffenson McElheran for outstanding research as-
sistance. All opinions and errors are ours alone.
References
Acemoglu, D., P. Aghion, R. Griffith, F. Zilibotti (2004). Vertical integration and technology: theory
and evidence. NBER Working Paper #10997.
Ang, S., D. Straub (1998). Production and transaction economies and IS outsourcing: a study of the
U.S. banking industry. MIS Quarterly 22, 535–552.
Antonelli, C. (2003). The digital divide: understanding the economics of new information and com-
munication technology in the global economy. Information Economics and Policy 15, 173–199.
Astebro, T. (1995). The effect of management and social interaction on the intra-firm diffusion of
electronic mail systems. IEEE Transactions on Engineering Management 42, 319–331.
Astebro, T. (2002). Noncapital investment costs and the adoption of CAD and CNC in U.S. metal-
working industries. RAND Journal of Economics 33, 672–688.
Astebro, T. (2004). Sunk costs and the depth and probability of technology adoption. Journal of
Industrial Economics 52, 381–399.
Athey, S., S. Stern (2002). The impact of information technology on emergency health care outcomes.
RAND Journal of Economics 33, 399–432.
Athey, S., S. Stern (2003). An empirical framework for testing theories about complementarity in
organization design. Working Paper, Northwestern University.
Atrostic, B.K. S.V. Nguyen (2002). Computer networks and U.S. manufacturing plant productivity:
New evidence from the CNUS data. Working Paper #02-01, Center for Economic Studies, U.S.
Census Bureau.
Attewell, P. (1992). Technology diffusion and organizational learning: the case of business computing.
Organization Science 3, 1–19.
Attewell, P., J. Rule (1984). Computing and organizations: what we know and what we don’t know.
Communications of the ACM 28, 1184–1192.
Au, Y. R. Kauffman (2001). Should we wait? Network externalities, compatibility, and electronic billing
adoption. Journal of Management Information Systems 18 47–63.
Augereau, A., S. Greenstein (2001). The need for speed in emerging communications markets: upgrades
to advanced technology at internet service providers. International Journal of Industrial Organization
19, 1085–1102.
Augereau, A., S. Greenstein, M. Rysman (2004). Coordination vs. differentiation in a standards war:
56 K modems. NBER Working Paper # 10334.
Austin, J.E. (1990). Managing in developing countries. Free Press, New York.
Autor, D., F. Levy, R.J. Murname (2002). Upstairs, downstairs: computer-skill complementarity and
computer-labor substitution on two floors of a large bank. Industrial and Labor Relations Review 55,
432–447.
Baily, M.H., R.Z. Lawrence (2001). Do we have a new economy?. American Economic Review 91, 308–
312.
Baker, G.P., T.N. Hubbard (2003). Make versus buy in trucking: asset ownership, job design, and
information. American Economic Review 93, 551–572.
Baker, G.P., T.N. Hubbard (2004). Contractibility and asset ownership: on-board computers and gov-
ernance in U.S. trucking. Quarterly Journal of Economics 119, 1443–1479.
Banker, R., I. Bardhan, H. Chang, S. Lin (2003). Impact of manufacturing practices on adoption of
plant information systems. Proceedings of the 24th International Conference on Information Systems,
pp. 233–245.
Barua, A., C.H. Lee, A.B. Whinston (1995). Incentives and computing systems for team-based organ-
izations. Organization Science 6, 487–504.
Barua, A., C.H. Lee, A.B. Whinston (1996). The calculus of reengineering. Information Systems Re-
search 7, 409–428.
Barua, A., T. Mukhopadhyay (2000). Information technology and business performance: past, present,
and future, in: R.W. Zmud (ed.), Framing the Domains of IT Management, Pinnaflex, Cincinnati,
pp. 65–84.
Barua, A., A.B. Whinston (1998). Decision support for managing organizational design dynamics.
Decision Support Systems 22, 45–58.
Bass, F.M. (1969). A new product growth for model consumer durables. Management Science 15, 215–227.
Battisti, G., A. Canepa, P. Stoneman (2004). The diffusion of e-business activities in the UK: produc-
tivity, profitability, externalities and policy. Working Paper, University of Warwick.
Battisti, G., P. Stoneman (2003). Inter- and intra-firm effects in the diffusion of new process technology.
Research Policy 32, 1641–1655.
Battisti, G., P. Stoneman (2005). The intra-firm diffusion of new process technologies. International
Journal of Industrial Organization 23, 1–22.
Beilock, R., D.V. Dimitrova (2003). An exploratory model of inter-country internet diffusion. Tele-
communications Policy 27, 237–252.
Bell, D.R., S. Song (2004). Neighborhood effects and trial on the internet: evidence from online grocery
retailing. Working Paper, Wharton.
Benaroch, M., R.J. Kauffman (1999). A case for using real options pricing analysis to evaluate in-
formation technology project investments. Information Systems Research 10, 70–86.
Benaroch, M., R.J. Kauffman (2000). Justifying electronic banking network expansion using real op-
tions analysis. MIS Quarterly 24, 197–225.
Bertschek, I., H. Fryges (2002). The adoption of business-to-business e-commerce: empirical evidence
for German companies. Discussion Paper No. 02-05, Centre for European Economic Research.
Bertschek, I., U. Kaiser (2004). Productivity effects of organizational change: econometric evidence.
Management Science 50, 394–404.
Bharadwaj, A. (2000). A resource-based perspective on IT capability and firm performance: an empirical
investigation. MIS Quarterly 24, 169–196.
Black, S.E., L.M. Lynch (2001). How to compete: the impact of workplace practices and information
technology on productivity. Review of Economics and Statistics 83, 434–445.
Bodensteiner. W.D. (1970). Information channel utilization under varying research and development
project conditions: an aspect of inter-oganizational communication channel usages. Unpublished
doctoral dissertation, University of Texas at Austin.
Borzekowski, R. (2004). In through the out door: the role of outsourcing in the adoption of
internet technologies by credit unions. Working Paper, Board of Governors of the Federal Reserve
System.
Bresnahan, T., E. Brynjolfsson, L. Hitt (2002). Information technology, workplace organization, and
the demand for skilled labor: firm-level evidence. Quarterly Journal of Economics 177, 339–376.
Bresnahan, T., S. Greenstein (1996). Technical progress in computing and in the uses of computers.
Brookings Papers on Economic Activity. Microeconomics 1996, 1–78.
Bresnahan, T., M. Trajtenberg (1995). General purpose technologies: ‘engines of growth’?. Journal of
Econometrics 65, 83–108.
Brousseau, E., B. Chaves (2004). Diffusion and impact of e-commerce. The French specific path.
Working Paper, Center for Research on Information Technology and Organizations (CRITO),
University of California, Irvine.
Brynjolfsson, E., L.M. Hitt (1996). Paradox lost? Firm-level evidence on the returns to information
systems spending. Management Science 42, 541–558.
Brynjolfsson, E., L.M. Hitt (2000). Beyond computation: information technology, organizational
transformation, and business performance. Journal of Economic Perspectives 14, 23–48.
Brynjolfsson, E., L.M. Hitt (2003). Computing productivity: firm-level evidence. Review of Economics
and Statistics 85, 793–808.
Brynjolfsson, E., L.M. Hitt, S. Yang (2002). Intangible assets: computers and organizational capital.
Brookings Papers on Economic Activity 2002, 137–199.
Brynjolfsson, E., T.W. Malone, V. Gurbaxani, A. Kambil (1994). Does information technology lead to
smaller firms?. Management Science 40, 1628–1644.
Brynjolfsson, E., A. Renshaw, M. Van Alstyne (1997). The matrix of change: a tool for business process
reengineering. Sloan Management Review 38, 37–54.
Cairncross, F. (1997). The Death of Distance. Harvard University Press, Cambridge.
Census (2002). Detailing tabulations of manufacturing e-business process use in 2000. Survey Release
from United States Department of Commerce. Available at www.census.gov/epcd/www/ebusi-
ness.htm.
Charles, S., Ivis, M., Leduc, A. (2002). Embracing e-business: does size matter? Research Report,
Statistics Canada.
Charlot, S., G. Duranton (2006). Cities and workplace communication: some French evidence. Urban
Studies 43, 1365–1394.
Chen, P., C. Forman (2006). Can vendors influence switching costs and compatibility in an environment
with open standards?. MIS Quarterly 30, 541–562.
Chen, P., L. Hitt (2006). Information technology and switching costs, in: T. Hendershott (ed.), Hand-
book on Economics and Information Systems, Elsevier, Amsterdam.
Chinn, M.D., R. Fairlie (2004). The determinants of the global digital divide: a cross-country analysis of
computer and internet penetration. NBER Working Paper #10686.
Choi, J.P., M. Thum (1998). Market structure and the timing of technology adoption with network
externalities. European Economic Review 42, 225–244.
Chwelos, P., I. Benbasat, A.S. Dexter (2001). Research report: empirical test of an EDI adoption model.
Information Systems Research 12, 304–321.
Clemons, E., S.P. Reddi, M. Row (1993). The impact of information technology on the organization of
production: the ‘‘move to the middle’’ hypothesis. Journal of Management Information Systems 10,
9–35.
Coleman, J., E. Katz, H. Menzel (1957). The diffusion of an innovation among physicians. Sociometry
20, 253–270.
Comin, D., B. Hobijn (2003). Cross-country technology adoption: making the theories face facts. Fed-
eral Reserve Bank of New York Staff Reports No. 169, June.
Cooper, R.B., R.W. Zmud (1990). Information technology implementation research: a technological
diffusion approach. Management Science 36, 123–139.
Corrado, C., J. Haltiwanger, D. Sichel (2005). Measuring Capital in the New Economy. University of
Chicago Press, Chicago.
Crandall, R.W., J.H. Alleman (2002). Broadband: should we regulate high-speed internet access? AEI-
Brookings Joint Center for Regulatory Studies, Washington, DC.
Daft, R.L., R.H. Lengel (1984). Information richness: a new approach to managerial behavior and
organization design. Research in Organization Behavior 6, 191–233.
Daft, R.L., R.H. Lengel, L.K. Trevino (1987). Message equivocality, media selection, and manager
performance: implications for information systems. MIS Quarterly 11, 355–366.
Dai, Q., R.J. Kauffman (2006). To be or not to B2B? An evaluative model for e-procurement channel
adoption. Information Technology and Management 7, 109–130.
Davenport, T. (1998). Putting the enterprise into the enterprise system. Harvard Business Review July–
August, 121–131.
David, P.A. (1969). A contribution to the theory of diffusion. Stanford Center for Research in Eco-
nomic Growth, Memorandum No. 71, Stanford University.
David, P.A. (1990). The dynamo and the computer: an historical perspective on the modern productivity
paradox. American Economic Review 80, 355–361.
Davis, F.D. (1989). Perceived usefulness, perceived ease of use, and user acceptance of information
technology. MIS Quarterly 13, 319–340.
Davis, F.D., R.P. Bagozzi, P.R. Warshaw (1989). User acceptance of computer technology: a com-
parison of two theoretical models. Management Science 35, 982–1003.
Debruyne, M., D.J. Reibstein (2005). Competitor see, competitor do: incumbent entry in new market
niches. Marketing Science 24, 55–66.
Devaraj, S., R. Kohli (2003). Performance impacts of information technology: is actual usage the
missing link?. Management Science 49, 273–289.
Dewan, S., D. Ganley, K.L. Kraemer (2005). Across the digital divide: a cross-country analysis of the
determinants of IT penetration. Journal of the Association of Information Systems 6, article 10.
Dewan, S., S.C. Michael, C. Min (1998). Firm characteristics and investments in information technol-
ogy: scale and scope effects. Information Systems Research 9, 219–232.
Dhar, V., A. Sundararajan (2004). Should the core information systems curriculum be structured
around a fundamental question. Proceedings of the Twenty-Fifth International Conference on Infor-
mation Systems, pp. 1021–1022.
Doms, M., T. Dunne, K. Troske (1997). Workers, wages, and technology. Quarterly Journal of Eco-
nomics 112, 253–290.
Dos Santos, B.L. (1991). Justifying investments in new information technologies. Journal of Manage-
ment Information Systems 7, 71–90.
Downes, T., S. Greenstein (2002). Universal access and local internet markets in the U.S.. Research
Policy 31, 1035–1052.
Dunlop, J.T., D. Weil (1996). Diffusion and performance of modular production in the U.S. Apparel
Industry. Industrial Relations 35, 334–355.
Duranton, G., D. Puga (2004). Micro-foundations of urban increasing returns: theory, in: J.V. Hend-
erson, J.-F. Thisse (eds.), Handbook of Regional and Urban Economics, Vol. IV. Elsevier, Amsterdam,
pp. 2063–2117.
Eveland, J.D., L. Tornatzky (1990). The development of technology, in: L. Tornatzky, M. Fleischer
(eds.), The Processes of Technological Innovation, Lexington Books, Lexington.
Farrell, J., P. Klemperer (2004). Coordination and lock-in: competition with switching costs and net-
work effects. Working Paper, Oxford University. Available at www.paulklemperer.org.
Farrell, J., G. Saloner (1986a). Standardization, compatibility, and innovation. RAND Journal of Eco-
nomics 16, 70–83.
Farrell, J., G. Saloner (1986b). Installed base and compatibility: innovation, product preannounce-
ments, and predation. American Economic Review 76, 940–955.
Fichman, R.G. (1992). Information technology diffusion: a review of empirical research, in: Proceedings
of the Thirteenth International Conference on Information Systems, Dallas, pp. 195–206.
Fichman, R.G. (2000). The diffusion and assimilation of information technology innovations, in: R.W.
Zmud (ed.), Framing the Domains of IT Management, Pinnaflex, Cincinnati, pp. 105–128.
Fichman, R.G. (2004). Real options and it platform adoption: implications for theory and practice.
Fichman, R.G., C.F. Kemerer (1997). The assimilation of software process innovations: an organiza-
tional learning perspective. Management Science 43, 1345–1363.
Fichman, R.G., C.F. Kemerer (1999). The illusory diffusion of innovation: an examination of assim-
ilation gaps. Information Systems Research 10, 255–275.
Fitoussi, D. (2004). So close and yet so far: information technology and the spatial distribution of
customer service. Proceedings of the Twenty-Fourth International Conference on Information Systems,
pp. 642–653.
Forman, C. (2005). The corporate digital divide: determinants of internet adoption. Management Sci-
ence 51, 641–654.
Forman, C., P. Chen (2005). Switching costs and network effects in the market for routers and switches.
Working Paper, Tepper School of Business, Carnegie Mellon University.
Forman, C., A. Goldfarb, S. Greenstein (2002). Digital dispersion: an industrial and geographic census
of commercial internet use. NBER Working Paper #9287.
Forman, C., A. Goldfarb, S. Greenstein (2003). The geographic dispersion of commercial internet use,
in: S. Wildman, L. Cranor (eds.), Rethinking Rights and Regulations: Institutional Responses to
New Communication Technologies, MIT Press, Cambridge, MA, pp. 113–145.
Forman, C., A. Goldfarb, S. Greenstein (2005). How did location affect adoption of the commercial
internet? Global village vs. urban leadership. Journal of Urban Economics 58, 389–420.
Forman, C., A. Goldfarb, S. Greenstein (2006). Understanding the inputs into innovation: do cities
substitute for internal firm resources? Working Paper, Tepper School of Business, Carnegie Mellon
University.
Forman, C., A. Gron (2005). Vertical integration and information technology adoption: a study of
the insurance industry. Proceedings of the 38th Hawaii International Conference on System
Sciences.
Fudenberg, D., J. Tirole (1985). Preemption and rent equalization in the adoption of a new technology.
Review of Economic Studies 52, 383–401.
Gandal, N. (1994). Hedonic price indexes for spreadsheets and an empirical test for network exter-
nalities. RAND Journal of Economics 25, 160–170.
Gandal, N. (2006). The effect of native language on internet use. The International Journal of the
Sociology of Language, forthcoming.
Gaspar, J., E. Glaeser (1998). Information technology and the future of cities. Journal of Urban Eco-
nomics 43, 136–156.
George, J.F., J.L. King (1991). Examining the computing and centralization debate. Communications of
the ACM 34, 62–72.
Gertner, R.H., R.S. Stillman (2001). Vertical integration and internet strategies in the apparel industry.
Journal of Industrial Economics 49, 417–440.
Goldfarb, A. (2006). The (teaching) role of universities in the diffusion of the internet. International
Journal of Industrial Organization 24, 203–225.
Goldfarb, A., J. Prince (2005). Internet adoption and usage patterns are different: implications for the
digital divide. Working Paper, University of Toronto.
Goolsbee, A. (2000). In a world without borders: the impact of taxes on internet commerce. Quarterly
Journal of Economics 115, 561–576.
Goolsbee, A., P. Klenow (2002). Evidence on learning and network externalities in the diffusion of home
computers. Journal of Law and Economics 45, 317–343.
Gowrisankaran, G., J. Stavins (2004). Network externalities and technology adoption: lessons from
electronic payments. RAND Journal of Economics 35, 260–276.
Greenstein, S. (1993). Did installed base give an incumbent any (measurable) advantages in federal
computer procurement?. RAND Journal of Economics 24, 19–39.
Greenstein, S. (2000). Building and delivering the virtual world: the commercial internet access market.
Journal of Industrial Economics 48, 391–411.
Greenstein, S. (2001). Technological mediation and commercial development in the early internet access
market. California Management Review 43, 75–94.
Griliches, Z. (1957). Hybrid corn: an exploration in the economics of technological change. Economet-
rica 25, 501–522.
Gurbaxani, V. (1990). Diffusion in computing networks: the case of BITNET. Communications of the
ACM 33, 65–75.
Gurbaxani, V., S. Whang (1991). The impact of information systems on organizations and markets.
Communications of the ACM 34, 59–73.
Hall, B., B. Khan (2003). Adoption of new technology, in: D.C. Jones (ed.), New Economy Handbook,
Academic Press, Amsterdam.
Hannan, T., J. McDowell (1984). The determinants of technology adoption: the case of the banking
firm. RAND Journal of Economics 15, 328–335.
Hargittai, E. (1999). Weaving the western web: explaining differences in Internet connectivity among
OECD countries. Telecommunications Policy 23, 701–718.
Harris, R. (1998). The internet as a GPT: factor market implications, in: E. Helpman (ed.), General
Purpose Technologies and Economic Growth, MIT Press, Cambridge, pp. 145–166.
Hart, P., C.S. Saunders (1997). Power and trust: critical factors in the adoption and use of electronic
data interchange. Organization Science 8, 23–42.
Hart, P., C.S. Saunders (1998). Emerging electronic partnerships: antecedents and dimensions
of EDI use from the supplier’s perspective. Journal of Management Information Systems 14,
87–111.
Helper, S. (1995). Supplier relations and adoption of new technology: results of survey research in the
U.S. auto industry. NBER Working Paper #5278.
Hempell, T. (2003). Do computers call for training? firm-level evidence on complementarities between ict
and human capital investments. Centre for European Economic Research (ZEW), Discussion Paper
#03-20.
Hinds, P., S. Kiesler (1995). Communication across boundaries: work, structure, and the use of com-
munication technologies in a large organization. Organization Science 6, 373–393.
Hitt, L.M. (1999). Information technology and firm boundaries: evidence from panel data. Information
Systems Research 10, 134–149.
Hitt, L.M., E. Brynjolfsson (1997). Information technology and internal firm organization: an explor-
atory analysis. Journal of Management Information Systems 14, 81–101.
Holland, W.E., B.A. Stead, R.C. Leibrock (1976). Information channel/source selection as a correlate of
technical uncertainty in a research and development organization. IEEE Transactions on Engineering
Management 23, 163–167.
Hubbard, T.N. (2000). The demand for monitoring technologies: the case of trucking. Quarterly Journal
of Economics 116, 533–560.
Hubbard, T.N. (2003). Information, decisions, and productivity: on-board computers and capacity
utilization in trucking. American Economic Review 93, 1328–1353.
Iacovou, C.L., I. Benbasat, A.S. Dexter (1995). Electronic data interchange and small organizations:
adoption and impact of technology. MIS Quarterly 19, 465–485.
Ireland, N., P. Stoneman (1985). Order effects, perfect foresight, and intertemporal price discrimination.
Recherche Economique de Louvain 51, 7–20.
Ito, H. (1996). Essays on investment adjustment costs. Ph.D. Thesis, Stanford University.
James, J. (2003). Bridging the Global Digital Divide. Edward Elgar, Cheltenham UK.
Jensen, R. (1982). Adoption and diffusion of an innovation of uncertain profitability. Journal of Eco-
nomic Theory 27, 182–193.
Jorgenson, D., K. Stiroh (2000). Raising the speed limit: U.S. economic growth in the information age.
Brookings Papers on Economic Activity, pp. 125–211.
Kagami, M., M. Tsuji (2001). The ‘IT’ revolution and developing countries: late-comer advantage.
Institute of Developing Economies, Japan External Trade Organization, Tokyo.
Karshenas, M., P.L. Stoneman (1993). Rank, stock, order, and epidemic effects in the diffusion of new
process technologies: an empirical model. RAND Journal of Economics 24, 503–528.
Katz, M.L., C. Shapiro (1986). Technology adoption in the presence of network externalities. Journal of
Political Economy 94, 822–841.
Kauffman, R.J., J. McAndrews, Y. Wang (2000). Opening the ‘‘Black Box’’ of network externalities in
network adoption. Information Systems Research 11, 61–82.
Kauffman, R.J., A.A. Techatassanasoontorn (2005). Is there a global digital divide for digital wireless
phone technologies? Journal of the Association of Information Systems 6, article 12.
Kemerer, C.F., G.L. Sosa (1991). Systems development risks in strategic information systems. Infor-
mation and Software Technology 33, 212–223.
Kenney, M. (2003). The growth and development of the internet in the United States, in: B. Kogut (ed.),
The Global Internet Economy, MIT Press, Cambridge.
Kiiski, S., M. Pohjola (2002). Cross-country diffusion of the internet. Information Economics and Policy
14, 297–310.
Kim, S.S., N. Malhotra (2005). A longitudinal model of continued is use: an integrative view of four
mechanisms underlying postadoption phenomena. Management Science 51, 741–755.
Kimberly, J.R., M.J. Evanisko (1981). Organizational innovation: the influence of individual, organ-
izational, and contextual factors on hospital adoption of technological and administrative innova-
tions. Academy of Management Journal 24, 689–713.
Kolko, J. (2000). The death of cities? The death of distance? Evidence from the geography of commercial
internet usage, in: I. Vogelsang, B.M. Compaine (eds.), The Internet Upheaval, MIT Press, Cam-
bridge.
Kolko, J. (2002). Silicon mountains, silicon molehills, geographic concentration and convergence of
internet industries in the US. Information Economics and Policy 14, 211–232.
Kraemer, K.L., J. Dedrick (2002). Information technology in Southeast Asia: engine of growth or digital
divide?, in: C.S. Yue, J.J. Lim (eds.), Information Technology in Asia: New Development Paradigms,
Institute of Southeast Asian Studies, Singapore.
Kraemer, K.L., J. Dedrick, D. Dunkle (2002). E-commerce in the United States: Leader or one of the
pack? Working Paper, CRITO, University of California, Irvine.
Kraemer, K.L., J.L.Gibbs, J. Dedrick (2005). Impacts of globalization on e-commerce use and firm
performance: a cross-country investigation. The Information Society 21, 1–18.
Kraut, R.E., R.E. Rice, C. Cool, R.S. Fish (1998). Varieties of social influence: the role of utility and
norms in the success of a new communication medium. Organization Science 9, 437–453.
Kremp, E., J. Mairesse (2004). Knowledge management, innovation and productivity: a firm level
exploration based on French manufacturing CIS3 data. NBER Working Paper #10237.
Krugman, P. (1991). Geography and Trade. MIT Press, Cambridge.
Kumar, R.L. (1996). A note on project risk and option values of investment in information technologies.
Journal of Management Information Systems 13, 187–193.
Kumar, R.L. (1999). Understanding DSS value: an options perspective. Omega 27, 295–304.
Lane, S. (1991). The determinants of investment in new technology. American Economic Review 81, 262–
265.
Leavitt, H.J., T.L. Whisler (1958). Management in the 1980’s. Harvard Business Review November–
December, 41–48.
Lee, M., C. Cheung (2004). Internet retailing adoption by small-to-medium sized enterprises (SMEs): a
multiple-case study. Information Systems Frontiers 6, 385–397.
Levy, F., R. Murname (2004). The New Division of Labor. Princeton University Press, Princeton.
Mahajan, V., E. Muller, F. Bass (1990). New product diffusion models in marketing: a review and
directions for research. Journal of Marketing 54, 1–26.
Mahajan, V., R. Peterson (1985). Models for Innovation Diffusion. Sage, Beverly Hills, CA.
Malone, T., J. Yates, R. Benjamin (1987). Electronic markets and electronic hierarchies. Communica-
tions of the ACM 30, 484–497.
Manski, C.F. (1993). identification of endogenous social effects: the reflection problem. Review of
Economic Studies 60, 531–542.
Margrabe, W. (1978). The value of an option to exchange one asset for another. Journal of Finance 22,
177–186.
Markus, M.L. (1994). Electronic mail as the medium of managerial choice. Organization Science 5, 505–
527.
Marschak, T. (2004). Information technology and the organization of firms. Journal of Economics and
Management Strategy 13, 473–515.
Marschak, T. (2006). Organizational structure, in: T. Hendershott (ed.), Handbook on Economics and
Information Systems, Elsevier, Amsterdam.
Marshall, A. (1890). The Principles of Economics 8th ed. Macmillan, London, 1920.
Mata, F., W. Fuerst, J. Barney (1995). Information technology and sustained competitive advantage: a
resource-based analysis. MIS Quarterly 19, 487–505.
McGrath, R.G. (1997). A real options logic for initiating technology positioning investments. Academy
of Management Review 22, 974–996.
Mendelson, H., R. Pillai (1998). Clockspeed and informational response: evidence from the information
technology industry. Information Systems Research 9, 415–433.
Milgrom, P., J. Roberts (1990). The economics of modern manufacturing: technology, strategy, and
organization. American Economic Review 80, 511–528.
Mukhopadhyay, T., S. Kekre (2002). Strategic and operational benefits of electronic integration in b2b
procurement processes. Management Science 48, 1301–1313.
Mukhopadhyay, T., S. Kekre, S. Kalathur (1995). Business value of information technology: a study of
electronic data interchange. MIS Quarterly 19, 137–156.
Nault, B.R., A.S. Dexter (1994). Adoption, transfers, and incentives in a franchise network with positive
externalities. Marketing Science 13, 412–423.
Oliner, S.D., D.E. Sichel (2000). The resurgence of growth in the late 1990s: is information technology
the story?. Journal of Economic Perspectives 14, 3–22.
Oxley, J., B. Yeung (2001). E-commerce readiness: institutional environment and international com-
petitiveness. Journal of International Business Studies 32, 705–724.
Parthasarathy, M., A. Bhattacherjee (1998). Understanding post-adoption behavior in the context of
online services. Information Systems Research 9, 362–379.
Png, I., B. Tan, K.-L. Wee (2001). Dimensions of national culture and corporate adoption of IT
infrastructure. IEEE Transactions on Engineering Management 48, 36–45.
Premkumar, G., K. Ramamurthy (1995). The role of interorganizational and organizational factors on
the decision mode for adoption of interorganizational systems. Decision Science 26, 303–336.
Premkumar, G., K. Ramamurthy, S. Nilakanta (1994). Implementation of electronic data interchange:
an innovation diffusion perspective. Journal of Management Information Systems 11, 157–186.
Raymond, L., G. Paré (1992). Measurement of information technology sophistication in small man-
ufacturing businesses. Information Resources Management Journal 5, 4–16.
Reddy, R., S. Reddy (2001). Supply Chains to Virtual Integration. McGraw-Hill, New York.
Reinganum, J.F. (1981). Market structure and the diffusion of new technology. Bell Journal of Eco-
nomics 12, 618–624.
Riggins, F.J., C.H. Kriebel, T. Mukhopadhyay (1994). The growth of interorganizational systems in the
presence of network externalities. Management Science 40, 984–998.
Rohlfs, J. (2001). Bandwagon Effects in High-Technology Industries. MIT Press, Cambridge.
Rogers, E. (1995). The Diffusion of Innovations 4th ed. Free Press, New York.
Rosenberg, N., M. Trajtenberg (2004). A general purpose technology at work: The Corliss steam engine
in late nineteenth century United States. Journal of Economic History 64, 61–99.
Rosenzweig, P. (1994). When can management science research be generalized internationally?. Man-
agement Science 40, 28–39.
Ryan, B., N.C. Gross (1943). The diffusion of hybrid seed corn in two Iowa communities. Rural
Sociology 8, 15–24.
Rysman, M., (2004). An empirical analysis of payment card usage. Working Paper, Boston University.
Sadowski, B.M., C. Maitland, J. van Dongen (2002). Strategic use of the internet by small- and medium-
sized companies: an exploratory study. Information Economics and Policy 14, 75–93.
Saloner, G., A. Shepard (1995). Adoption of technologies with network effects: an empirical exam-
ination of the adoption of automated teller machines. RAND Journal of Economics 26, 479–501.
Saloner, G., A.M. Spence (2002). Creating and Capturing Value: Perspectives and Cases on Electronic
Commerce. Wiley, New York.
Schwartz, E., C. Zozaya-Gorostiza (2003). Investment under uncertainty in information technology:
acquisition and development projects. Management Science 49, 57–70.
Scott Morton, M. (1991). The Corporation of the 1990s: Information Technology and Organizational
Transformation. Oxford University Press, New York.
Shapiro, C., H.R. Varian (1999). Information Rules: A Strategic Guide to the Network Economy. Ha-
rvard Business School Press, Boston.
Sinai, T., J. Waldfogel (2004). Geography and the internet: is the internet a substitute or a complement
for cities?. Journal of Urban Economics 56, 1–24.
Spulber, D. (2006). Networks and two-sided markets, in: T. Hendershott (ed.), Handbook on Economics
and Information Systems, Elsevier, Amsterdam.
Srinivasan, K., S. Kekre, T. Mukhopadhyay (1994). Impact of electronic data interchange technology
on JIT shipments. Management Science 40, 1291–1304.
Steinmuller, W.E. (2001). ICTs and the possibilities of leapfrogging by developing countries. Interna-
tional Labour Review 140, 193–210.
Stiroh, K. (2002a). Information technology and the U.S. productivity revival: what do the industry data
say?. American Economic Review 5, 1559–1576.
Stiroh, K. (2002b). Are ICT spillovers driving the new economy?. Review of Income and Wealth 48, 33–57.
Stoneman, P. (2002). The Economics of Technological Diffusion. Blackwell Publishers, Oxford, UK.
Szajna, B. (1996). Empirical evaluation of the revised technology acceptance model. Management Sci-
ence 42, 85–92.
Tachiki, D., S. Hamaya, K. Yukawa (2004). Diffusion and impacts of the internet and e-commerce in
Japan. Working Paper, Center for Research in Information Technology and Organization (CRITO),
University of California Irvine.
Tan, Z., W. Ouyang (2004). Diffusion and impacts of the internet and e-commerce in China, electronic
markets. The International Journal of Commerce and Business Media 14, 25–35.
Taudes, A., M. Feurstein, A. Mild (2000). Options analysis of software platform decisions: a case study.
MIS Quarterly 24, 227–243.
Teo, H.H., K.K. Wei, I. Benbasat (2003). Predicting intention to adopt interorganizational linkages: an
institutional perspective. MIS Quarterly 27, 19–49.
Tigre, P.B. (2003). E-commerce readiness and diffusion: the case of Brazil. Working Paper, CRITO,
University of California, Irvine.
Tornatzky, L.G., M. Fleischer (1990). The Processes of Technological Innovation. Lexington Books,
Lexington.
Tucker, C. (2005). Network effects and the role of influence in technology adoption. Working Paper,
MIT Press.
U.S. Department of Agriculture (2001). Farm computer usage and ownership report, 07.30.01. Avail-
able at http://usda.mannlib.cornell.edu/reports/nassr/other/computer/
Venkatesh, V., F.D. Davis (2000). A theoretical extension of the technology acceptance model: four
longitudinal field studies. Management Science 46, 186–204.
Wallsten, S. (2003). Regulation and internet use in developing countries. Working Paper, AEI and
Brookings Institution, Washington, DC.
Wang, E.T.G., A. Seidmann (1995). Electronic data interchange: competitive externalities and strategic
implementation policies. Management Science 41, 401–418.
Wheeler, C.H., 2005. Technology and industrial agglomeration: evidence from computer usage. Work-
ing Paper #2005-16A, Federal Reserve Bank of St. Louis.
Williamson, O. (1975). Markets and Hierarchies: Analysis and Antitrust Implications. Free Press, New
York.
Wong, P. (2002). ICT production and diffusion in Asia: digital dividends or digital divide?. Information
Economics and Policy 14, 167–187.
Wong, P., Ho (2004). E-commerce in Singapore: impetus and impact of globalization. Working Paper,
CRITO, University of California, Irvine.
Xu, S., K. Zhu, J. Gibbs (2004). Global technology, local adoption: a cross-country investigation of
internet adoption by companies in the United States and China. Electronic Markets 14, 13–24.
Zhu, K., K.L. Kraemer (2005). Post-adoption variations in usage and value of e-business by organ-
izations: cross-country evidence from the retail industry. Information Systems Research 16, 61–84.
Zhu, K., K.L. Kraemer, V. Gurbaxani, S. Xu (2006). Migration to open-standard interorganizational
systems: network effects, switching costs, and path dependency. MIS Quarterly 30, 515–539.
Zhu, K., K.L. Kraemer, S. Xu (2002). A cross-country study of electronic business adoption using the
technology-organization-environment framework. Proceedings of the Twenty-Third International
Conference on Information Systems, pp. 337–348.
Zhu, K., K.L. Kraemer, S. Xu (2003). Electronic business adoption by European firms: a cross-country
assessment of the facilitators and inhibitors. European Journal of Information Systems 12, 251–268.
Chapter 2
Economics of Data Communications
Philipp Afèche
University of Chicago, Chicago, IL, USA
Abstract
This chapter reviews research on data transport pricing decisions for a single
provider who pursues one of two economic performance objectives: welfare
maximization or profit maximization. It largely focuses on studies that con-
sider these issues in the context of mathematical models of data transport
service demand and production. The survey aims to provide a broad road-
map for important questions, modeling approaches and results, and to out-
line open questions. To this end it develops an unifying framework that
classifies data transport service contracts based on the types of their quality
of service (QoS) promises into guaranteed, best effort and flexible bandwidth-
sharing services. It highlights and compares their key features and discusses
their pricing decisions along common dimensions: basic pricing and alloca-
tion principles; incomplete information and adaptive pricing; more differen-
tiation; optimal dynamic pricing; and conclusions and directions. The survey
then examines benefits and challenges of auctions versus posted prices; the
debate on flat-rate versus usage-based pricing; and the merits and challenges
of alternative QoS designs and their delivery mechanisms.
1 Introduction
The last few years have witnessed a revolution in the market for data
communications, marked by the transformation of the Internet from an
academic and research network with restricted access to a commercial
platform used by millions of individuals and organizations to communicate,
access information and conduct business electronically. Key enablers of this
growth have been the development of cost-effective optical network tech-
nologies, the widespread acceptance of Internet protocols as the standard
for transporting any kind of digitized information, and large investments in
fibre infrastructure and startup companies during the e-commerce boom of
53
54 Philipp Afèche
the late 1990s. Over the past decade, the number of Internet service pro-
viders has grown from dozens to thousands and the worldwide Internet
population from 40 million to about one billion people who use an in-
creasingly diverse set of applications ranging from basic email, real-time
audio and video to online auctions and distributed computing. This in-
crease in the number of users and the growth in usage per capita, spurred by
the emerging deployment of broadband access connections, have resulted in
dramatic traffic growth and network capacity expansion. Annual data
traffic and bandwidth growth rates, roughly 1000% in 1995, remain sub-
stantial at close to 100% and Internet traffic recently surpassed voice traffic
in volume, cf. Odlyzko (1999a, 2003).
However, this enormous growth in data traffic volume and variety has also
created substantial problems, most notably the persistence of intermittent
network congestion leading to poor service quality in the form of transmis-
sion delays and data losses. The ensuing debate has centered around several
problem causes and challenges. Flat connection fees that only depend on the
access capacity but not on actual usage are prevalent in the consumer and
parts of the interconnection market, contributing little to discourage overuse
at peak times. The Internet was designed to offer undifferentiated ‘‘best
effort’’ service to all applications; it is ill-equipped for tailoring service quality
attributes and guarantees to specific applications or users, making it difficult
to offer preferential treatment to those who value and are willing to pay for
it. This absence of strict economic and technological usage controls has led
many to view Internet congestion as a classical ‘‘tragedy of the commons.’’
Capacity levels may be too low in certain network segments and excessive in
others, and limitations of currently deployed technologies constrain the abil-
ity of providers to offer premium services, leaving them to compete in a
commodity market which hurts their revenues.
These challenges have raised many questions and sparked much research
on how to better manage data communication services and the infrastructure
to deliver them. Can congestion simply be eliminated by deploying ample
capacity to simultaneously accommodate all users? Or, is there a need for
better control mechanisms to regulate usage and allocate scarce capacity
among competing traffic flows? How should differentiated services be de-
signed and priced to increase customer value and provider profits? What are
the relative strengths and weaknesses of flat and usage-based pricing
schemes? How should price and resource allocation mechanisms be designed
to account for the decentralized nature of the Internet whose performance
depends on the decisions of many independent self-interested parties with
conflicting incentives and private information? This chapter surveys research
on these questions and offers an outlook on future directions.
Scope. The economics of data communications is a vast subject area.
This survey must therefore be selective. It makes no attempt to discuss all
the important topics, models and questions that might have been included
in a chapter with this title. Its scope is defined as follows.
Ch. 2. Economics of Data Communications 55
The operation of a data communications network involves three basic

functions: transport or communication of data among interconnected net-
work nodes, storage and processing at these nodes. This survey focuses on
data transport services at the exclusion of storage and processing services
such as those offered by application service providers. Although the Inter-
net accounts for much of the growth in data communications, the discus-
sion is only in part specific to the Internet. For the most part it emphasizes
economic principles that are valid for a range of data network technologies.
The survey is very much focused on the economics of traffic sources and
resource allocation mechanisms that are particular to modern data com-
munication networks; it is not concerned with voice communications over
traditional telephone networks. While this chapter focuses on the econom-
ics, not the technology, of data transport services, it does outline and
compare the resource allocation principles that are of particular importance
for their economic analysis, but it ignores implementation details.
From an economics perspective managing data transport services can be
viewed as a complex resource allocation problem that involves a range of
decentralized decisions by independent and self-interested agents. Providers
decide how to design, price and produce these services taking into account
the demand characteristics that drive customers’ purchase and usage de-
cisions. This chapter surveys various aspects of this resource allocation
problem, focusing on various dimensions of usage-based pricing decisions
for a single provider who pursues one of two economic performance objec-
tives: welfare maximization or profit maximization.
Research on data transport service pricing draws on tools and contri-
butions from several disciplines, including economics, computer science and
electrical engineering, information systems, management science/operations
research, and applied probability. These efforts are also characterized by a
striking variety in terms of their objectives (optimization of a given system
or design of new pricing architectures), methodologies (mathematical
modeling and analysis, simulation or empirical study) and abstraction levels
(stylized models of generic systems or detailed representations of specific
systems). This survey largely focuses on research that studies pricing issues
in the context of mathematical models of data transport service demand
and production. To a lesser extent, the survey also outlines implementation
issues and architectural proposals for pricing and service design. The survey
does not discuss, but references a few empirical analyses of data transport
service demand and pricing.
Goals and approach. This chapter aims to provide a broad roadmap
rather than an in-depth tutorial. The main goals are to identify and com-
pare important questions and modeling approaches, existing answers and
proposals, and outline open questions. To this end the survey emphasizes
the development of an unifying framework that classifies economic analyses
of data transport services based on the types of QoS guarantees they make
and the pricing decisions they consider. It uses this framework to discuss and
56 Philipp Afèche
compare the partly disjointed research streams through a common lens. The
discussion is mainly conceptual in nature, but mathematical notation is
used selectively where doing so may help clarify the key ideas.
Organization. The chapter is organized as follows. Section 2 defines the
problem and develops the discussion framework. It categorizes data transport
service contracts based on the types of QoS guarantees they make into guar-
anteed, best effort and flexible bandwidth-sharing services, and characterizes
the salient features of service demand and production that shape their anal-
ysis. It then describes major dimensions of pricing decisions, outlines for each
dimension the relevant issues for this survey, and outlines the framework for
their discussion. The core Sections 3–5 focus on one of the three service types,
but follow the same organization outlined in the framework: features, mode-
ling and overview; basic pricing and allocation principles; incomplete infor-
mation and adaptive pricing; more differentiation; optimal dynamic pricing;
and conclusions and directions. The discussion in each of the core sections
focuses on the paradigm of economic optimality. Section 6 goes beyond the
optimality criterion and considers implementation issues in pricing and service
design, including the benefits and challenges of auctions versus posted prices;
the debate on flat-rate versus usage-based pricing; and the merits and chal-
lenges of alternative QoS designs and their delivery mechanisms. Section 7
concludes and outlines some overall research directions.
2 Problem definition and discussion framework
From an economics perspective managing data transport services can be

viewed as a complex resource allocation problem that involves a range of
decentralized decisions by independent and self-interested agents. Providers
decide how to design, price and produce these services taking into account the
demand characteristics that drive customers’ purchase and usage decisions.
This chapter surveys various aspects of this resource allocation problem,
focusing on various dimensions of usage-based pricing decisions for a single
provider who pursues one of the economic performance objectives: welfare
maximization or profit maximization. Welfare maximization or economic
efficiency, also referred to as net value maximization or social optimization,
serves as the key benchmark for any allocation since it captures the sum of
customer surplus and provider surplus. It is also the relevant optimization
criterion for the provider of an internal data network who should maximize
the overall value to the entire organization rather than her own profit.
Profit maximization is the key optimization criterion for a commercial
provider with external customers.
Users or customers—this survey uses the terms interchangeably—gener-
ate demand for data transport services through a variety of applications
such as email, file downloads, distributed computing and e-commerce
transactions. These applications are quite heterogenous in terms of their
traffic source properties and quality of service requirements, which specify

data transfer performance targets or preferences. Data transport services
are sold in the context of service contracts that specify three sets of at-
tributes: QoS parameters or metrics that the provider guarantees, con-
straints on traffic parameters that the user must meet, and a tariff that
determines the user’s charge. The provider ‘‘produces’’ services with the
desired attributes by means of a communication network. It can be viewed
as a service factory consisting of a resource network with the capacity and
connectivity to carry traffic and control mechanisms to manage traffic and
resource allocation.
To organize the discussion, this section develops a simple framework that
classifies economic analyses of data transport service contracts based on the
types of QoS guarantees they make and the pricing decisions they consider.
Sections 2.1–2.3 lay the groundwork for the framework. They describe three
types of service contracts, guaranteed, best effort and flexible bandwidth-
sharing services, and characterize the salient features of service demand and
production that shape their analysis. Along the way they introduce key
concepts and terminology. Section 2.4 highlights differences between mod-
ern data communications and traditional telephone networks. Section 2.5
discusses the role of pricing, describes major dimensions of pricing decisions
and outlines, for each dimension, the relevant issues for this survey. Section
2.6 outlines the organizing framework for discussing these decisions.
2.1 Service demand
Users and their applications generate traffic sources that drive the de-
mand for data transport services. Key demand characteristics are the prop-
erties of traffic sources, their QoS requirements, and user preferences over
the quantity, quality and time of service.
Traffic sources
The properties of traffic sources determine their bandwidth demand and
the relationship between the traffic mix a network can carry and the QoS it
can deliver. As detailed below the data flow generated by a traffic source is
typically broken into packets or cells. Based on their data flow pattern one
can broadly distinguish between bulk transfer, constant bit rate (CBR) and
variable bit rate (VBR) sources. Bulk transfers are triggered by web file
downloads, network print jobs or email transmissions. Each generates a
data ‘‘burst’’ at a single point in time which is fully characterized by its data
volume. CBR sources are generated by certain voice and video applications;
each is characterized by its duration or holding time and constant data rate
which equals its bandwidth requirement. CBR source models are also rep-
resentative of leased virtual circuits or dialup access connections that need a
fixed amount of capacity over a given time horizon. VBR sources are gen-
erated by interactive voice, audio and video applications. They exhibit a
58 Philipp Afèche
range of traffic patterns and are characterized by their duration, mean and
peak data rate, and measures of their ‘‘burstiness’’ or data rate fluctuation.
See Kelly (1996) and Kumar et al. (2004) for specific source models.
QoS requirements
Customers and their applications have QoS requirements that are expressed
in terms of any combination of three key performance metrics: throughput
(bandwidth or data rate), delay (or latency), and loss. These QoS requirements
may depend both on the technical features of the hardware and software
involved and on user preferences. For example, order entry processing trans-
actions may be time-sensitive due to technologically imposed time-outs but
also because their users require low delays to maintain their productivity.
Based on the nature of their QoS requirements one can roughly distin-
guish between elastic applications that have flexible QoS preferences and
real-time applications that require relatively strict QoS bounds, cf. Shenker
(1995). Elastic applications are inherently (technologically) tolerant of de-
lay, delay variation (jitter) and throughput limitations: quality degradation
due to network congestion hurts performance but does not compromise the
integrity of the transmitted data. When transmissions slow down they keep
working, just not as quickly as usual—performance ‘‘degrades gracefully.’’
Examples include bulk transfers such as email, web file downloads or
streaming audio that can adapt its compression and data rate to available
resources. Depending on the desired level of interactivity, elastic applica-
tions may differ in their delay tolerance: transactions initiated during an
interactive remote login session are more delay-sensitive than interactive
bulk file transfers, which in turn are more sensitive than asynchronous
(non-interactive) bulk file transfers such as electronic mail. Delay tolerant
text-based transmissions must be loss- and error-free whereas audio or
video transmissions may tolerate some data loss.
Real-time applications require relatively strict QoS bounds such as mini-
mum throughput, maximum delay, or maximum packet loss ratio. Violating
these bounds compromises the interactive nature of these applications and
may distort their data signals to the point where they become incomprehen-
sible. Examples include interactive CBR or VBR voice and video applications.
User preferences
User preferences are commonly modeled through utility functions that
measure their willingness to pay as a function of service attributes such as the
quantity, quality and time of service. In making their service purchase and
usage decisions users act as self-interested agents who maximize their ex-
pected utility minus their cost of service for a given transmission or over
some time horizon. This utility-maximizing behavior determines inverse de-
mand functions that map service attributes to prices. An alternative modeling
approach is to specify user utility implicitly via demand functions that specify
demand depending on the quality, time and price of service. This survey uses
utility functions to illustrate model features and certain results.
The heterogenous nature of data communications demand has given rise
to a range of demand models. This survey does not discuss empirically
derived demand models. See Edell and Varaiya (1999), Altmann et al.
(2000), Varian (2002) and Beckert (2005) for empirical analyses of demand
for Internet-based data transport services. Utility functions may have any
combination of the following attributes and features. Modeling details are
discussed later in conjunction with the respective studies.
Throughput—quality or quantity: Throughput is a measure of quantity
per unit time. Depending on the context, it measures the throughput of data
units belonging to an individual flow (e.g., bits, packets or cells of a file
downloaded per second) or that of entire individual flows (e.g., files trans-
mitted per second). In the former case throughput may be more naturally
interpreted as a quality attribute, in the latter case it may equally well be
understood to measure service quantity per unit time.
Delay or loss quality: These QoS metrics are typical utility function at-
tributes in settings with elastic applications; they capture the users’ delay or
loss sensitivity.
Congestion externalities and network load: An important distinction is
between models with and without (negative) congestion externalities. In
models with congestion externalities, appropriate for services without QoS
guarantees, a user’s utility depends not only on her own service quantity but
also on (all) other users’ consumption, captured by some measure of net-
work load, since more traffic overall leads to congestion and reduces the
quality of transmissions. This chapter does not consider positive network
externalities where user utility increases in the number of potential con-
nection partners. Rohlfs (1974) and Oren and Smith (1981) study pricing
for communication services with network externalities. Westland (1992)
and Giridharan and Mendelson (1994) study settings with network and
congestion externalities.
Time of service: In models that capture non-stationary demand, utility
functions also depend on when a particular transmission occurs. The vast
majority of studies consider stationary demand.
Market segmentation: Models of heterogenous user populations typi-
cally segment users into classes or types based on their traffic source prop-
erties and/or utility function attributes.
2.2 Service contracts: QoS guarantees and service designs
As noted above data transport services are bought and sold via contracts
between customers and providers. A contract specifies three sets of service
attributes: QoS parameters or metrics that the provider guarantees, con-
straints on user traffic parameters, and a tariff that determines the user’s
60 Philipp Afèche
charge. This survey uses the term ‘‘contract’’ broadly to refer to any agree-
ment between a user and the network that specifies the terms for performing
a transmission. A contract may apply to services at various levels of re-
source and time aggregation, e.g., from sending a single file at a particular
point in time to transmitting multiple files over a fixed time horizon.
The heterogenous nature of traffic sources and QoS requirements calls for
a variety of data transport services, each tailored to suit a particular kind of
application. This survey categorizes data transport service contracts into
those for guaranteed, best effort and flexible bandwidth-sharing services,
based on the type of QoS guarantees they offer and the traffic constraints
they impose. It is important to distinguish between the contract that defines
the attributes of a service and the resource network and control mechanisms
used to produce it. This section focuses on the QoS and traffic attributes of
these contract types. Section 2.3 describes how these services are produced.
The discussion of tariff and pricing issues starts in Section 2.5.
Guaranteed services
Guaranteed services specify QoS parameters the network promises to de-
liver provided the user’s traffic profile conforms to certain traffic parameters.
QoS parameters specify the maximum tolerable congestion level in terms of
deterministic or probabilistic traffic flow performance bounds such as min-
imum bandwidth, maximum delay, jitter or packet loss rate. Common traffic
parameter constraints concern the peak rate, mean rate and burstiness. QoS
guarantees can be fixed (static) or flexible (dynamic) over the contract lifetime.
In this survey the term ‘‘guaranteed services’’ refers exclusively to contracts
with fixed guarantees. They are discussed separately (Section 3) from ‘‘flexible
bandwidth-sharing services’’ (Section 5) that are introduced below. The two
rely on different control mechanisms and pricing paradigms. Fixed guarantees
are best suited for real-time applications with strict QoS requirements. Ex-
amples of such guaranteed services include CBR and real-time VBR in the
Asynchronous Transfer Mode (ATM) framework and the Guaranteed Qual-
ity services in the Integrated Services (IntServ) architecture proposal of the
Internet Engineering Taskforce (IETF).
Flexible bandwidth-sharing services

Flexible bandwidth-sharing services offer dynamic bandwidth guarantees
and give users some control over and the ability to adapt to their allocation
while their transmissions are in progress. At each point during a transmis-
sion the contract guarantees a certain peak data rate to the user, in response
to changing demand and supply conditions, which represents the user’s
flexible QoS guarantee and traffic parameter constraint. Flexible band-
width-sharing services are best suited for elastic applications whose per-
formance easily adapts to data rate fluctuations. Examples include the
Internet’s Transmission Control Protocol (TCP) which supports email and

the World Wide Web, and the Available Bit Rate (ABR) service in the
ATM framework.
Best effort services

Best effort services make no explicit quality guarantees. The network
promises to provide the best quality it can to each customer, attempting to
deliver data as quickly and error-free as possible without committing to any
quantitative bounds. In turn, best effort services do not constrain users’
traffic inflow other than through the capacity of their access link. Users do
not request permission before transmitting and may do so at their max-
imum rates. They may experience delays or losses during congestion periods
since performance is determined not only by the network itself but also by
other users’ offered load. While the network does not guarantee QoS for
individual flows, the notion is that users estimate or receive information on
aggregate performance statistics such as the delay distribution or packet
loss ratio across all flows. Best effort services are suited for elastic appli-
cations that are somewhat tolerant of delay and loss or if the capacity is so
abundant that performance degradation is unlikely to reach critical levels.
Examples are the Internet’s Internet Protocol (IP) and User Datagram
Protocol (UDP), the Unspecified Bit Rate (UBR) service in the ATM
framework and the Best Effort service in the IntServ Architecture proposal.
Loss-sensitive applications that do use best effort service, e.g., streaming
media that use UDP, may also provide their own mechanisms to adjust
their sending rates and avoid data loss.
Remarks: This survey introduces the term ‘‘flexible bandwidth-sharing
services’’ to highlight the most striking features of TCP-like services from
the users’ perspective. It adopts a literal definition of best effort services and
distinguishes them from ‘‘flexible bandwidth-sharing services,’’ even though
the Internet’s TCP transport service is often considered part of the ‘‘best
effort’’ Internet. As understood here, best effort services offer no quality
guarantees whatsoever and impose no constraints on the traffic that users
may send into the network, irrespective of congestion conditions. Unlike
TCP, they have no built-in traffic flow control and feedback mechanisms
for avoiding congestion. Thus, users have little information on precisely
when and how to adjust their sending rates to match fluctuations in band-
width availability. The Internet transport service provided by the UDP
protocol is a typical example: packets may be delayed, lost or arrive out of
order without notice. By contrast, the Internet transport service provided
by the TCP protocol guarantees zero packet loss by retransmitting lost
packets and dynamically controls the peak rates at which competing con-
nections are allowed to transmit via adaptive flow control and feedback
mechanisms. Its distinctive features are closed-loop control, on fast time
scales and at the level of individual flows. The distinction between best effort
62 Philipp Afèche
and flexible bandwidth-sharing services made here is useful, since the re-
spective pricing and resource allocation decisions occur on different time
scales that call for somewhat different analytical frameworks and perspec-
tives. In the case of flexible bandwidth-sharing services the fast feedback
loops afford users (or agents on their behalf) some control over their trans-
missions while they are in progress. In the case of best effort services the
notion is that feedback loops between sending rates and congestion levels
are relatively slow—so slow that users have no control over a given indi-
vidual flow through the network once it is in progress. Thus, users make all
decisions concerning an individual flow or a collection of such flows cov-
ered by their contract at the start of the respective transmissions, based on
the contract parameters and performance information available at that
time. In this sense best effort service contracts only have static parameters.
Since the network does not make any strict QoS guarantees per user, the
notion is that users, prior to sending data, estimate or receive information
on aggregate QoS statistics, e.g., the delay distribution or packet loss ratio
over all flows. One can think of these estimates as the expected QoS metrics
of a given best effort service.
Service designs
In this chapter the term ‘‘service design’’ refers to the key features that
characterize a set of service classes offered by a provider. The survey
broadly distinguishes among service designs based on three dimensions:
1. Number of service classes. Under uniform service all traffic sources—
whether homogenous or heterogenous—are given identical or no QoS
guarantee(s) and subjected to the same traffic constraints. The ca-
nonical example is a single class of best effort service that offers no
QoS guarantees and imposes no constraints on traffic inflows. Differ-
entiated services differ in at least one QoS metric or traffic constraint.
For example, multiple classes of guaranteed service designed to sup-
port heterogenous traffic sources may share the same QoS guarantee,
e.g., the packet loss probability, but differ in their traffic parameter
constraints, e.g., the peak data rate. Services may also be differentiated
in terms of the type of QoS guarantee with respect to a single QoS
metric such as delay. In multi-link networks, which have received less
attention than the single link case, otherwise identical service classes
may also differ in their source-destination or route.
2. Type of QoS guarantee per service class. Each class may offer guar-
anteed, best effort, or flexible bandwidth-sharing service. The majority
of papers consider service designs with a single type of QoS guarantee
for all classes.
3. Metrics per service class. Based on the metrics of the different classes
this survey distinguishes between service designs with bandwidth
differentiated services, which only differ based on bandwidth (or
throughput) related QoS metrics or traffic parameter constraints but

not based on delay or loss, and delay or loss differentiated services
.
2.3 Service production
A communication network can be viewed as a service factory that pro-

duces and delivers services with the desired attributes. It is a collection of
interconnected resources and control mechanisms. Resources provide the
capacity and connectivity to carry traffic, and control mechanisms manage
traffic demand and resource allocation. There are many valuable references
on the technology, operation, analysis and control of data communication
networks. A seminal work on communication networks is Kleinrock (1964).
Recent textbooks include Walrand and Varaiya (2000), Tanenbaum (2003)
and Kumar et al. (2004). Courcoubetis and Weber (2003) provide a con-
ceptual overview of network services and technology. Gevros et al. (2001)
discuss congestion control mechanisms for best effort services.
This section describes the structural and functional characteristics of data
communication networks. It focuses on the conceptual principles that are of
particular importance for their economic analysis and ignores implemen-
tation details. It first outlines the main features of the resource network, its
cost structure and basic operating principles. It then describes the network
control mechanisms that are instrumental in delivering services with certain
QoS metrics. The section then compares guaranteed, best effort and flexible
bandwidth-sharing services in terms of their reliance on these control
mechanisms for service delivery.
Resource network
Network resources are grouped into nodes, computers where traffic orig-
inates and terminates and switches or routers that forward traffic, and
connecting links. Nodes and links have buffers in which data packets queue
in the presence of congestion. Network design and investment decisions
determine the numbers, types and capacities of these resources and their
interconnections. The performance of a network crucially depends on its
capacity. The processing capacity or bandwidth of a resource measures its
maximum throughput or transmission rate per unit time, in packets, cells or
bits per second (bps). The capacity along a network route is determined by
the bottleneck, the resource with the smallest capacity. The buffer size or
storage capacity is an important factor in determining at what point con-
gestion translates to data loss as opposed to queueing delays. Systems
without buffers are referred to as loss systems, those with buffers as queue-
ing or queueing-loss systems. The network topology describes the pattern of
links connecting the nodes. Large networks such as the Internet consist of
multiple interconnected networks at different hierarchical levels: end-users
are connected via access links to Internet Service Providers (ISPs) which in
turn interconnect through bilateral and multilateral agreements.
64 Philipp Afèche
Network cost structure

The network cost has several components: the cost for setting up, op-
erating and maintaining the network resources and technologies, the cost of
capacity expansion and the connection cost for access lines and customer
premises equipment. An important network cost characteristic is that the
fixed cost is a major percentage of total cost. As a result, for a given
network capacity level the direct marginal cost of a transmission is essen-
tially zero. However, transmissions cause an indirect congestion cost if their
network use reduces the service quality of other users, e.g., by increasing
their delay or data loss. The congestion cost is negligible at low network
utilization but very significant when traffic demand levels approach the
network capacity.
Basic operating principles

The fundamental functions multiplexing, switching and routing, define the
context for network control mechanisms. Multiplexing and switching spec-
ify the principles for allocating bandwidth to multiple individual traffic
sources such as voice conversations, email or streaming videos. Routing
determines the data path through the network. Traditional telephone net-
works significantly differ from data communication networks in how they
perform these functions.
Telephone networks use circuit switching and deterministic multiplexing.
When a call request arrives the admission control function decides whether to
accept or reject it by comparing the network’s resource availability with the
call’s requirements. Circuit switching establishes for each accepted voice call
a dedicated physical circuit, fixing a route and reserving a fixed bandwidth
amount equal to the data flow’s peak rate for the circuit lifetime. All data
then travels along this connection. Deterministic multiplexing manages the
static bandwidth shares at each link by letting active circuits only transmit in
their pre-assigned time slots, regardless of need. This works great for voice
and other CBR sources. Since resource needs of individual flows are constant
over time, it is easy to simultaneously deliver QoS guarantees and attain high
resource utilization by performing admission control and reserving matching
bandwidth. However, circuit switching can waste a lot of capacity if used for
bursty traffic sources with large ratios of peak to mean rate.
Data communication networks such as the Internet use packet or cell
switching and statistical multiplexing to mitigate this problem. Packet switch-
ing divides data flows into packets that travel individually to their destination
where they are reassembled. Statistical multiplexing dynamically allocates
bandwidth at each link to competing data packet streams based on their
probability of need, QoS and fairness considerations. Packet switching net-
works are connectionless if packets of the same flow are routed independently
of each other, or connection-oriented if they all travel along the same a priori
established route. In packet networks the term ‘connection’ refers to a set of
logically linked end-points for data exchange. In contrast to circuit switched

phone networks, traditional packet networks neither use admission control
nor dedicate fixed bandwidth for a data flow or connection during its entire
lifetime. Rather they allocate bandwidth on demand when and where needed
by a packet. Packet switching therefore achieves higher bandwidth utilization
for bursty traffic compared to circuit switching, but it is vulnerable to quality
degradation in the form of data delay and loss when the network is con-
gested. A range of mechanisms, including admission control, are designed to
control congestion in modern packet networks.
Network control mechanisms

There is a daunting array of network technologies that implement control
mechanisms in the hardware and software of nodes and at all layers of the
network architecture. This survey describes a selection of these mechanisms
at a conceptual level, focusing on their role in delivering certain service
performance attributes and in creating differentiated services. It classifies
these control mechanisms into demand regulation or supply allocation
mechanisms depending on whether they regulate the traffic inflow or the
allocation of a resource network with fixed capacity to admitted traffic
flows. Their key functions and characteristics are as follows.
Demand regulation mechanisms include admission control, traffic shaping
and policing and flow control. (This chapter ignores content replication,
caching and data compression techniques that reduce the potential demand
for data transmissions.) Admission control decides whether to accept or
reject a connection request with certain traffic constraints and QoS re-
quirements after checking whether the network has the necessary resources.
It operates on the time scale of connection interarrival times. Traffic po-
licing and shaping mechanisms ensure that the data flow of an admitted
connection conforms to contractually established traffic parameter con-
straints. They operate on individual packets and cells and are commonly
implemented via leaky bucket descriptors. Nonconforming traffic may be
shaped to conform by delaying cells or allowed into the network but
dropped if the network becomes congested. Flow control mechanisms reg-
ulate traffic inflow by controlling the sending rates of whatever sources
choose to be active, rather than via the number of traffic sources as does
admission control. Flow control dynamically manages bandwidth demand
and sharing via fast feedback loops between network nodes that send con-
gestion indication signals and end user applications which adapt their
sending rates accordingly. It operates on the time scale of round trip signal
propagation delays which are shorter than connection interarrival times,
but typically longer than packet queueing delays.
Supply allocation mechanisms considered here include link partitioning,
routing, resource reservation and packet scheduling. (This survey distinguishes
between supply allocation for given capacity and capacity expansion.) Link
66 Philipp Afèche
partitioning slices the bandwidth of a link into multiple parallel channels.

Routing assigns individual data packets or entire data flows to channels and
links from source to destination. Data flows may be assigned to resources
using a combination of exclusive access via reservations and shared access via
packet scheduling and routing. Resource reservation a priori fixes a route
along which it dedicates a set of buffers and channels to an individual data
flow; it is used at connection setup in conjunction with admission control.
Packet scheduling disciplines allocate resources dynamically: they regulate
the access of multiple individual data flows to a shared set of buffers and
channels by specifying the order in which queued data packets are transmit-
ted. Common packet scheduling disciplines are first-in-first-out (FIFO); strict
or absolute priority disciplines, where higher priority traffic waiting for
transmission always gets exclusive bandwidth access; and processor sharing
disciplines where all priority levels get some bandwidth access based on their
relative priority index. Variations of the latter include egalitarian, discrim-
inatory and generalized processor sharing, which is an idealized model of
Weighted Fair Queueing.
Guaranteed, best effort and flexible bandwidth-sharing services

Guaranteed services rely on admission control, traffic shaping and po-
licing for demand regulation. In terms of supply allocation tools their
analysis typically assumes fixed link partitions and routes, which are as-
signed to individual data flows via reservations and/or packet scheduling
mechanisms—FIFO if connections share and priority disciplines if they
differ in their QoS requirements. These supply allocation mechanisms are
insignificant from the perspective of users since the network guarantees
their transmission performance. A key performance issue for pricing is to
quantify the bandwidth requirement of each connection. Guaranteed serv-
ices rely on open-loop control and require extensive network state infor-
mation at the start of a transmission contract to ensure strict flow-level QoS
guarantees.
Best effort services, unlike guaranteed services, hardly rely on technolog-
ical demand regulation mechanisms, except through the bandwidth con-
straints of users’ access links. They do not have built-in flow control and
feedback mechanisms for congestion avoidance. Pricing is the only direct
demand regulation tool, but demand is also regulated indirectly based on
customers’ response to and anticipation of congestion. Best effort services
may rely on a range of (reservationless) supply allocation mechanisms to
manage QoS performance, cf. Gevros et al. (2001). This survey considers
the role of packet scheduling, link partitioning and routing mechanisms in
controlling congestion and creating quality differentiated services. These
tools may play a larger role in a future Internet with multiple classes of best
effort or ‘‘better-than best effort’’ service, somewhat similar to the Differ-
entiated Services (DiffServ) architecture proposal of the IETF. A key per-
formance issue for pricing is to quantify the QoS of each class in terms of
queueing delay or data loss. Best effort services rely on open-loop and slow
user-driven closed-loop control to try and meet certain average class-level
QoS metrics over longer time periods.
Flexible bandwidth-sharing services rely only on flow control for demand
regulation and congestion avoidance. In terms of supply allocation their
analysis typically assumes fixed routes of nonpartitioned links, which are
assigned to data flows without resource reservation via FIFO packet sched-
uling. Flow control leads users to share the link bandwidth so their cumu-
lative data rates equilibrate around the amount of available capacity, keeping
queueing delays small. The result is a set of services with arbitrary quality
differentiation in terms of their bandwidth allocations. Hence, the relation-
ship between bandwidth and quality of each flow is immediate. Flexible
bandwidth-sharing services rely on fast closed-loop control, while transmis-
sions are ongoing to deliver flexible flow-level bandwidth guarantees.
2.4 Distinctive features in comparison to telephony
The context of data communications differs in important ways from that

of ‘‘plain old telephone service’’ which has been studied extensively.
On the demand side, the traffic properties and QoS requirements of voice
conversations are uniform, simple and stable, while those of data sources
are heterogenous, complex and keep evolving as new applications emerge.
The demand for data transport is therefore less well understood than te-
lephony demand. For example, even users may not fully know their traffic
source properties ex ante. On the supply side, the need for data networks to
simultaneously carry a wide variety of traffic types with a multitude of QoS
requirements has spawned new types of service contracts along with new
technologies and control mechanisms that use fundamentally different re-
source allocation principles, including flow control, statistical multiplexing
and discriminatory packet scheduling disciplines.
These differences have posed new challenges and generated new research
on the economics of data communication services and on closely related
issues of operational network performance, specifically, on the bandwidth
requirements of traffic sources and on the relationship between traffic de-
mand, network capacity, control mechanisms, and QoS.
2.5 Pricing decisions
The preceding sections have introduced and characterized three funda-

mentally different types of data transport service contracts and the features
of service demand and production that shape the context for their analysis.
This section discusses the role of pricing, identifies major dimensions of
pricing decisions and outlines for each dimension the relevant issues for this
survey.
68 Philipp Afèche
The role of pricing in managing data transport services

From an economics perspective the performance of a data communica-
tions network is measured by its total surplus, i.e., total customer utility
minus total cost, and its division among customers and providers, which
depends on the payment flow from the former to the latter. Utility and cost
in turn depend on the specifics of network deployment and allocation to
customer traffic.
Pricing plays a key role in optimizing network deployment and allocation.
Early debates on Internet congestion considered and rejected several non-
pricing approaches, including voluntary user restraints and administratively
mandated and technologically enforced usage quotas (cf. MacKie-Mason
and Varian, 1995a). These approaches are inappropriate because of key fea-
tures of the data communications environment: users and the provider(s) are
independent self-interested decentralized decision makers with conflicting in-
centives and they have private information—users about their preferences, the
provider(s) about network characteristics. As a result, neither users nor the
provider(s) are well-equipped to make all allocation decisions. The use of
pricing serves several functions that achieve a more appropriate distribution
of decision making. Prices generate revenues for providers that allows them
to cover costs and make profits. They shape the incentives of providers to
invest in capacity. They shape the incentives of customers to choose certain
quantities, types and times of service. This regulates the amount and com-
position of traffic admitted to the network over time and its allocation to
bandwidth via specific routes or priority classes. As such pricing is an im-
portant economic congestion control tool that complements the technolog-
ical network controls discussed above. It also makes network control more
scalable, since a single or a few prices are sufficient to control the traffic
demand of millions of users. By observing user reactions to different price
levels providers may also gain important demand information that allows
them to further improve their profits and/or social surplus.
Pricing decisions and issues

Managing transport services involves service design, pricing and produc-
tion decisions. Reflecting the emphasis in the literature, this survey prima-
rily focuses on pricing decisions but also considers capacity decisions. As
noted earlier the chapter focuses on the case of a single provider with one of
two performance objectives: welfare maximization or profit maximization.
Pricing defines a set of tariffs for the offered set of service classes. A tariff
for a given service class is a function that specifies the user’s total payment
depending on two sets of usage attributes, the expected or actual bandwidth
consumption per service unit and the quantity of service units. Depending on
the context a service unit may be defined as the transmission of a data
packet, a bulk transfer, a CBR or a VBR traffic source. In each case, one
can quantify the expected or actual bandwidth consumption based on a
priori known or measured parameters such as packet or file size, peak rate,
mean rate, duration and network route that are inherent to the source and/
or specified by the service contract.
Virtually all pricing analyses assume a communication network with a
given structure and set of control mechanisms. This section categorizes
pricing analyses further along several dimensions and outlines, for each
dimension, the relevant issues for this survey. The number of network links
and the information structure characterize the decision environment. The
usage-sensitivity of tariffs, price and service differentiation, dynamic tariff
fluctuations, and posted prices versus auctions refer to important issues in
structuring and computing tariffs.
Number of network links. The studies considered here assume networks
with a fixed number of resources and topology. One can distinguish be-
tween services based on whether they are delivered over a single link or a
multi-link network. In a multi-link network the bandwidth consumption and
price of a service unit also depends on the set of links along its route. The
studies of multi-link networks considered here typically assume a given set
of fixed routes.
Information structure. Pricing and capacity decisions hinge on demand
information about utility functions and traffic source properties, and serv-
ice usage decisions require information about QoS metrics such as delay or
network load. Many pricing studies assume complete information about all
these characteristics. Studies with incomplete information assume at least
some of these characteristics to be the private information of users or the
provider. One can distinguish between several cases of incomplete infor-
mation. Cases where the provider has no or incomplete information on de-
mand functions or users have no or incomplete system QoS information
motivate the study of adaptive and auction-based price mechanisms. Cases
where the provider has incomplete information on traffic source parameters
or only aggregate information but no information on QoS preferences of
individual users motivate the design of incentive-compatible tariffs that dis-
criminate based on actual usage characteristics or service quality attributes.
Usage-sensitivity of tariffs. A fundamental question is whether tariffs
should at all be usage-sensitive. ‘‘Flat-rate pricing’’ has been and still is the
prevalent tariff structure in the Internet retail market. A user pays a flat
periodic subscription fee that only depends on the bandwidth of her access
link. Thus her charge is related to her maximum sending rate but independ-
ent of actual usage, quality or time of service. Like much of the research
activity on data pricing this survey, including the following discussion of the
framework, focuses on usage-based pricing. It briefly turns to the question of
flat-rate versus usage-based pricing in Section 6.2.
Price and service differentiation. A central research issue focuses on how
to tailor tariffs and services to users with heterogenous preferences and
traffic properties. Most studies consider tariff design for a given service
design, defined by a fixed number of service classes with a single type of QoS
guarantee for all classes, and delivered by prespecified control mechanisms.
70 Philipp Afèche
A few of these papers investigate the economic value of differentiation by

comparing optimally priced systems with different numbers of service
classes, or jointly optimize tariff design and the number of service classes.
Sections 3–5 consider these studies by type of QoS guarantee. Section 6.3
briefly outlines the merits and challenges of alternative QoS designs and
their delivery mechanisms.
For each type of QoS guarantee the discussion in this survey distinguishes
between pricing studies for uniform service or bandwidth differentiated
services and those for delay or loss differentiated services.
In general, tariffs may discriminate based on service class, usage and
customer attributes. This survey does not consider tariffs that discriminate
solely based on customer attributes (their preferences, identity, business
versus home, etc.). It focuses on tariffs that are linear in the quantity of
service units (e.g., throughput) and that may discriminate based on service
class, and the expected or actual bandwidth consumption per service unit.
(See Wilson, 1993, for a classic treatment of nonlinear pricing.)
An important issue in price and service differentiation arises when the
provider does not have complete information about demand characteristics.
The case where the provider has incomplete information on traffic source
parameters is particularly relevant for guaranteed services. The case where
the provider has aggregate information but no information on QoS prefer-
ences of individual users is particularly relevant for delay or loss differentiated
best effort services. Both cases call for the design of incentive-compatible
tariffs that take into account users’ choice behavior.
Dynamic tariff fluctuations. This survey distinguishes between static pric-
ing and two types of mechanisms with dynamic tariff fluctuations, adaptive
pricing and optimal dynamic pricing. Adaptive pricing mechanisms have the
objective of finding the optimal static equilibrium prices and allocation in
environments with no or incomplete information about demand functions.
They resemble the classical tatonnement process. Price fluctuations serve the
purpose of learning demand, and the analysis focuses on how to adjust prices
to reach an equilibrium for a given stationary demand environment and
issues of convergence speed and stability. Optimal dynamic pricing mecha-
nisms seek to optimally adjust prices to the state of network load or demand
to improve performance over static (state-insensitive) pricing. In contrast to
adaptive pricing, here the equilibrium prices are dynamic, not static, over the
time horizon, and demand functions are typically known. Research issues
focus on characterizing the optimal price policy for a given service design and
on quantifying the benefits of dynamic over static pricing. This survey further
distinguishes between two forms of dynamic pricing. Load-sensitive dynamic
pricing seeks to optimally adjust prices based on the network load to improve
performance over static load-insensitive prices. The analysis typically con-
siders stationary known demand functions that are independent of the time
of service. Demand-sensitive dynamic pricing is closely related to peak-load
pricing that has been important for telephone tariffs. It seeks to optimally
adjust prices based on the state of demand at different times of service, to

reduce traffic peaks and improve performance over static demand-insensitive
prices. The analysis typically considers nonstationary known demand func-
tions that depend on the time of service, e.g., the time of day. In an empirical
study Edell et al. (1995) show that time of day prices may move 33% of
traffic from peaks to less busy times.
Posted prices versus auctions. The majority of papers analyze centralized
pricing where the provider designs and posts tariffs. A fundamentally differ-
ent approach is via auctions where pricing decisions are at least partly de-
centralized: users submit bids and the provider determines prices and
allocations based on bids and previously announced payment and allocation
functions. Starting with the ‘‘smart market’’ proposal of MacKie-Mason and
Varian, (1995a), which was inspired by Vickrey’s (1961) second-price sealed-
bid auction, auction-based pricing approaches have received considerable
attention in data pricing research, for services with different types of QoS
guarantees. An important motivating factor for these efforts is the demand-
revealing property of the classic second-price auction which allows the pro-
vider to attain the efficient allocation even if she has no or incomplete in-
formation on demand functions. A reasonable question to ask is whether this
demand-revealing property is preserved in the data communications context
and under what conditions. Section 6.1 surveys auction-based approaches
and considers this question.
2.6 Discussion framework
After describing major dimensions and issues of pricing decisions this

section outlines the framework for their discussion. The core Sections 3–5
focus on the paradigm of economic optimality. Section 6 goes beyond the
optimality criterion and also considers some implementation issues in pric-
ing and service design.
Guaranteed, best effort and flexible bandwidth-sharing services

Sections 3–5 consider usage-based pricing decisions by type of QoS guar-
antee, for guaranteed, best effort, and flexible bandwidth-sharing services,
respectively. These sections share the following organization. Not all
parts are equally relevant for all service types but the common structure
aims to provide an unifying framework for comparing the different research
streams.
1. Features, modeling and overview. This introductory part reviews the
main contract features and model characteristics that shape the
context for their analysis, and outlines the issues discussed in the
section.
2. Basic pricing and allocation principles. This part considers environ-
ments with complete information. It focuses on static socially optimal
72 Philipp Afèche
pricing for fixed capacity but also discusses the relationship between
socially optimal prices and capacity levels. Throughout it assumes a
fixed service design that reflects in some sense the ‘‘basic level of serv-
ice differentiation’’ for each service type: a single service class (best
effort) or multiple bandwidth differentiated services (guaranteed serv-
ices or flexible bandwidth-sharing services). The discussion emphasizes
the single link case but also outlines how the results extend to a multi-
link network, typically with fixed routes. Where applicable this
part also outlines how the results extend to a profit-maximizing mo-
nopoly.
3. Incomplete information and adaptive pricing. This part considers is-
sues that arise due to incomplete information. All three contract types
consider the case where the provider has no or incomplete information
on demand functions and study adaptive price mechanisms to find the
right prices. The case where users have no or incomplete system QoS
information is relevant for best effort and flexible bandwidth-sharing
services since the network does not a priori commit to a certain QoS
level, which calls for user adaptation to QoS changes.
4. More differentiation. This part discusses work that goes beyond the
basic forms of price and service differentiation considered in Part 2.
Issues include pricing for service designs with delay or loss differen-
tiated services; pricing for service designs with two different types of
QoS guarantees (e.g., best effort and strict guarantees) and incentive-
compatible tariff design under incomplete information on traffic source
parameters (for guaranteed services) or on individual users’ QoS pref-
erences (for multi-class best effort services).
5. Optimal dynamic pricing. This part surveys studies of load- and de-
mand-sensitive dynamic pricing. Maybe in part due to the mathemat-
ical challenges involved in their analysis, these pricing approaches
have received comparatively little attention.
6. Conclusions and directions. This part summarizes and suggests di-
rections for future research.
Discussion
After considering each contract type in isolation the survey then turns to
broad questions that concern all contract types.
1. Auctions versus posted prices. This part surveys auction-based ap-

proaches and discusses their benefits and challenges, including their
potential for demand revelation depending on the level of resource and
time aggregation of the service units sold, dynamic effects and the
presence of delay externalities.
2. Flat-rate versus usage-based pricing. This part outlines the debate on flat-rate
versus usage-based pricing, outlines some results on two-part tariffs that
combine a flat and a usage-sensitive component, and characterizes the
fundamental trade-off between optimality and implementability criteria that

should be considered in the design and evaluation of any pricing mechanism.
3. Providing QoS: overprovisioning versus control. This part outlines the
merits and challenges of alternative QoS designs and their delivery
mechanisms.
3 Pricing guaranteed services
3.1 Features, modeling and overview
Features
Guaranteed service contracts specify QoS parameters the network promises
to deliver, traffic parameters that constrain the user’s traffic profile, and a
tariff that determines the user’s charge. Guaranteed services are suited for
real-time applications with CBR or VBR traffic sources. QoS parameters
express the maximum tolerable congestion level in terms of deterministic or
probabilistic traffic flow performance statistics such as minimum throughput
rate; average, maximum delay or delay variation (jitter) and packet loss
probability. This section focuses on static QoS parameters that are agreed
upon at contract inception and remain fixed over its lifetime. Traffic pa-
rameters may include metrics such as the peak rate, mean rate and burstiness
of the source. In the basic case all traffic parameters are a priori known to the
provider and statically defined at the start of the contract, serving as the basis
for tariffs and network control decisions. If some traffic properties are a
priori not known, contracts may give users more flexibility in choosing traffic
parameters and tariffs that depend not only on a priori known (declared)
traffic parameters but also on a posteriori information based on actual traffic
measurements, cf. Kelly (1994) and Courcoubetis et al. (2000a, 2000b).
Since the network strictly guarantees quality there are no congestion
externalities: a user’s net utility function depends only on the attributes of
her own service contracts—quantities, QoS and traffic parameters, and
tariffs, but not on other users’ network usage.
The network uses several control mechanisms to ensure the promised
quality levels. Among these the ones directly visible to users are admission
control, which determines whether the network has the required resources
to meet contractual QoS parameters for traffic with a certain profile, and
traffic policing, which holds admitted calls to contractual traffic parameters.
A key issue in managing guaranteed services is to estimate the bandwidth
consumption of a source at each link along its route. This measure of
resource consumption forms the basis for admission control, resource res-
ervation and pricing decisions. Pricing can regulate demand for connection
requests and the traffic properties of accepted connections. The control
mechanisms operate at the level of individual data flows which requires
considerable complexity and ‘‘intelligence’’ in the core of the network.
74 Philipp Afèche
Modeling
There is a fixed set of CBR or VBR traffic source types, each with its own
statistical properties, pre-specified QoS requirements and traffic parame-
ters. Typically there is a fixed set of service contracts, each tailored to one of
the traffic source types. A customer’s utility is a function of the quantities of
traffic contracts she buys. The bandwidth requirement of an active CBR
source equals its constant data rate. The bandwidth requirement of an
active VBR source is much more complex; it depends not only on its own
traffic parameters and QoS requirements but also on the traffic mix and
control policies on the link, and on link and buffer capacity levels. The
effective bandwidth of a traffic source maps these multiple factors into a
one-dimensional measure of resource consumption for the connection. In
some cases the traffic properties of a VBR source, and hence its effective
bandwidth, may not be fully known a priori but only a posteriori after
actual usage measurements.
Pricing analyses of guaranteed services consider primarily two environ-
ments. Studies of VBR sources typically focus on a static environment with a
given connection mix, since the effective bandwidths of VBR sources are
highly dependent on the traffic mix. Studies of CBR sources typically con-
sider a dynamic environment where individual connections arrive and depart
over a time horizon much longer than their own duration. These dynamics
are easier to analyze for CBR than for VBR sources since their resource
consumption is invariant to the overall traffic mix.
The key network control mechanisms experienced by the user are admis-
sion control coupled with resource reservation and traffic policing. Other
network control mechanisms, including routing at connection acceptance and
dynamic bandwidth allocation through scheduling of admitted data packets
at network nodes, are essentially invisible to users since the network guar-
antees quality.
Overview
Section 3.2 first reviews the notion of effective bandwidth, a key measure
of bandwidth consumption for traffic sources with guaranteed service. It
then discusses the basic principles for optimal static pricing in a static
environment for bandwidth differentiated services: the connections they
serve have different but fully known traffic characteristics, a single common
QoS parameter and are given FIFO service. The analysis focuses on social
optimization for the single-link case and discusses extensions to a monopoly
and a multi-link network. In this static environment the socially optimal
linear prices for guaranteed service classes are proportional to the effective
bandwidths of their connections, whereas the monopoly prices also depend
on the demand elasticity. The section then outlines the difference to static
analyses of dynamic environments, where socially optimal static prices are
not simply proportional to effective bandwidths but also depend on the
blocking probability. Section 3.3 reviews work on iterative adaptive price
mechanisms that are designed to attain the socially optimal bandwidth

allocation when the network lacks information on utility functions. Section
3.4 surveys several cases with more service differentiation compared to the
basic amount of service differentiation considered in Section 3.2: First,
multiple delay or loss differentiated QoS classes supported by priority-serv-
ice at network nodes; second, settings where some traffic source properties
are a priori not known to the provider who therefore designs a menu of
multi-part tariffs to elicit this information from users. Each tariff is linear in
measured usage parameters of the connection, time and data volume in the
simplest case, with coefficients that vary based on the user’s a priori de-
clared traffic parameters. Third, the case where a provider offers guaranteed
and best effort service over the same infrastructure. Section 3.5 discusses
analyses of load-sensitive and demand-sensitive dynamic pricing. Section
3.6 summarizes and outlines directions for future research.
3.2 Basic pricing and allocation principles
Consider the pricing problem for a single capacitated link in a static

environment with a fixed mix of potential connections. How to allocate the
link to different types of traffic sources and at what prices? Solving this
problem requires two steps. First, the network needs to understand the
resource requirements of traffic sources that request connections so it can
characterize the admission region, defined as the set of feasible contract
quantities that it can simultaneously offer without violating the QoS con-
straints. Second the network needs to choose the desired contract mix
within the admission region and prices that support the corresponding de-
mand.
Resource consumption and admission region

Consider a provider offering K types of service contracts, each tailored to
a traffic source with certain statistical properties and QoS parameters, and
let x ¼ (x1,x2, y, xK) denote the number of active contracts of each class.
The admission control problem is to ensure that the traffic mix always re-
mains within the admission region. The admission control problem is sim-
ple in the case of CBR sources such as certain voice and video applications
(leased virtual circuits or dialup access connections may also be modeled as
CBR sources): the network simply reserves a bandwidth amount equal to
the constant source data rate for each admitted contract. Traffic streams do
not interact, the QoS of a traffic stream directly corresponds to its resource
requirement—defined at any point in time by a known scalar quantity, and
the resource consumption of a traffic mix is additive in those of its com-
ponent streams. However, the admission region can be difficult to evaluate
for VBR traffic sources due to their stochastic and heterogenous nature,
especially if QoS constraints are multi-dimensional. When traffic sources
have randomly fluctuating bandwidth requirements the network reserves
76 Philipp Afèche
for each source less than its peak data rate and uses statistical multiplexing
for multiple sources to dynamically share the link capacity. This dramat-
ically increases the complexity of the admission control problem: traffic
sources now interact through the sharing mechanism, and the relationship
between traffic source properties, QoS requirements and the acceptable
capacity utilization level is far from obvious. For example, to provide ac-
ceptable QoS to bursty sources with tight delay and loss requirements it
may be necessary to keep the required average utilization of a link below
10%, while for constant rate sources the average utilization may be close to
100%. The concept of effective bandwidth, based on the seminal paper of
Hui (1988), has given rise to elegant and powerful solutions to this resource
measurement problem. It provides a one-dimensional measure of resource
usage which adequately represents the trade-off between sources of differ-
ent types taking proper account of their varying statistical characteristics
and multi-dimensional QoS requirements.
The main result on effective bandwidths for the single-link case is out-
lined as follows. (See Kelly, 1996, and Chapter 4 in Courcoubetis and
Weber, 2003, for reviews of effective bandwidths.) Suppose that the K
traffic classes are multiplexed over one FIFO link with bandwidth C and
buffer size B. The traffic sources may have different, a priori fully known,
statistical properties but share a single common QoS parameter, e.g., that
the cell loss probability of any source not exceed 10–8. The main result
defines and characterizes the effective bandwidth of traffic source k by a
function ak (s, t) and the link’s effective capacity as a function C ðs; tÞ: The
parameters s and t depend on the link capacity C, the buffer size B, and on
the operating point x of traffic sources and their statistical properties and
QoS parameter. Based on these quantities the admission control problem
simplifies to verifying whether a traffic mix x satisfies the linear constraint
X
K
xk ak ðs; tÞ C n ðs; tÞ. (1)
k¼1
Thus, effective bandwidths can be interpreted as defining a local linear

approximation to the boundary of the admission region at an operating
point x.
Socially optimal prices in static environments

The basic problem of socially optimal static linear pricing of guaranteed
services for VBR sources hinges entirely on the characterization of their
effective bandwidths. Since the effective bandwidth of a VBR source is
sensitive to the traffic mix, most effective bandwidth models assume a fixed
set of contract types or service classes. Under this assumption the char-
acterization of resource consumption by effective bandwidths reduces the
analysis of static pricing for such contracts to a classic multi-product pric-
ing problem with linear capacity constraints.
For illustration, let U (x) denote the total system utility as a function of
the traffic mix x. Assume for now that it is known to the provider and that
the direct marginal cost of a contract is zero. Let b denote the Lagrange
multiplier of the constraint (1). The problem of finding the socially optimal
allocation of contracts is
!
XK

max UðxÞ þ b C ðs; tÞ xk ak ðs; tÞ . (2)
x0;b0
k¼1
At the socially optimal traffic mix x ; the marginal utility of each contract
type must be equal to the product of its effective bandwidth by the shadow
price of the admission region constraint
@Uðx Þ
¼ b ak ðs; tÞ; k ¼ 1; 2; . . . ; K. (3)
@xk
Suppose, a linear pricing scheme that charges a unit price Pk per class-k
contract. Given these prices users demand contracts up to the point where
their marginal utility equals the prices
@UðxÞ
¼ Pk k ¼ 1; 2; . . . ; K. (4)
@xk
Hence, if the socially optimal traffic mix is on the boundary of the ad-
mission region, i.e., the constraint (1) is binding, then b 40 and the socially
optimal unit prices of contracts are proportional to their effective bandwidths
Pk ¼ b ak ðs; tÞ; k ¼ 1; 2; . . . ; K. (5)
Network. The single-link social optimization problem (2) naturally gen-
eralizes to the multi-link network case by adding for each link one capacity
constraint of the form (1).
Price and capacity decisions. The joint problem of socially optimal pricing
and capacity investment is discussed in Section 4.2 together with the case of
best effort services.
Profit-maximizing prices in static environments

Of course the profit-maximizing prices in a static environment also de-
pend on the demand elasticity, not just on resource shadow prices and
effective bandwidths. The respective analysis and results follow immediately
from the classical multi-product monopoly problem with linear technology.
Optimal static prices in dynamic environments

The principle for socially optimal pricing in static environments given by
(5) ceases to hold in dynamic environments where connections arrive and
depart over time. Specifically, the socially optimal static linear prices de-
pend not only on the effective bandwidths of individual sources, but also on
the system’s blocking probability which measures the fraction of time the
78 Philipp Afèche
system spends on the boundary of the admission region. See Kelly (1991)
and Ross (1995) for theoretical and computational issues in the analysis of
blocking probabilities and their approximations. Courcoubetis and Reiman
(1999) and Paschalidis and Tsitsiklis (2000) study the problem for a single
link, and Paschalidis and Liu (2002) tackle the network case. They all model
the service provider as a multi-server loss system and consider the case of
CBR sources, which is more tractable since the effective bandwidth of each
source is independent of the traffic mix. Courcoubetis and Reiman (1999)
consider static pricing under revenue and welfare maximization using an
asymptotic analysis that simplifies the calculation of blocking probabilities.
3.3 Incomplete information and adaptive pricing
The analysis discussed so far assumes that the provider knows the cus-
tomer demand functions. Low and Varaiya (1993) and Thomas et al. (2002)
relax this assumption. They propose and study decentralized tatonnement-
style iterative adaptive price mechanisms to attain the socially optimal al-
location in a static environment (with a fixed set of possible connections)
when the network does not know the customer demand functions. Rather
than relying on the notion of effective bandwidths, they assume that the
relationships between QoS and resource requirements are explicitly given as
opposed to derived from first principles based on traffic source properties.
Low and Varaiya (1993) model ATM networks as pools of capacitated
transmission links and buffers. They rule out resource sharing across con-
nections through statistical multiplexing. They specify the QoS constraints
of each service class in terms of a minimum bandwidth and a ‘‘burstiness
curve’’ that captures acceptable bandwidth-buffer size substitutions. The
socially optimal unit prices for buffer and link capacity equal their shadow
prices. Their decentralized algorithm has the network set resource prices
and users respond with resource requests in iterative fashion. Thomas et al.
(2002) consider more general QoS resource allocation relationships and a
market mechanism for a competitive economy with three price-taking agent
types, users, resource providers and service providers, and an auctioneer
who updates prices based on the observed aggregate resource supply and
demand. Their distinctive contribution is to prove convergence of their
algorithm in a finite number of iterations to a resource allocation that is
arbitrarily close to the socially optimal solution. Similar iterative mecha-
nisms for settings with congestion externalities are discussed in Section 4.3.
3.4 More differentiation
The pricing analysis of Section 3.2 focuses on bandwidth differentiated

services. It assumes that all traffic sources share a single QoS bound, that
the traffic properties of each source are a priori fully known, and that the
network only offers guaranteed services. This section outlines work that
relaxes these assumptions.
Delay or loss differentiated services: multiple QoS parameter constraints

If there are multiple constraints on QoS parameters that differ by traffic
source the admission region can be reasonably well approximated by mul-
tiple linear constraints of the form (1), one for each QoS constraint rep-
resented at that link. In the network case there is one set of such constraints
for each link. Effective bandwidth characterizations span a range of models
and service disciplines. Zhang (1993), Elwalid and Mitra (1995), de Veciana
and Walrand (1995), Kulkarni and Gautam (1997), Zhang (1997), and Be-
rger and Whitt (1998a,1998b) study effective bandwidths for systems that
implement service differentiation through priority scheduling disciplines.
Elwalid and Mitra (1992) and Kulkarni et al. (1995) consider an alternative
priority mechanism with loss priorities: it uses FIFO scheduling but rejects
or discards low priority data packets if the workload is above a threshold. It
appears that the pricing of such differentiated guaranteed services with two
or more QoS bounds has not received significant attention. However, from
the perspective of economics what matters most about these service designs
is their output, i.e., the characterization of resource consumption for each
traffic mix, which determines the admission region and the optimal allo-
cation and prices. The implementation details of these network control
mechanisms are hidden from users; all that matters to them is that their
QoS requirements be met.
A priori unknown traffic parameters

In practice the network may not a priori have the necessary information
on traffic source properties with given QoS parameters to compute their
effective bandwidths. It must instead rely on traffic parameters announced
at contract inception which raises the following questions: how to estimate
effective bandwidths for admission and pricing decisions, and how to
structure price tariffs to set the right user incentives for truthfully declaring
their best estimate of traffic properties? For given QoS parameters users’
traffic parameter choices affect both the value and the effective bandwidth
of their transmissions. Charging contracts based only on a priori informa-
tion or only on a posteriori measurements is problematic. Charging only
based on a priori estimates of their average effective bandwidths, using
historic data for similar traffic, gives users the incentive to use the maximum
that their contractual traffic parameters allow. This in turn leads the op-
erator to calculate and charge users the maximum effective bandwidths
possible given their contract parameters, which discourages users with a
lower effective bandwidth from buying such contracts. At the other ex-
treme, charging only based on a posteriori measurements may leave the
network with too much of the risk associated with the opportunity cost of
80 Philipp Afèche
admitting contracts with high a priori, but low actual effective bandwidth
and revenue.
Kelly (1994) and Courcoubetis et al. (2000a, 2000b) develop the following
solution to this incentive issue for sources with a priori known (static) traffic
parameters specified in the contract such as the peak rate, and a priori
unknown and measured (dynamic) traffic parameters such as the mean rate.
They construct charges based on the effective bandwidth computed from
the static parameters and measurements of the dynamic parameters. Users
are offered a menu of multi-part tariffs. Each tariff is linear in measured
usage parameters of the connection with coefficients that depend both on
the static parameters and on users’ declared estimates of the dynamic traffic
parameters. This scheme gives rise to charges proportional to effective
bandwidths and is incentive-compatible, giving risk-neutral users who min-
imize their expected charge the incentive to truthfully report their expected
value of their a priori unknown traffic statistic. In the simplest case this
approach gives rise to time-volume charging whereby the total charge is a
linear function of the measured connection time and data volume with
coefficients that depend on the known static peak rate and the user’s de-
clared mean data rate.
Guaranteed and best effort service

Wang et al. (1997) and Maglaras and Zeevi (2005) study pricing for the
case where a provider offers guaranteed service (for real-time CBR sources)
and best effort service over the same infrastructure.
3.5 Optimal dynamic pricing
Dynamic pricing studies of guaranteed services appear to have focused on

CBR sources so far.
Load-sensitive dynamic pricing

Paschalidis and Tsitsiklis (2000) and Paschalidis and Liu (2002) derive
and compare the performance of the optimal static and load-sensitive dy-
namic prices under revenue and welfare maximization. They consider a
market with a fixed set of service classes, each defined by its fixed band-
width requirement, an exponentially distributed holding time, and a price-
sensitive Poisson arrival process. They assume that demand is stationary,
i.e., the arrival rates only depend on prices, not on the time of service. The
provider is assumed to know all demand characteristics. Paschalidis and
Tsitsiklis (2000) model the provider as a single-node multi-server loss sys-
tem who chooses class-specific prices. They show that in an asymptotic
regime of a large system with many relatively small users, the performance
of optimal dynamic load-sensitive prices is closely matched by suitably
chosen class-dependent but static, i.e., load-insensitive, prices. Paschalidis
and Liu (2002) extend this result to a network with fixed routing and allow
for demand substitution across service classes. This result is quite intuitive:
as the number of users and the capacity increase while the utilization re-
mains constant, the law of large numbers takes over, eliminating statistical
fluctuations so that loss probabilities become small; this in turn reduces the
value of optimizing prices based on network load. This result suggests that,
at least in settings with many small users, time-of-day prices that adjust to
slow changes in average network load may match the performance of prices
that fluctuate more rapidly based on instantaneous changes in network
load.
These studies assume that prices for a contract are fixed during its life-
time, even though prices may vary across contracts based on congestion
levels. In contrast, Semret and Lazar (1999) analyze a proposal to price loss
system services in a stationary demand environment using load-sensitive
spot prices and a derivative market mechanism. To avoid being dropped
before service completion, a user must either continuously pay the spot
price, which fluctuates based on auctions held among recent requests, or
purchase a hold option up-front which gives her the right to buy service for
a fixed price at any time up to a specified future date. They characterize the
equilibrium spot price and the reservation fee as the fair market price of the
hold option.
Demand-sensitive dynamic pricing

Wang et al. (1997) develop a procedure to determine optimal demand-
sensitive dynamic prices. They assume given time-varying demand functions
that are known to the provider. They consider an integrated-services net-
work which offers multiple guaranteed service classes to CBR sources and a
single class of best-effort service. Requests for guaranteed service are lost if
not admitted immediately while those for best-effort service are queued
until they are transmitted.
3.6 Conclusions and directions
The economic analysis of guaranteed services crucially hinges on the

ability to adequately quantify key measures related to resource consump-
tion: the effective bandwidths of individual traffic sources, and the admis-
sion region and blocking probabilities of the network. Much of the
technical mechanisms and analysis underlying how the network delivers a
set of QoS requirements is hidden from the users’ decision problem; they
only care about what the service class attributes are. Pricing in static en-
vironments closely follows the classic analysis of multi-product pricing with
linear capacity constraints, the main challenge is to characterize the effec-
tive bandwidth of each ‘‘product.’’ Hence, the socially optimal prices of
connections in a static environment with perfect traffic source information
are proportional to their effective bandwidths whereas the corresponding
monopoly prices also depend on demand elasticities. In the absence of
82 Philipp Afèche
perfect a priori traffic source information, tariffs with components based on

both a priori declared and ex post measured traffic source properties set the
right incentives for users to declare their best estimate of traffic parameters.
Studies of static and load-sensitive dynamic pricing in dynamic environ-
ments have so far focused on CBR sources and suggest that the relative
benefit of load-sensitive dynamic pricing may be minor in settings with
many small users.
Several issues appear to have received little attention so far and to merit
further research.
Dynamic environments of VBR sources. It appears that pricing studies for
VBR sources have so far focused on static environments. Determining the
optimal (static or dynamic) prices for VBR sources in dynamic environ-
ments faces the additional challenge that their effective bandwidths depend
on the traffic mix, whereas the bandwidth requirements of each CBR source
is independent of the traffic mix. One approach is to define an ‘‘average’’
effective bandwidth for each VBR source and to apply the analytical tech-
niques for loss networks with CBR sources to those with VBR sources. Are
there suitable approximations of such ‘‘average’’ effective bandwidths for
use in dynamic environments? If so, how good are these approximations
relative to more precise approaches that re-compute the effective band-
widths as the traffic mix changes? What are the pricing implications of
different effective bandwidth characterizations?
Value of load-sensitive dynamic pricing. Paschalidis and Tsitsiklis (2000)
and Paschalidis and Liu (2002) show that the value of load-sensitive pricing
over static pricing is negligible in an asymptotic regime with many small
users. Under what conditions does load-sensitive pricing generate a signifi-
cant performance improvement?
Demand-sensitive dynamic pricing. Existing analyses, e.g., Wang et al.
(1997), assume that the nonstationary demand rate is a function of time t
and of prices only at that time, but does not depend on prices at other times.
In other words, customers are nonstrategic with respect to the time of their
transmission request. How should prices be set to smooth demand peaks
when customers substitute over time? Analyses of peak-load pricing with
substitution effects has received some attention outside the data commu-
nications context, e.g., by Bergstrom and MacKie-Mason (1991) for a static
environment and by Shy (2001) for a dynamic environment.
Service design. Much of the analysis of guaranteed services takes as a
starting point a given set of pre-defined traffic sources and service classes.
However, customers may choose to aggregate and split traffic. What are the
incentives of customers for doing so and how should providers define and
price service bundles of multiple connections? Courcoubetis and Weber
(2003, p. 212) observe that the incentive to split traffic may arise in certain
situations and that it can be discouraged by adding a fixed charge per
connection. However, the broader issue of how to design the mix and pric-
ing of service classes appears to have received little systematic attention.
4 Pricing best effort services
Features
Congestion occurs whenever demand for resources exceeds their supply.
Since data networks face significant demand variability in terms of the
number of active users and their data transmissions, they are prone to
intermittent congestion even if on average underutilized. The guaranteed
services just discussed commit to strict quantitative QoS bounds for each
individual transmission flow before it starts which the network then ensures
by tightly controlling congestion via admission control, resource reserva-
tion and related mechanisms. In contrast, best effort services make no such
quality promises and do not constrain users’ traffic inflow to the network.
As a result the QoS performance of a given user’s transmission is deter-
mined not only by her own behavior and the network control mechanisms,
but also depends on other users’ offered load—users create and are subject
to congestion externalities. Congestion may lead to performance degrada-
tion in the form of data delays, losses and throughput limitations that
reduce customer utility. Best effort services are suited for elastic applica-
tions which tolerate this QoS degradation.
Recall from Section 2.2 that this survey adopts a literal definition of best
effort services and distinguishes them from flexible bandwidth-sharing
services that are discussed in Section 5. As understood here, best effort
services offer no quality guarantees whatsoever and impose no constraints
on the traffic that users may send into the network, irrespective of con-
gestion conditions. They have no built-in traffic flow control and feedback
mechanisms for avoiding congestion. Thus, users have little information on
precisely when and how to adjust their sending rates to match fluctuations
in bandwidth availability. As a result there can be a significant mismatch
between a user’s sending rate and her realized bandwidth at any given time,
which translates into delays and/or losses. The Internet transport service
provided by the UDP protocol is a typical example: packets may be de-
layed, lost or arrive out of order without notice. It is used for transmissions
where more control is not worthwhile, e.g., for small bursts of data or for
high-speed multimedia applications where resending data is of no value. By
contrast, flexible bandwidth-sharing services offer dynamic bandwidth
guarantees and give users some control over and the ability to adapt to their
allocation, while their transmissions are in progress. The Internet transport
service provided by the TCP protocol is a prime example. It guarantees zero
packet loss by retransmitting lost packets and dynamically controls the
peak rates at which competing connections are allowed to transmit via
adaptive flow control and feedback mechanisms. Its distinctive features are
closed-loop control, on fast time scales and at the level of individual flows:
individual end user applications receive congestion signals from network
84 Philipp Afèche
nodes every few tens of milliseconds that vary with the state of packet
queues and adapt their sending rates correspondingly quickly to track
changes in their bandwidth allocations.
Keeping in mind that the current Internet—including the TCP protocol—
is often broadly referred to as a ‘‘best effort network,’’ the distinction be-
tween best effort and flexible bandwidth-sharing services made here is use-
ful, since the respective pricing and resource allocation decisions occur on
different time scales that call for somewhat different analytical frameworks
and perspectives. In the case of flexible bandwidth-sharing services the fast
feedback loops afford users (or agents on their behalf) some control over
their transmissions while they are in progress. In the case of best effort
services the notion is that feedback loops between sending rates and con-
gestion levels are relatively slow—so slow that users have no control over a
given individual flow through the network once it is in progress. Thus, users
make all decisions concerning an individual flow or a collection of such
flows covered by their contract at the start of the respective transmissions,
based on the contract parameters and performance information available at
that time. In this sense, best effort service contracts only have static pa-
rameters. Since the network does not make any strict QoS guarantees per
user, the notion is that users, prior to sending data, estimate or receive
information on aggregate QoS statistics, e.g., the delay distribution or
packet loss ratio over all flows. One can think of these estimates as the
expected QoS metrics of a given best effort service.
Best effort services—unlike guaranteed services—hardly rely on techno-
logical demand regulation mechanisms, except through the bandwidth con-
straints of users’ access links. Pricing is therefore the only direct demand
regulation tool. Demand is also regulated indirectly based on customers’
response to and anticipation of congestion. Best effort services may also
rely on a range of supply allocation mechanisms to manage QoS perform-
ance, cf. Gevros et al. (2001). This survey considers the role of packet
scheduling, link partitioning and routing mechanisms in controlling conges-
tion and creating quality differentiated services. These tools may play a
larger role in a future Internet with multiple classes of best effort or ‘‘better-
than best effort’’ service, somewhat similar to the DiffServ architecture
proposal of the IETF.
Modeling
The following generic model of congestion externalities illustrates the
analysis framework for best effort services. It is worth noting that several
model features and results presented here are developed in congestion pric-
ing studies that are not specifically targeted to data communications. Nev-
ertheless they consider a collection of issues and models that are quite
relevant in this setting.
For now consider a single congestible resource with capacity K and N users
(or user segments), indexed by n ¼ 1,2, y, N. Let xn be user n’s average usage
or transmission rate, i.e., the number of requests per unit time, and let x
denote the vector of usage rates. Depending on the context, the usage rates x
may measure the throughput of bits, data packets or cells, or the throughput
of higher level data flows such as email messages or file downloads. Let s̄n
denote the average size of user n’s data files or packets. For now suppose that
all file or packet sizes are i.i.d., with average resource requirement equal to
one unit of capacity ðs̄n ¼ 1Þ and let X :¼ SNn¼1 xn be the average system load
or resource requirement per unit time. An important issue in the economic
analysis of best effort services is how users forecast the QoS performance of
their transmissions and based on what network state information. For now
assume that users lack dynamic system state information and instead forecast
or receive an announcement of aggregate performance statistics. Further
assume that the expected utility of user n, denoted by the differentiable
function un (xn, X, K), depends on the system performance only through the
average system load X and capacity K. It is increasing and concave in her
own usage rate xn ð@un =@xn 40 and @2 un =@x2n o0Þ: In the presence of conges-
tion cost the utility decreases in the load X ð@un =@X o0Þ and increases in the
capacity K ð@un =@K40Þ: This reflects the negative quality impact of conges-
tion and the fact that in the presence of demand variability, congestion effects
such as queueing delays or data loss occur intermittently even if the system is
on average underutilized, i.e., XoK. In summary, customer segmentation or
differentiation may by along several attributes: user (segment) n is charac-
terized by an (typically Poisson) arrival process for transmission requests at
rate xn that depends on prices and congestion levels, a random packet or file
size with mean s̄n and an utility function un (xn, X, K).
Congestion may impact QoS and customer utility via three effects: data
delay, loss and throughput or bandwidth limitations. Based on these at-
tributes one can identify the following types of utility models.
Delay-sensitive utility. Most studies focus on delay (response or through-
put time) as the quality attribute linking congestion and customer utility. It
measures the time from request to completion of a packet or flow trans-
mission, including time in buffers and in transmission. Delay adversely
affects the value of interactive applications such as online data bases and
also the value of time-sensitive information. Following Naor (1969) the
canonical delay cost model assumes that utility is additive in service values
and delay cost, i.e., a transmission of user n with value u generates utility u –
dn (t) where the delay cost function dn (t) increases in the total delay t. The
delay cost functions are typically assumed to be linear: dn (t) ¼ dnt, where
dn>0 is a constant marginal delay cost. Dewan and Mendelson (1990), Van
Mieghem (2000) and Afèche and Mendelson (2004) also consider increasing
convex delay costs. An important question is how users estimate the delay t.
For now assume that users lack dynamic system state information and
instead compute their expected utility based on their forecast or the pro-
vider’s announcement of the steady-state delay distribution. This is a com-
mon assumption in many queueing models that focus on static equilibrium
86 Philipp Afèche
analyses: the system state evolves dynamically but users’ transmission de-
cisions are static. This time-averaging sacrifices the ability to study the
dynamic relationship between congestion and pricing but yields tract-
ability and insights into system properties that are pertinent on longer time
scales.
In this case, a user’s expected utility per unit time has the form
un ðxn ; X ; KÞ ¼ V n ðxn Þ xn Dn ðX ; KÞ; n ¼ 1; 2; . . . ; N, (6)
where Vn (xn) is the value of usage rate xn under delay-free service and
Dn(X, K) is the expected delay cost of a transmission (the time average in
steady-state), which increases in the average load ð@Dn =@X 40Þ and de-
creases in the capacity ð@Dn =@Ko0Þ: The expected delay cost is defined as
Dn ðX ; KÞ :¼ E½d n ðT n ðX ; KÞÞ where the random variable Tn is the steady-
state delay and represents the QoS. In the case of uniform or undifferentiated
service, i.e., if all transmissions are treated the same, Tn ¼ T for all n.
The analysis of usage and pricing decisions hinges on a quantitative un-
derstanding of the relationship between the QoS performance of each class
and the characteristics of traffic demand (arrival processes and resource
requirements of transmission requests), network resources (number, topol-
ogy and capacity of links and buffers) and control mechanisms (e.g.,
scheduling disciplines such as FIFO, strict priorities, processor sharing,
etc.). For example, many studies that consider delay-sensitive customers
focus on queueing systems with infinite buffer size. For a resource modeled
as an M/M/1 system with infinite buffer and FIFO service, the delay dis-
tribution is
PrfTðX ; KÞ tg ¼ 1 expftðK X Þg (7)
and the expected utility under linear additive delay costs satisfies
1
un ðxn ; X ; KÞ ¼ V n ðxn Þ xn Dn ðX ; KÞ ¼ V n ðxn Þ xn d n .
K X
(8)
Under the common assumptions that the delay cost functions dn (t) are
unbounded and buffer size limitations are insignificant it is optimal to un-
derutilize the system since the expected delay cost is unbounded in the
utilization and the expected utility is negative at close to full utilization, i.e.,
Dn ðX ; KÞ ! 1 and un ðxn ; X ; KÞ ! 1 as X ! K:
The additive delay cost model (6) is inadequate if the delay cost and value
of a transmission are correlated, for example, as in financial or industrial
markets where delays in the execution of electronic trades may deflate the
investor’s expected profit, or when delays in video transmissions reduce the
image quality. Afèche and Mendelson (2004) introduce a generalized delay
cost structure that captures this interdependence by augmenting the stand-
ard additive model with a multiplicative component. In this model, a
transmission with value u and delay t has utility u qn ðtÞ d n ðtÞ where
the delay discount function qn (t) is decreasing and qn(0) ¼ 1. The expected
delay discount function satisfies Qn ðX ; KÞ :¼ E½qn ðTðX ; KÞÞ 1: It
decreases in the load X and increases in the capacity K. The expected util-
ity in this model is given by
un ðxn ; X ; KÞ ¼ V n ðxn Þ Qn ðX ; KÞ xn Dn ðX ; KÞ. (9)
As discussed below the structure of delay costs can significantly impact
the efficiency of a monopoly.
Loss-sensitive utility. Only a few studies focus on loss as the quality at-
tribute linking congestion and customer utility, which occurs in systems
with finite buffers. Loss may measure packet loss, which may impact audio
and video transmission quality (text does not tolerate data loss), or block-
ing of higher level service requests such as dialup connections. Marbach
(2004) models a system where loss is naturally interpreted as packet loss.
Courcoubetis and Reiman (1999) consider a system where loss represents
blocking. The system serves transmissions with a fixed bandwidth require-
ment over a random holding time. A service request is blocked whenever its
bandwidth requirement exceeds the supply. If all transmissions have equal
resource requirements then they face the same steady-state blocking prob-
ability B(X, K), which is the QoS measure in this context. It increases in the
average load X and decreases in the capacity K. Expected utility per unit
time is of the form
un ðxn ; X ; KÞ ¼ f n ðxn ; BðX ; KÞÞ; n ¼ 1; 2; . . . ; N (10)
where the function fn increases in the offered usage rate xn and decreases in
the blocking probability B. The effective usage rate of user n is xn ð1
BðX ; KÞÞ: For example, if there are K circuits without buffers, arrivals of
transmissions follow a Poisson process with rate K and each requires one
circuit for an exponentially distributed time, then the blocking probability
satisfies Erlang’s classic formula
ðX =KÞK =K!
BðX ; KÞ ¼ . (11)
P
K
k
ðX =KÞ =k!
k¼0
Studies with delay- or loss-sensitive utility functions operate on the

premise that congestion may be significant on the time scales where cus-
tomers make their usage decisions. This is natural for the best effort service
context as defined above where congestion avoidance mechanisms are in-
herently lacking.
Throughput-sensitive utility. The work discussed in Section 5 takes a
different approach. It hinges on the premise that congestion is and can be
maintained at low levels by letting users dynamically adapt their sending
rates, while transmissions are in progress in response to rapid congestion
feedback signals. These studies typically model the throughput or bandwidth
88 Philipp Afèche
as the only explicit utility function attribute. User n has utility un(xn) per
unit time where xn is naturally interpreted as the throughput of a flow in
packets or cells. Delay or losses play a role as a basis for congestion feed-
back signals, but they are not explicit attributes of utility functions.
User decisions. Economic analyses of best effort services consider various
user transmission decisions that mainly depend on the service design and
the availability of system state information. The vast majority of studies
assume that users do not strategize over the arrival times of their trans-
mission demands; random arrival processes are exogenous to the models,
with arrival rates depending on prices and expected QoS levels. The basic
user decision is whether or not to buy transmission service and if so, how
much. In studies of differentiated services (two or more service classes),
users also get to choose among all service menu options unless the provider
can distinguish among users and so limit their options. In systems where
users lack dynamic state information they make static transmission deci-
sions based on the forecast of steady-state QoS statistics; in static equilib-
rium analyses once for an infinite horizon, in adaptive pricing studies
repeatedly—once in each of successive demand periods. If users do have
dynamic state (queue length) information they make transmission decisions
repeatedly on a case-by-case basis.
Overview
The remainder of this section mostly focuses on delay as the explicitly
modeled utility function attribute, reflecting the emphasis in the literature,
although some results also hold in settings where congestion leads to data
loss. Most of the references assume the above described additive delay cost
model, but to provide an unified framework the discussion presents the
results for generic utility functions un ðxn ; X ; KÞ: Section 4.2 discusses the
basic principles of socially optimal and profit-maximizing static pricing and
capacity decisions for uniform service, assuming that users have perfect
steady-state delay forecasts and the provider knows demand. It starts with
the single-link case and transmissions with identical resource requirements
and later relaxes these assumptions. Section 4.3 considers two approaches
that go beyond the static framework to handle users’ and the provider’s
information constraints: providing congestion information for users and
iteratively updating price and delay forecasts over successive demand pe-
riods. Section 4.4 discusses decisions involved in creating differentiated best
effort services, including how many service classes to offer (service design),
how to allocate bandwidth to these classes (service production), and how to
price and target these classes to users (pricing), considering users’ choice
behavior in designing the price-QoS menu. Section 4.5 outlines studies on
optimal dynamic pricing, in particular on load-sensitive pricing, and ob-
serves the relative sparsity of work in this area. Section 4.6 concludes with
suggestions for future research.
The following discussion considers static pricing for uniform (undiffer-

entiated) service. It builds on the above model framework. It starts with the
case of a single link and transmissions with identical resource requirements
and later relaxes these assumptions. The mathematical analysis assumes
that users do not have dynamic system state information—they make static
decisions based on static forecasts of system performance—but several re-
sults apply in settings where they do.
Socially optimal congestion prices

The basic principle for socially optimal congestion pricing is simple and
well known: charge customers their congestion externality, i.e., the marginal
congestion cost they impose on the system. This ‘‘internalizes the external-
ity’’ and thereby aligns the system and individual user incentives, yielding
the socially optimal allocation. The system’s social welfare rate, W, is de-
fined as the rate of aggregate utility minus capacity cost C(K), which is
assumed to be increasing and convex
X
N
W ðx; KÞ :¼ ui ðxi ; X ; KÞ CðKÞ. (12)
i1
The problem of finding the socially optimal allocation for fixed capacity
level K is
max W ðx; KÞ s:t: X K. (13)
x0
Let lðx; K; bÞ ¼ W ðx; KÞ þ bðK X Þ be the Lagrangian where b is the

shadow price of the capacity constraint. The socially optimal usage rates x
and shadow price b satisfy
@un ðxn ; X ; KÞ XN
@ui ðxi ; X ; KÞ
¼ þ b ; n ¼ 1; 2; . . . ; N, (14)
@xn i¼1
@X
b ðK X Þ ¼ 0; b 0, (15)
X K. (16)
The marginal utility of a transmission equals its marginal congestion
externality cost plus the resource shadow price. Even if the network is
operated by a single provider, the solution of (14)–(16) cannot be imposed
in centralized fashion since usage decisions and information about customer
preferences are typically decentralized. The basic objective of congestion
pricing is to achieve the socially optimal usage rates in such a decentralized
environment in which self-interested customers make their own usage de-
cisions. The following discussion focuses on the structural properties of the
90 Philipp Afèche
socially optimal prices. Section 4.3 considers mechanisms for computing

these prices in the presence of informational constraints on utility functions
and network load.
Suppose a linear pricing scheme that charges each user a unit price p per
transmission. Given the price p and taking other user demands and the
capacity as fixed, user n solves
max un ðxn ; X ; KÞ pxn ; n ¼ 1; 2; . . . ; N. (17)
xn0
She increases xn up to the point where her marginal utility equals the
price plus the incremental congestion cost she imposes on her own traffic
@un ðxn ; X ; KÞ @un ðxn ; X ; KÞ
¼p ; n ¼ 1; 2; . . . ; N. (18)
@xn @X
If the social planner and customers know all the utility functions, the
social planner can choose prices that induce users to choose the socially
optimal usage rates x : If there are many users (N is large) whose individual
demands are small relative to the total load X, as in the Internet, the socially
optimal unit price is (approximately) the same for all users and satisfies
X
N
@ui ðx ; X ; KÞ
i
p ¼ þ b (19)
i¼1
@X
where the first term on the right-hand side is the marginal congestion ex-
ternality and the second is the resource shadow price at the socially optimal
usage vector. Combining (18) and (19) yields
@un ðxn ; X ; KÞ @un ðxn ; X ; KÞ
¼ p p
@xn @X
XN
@ui ðxi ; X ; KÞ
¼ þ b . ð20Þ
i¼1
@X
If users are relatively small, then (14) and (20), the optimality conditions
of the system and those of individual users, agree and p supports x as a
Nash equilibrium: no user n has an incentive to unilaterally choose a usage
rate other than xn given the price p and the anticipated expected delay cost
resulting from all other usage rates xn ; and the ex-ante anticipated ex-
pected delay costs are correct ex-post given the equilibrium usage rates. As
discussed below the Nash equilibrium concept makes strong assumptions
on information availability.
It is insightful to compare (19) for the cases with and without congestion
externalities. For the case with congestion externalities ð@un =@X o0Þ; as-
sume that un ðxn ; X ; KÞ 0 as X ! K for all n. In this case, the capacity
constraint is not binding at the optimal allocation, b ¼ 0; and the optimal
price (19) equals the marginal congestion externality. For example, in

the case of linear additive delay cost as in (6) the socially optimal price
satisfies
X
N
@ui ðx ; X ; KÞ X
N
@E½TðX ; KÞ
i
p ¼ ¼ xi d i . (21)
i¼1
@X i¼1
@X
Edelson and Hildebrand (1975), Mendelson (1985), Dewan and Mendelson

(1990), Ha (1998), Afèche and Mendelson (2004) are among studies that
derive this relationship for FIFO scheduling, Ha (2001) and Maglaras and
Zeevi (2003) consider (egalitarian) processor-sharing systems, and MacKie-
Mason and Varian (1995b) consider a generic nondiscriminatory scheduling
discipline. Papers that characterize the socially optimal static price for
systems where customers observe the queue length include Naor (1969),
Yechiali (1971, 1972), Knudsen (1972), Edelson and Hildebrand (1975) and
Lippman and Stidham (1977).
In the absence of congestion externalities ð@un =@X 0Þ the optimal price
equals the resource shadow price, i.e., p ¼ b ; which is positive only if it is
socially optimal to fully utilize the network. This corresponds to the case of
guaranteed services; compare with (5) in Section 3 and set the effective
band-widths ak ðs; tÞ ¼ 1: Thus, the socially optimal congestion price is sim-
ilar to a resource shadow price in the absence of congestion cost, except that
it is positive at lower than full capacity utilization.
Variations. The above analysis finds that a single linear usage price sup-
ports the socially optimal allocation, assuming that users (i) decide how
much to consume, (ii) are so small that the marginal congestion cost they
cause for theirPown traffic is negligible relative to that imposed on all users
ð@un =@X
N i¼1 @ui =@X Þ and (iii) have transmissions with identical
resource requirements.
This result still holds in the limiting case of a continuum of users that
have unit demand. Simply interpret un ðxn ; X ; KÞ as the aggregate utility of
user segment n where @un ðxn ; X ; KÞ=@xn is the utility of that group’s mar-
ginal user at given load and capacity, and the congestion cost term
@un =@X vanishes from the individual user’s optimality condition (18).
If individual users account for a significant fraction of network load and
are large enough to affect congestion levels, then a single price will not align
individual and system incentives. Based on (14) and (18) users should only
be charged the congestion cost they impose on all other users
X @ui ðxi ; X ; KÞ
pn ¼ ; n ¼ 1; 2; . . . ; N, (22)
ian
@X
which differs across users except if they are identical.

If users are small but consume services with different resource require-
ments, e.g., different file sizes, then congestion externalities, and hence the
92 Philipp Afèche
socially optimal prices also vary across customers. For example, if customer
n requires on average s̄n units of resources per transmission andPthe utility
depends on the usage vector only through the total load X ¼ N n¼1 xn s̄n
then the socially optimal prices based on (19) are
!
X N
@ui ðxi ; X ; KÞ

pn ¼ s̄n ; n ¼ 1; 2; . . . ; N. (23)
i¼1
@X
If the provider cannot a priori tell apart different users, then prices should
be aPfunction of actual data volumes. For example, given the tariff p ðsÞ ¼
sð N
i¼1 @ui ðxi ; X ; KÞ=@X Þ where s is the actual data volume transmitted,
the expected price of user n is pn as in (23).
Time scales for congestion price fluctuations. The above analysis computes
the optimal static price based on the average congestion impact of trans-
missions. This raises several issues pertaining to the time scale for price
changes. For one, it may be desirable for prices to more accurately reflect
the differential congestion impact of different transmissions, even while
keeping the average price constant. Gibbens and Kelly (1999a) propose and
show how to charge each packet individually its sample path shadow price,
which is the exact amount of congestion it inflicts on the system at each
resource. They show conditions for when the rate of sample path shadow
prices so derived equals the average congestion impact. Another issue is
that the computation of optimal congestion prices also requires information
on users’ utility functions, specifically on their sensitivity to delay or other
QoS metrics. In the setting of Gibbens and Kelly (1999a) the cost of con-
gestion is incurred by and known to the system, unlike here where it is
borne by users and likely unknown to the system. Thus, even if sample path
shadow prices correctly reflect the congestion impact of individual packets,
the problem of correctly evaluating the resulting cost impact on users still
remains. The adaptive price mechanisms discussed in Section 4.3 may help
the provider learn the utility functions and find the right prices iteratively.
Congestion prices may also fluctuate under optimal dynamic price policies,
discussed in Section 4.5.
Network. Mendelson and Whang (1994) and Masuda and Whang (2002)
study the natural network extension of this problem, finding that for fixed
capacity the socially optimal congestion prices are additively separable in
the node prices and that they induce the desired usage rates and routes if
customers can choose their routes.
Socially optimal congestion prices and capacity investment

Now consider the joint problem of choosing the efficient use and capacity
level. Let x ðKÞ; X ðKÞ; p ðKÞ and b ðKÞ be the socially optimal usage vec-
tor, network load, unit price and resource shadow price, respectively, as a
function of capacity, and let LðKÞ :¼ W ðx ðKÞ; KÞ þ b ðKÞ ðK X ðKÞÞ
be the maximum value of the Lagrangian as a function of capacity. The

marginal change in social welfare as a function of capacity satisfies
@W ðx ðKÞ; KÞ
L0 ðKÞ ¼ þ b ðKÞ
@K
XN
@ui ðxi ; X ðKÞ; KÞ
¼ C 0 ðKÞ þ b ðKÞ. ð24Þ
i¼1
@K
In the absence of congestion cost utility is independent of capacity
ð@ui =@K 0Þ and p ðKÞ ¼ b ðKÞ; so it is optimal to expand capacity if and
only if the price exceeds the marginal capacity cost
L0 ðKÞ ¼ p ðKÞ C 0 ðKÞ. (25)

If there is a congestion cost, then b ðKÞ 0 and the marginal value of
capacity depends on the impact of network load X and capacity K on
customer utility. If the utility depends on X and K only through the average
excess capacity K–X, as in (8) for an M/M/1 system, then @ui =@K ¼
@ui =@X and (25) holds, i.e., the optimal price equals the marginal capacity
cost (Mendelson, 1985; Dewan and Mendelson, 1990). If the utility depends
on X and K only through the average utilization X/K, then @ui =@K ¼
@ui =@X X =K; so (19) and (24) yield
X ðKÞ
L0 ðKÞ ¼ p ðKÞ C 0 ðKÞ, (26)
K
which implies that the optimal price at the socially optimal capacity exceeds
the marginal cost: expanding capacity increases welfare if and only if the
price, multiplied by the utilization, exceeds the marginal capacity cost (Mac-
Kie-Mason and Varian, 1995b).
Network. Masuda and Whang (2002) investigate the capacity expansion
problem for a decentralized system with general network topology. They
show that the optimal capacity investment principles carry over to a de-
centralized network with fixed demand and routing whereas this need no
longer hold under routing choice; they propose a modified capacity analysis
for such systems. Korillis et al. (1995) consider pricing and capacity de-
cisions for decentralized parallel link networks where customers have rout-
ing choices; they show that the optimal decentralized solution coincides
with that under centralized control.
In summary, the socially optimal congestion price plays a dual role: it
measures the social cost of increased usage at any given capacity level and
serves as a signal to determine whether to increase capacity.
Perfect competition
MacKie-Mason and Varian (1995b) find that in a competitive market in
which many small providers charge connection fees and usage prices the
equilibrium yields the social welfare maximizing usage and capacity levels.
94 Philipp Afèche
Free usage
It is well known that the absence of usage prices leads to resource overuse
because users do not internalize the congestion externality. Studies that de-
rive this classic result under FIFO scheduling for the case of unobservable
system state include Mendelson (1985) and for the case of observable system
queue include Naor (1969), Knudsen (1972), Yechiali (1971, 1972), Lippman
and Stidham (1977). Two exceptions to this result are noteworthy. First,
Hassin (1985) shows that social optimality can be achieved without usage fees
if the service discipline is last come first served preemptive resume (users can
observe the queue length). The key customer decision under this mechanism
is when to renege (leave the system without obtaining service) rather than
whether to request service, a decision made by the customer who is last in
queue and who therefore imposes no externality. As a result, the last cus-
tomer’s incentive is aligned with that of the system: the queue only grows as
long as the last customer’s service value exceeds her expected delay cost.
While this mechanism is probably not of practical importance for data
communication services, it nicely highlights the link between the service dis-
cipline and the distortion under individual optimization. Second, Giridharan
and Mendelson (1994) show that in the presence of both negative (conges-
tion) and positive (network) externalities usage levels may be close to socially
optimal even if usage is free, since the positive at least partially offset the
negative externalities. See Westland (1992) for a further study of socially
optimal pricing under congestion and network externalities.
Monopoly price and capacity

The pricing and capacity decisions of a profit-maximizing monopoly
provider compare as follows to the socially optimal ones. The profit-max-
imizing prices in the case of users with identical resource requirements—the
analog of (19), typically do depend on users through their demand elasticity.
If the provider cannot tell apart individual users, she can only charge a
single price that may result in some users transmitting more and others less
traffic than what is optimal if the provider can price discriminate. The
respective analysis follows from the classic multi-product monopoly anal-
ysis with the direct cost replaced by the delay cost functions. For a fixed
capacity level and linear pricing a monopoly serving a single user segment
will typically charge a higher price and reduce usage (quantity) compared to
socially optimal levels, consistent with the standard result in the absence of
congestion externalities. Naor (1969) and Mendelson (1985) establish this
result for the case of customers with identical delay cost under the additive
delay cost structure given by (6). However, Edelson and Hildebrand (1975)
give counterexamples that show this property need not hold if customers
differ in their cost of delay. Afèche and Mendelson (2004) provide analyt-
ical conditions on the delay cost structure and the price elasticity for when
the monopoly price exceeds, equals or is lower than the socially optimal
price.
Under linear pricing a monopoly typically restricts capacity below the

socially optimal level. Moreover, since the optimal capacity utilization is
below 100% in the presence of congestion, (25) and (26) imply that the
provider incurs a loss at the socially optimal price and capacity levels if the
capacity cost function is linear (or more generally, not ‘‘too convex’’).
Mendelson (1985), Dewan and Mendelson (1990) and Maglaras and Zeevi
(2003) observe this point. Dewan and Mendelson (1990) study how the
budget deficit at the socially optimal capacity depends on the delay cost
function. This implies that internal data communication services with the
goal of maximizing the overall net value to the organization should not be
evaluated as profit centers or not use linear pricing.
There are various approaches to mitigate the problem of inefficient use
and capacity investment under a monopoly provider. Under Ramsey pricing
linear prices are chosen to maximize social welfare subject to a constraint
on the provider’s profit, usually taken to be a balanced-budget constraint.
Ramsey pricing is mostly relevant for regulated firms. See Mitchell and
Vogelsang (1991) for a survey of Ramsey pricing in traditional telecom-
munication markets. Another approach is to use a two-part tariff of the
form Pðxn Þ ¼ a þ p xn where the usage-insensitive access or subscription
fee a serves to recover the cost and the linear usage charge p regulates
congestion. If users are homogenous then the monopoly can extract all
consumer surplus and maximizes total welfare by choosing the socially
optimal p equal to the marginal congestion cost, and by setting a to extract
the remaining surplus. This is the classic result of Oi (1971). In the more
realistic case of heterogenous customers, the monopoly can typically not
extract all surplus. In this case the optimal usage price is below the marginal
congestion cost if the marginal customer’s bandwidth demand exceeds that
of the average customer, e.g., demand for low-value video vs. high-value
email, and vice versa (see MacKie-Mason and Varian, 1995b). The provider
can typically improve upon a single two-part tariff by offering a more
general non-linear tariff P(x) or a menu of multiple ðM NÞ two-part
tariffs Pm ðxn Þ ¼ am þ pm xn from which customers select the option that
suits them best. Masuda and Whang (2006) identify conditions where the
optimal menu of tariffs with a simple ‘‘fixed-up-to’’ structure perform as
well for a monopoly as any nonlinear pricing schedule. See Mitchell and
Vogelsang (1991) and Wilson (1993) for surveys of nonlinear pricing.
The static posted price mechanisms of the preceding section involve two
steps. The provider first determines the optimal capacity and usage rates
and the prices that induce customers to choose these rates. Each customer
then determines her utility-maximizing usage given the posted prices and
her expected delay cost, which depends on system properties and other
users’ anticipated usage rates. The notion that an optimal allocation can be
96 Philipp Afèche
supported as a Nash equilibrium in user decisions assumes that the provider

and users are omniscient and have unbounded rationality. To compute the
optimal price and capacity levels the provider must have complete infor-
mation on demand (the number, preferences and traffic characteristics of
customers) and system performance characteristics (the relationship be-
tween usage levels and QoS metrics such as delay). To determine her utility-
maximizing transmission rates each user must have complete information
on system parameters and other users’ preferences; they need this infor-
mation to form rational expectations about equilibrium usage rates and
delay cost since the network does not guarantee the QoS. By contrast,
customers do not require this information if the provider delivers QoS
guarantees to each user regardless of overall network conditions.
These information requirements pose significant implementation chal-
lenges. The rapidly evolving data communications market makes it difficult
for providers to accurately estimate utility functions, particularly delay-sen-
sitivity and other quality preferences, except perhaps in internal networks
with known users. It is also unrealistic to require from customers the ability
to accurately forecast their congestion cost over their usage horizon. While
the optimal congestion prices can be formulated in theory, it is not clear how
or even whether they can be computed and implemented in practice.
A couple of approaches that may help overcome these informational con-
straints have been considered. One is to use auction mechanisms, where users
submit price bids before the provider determines payments and allocations.
Section 6.1 discusses auctions and their potential for revealing demand in-
formation. Two approaches that have mainly been considered for posted price
mechanisms include providing congestion information for users and adaptive
pricing and delay forecasting. The former focuses on improving the informa-
tion available for any single user-network interaction, the latter focuses on
adaptation based on information gained through repeated interactions.
Congestion information for users

This approach focuses on giving users the required forecast information on
system state and/or performance statistics. In models where users do not
have dynamic queue length information the assumption that users can per-
fectly forecast steady-state delay statistics on their own can alternatively be
interpreted as the provider posting information on the system’s (uncondi-
tional) steady-state delay distribution. The only requirement for this alter-
native interpretation is that the ex ante information be consistent with the ex
post realized performance. If users do receive queue-length information be-
fore submitting a transmission, they can improve their delay forecast accu-
racy for any given transmission by basing their QoS estimate on the
conditional delay distribution. Studies that consider various questions for this
observable queue case include Naor (1969), Yechiali (1971,1972), Knudsen
(1972), Edelson and Hildebrand (1975), Lippman and Stidham (1977), and
Afèche and Ata (2005) for uniform FIFO service; Balachandran (1972), Adiri
and Yechiali (1974), Dolan (1978) and Alperstein (1988) for strict priority
scheduling; Hassin (1985,1986) for last-come-first-served preemptive repeat;
Altman and Shimkin (1998) for processor sharing; and Maglaras and Zeevi
(2005) for (egalitarian) processor sharing of a randomly fluctuating amount
of bandwidth. Most of these papers study static price policies, just like the
one considered above, but they allow customers to react to a given set of
prices based on dynamic congestion information. Section 4.5 considers stud-
ies of optimal dynamic price policies. Hassin (1986) studies whether making
dynamic queue-length information available to customers increases or re-
duces the total surplus and the provider revenue, compared to the case where
customers do not have this information. The answer is ambiguous and de-
pends on the system utilization. Maglaras and Zeevi (2005) show that in large
scale systems, i.e., that process large volumes of data and have proportionally
large processing capacity, real-time delay information increases system rev-
enues and the resource utilization rate while simultaneously decreasing the
average delay experienced by the users.
Adaptive pricing and delay forecasting

This approach focuses on users and the provider adapting their decisions
based on information they gain over repeated interactions, assuming users
cannot observe the current queue length before deciding whether or not to
transmit. Users adjust their delay forecast and demand based on their past
delay experience and past prices, and the provider learns about demand and
adjusts prices based on the observed user response. Analyses of this ap-
proach examine the stability of dynamic adaptation algorithms under al-
ternative update rules for prices and/or delay forecasts.
In Stidham (1992) and Rump and Stidham (1998) the provider does know
user demand attributes and fixes a price. Both studies focus on users up-
dating their delay cost expectation and arrival rate based on past experi-
ence, in Stidham (1992) the provider also updates the capacity. Their basic
approach is similar to the classic tatonnement process. Time is partitioned
into successive periods; in each period users update their delay forecast and
base their usage decisions on the updated forecast. Demand is assumed to
be stationary for a sufficient number of time periods so that the system
converges to steady-state before conditions change.
Gupta et al. (1997) and Masuda and Whang (1999) analyze how the
socially optimal prices and usage rates can be attained in a resource network
through an adaptive algorithm when both the provider and customers have
imperfect information. The provider does not know the demand functions
and customers do not know the system delay. Their basic approach is the
same as in Stidham (1992) and Rump and Stidham (1998), but the provider
also updates prices, in addition to customers adjusting their usage rates.
The pricing scheme is linear with node-specific unit prices for each trans-
mission. Low and Varaiya (1993) and Thomas et al. (2002) also study such
adaptive mechanisms, but they focus on guaranteed services where each
98 Philipp Afèche
user’s utility is independent of others’ consumption and user decisions only

depend on prices.
Gupta et al. (1997) and Masuda and Whang (1999) consider a capacity-
constrained queueing network that serves multiple delay-sensitive user
classes, each characterized by a network route and an aggregate utility func-
tion with a linear delay cost parameter. The provider does not know the utility
functions but knows delay cost parameters. The studies differ in several as-
sumptions. Gupta et al. (1997) consider delay cost parameters drawn from a
distribution, generic transmission time distributions and multiple predefined
priority classes at each node with priority class-dependent prices, which re-
quires simulation and approximations to evaluate delays. Masuda and
Whang (1999) consider an identical delay cost parameter for all users, FIFO
service at each node and assume Poisson arrivals and exponential transmis-
sion times which yields analytical expected steady-state delay formulae. As a
benchmark the studies characterize the socially optimal arrival rates and the
prices and expected delays that support these arrival rates as a stochastic
equilibrium (cf. Stahl and Whinston, 1994), which is defined by two prop-
erties. Best-reply, i.e., the arrival rates maximize users’ expected utility given
the posted prices and anticipated delays, and accurate expectation, i.e., the
anticipated delays are the correct ex-ante expected delays given the resulting
arrival rates. (Every stochastic equilibrium is a Nash equilibrium and vice
versa. The concepts differ in their informational assumptions.)
The following measurement-based rules govern price and delay forecast
updates. The provider updates prices using exponential smoothing. For
example, for a single node with FIFO service
^
pðt þ 1Þ ¼ a pðtÞ þ ð1 aÞ pðxðtÞÞ, (27)
where a 2 ½0; 1 is the exponential smoothing parameter, p (t) and x (t) are the
^
period-t posted price and usage rates, and the period-t price estimate pðxðtÞÞ
evaluates the marginal delay cost, the right-hand side of (21), at the actual
time-average usage rates. The price adjusts upwards (downwards) if the prior
period price is lower than (exceeds) the marginal delay cost estimate. In each
period users choose their transmission rates based on the posted prices and
their delay forecasts, and the resulting usage rates and delays are used for the
next round of updates. The studies differ somewhat in how users forecast
delays. In Gupta et al. (1997) the network estimates and announces the next
period expected delays based on current time-average estimates. In Masuda
and Whang (1999) users estimate the next period expected delays based on
their estimate of next period arrival rates and the analytical steady-state delay
formulae. For example, for a single M/M/1 node with FIFO service
1
X^ ðt þ 1Þ ¼ b X^ ðtÞ þ ð1 bÞ X ðtÞ and Tðt
^ þ 1Þ ¼ ,
K X^ ðt þ 1Þ
(28)
where X^ ðt þ 1Þ and Tðt

^ þ 1Þ are the arrival rate and expected delay estimates
for period t+1, X (t) is the actual current period average arrival rate and
b 2 ½0; 1 is an exponential smoothing parameter.
The following insights on the dynamic network behavior and its eco-
nomic performance emerge under these update rules. Gupta et al. (1997)
develop and run several simulation models that show prices and arrival
rates generally converging quickly, also in scenarios where demand alter-
nates between low and high states (Gupta et al., 1999), to what they suggest
are approximately optimal levels. Their simulations also show that cus-
tomer and total benefits of priority pricing over uniform pricing and free
access may be significant. Masuda and Whang (1999) focus on the issue of
network stability. They show that in the absence of exponential smoothing
(a ¼ b ¼ 0) the network may be unstable and continuously oscillate be-
tween low- and high-utilization periods, with the low-utilization periods
reminiscent of Yogi Berra’s famous quote, ‘‘Nobody goes there any more
since it is so crowded.’’ They establish analytical conditions on the demand
function, delay cost and smoothing parameters for the network to be locally
stable at the socially optimal equilibrium, and they prove that a single node
M/M/1 ‘‘network’’ is stable for a sufficiently large smoothing parameter.
These mechanisms resemble the adaptive flow control and pricing mech-
anisms considered for flexible bandwidth-sharing services in Section 5, e.g.,
Kelly et al. (1998). A main difference is that there flow rates and prices are
updated perhaps every tens of milliseconds whereas here these updates oc-
cur on a longer time scale, perhaps on the order of hours or days. The
setting here implicitly assumes that users react slowly to perceived conges-
tion, inasmuch as the system operates long enough at a particular set of
load values for steady-state to be (approximately) achieved in each time
period.
These adaptive pricing mechanisms have the potential to dramatically
reduce the information requirements: the provider need not know the de-
mand functions and users need not know each others’ preferences. In ad-
dition, prices can be calculated in a decentralized manner at each network
node based on local flow rate and delay measurements. However, several
theoretical and practical issues require further study. For one, the conver-
gence behavior and its sensitivity to the network characteristics are not fully
understood. Under what conditions is the system globally stable, and if
there are multiple equilibria which one will be reached? Kelly et al. (1998)
and Thomas et al. (2002) prove that their adaptive algorithms converge.
However, they consider systems without congestion externalities that have
unique socially optimal allocations; systems without these features need not
have the same stability properties. The system considered by Stidham
(2003) considers a generic delay cost model for a network with undiffer-
entiated best effort service that serves users who differ in their delay sen-
sitivity. His analysis suggests that the total system benefit function may
have multiple stationary points and that adaptive algorithms may converge
100 Philipp Afèche
to a local rather than global maximum, depending on the starting point.

Stidham (2003) develops this result for a continuous-time adaptive flow
control algorithm very similar to Kelly et al. (1998), but his model frame-
work is general enough to suggest similar stability issues are likely to be
prevalent in the setting considered in this section.
Another set of issues concern the convergence speed of such adaptive
mechanisms. How long does it take for prices to converge in stationary and
nonstationary environments? How do the answers depend on the update
period length and update rules? How should these mechanisms be modified
and how would they perform if the provider lacks information not only on
the value functions, as in Gupta et al. (1997) and Masuda and Whang
(1999), but also on delay cost parameters and other quality preferences?
Gupta et al. (2000) contribute in this direction: they propose a nonpara-
metric statistical technique for estimating users’ delay cost parameters
based on their observed choices and show via simulation that substituting
these estimates for the ‘true’ parameters in the price updates only results in
a minimal efficiency loss. Afèche and Ata (2005) analytically characterize
the optimal dynamic pricing policy under Bayesian updating for a stylized
single-node system that serves customers with unknown delay cost param-
eter distribution in a stationary environment. They characterize the time to
learn and the probability that the provider eventually learns the correct
distribution.
From a practical perspective, customers may find it difficult to plan
communication service purchases given the price fluctuations inherent in
these adaptive approaches. It may be useful to run such mechanisms in
representative test markets to determine ‘‘appropriate’’ prices for low- and
high-demand conditions which are then to be used in actual markets.
The discussion in this section has so far focused on uniform or undiffer-

entiated service—the network treats all transmissions the same. For uniform
service differentiated pricing is required for social optimality only if trans-
missions have different resource requirements, e.g., as in (23), whereas the
profit-maximizing pricing scheme may involve price discrimination even for
homogenous transmissions.
If users are heterogenous in terms of their traffic characteristics and/or
QoS preferences it may be beneficial to offer them a menu of delay or loss
differentiated service classes and to differentiate prices not only based on
resource requirements but also in function of the QoS. Such differentiation
typically attains a more efficient resource allocation; it also allows the pro-
vider to increase her revenue by better segmenting the market and giving
users the option to pay more for higher QoS. Chao and Wilson (1987) and
Wilson (1989) are classic references on priority service in the economics
literature.
Service differentiation and best effort service

The Internet presently offers only a single class of best effort service.
While best effort was conceived with nondiscrimination in mind ‘‘service
differentiation and best effort service are conceptually orthogonal’’ (Gevros et
al., 2001). In the context of the best effort service model it is possible to
differentiate services by defining two or more traffic classes that are treated
differently within the network, have qualitatively different QoS character-
istics (e.g., average delay or loss ratios) tailored to elastic applications with
different QoS preferences and are priced accordingly. Analyses of differ-
entiated best effort services consider a range of control mechanisms for
achieving differential QoS metrics, including link partitioning, routing and
packet scheduling mechanisms. The notion is to have only a few classes,
each chosen by multiple individual flows, in contrast to guaranteed services
where QoS is managed for each individual flow via admission control,
resource reservation and traffic enforcement mechanisms. In differentiated
best effort services the QoS of each service class is not expressed in terms of
strict performance bounds that are pre-defined for each individual flow, as
in guaranteed services, but rather in terms of expected performance metrics
averaged over all flows in a service class. Depending on how bandwidth is
assigned to a service class its QoS metrics may depend on all flows within
the network or only on its own flows.
Proposals
A couple of QoS and pricing proposals build on the notion of a multi-
class best-effort network. The DiffServ architecture proposal of the IETF
aims to provide statistical QoS guarantees for traffic aggregates, as opposed
to strict guarantees for individual data flows. As such it can be viewed as
augmenting the basic elements of a multi-class best effort network—traffic
is classified into a few categories and all packets within a class are handled
the same—with demand regulation mechanisms that enforce certain traffic
inflow constraints and help achieve statistical class-level QoS targets.
DiffServ does not by itself say anything about the assignment of priorities
to different traffic classes and pricing. Odlyzko (1999b) proposes an
approach called Paris Metro Pricing (PMP) for Internet pricing, with the
name due to its resemblance to the pricing structure of the Paris Metro.
The concept is to partition the Internet into several logically separate
‘‘channels’’ for different service classes with higher prices for using the
better provisioned and therefore (allegedly) less congested classes. PMP is
similar to DiffServ in that it considers multiple best effort service classes,
but unlike DiffServ, it does not consider the use of technological demand
regulation mechanisms to attain some (class-level) QoS targets. In PMP
pricing is the primary demand regulation mechanism and it focuses on
how to price and assign bandwidth to the different classes whereas DiffServ
does not.
Issues
The economic analysis of multi-class best effort services should address
several interrelated provider decisions: how many service classes to create
(service design), how to allocate bandwidth to these classes (service pro-
duction), and how to price and target these classes to users (pricing). An
important consideration for these provider decisions is how much infor-
mation she has on user preferences. In particular, if the provider only has
aggregate information about user preferences, but no information on QoS
preferences of individual users, then she cannot distinguish among them ex
ante; all users can and will choose among all options on the provider’s
service class menu in line with their own self-interest. This behavior gives
rise to incentive-compatibility constraints, which the provider must consider
in designing her price-QoS menu.
The vast majority of papers consider pricing decisions for a given service
design, defined by a fixed number of service classes, and for given supply
allocation mechanisms at a single network link, for example: how to price
two service classes delivered over a single server with a strict nonpreemptive
priority discipline? A few of these papers investigate the economic value of
differentiation for given supply allocation mechanisms by comparing op-
timally priced systems with different numbers of service classes, or jointly
optimize tariff design and the number of service classes. Hardly any papers
jointly optimize or compare alternative combinations of service design,
service production and pricing.
Number of service classes

Adiri and Yechiali (1974), Alperstein (1988), Cocchi et al. (1993), Gupta
et al. (1997), Bajaj et al. (1998), Odlyzko and Fishburn (1998), Van Mieghem
(2000), Afèche and Mendelson (2004), Afèche (2004) and Zhang et al. (2006)
among others study the economic benefit of two or more service classes
compared to uniform service. All demonstrate revenue and/or total surplus
gains as a result of adding a class, in various settings. The marginal benefit
from an additional class typically drops (Afèche and Mendelson, 2004), sug-
gesting that a few classes may generate most of the benefits. However, ‘‘more
differentiation’’ need not always be beneficial: Afèche (2004) shows that
pooling different user segments in a single class may increase revenues if users
have private information about their preferences.
Bandwidth allocation to service classes

In multi-class best effort services the specification of bandwidth alloca-
tion has a significant impact on the user-perceived quality and therefore
plays an important role in pricing studies. Bandwidth allocation occurs at
multiple links along a route. Reflecting the emphasis in the literature this
survey focuses on a single link ‘‘network’’. Bandwidth can be allocated to
service classes using a combination of two basic methods: reserving for each
class exclusive access to a dedicated (set of) channel(s); or letting two or
more classes share a common pool of communication channels according to

certain packet scheduling and routing mechanisms. Thus, exclusive access
results in a static and shared access yields a dynamic channel allocation.
Exclusive access is typically simpler to manage but offers less flexibility and
has inferior delay performance compared to shared access with appropriate
scheduling and routing control.
Studies of price and service differentiation have so far focused either on
link partitioning and routing issues for multi-channel systems, or on pri-
ority scheduling policies for single-channel systems; the joint problem of
pricing, routing, and priority scheduling for multi-channel systems has
hardly received any attention.
Priority scheduling mechanisms play a major role in delivering differenti-
ated services using pooled capacity. The most common are strict or absolute
priority disciplines, where higher priority traffic waiting for transmission al-
ways gets exclusive bandwidth access, and processor sharing disciplines where
all priority levels get some bandwidth access based on their relative priority
index. The majority of pricing studies focus on strict priority disciplines, which
minimize the system’s average delay cost in single-node systems and are
therefore more efficient than processor-sharing disciplines when users are de-
lay-sensitive. However, processor-sharing systems typically exhibit a lower
delay variance across service classes, which makes them attractive for fairness
objectives.
As noted above most papers consider pricing for a given bandwidth al-
location scheme. Exceptions include Odlyzko and Fishburn (1998) who
compare the revenues for dedicated and pooled allocations of two service
classes; Van Mieghem and Van Mieghem (2002) who consider generalized
strict priority and processor-sharing disciplines for a single channel; and
Afèche (2004) who jointly optimizes over prices and scheduling policies for
a provider who serves two customer segments over a single channel.
Multiple channels: link partitioning and routing. An appealingly simple
approach is to partition the capacity of a link into multiple channels and
statically allocate each service class to a dedicated channel pool with FIFO
queueing at each channel pool. This corresponds to the basic design for
Odlyzko’s (1999b) PMP proposal. A major question is how much capacity
to allocate for each class. Odlyzko assumes a monopolistic setting. Gibbens
et al. (2000) address the issue of pricing and capacity allocation also for a
duopoly market.
Other studies consider how to route multiple service classes that share
access to multiple channels with given capacities. Bell and Stidham (1983)
and Bradford (1996) characterize the equilibrium behavior of customers in
such multi-server queueing systems.
One channel: processor sharing. As noted above processor sharing sys-
tems with differentiated services have received relatively little attention in
pricing studies. Haviv and van der Wal (1997) consider equilibria for users’
priority choice problem in a discriminatory processor sharing system. Van
Mieghem and Van Mieghem (2002) study and compare generalized strict
priority and processor-sharing disciplines. Maglaras and Zeevi (2005) study
profit-maximization for a system with guaranteed and best effort service
where best-effort traffic shares the dynamically fluctuating leftover capacity
not tied up by guaranteed service.
One channel: strict priorities. The majority of papers consider systems
with strict or absolute priorities. Bohn et al. (1994) propose how to im-
plement priority levels based on existing Internet protocols without con-
sidering pricing and incentive issues. Cocchi et al. (1993) simulate example
networks with strict priority service and different traffic types, and they
determine prices that make all users better off than under uniform service.
The following studies focus on social optimization for the linear delay cost
case. Kleinrock (1967) first studied priority pricing by ignoring customer
incentives. Marchand (1974), Ghanem (1975), Dolan (1978), and Mendel-
son and Whang (1990) focus on and provide a thorough understanding of
the incentive-compatible and socially optimal price-scheduling mechanism:
the socially optimal and incentive-compatible scheduling policy is the tra-
ditional cm rule—known to minimize the system’s average delay cost rate—
and each customer class is charged her externality. If customer types have
private information on their service requirements, e.g. file size, the socially
optimal charging scheme consists of a static a priori payment and an ex post
payment based on the actual file size. Gupta et al. (1997) consider a network
with priorities (see Section 4.3). Van Mieghem (2000) generalizes the anal-
ysis of Mendelson and Whang (1990) to the case of convex delay cost, and
for the same cost structure Van Mieghem and Van Mieghem (2002) study
and compare pricing for generalized strict priority and processor-sharing
disciplines.
Interestingly, the design of revenue-maximizing and incentive-compatible
price-scheduling mechanisms has only received limited attention so far. Rao
and Petersen (1998) consider incentive-compatible pricing for welfare and
revenue-maximization but implicitly assume, as opposed to optimize over,
the scheduling policies. Afèche (2004) jointly optimizes over prices and
scheduling policies for a revenue-maximizing provider who serves two cus-
tomer segments. He shows that the revenue-maximizing and incentive-
compatible scheduling policies have novel features and may significantly
differ from those under social optimization. In particular, one such policy
involves strict priorities, but with the insertion of optimal strategic delay and
another reverses the priority ranking compared to the socially optimal
ranking. Afèche’s (2004) analysis assumes linear delay cost. Yahalom et al.
(2006) extend his treatment by allowing for convex delay cost and show that
the use of strategic delay may also be optimal in this case.
Papers that study auctions for strict priority service include Balachandran
(1972), Glazer and Hassin (1985), Lui (1985), Hassin (1995), Stahl (2002),
Afèche and Mendelson (2004), and Kittsteiner and Moldovanu (2005); they
are discussed in Section 6.1.
The above studies all measure QoS by delay. Marbach (2004) considers
static pricing and bandwidth allocation for static loss priorities for the case
where QoS is measured by throughput and packet loss.
Optimal dynamic pricing policies have received little attention, maybe in

part due to the mathematical challenges inherent in their analysis.
Load-sensitive dynamic pricing

Low (1974) and Chen and Frank (2001) characterize optimal dynamic
policies where prices depend on the queue length, which is visible to users.
Demand-sensitive dynamic pricing

Demand-sensitive dynamic pricing appears not to have received any ex-
plicit attention. Time-of-day pricing could be studied in a similar frame-
work as discussed in Section 3.5 by considering an exogenous time
dependent arrival process and ignoring intertemporal effects. However, this
fails to capture users’ choices of when to transmit as they anticipate price
and delay fluctuations over time. The analysis of this issue is likely quite
complicated due to the presence of congestion externalities.
The basic principle for socially optimal pricing is to charge customers the
marginal congestion cost they impose on the system. This price also serves
as a signal for optimal capacity expansion. For uniform service differen-
tiated pricing is socially optimal only if transmissions have different re-
source requirements whereas the profit-maximizing pricing scheme may
involve price discrimination even for homogenous transmissions. If users
are heterogenous in terms of traffic characteristics and/or QoS preferences
it may be beneficial to offer delay or loss differentiated best effort service
classes with appropriately differentiated prices. Doing so involves decisions
on service design and bandwidth allocation to the service classes. If the
provider has no information on QoS preferences of individual users, then
she must consider incentive-compatibility constraints that reflect users’
choice behavior in designing her price-QoS menu. In this case, it may be
optimal for a profit-maximizing provider to ‘‘pool’’ multiple segments in a
single service class or, conversely, to increase the amount of differentiation
between classes relative to the socially optimal levels.
An important issue in the economics of best effort services is the infor-
mation available to users and the provider. Users have to form QoS fore-
casts since the network inherently provides neither quality guarantees nor
congestion feedback signals, and the provider needs to understand users’

utility functions, and in particular their QoS sensitivity. Static equilibrium
analyses assume that users have perfect steady-state delay forecasts and the
provider knows demand. Two approaches that go beyond this static frame-
work include providing congestion information for users and iteratively
updating price and delay forecasts over successive demand periods.
further research.
Service design. Most best effort service pricing studies take as a starting
point a given set of service classes and supply allocation mechanisms. Since
users of best effort services are inherently flexible in terms of their QoS
requirements, there is a research opportunity for jointly optimizing pricing,
service design and supply allocation mechanisms. Afèche (2004) illustrates
this approach for a single server system with two customer segments and
develops a stepwise solution approach that can be used to study systems
and markets with more general characteristics.
Loss as QoS. Virtually all studies consider delay as the only utility
function attribute, whereas data loss has hardly received any attention as a
measure of QoS. While some best effort applications, certainly text-based
ones, may not be loss-tolerant, others may be, and it would be interesting to
identify similarities and differences in the analysis and results for delay- and
loss-tolerant applications.
Network. The majority of studies consider a single-link ‘‘network.’’
While some results naturally extend to multi-link networks, e.g., the op-
timal prices for a given transmission may be additive in the prices at each
link along its route, issues relating to service differentiation in a network
and routing control call for further study.
Value of load-sensitive and demand-sensitive dynamic pricing. The prevalent
static equilibrium analysis framework generates a lot of insights into structural
properties of prices and capacity levels, but it sacrifices the ability to study the
dynamic relationship between user decisions, congestion and pricing. The an-
alytical difficulty notwithstanding, it would be interesting to shed more light on
the structural properties and the value of optimal dynamic policies.
5 Pricing flexible bandwidth-sharing services
Features
The guaranteed and best effort service contracts considered so far typically
have only static parameters: guaranteed or expected QoS parameters, traffic
parameter constraints and tariffs are fixed at the start of the contract. Users
make all decisions concerning an individual flow or a collection of flows
covered by their contract at the start of the respective transmission; once a
user submits a flow she has no further control over its progress through the
network.
This section turns to services that offer dynamic bandwidth guarantees
and give users some control over and the ability to adapt to their allocation
while their transmissions are in progress. This flexibility is ideally suited for
elastic applications such as email, Web browsing, or streaming video that
can adapt their data rates to match the available bandwidth while achieving
a graceful degradation in the perceived quality of service. As further dis-
cussed below, the survey uses the phrase flexible bandwidth-sharing services
to emphasize its focus on a particular type of bandwidth allocations that
determine the shares of a given set of concurrent flows based on user utility
functions. While one can think of contracts with other dynamic guarantees,
in practice the peak rate is the most common one. The Internet transport
service provided by the TCP protocol is the most prominent example. It
guarantees zero packet loss by retransmitting lost packets and dynamically
controls the peak rates at which competing connections are allowed to
transmit via adaptive flow control and feedback mechanisms. The network
nodes transmit congestion signals to end user applications which adapt
their sending rates accordingly. The distinctive features of this approach are
closed-loop control, on fast time scales and at the level of individual flows:
individual end user applications receive feedback signals every few tens of
milliseconds and adapt their sending rates correspondingly quickly to track
changes in their bandwidth allocations. In contrast to this approach the
guaranteed services discussed above use open-loop control—admission de-
cisions are based solely on a priori available traffic and network informa-
tion—and the self-optimizing behavior of users of best effort services allows
at best for much slower feedback between demand and congestion levels.
Bandwidth sharing mechanisms generally pursue multiple allocation ob-
jectives: efficiency (using all available bandwidth to the fullest), feasibility and
congestion avoidance (small packet delays and losses) while achieving and
maintaining a certain fairness in the bandwidth shares attributed to different
flows. There are several definitions of fairness (cf. Mo and Walrand, 2000,
Massoulié and Roberts, 2002) and the relationship between fairness and
network performance is a topic of intense research activity (cf. Roberts,
2004.)
The body of work reviewed here is inspired by adaptive flow control
procedures as implemented by TCP. To embed these procedures in an eco-
nomic environment it associates congestion signals with per-packet prices,
endows users with utility functions and lets them adapt their data rates to
maximize their utility net of payments. The approach proposes and anal-
yzes the potential of such distributed adaptive price and flow control al-
gorithms for achieving proportionally fair bandwidth sharing among
competing elastic flows. The proportional fairness criterion is of particu-
lar interest from an economics perspective since it is closely related to
welfare maximization when utility functions depend on throughput only.
This proposal takes a novel approach to quality of service differentiation

and provisioning. A key feature is that it delivers a set of arbitrarily differ-
entiated and evolving services that are defined by users based on their
personal needs—hence the term flexible bandwidth-sharing services used
here. The approach assumes that capacity is adequate and queueing delays
are small in comparison with propagation delays, and that a simple network
with a single packet class provides differentiation through algorithms lo-
cated at the end nodes, in line with the ‘‘end-to-end principle.’’ This ap-
proach stands in contrast with conventional service differentiation solutions
for guaranteed and best effort services, which operate on the premise that
congestion can be significant on the time scales where customers make their
usage decisions. They offer a fixed set of service classes with pre-defined
quality levels and prices, and they require intelligence embedded within
the network and sophisticated control mechanisms such as admission
control, resource reservation and discriminatory packet scheduling disci-
plines, to manage quality in the presence of potentially significant queueing
delays.
Modeling
There is a discrete economy with a fixed set of users, each generating a
single rate adaptive flow over a fixed network route. Like much of the work
on fairness and flow control, bandwidth sharing is mostly considered in a
static regime where users share network resources for the transfer of infinite-
sized documents. In other words, the number of users and their preferences
are constant throughout the horizon, and it is long enough for the system to
converge to the desired bandwidth allocation. This setting is useful when the
network carries large file transfers and the number of flows changes infre-
quently. A user’s utility at a given time only depends on her data rate or
bandwidth at that time, which is the only explicit QoS metric. In contrast to
best effort services, here delays or losses typically are not explicit attributes of
customer utility but merely act as feedback signals for flow control. Con-
gestion externalities are thus not modeled directly, but congestion impacts a
user’s utility indirectly by reducing her future bandwidth.
The key network control mechanism is flow control, which describes the
dynamic feedback process of price adaptation at routers and sending rate
adaptation at end nodes. Once in the network all flows are treated the same;
thus services are differentiated only based on bandwidth, and this differ-
entiation is the result of different sending rates.
The fundamental results model traffic sources as deterministic fluid flows
and ignore details of specific flow control algorithms, packet scheduling
and congestion signalling. Studies of implementation issues also consider
stochastic fluctuations, packet level dynamics and operational details. In
its original form this approach uses non-discriminatory FIFO packet sched-
uling.
Overview
Section 5.2 discusses static bandwidth sharing in a static regime with
complete information. It introduces the weighted proportional fairness cri-
terion, which roughly specifies that the network allocate flow rates in pro-
portion to how much users choose to pay. Under appropriate weights (user
payments) this sharing criterion yields the socially optimal bandwidth al-
location. Section 5.3 discusses distributed adaptive flow control and price
algorithms that are designed to attain the socially optimal proportionally
fair flow allocation when the network has no or incomplete information on
utility functions and users are unaware of capacity constraints and con-
gestion. Section 5.4 outlines extensions of this approach to networks with
more than a single packet class serving delay- or loss- sensitive flows, to
bandwidth sharing among real-time applications that have flexible peak
rates but are also subject to a strict QoS bound, and to distributed admis-
sion control via congestion signals. Section 5.5 observes that this work has
so far focused on sharing under a static regime with a fixed number of flows
and outlines basic questions that may arise if one wants to investigate
optimal dynamic pricing policies in dynamic environments where the
number of flows is in flux. Section 5.6 concludes with suggestions for future
research.
Socially optimal allocation and prices

Consider the social welfare maximization problem in the basic model
where a fixed set of users each transmit a single long-lived data flow over a
capacitated network. Let L be the set of network links, and R denote the set
of routes, each comprising a collection of links. Users are associated with
fixed routes. User r A R gets utility ur (xr) per unit time when sending at rate
xr, and x is the vector of flow rates. Following Shenker’s (1995) notion of
elastic applications, utility functions are increasing and concave in through-
put. The system incurs a congestion cost, e.g., measured by the cumulative
packet loss rate at all resources. Let Cl (yl) denote the expected packet loss
rate at resource l when its total flow rate is yl ¼ Ss:lAs xs where Cl ( ) is
convex, strictly increasing and differentiable. (The congestion cost functions
could be chosen to represent hard capacity constraints.) This network
model is relatively coarse in that it captures only fluid flow rates and ab-
stracts from packet details and queue lengths. All packets are treated
equally by the network.
The system objective is to maximize total welfare W over flow rates x
!
X X X
SYSTEM : max W ðxÞ :¼ ur ðxr Þ Cl xs : (29)
x0
r2R l2L s:l2s
At the socially optimal flow rates x ; the marginal utility of each user
equals the marginal congestion cost
!
X X
u0r ðxr Þ ¼ C 0l xs ; r 2 R. (30)
l2r s:l2s
The main challenge in solving SYSTEM and finding x is that neither the
network nor the users have the required information. The network does not
know the utility functions ur and users know neither each other’s utilities
nor the capacity constraints or congestion cost functions Cl.
Proportionally fair bandwidth sharing

However, it follows from convex optimization theory that SYSTEM may
be decomposed into subsidiary optimization problems, one for each user
and one for the network, that are linked by the appropriate resource
shadow prices and information exchange between the optimizing agents.
Consider the following decomposition. Given the network sets linear prices
Pr per flow unit, user r determines the total payment br (Pr) per unit time
that maximizes her net utility:
USERr ðPr Þ : max ur ðbr =Pr Þ br . (31)
br 0
Individual users are assumed too small to anticipate the effect of their
actions on the packet loss rate and related prices, hence they simply max-
imize their utility minus payment rate. Clearly br (Pr) satisfies:
u0r ðbr =Pr Þ ¼ Pr . (32)
The network in turn takes the vector of user payments b per time unit as
given and computes the weighted proportionally fair allocation of flow rates
x (b):
!
X X X
NETWORKðbÞ : max br log xr Cl xs , (33)
x0
r2R l2L s:l2s
obtaining the prices Pr per flow unit on route r as the sum of the respective
resource shadow prices:
!
br X 0 X
¼ Cl xs ¼ Pr . (34)
xr l2r s:l2s
This decomposition, due to Kelly (1997), only imposes local information

requirements on the network and users. Kelly (1997) proves existence of a
social welfare maximizing equilibrium, i.e., there exist vectors P ; b and x
with xr ¼ br =Pr that simultaneously solve USERr (Pr) for all r, NET-
WORK(b) and SYSTEM. The proportional fairness criterion specifies,
loosely speaking, that the network allocate flow rates in proportion to how
much users choose to pay. It is appealing for a couple of reasons. First, it

yields the socially optimal bandwidth allocation under appropriately chosen
user payments b : As noted above this survey refers to this bandwidth
allocation as flexible since it takes into account user utility functions. Spe-
cifically, unlike under max–min fairness, it does not just depend on the
number of flows and their network routes. (Mo and Walrand, 2000, define
the notion of a-fairness that encompasses all these fairness objectives.) It
also differs from the predefined bandwidth sharing rules for single-node
egalitarian processor sharing systems discussed in Section 4, cf. Altman and
Shimkin (1998), Ha (2001), and Maglaras and Zeevi (2003). Second the
weighted proportionally fair allocation arises naturally as a result of ex-
isting flow control algorithms embedded in the Internet, making it feasible
to solve the network problem in a distributed fashion, as discussed next.
Price and flow adaptation

While the decomposition of (31) and (33) only requires locally available
information there is no guarantee for iterative solutions of the user and
network problems to converge to the welfare maximizing equilibrium. The
network problem is mathematically tractable but difficult to implement by a
centralized processor, even one that could reliably handle the computa-
tional complexity, since the required information exchanges with users are
asynchronous and vulnerable to random link delays and failures which
differ from user to user. One should also keep in mind that the socially
optimal bandwidth allocation changes as soon as the number of flows in the
system or one of their utilities changes, requiring a new set of iterations.
Kelly et al. (1998) propose to control the system with distributed rate and
price control algorithms through a feedback loop between end users and
resources (routers, switches) so that the overall network reacts intelligently
to perturbations. Golestani and Bhattacharyya (1998) and Low and Laps-
ley (1999) study related approaches. Each end user r chooses a payment rate
br(t) and a data rate xr(t) at time t. Resource l marks a proportion of
packets pl ðyl ðtÞÞ ¼ C 0l ðyl ðtÞÞ with feedback signals, where pl (yl (t)) can be
interpreted as the resource shadow price, the marginal increase in expected
packet loss rate at link l in response to an infinitesimal increase in the total
flow rate. The packet marks are sent back to users at the rate they produce
them; they can be viewed P as prices and/or congestion signals. User r receives
marks at a rate of xr ðtÞ l:l2r pl ðyl ðtÞÞ and adjusts her instantaneous flow rate
according to
!
dxr ðtÞ X
¼ kr br ðtÞ xr ðtÞ pl ðyl ðtÞÞ ; r 2 R (35)
dt l2r
where kr>0 is a constant gain parameter. Thus a user increases or decreases

her sending rate depending on whether the rate of packet marks falls short
of or exceeds her willingness to pay. In addition to adjusting their flow rates
xr (t), users also update their willingness to pay br (t) to maximize their net
utilities at the implied prices, either continuously or on a slower time scale.
Kelly et al. (1998) establish stability of this algorithm at a proportionally
fair and socially optimal allocation. First, for given fixed payment vector b,
the system (35) has an unique stable point of flow rates x ðbÞ which solves
NETWORK(b) and to which all trajectories x (t) converge. Second, if users
also adjust their payment rates br (t) in response to fluctuations in their data
rate xr (t) to continuously track the optimum to their local net utility max-
imization problems, then all data rate trajectories x (t) converge to the
socially optimal allocation x ; the solution of SYSTEM. They also extend
these results to the case where users have routing choices, and for the dual
of this algorithm where shadow prices vary gradually, and flow rates are
given as functions of the shadow prices.
The adaptive nature of this mechanism bears some resemblance to those
for best effort services discussed in Section 4.3, cf. Gupta et al. (1997) and
Masuda and Whang (1999). A main difference is that here flow and price
adaptation decisions occur on a much shorter time scale, perhaps tens of
milliseconds versus minutes, hours or days.
The stylized model described so far only captures the macroscopic fluid
flow characteristics and ignores microscopic details of packet level marking
and flow control algorithms. Several issues need to be addressed for an
implementation to achieve the desired macrobehavior.
Delays and randomness in feedback signals

The basic stability results assume that the negative feedback is instan-
taneous. A number of studies, e.g., Kelly et al. (1998), Johari and Tan
(2001), Vinnicombe (2002), Massoulié (2002) and Kelly (2003) consider
how queueing or propagation delays in the communication between end
users and resources and stochastic fluctuations in packet loss or marking
affect the stability and speed of convergence to the system equilibrium,
depending on the sensitivity of end-user responses to congestion signals, the
gain parameters kr, and of congestion marks to resource load, the deriv-
atives of the marking functions pl ( ).
Packet marking
System stability requires that the packet marking rate at each resource
equal its shadow price where both are functions of the total flow rate, i.e.,
pl ðyl ðtÞÞ ¼ C 0l ðyl ðtÞÞ for all l A L. Given the stochastic nature of traffic loads
at resources, the congestion cost randomly fluctuates from packet to packet.
The challenge is to design packet marking algorithms that result in a rate of
packet marks that indeed equals the resource shadow price. Gibbens and
Kelly (1999a) propose to charge each packet individually the exact amount
of congestion cost it inflicts on the system at each resource, defined by its

sample path shadow price: it equals one if and only if deleting the packet
from the sample path of arrivals at that resource would result in one less
packet loss. They show for Poisson traffic that the packet marking rate
under this principle precisely equals the resource shadow price.
The sample path shadow price is intuitively appealing but hard to measure
in practice. By definition it can only be determined once packet loss has
occurred, at which point some of the packets contributing to this loss have
already left the queue and routers do no longer have information to track
them. Furthermore, signalling congestion only after packets are dropped
limits the ability to anticipate or prevent packet losses, especially in the pres-
ence of communication delays. These considerations have led to proposals for
congestion marking algorithms that generate marks before packet losses occur
based on locally available queue length information at routers. These algo-
rithms use either actual statistics (e.g., Random Early Marking by Lapsley
and Low, 1999, a variant of Random Early Detection by Floyd and Jacobson,
1993; see also Floyd, 1994, on Explicit Congestion Notification) or simula-
tions (the virtual queue algorithm of Gibbens and Kelly, 1999a). Wischik
(2001) analyzes under what conditions marking algorithms achieve propor-
tional fairness taking into account the impact of traffic burstiness. As long as
queueing delays are small relative to signal propagation delays—the premise
of the approach described here—randomness associated with whether or not a
particular packet is marked under a given algorithm becomes less relevant
than the proportion of packet marking since user adaptation decisions occur
on the time scale of propagation delays. Results suggest that simple algo-
rithms yield marking rates that equal or closely approximate the resource
shadow prices, inducing the desired system stability.
Flow control
The adaptation algorithm (35) specified by Kelly et al. (1998) raises a
couple of issues, on its consistency with user behavior and on its potential
for implementation in the current Internet.
What if users do not voluntarily adapt their rates according to (35)? This
could occur even if the algorithm were implemented at end nodes. Ganesh
et al. (2006) and Wu and Marbach (2003) show that even if users behave
selfishly, choosing their net benefit maximizing transmission rates at each
time step based on adaptive price estimates and unconstrained by flow rate
algorithms, then the system still converges to the equilibrium allocation of
Kelly et al. (1998). Another point concerns the assumption that a single user
does not anticipate the effect of its own actions on prices. This is appro-
priate if individual data rates are small relative to the total flow rate as in
the case of a network with a large user population. Gibbens and Kelly
(1999a) observe that a user who accounts for a significant fraction of the
network load and who does anticipate her impact on prices will tend to
reduce its data rate, relative to the case where she does not, since she
internalizes the congestion cost she imposes on her own flow.
How can the adaptation algorithm specified by (35) be implemented
based on Jacobson’s (1988) rate control algorithm TCP that operates in the
current Internet? Crowcroft and Oechslin (1998) propose MulTCP, a TCP
variation in which users set certain parameters so that it attains weighted
proportional fairness. Gibbens and Kelly (1999a) explore how various
TCP-like algorithms and packet marking mechanisms can support het-
erogenous applications by tailoring them to a mix of rate adaptive real-time
traffic and to file transfers, and they compare their mechanisms to those of
the Internet. Further comparisons of TCP-like flow control algorithms via
distributed network simulations and games are developed by Key et al.
(1999) and Key and McAuley (1999). Kelly (2001) proves that MulTCP
maximizes system welfare for particular utility and cost functions. He also
shows that if users have routing choices, then the equilibrium allocation
under a TCP-like algorithm suffers from two problems that do not occur in
the stylized system model described above: it need not be pareto-efficient
and the addition of a network link may in fact decrease welfare, an instance
of Braess’ paradox, see Braess (1968).
Intelligent user agents

In the context of this bandwidth sharing approach a user will typically find
it optimal to frequently adapt her data rate, to respond to randomly chang-
ing demand and supply conditions and also perhaps as a means to learn her
true utility function. The burden of these decisions may be prohibitive,
prompting studies on how intelligent agents can perform these tasks on behalf
of users, e.g., see Courcoubetis and Weber (2003, pp. 250–254).
Price uncertainty and bandwidth contracts

Flexible contracts and their inherent fluctuations in prices and bandwidth
availability may not appeal to customers who cannot or do not wish to
absorb this uncertainty, creating a market opportunity for bandwidth fu-
tures contracts that offer the use of a pre-specified amount of bandwidth
over some future time interval at a given price. Since bandwidth cannot be
stored—capacity available at a given time is either used or lost forever,
bandwidth contracts cannot be hedged in the same way as storable com-
modities such as silver, i.e., by purchasing the underlying asset for sale at a
later date. The pricing of bandwidth derivative contracts differs from the
traditional analysis of financial derivatives. Kenyon and Cheliotis (2001)
describe the special features of bandwidth in comparison with other com-
modities, primarily electricity, and propose a method for constructing spot
prices. Upton (2002) prices forward bandwidth contracts in a market for
short term single-link data transfers over a time scale of seconds to hours
and Keppo (2005) analyzes the pricing of bandwidth derivatives under
network arbitrage conditions. The main task in pricing such contracts is to
determine the expected spot price over the relevant time horizon. Anderson
et al. (2006) take a different perspective: they propose a contract and bal-
ancing mechanism for a set of users to choose how much bandwidth to buy
up-front over a fixed contract horizon at given unit prices, but at quantities
that depend on the expected time-average spot price which fluctuates in
function of usage-sensitive packet marks. The packet marks serve as a basis
for computing balancing payments that users make or receive at the end of
the contract.
The approach discussed above focuses on bandwidth differentiated serv-

ices (throughput is the only attribute of QoS, customer utility and service
differentiation), a single packet class in the network, and on long-lived rate-
adaptive elastic flows. The following studies consider extensions.
Key and Massoulié (1999) and Gibbens and Kelly (1999a) consider a mix
of users with rate-adaptive traffic and utility functions that only depend on
throughput, as modeled above, and users with finite-size file transfers whose
utility is a function of total transfer time, the ratio of data volume by
throughput.
Alvarez and Hajek (2002) and Gibbens and Kelly (2002) study marking
mechanisms for priority scheduling policies for two packet classes, which
may be beneficial in the presence of significant queueing delays and if some
users value the ability to choose among delay or loss differentiated service
classes. These studies make no attempt to optimize performance; they focus
on the differences in throughput-, loss- and delay-performance for given
marking schemes.
Gibbens and Kelly (1999b) describe how a network can accommodate
competing rate-adaptive traffic and nonadaptive traffic such as traditional
telephone calls requiring low packet loss by using packet marks to make
distributed admission control decisions.
Courcoubetis and Weber (2003, p. 250) discuss how to apply ideas from
the approach discussed here to allocate and price bandwidth to flows of
real-time applications that have flexible peak rates but are also subject to a
strict QoS bound.
Recall that the analysis in this section focuses on a static regime with a
fixed set of users with constant utility functions who transfer infinite-sized
flows over fixed routes. The weighted proportional fairness criterion spec-
ifies the desired allocation for this fixed set of flows and the dynamic el-
ement of the approach shows how to attain this equilibrium via
decentralized price and flow rate adaptation in the presence of incomplete
information. While bandwidth sharing and flow control has until recently
typically been studied in this static context, in reality bandwidth sharing
takes place in a dynamic stochastic environment: finite-sized documents
arrive at random instants and depart when their transfer is complete.
However, the topic of bandwidth sharing in such dynamic environments
has only recently received research attention, and the study of optimal
dynamic pricing seems completely open. Therefore the following discussion
merely outlines basic questions that may arise if one wants to investigate
optimal dynamic pricing in the context of dynamic bandwidth sharing.
The topic of bandwidth sharing in dynamic stochastic networks has re-
cently become the subject of intense research, e.g., Massoulié and Roberts
(2000), de Veciana et al (2001), Bonald and Massoulié (2001), Roberts
(2004). However, unlike the approach discussed above, these papers ignore
the time scale on which flow control processes converge to equilibrium,
assuming that the new bandwidth allocation is immediately attained as the
set of ongoing flows changes. They focus instead on the dynamic effects
occurring on the slower time scale of flow arrivals and departures, consid-
ering the stability and throughput and response time performance of net-
works that operate under different fairness rules. It is unclear how much
insight can be gained on the interaction between these effects that occur on
different time scales. However, several questions may be worth investigating
for such stochastic dynamic environments. What is the right criterion for
user-perceived QoS performance in such settings? Key and Massoulié
(1999) and Roberts (2004) among others argue that the utility for the
transfer of fixed-size documents may be more reasonably a function of
response time than flow throughput at any given time. In the single-node
case strict priority scheduling disciplines often outperform those that aim
for ‘‘fairness’’ in terms of total system delay cost. Unfortunately, it is not
clear which sharing criterion optimizes overall response time performance
in a network, but max–min fair sharing has been shown to outperform
proportional fair sharing in some settings, e.g., Massoulié and Roberts
(2000). The issue of utility function formulation aside, how does the net-
work behave under the proposed flow control and pricing mechanisms in a
dynamic environment where flows arrive and terminate over time? Should
the mechanisms be modified for such dynamic environments, and if so,
how? Considering proportional fairness and alternative fairness criteria,
how long does convergence to successive equilibrium bandwidth allocations
take on average relative to the time scale of flow arrivals and departures?
How does this adaptive ‘‘rebalancing’’ process affect operational and eco-
nomic performance? What should be the relationship between flow level
prices that users might pay prior to requesting a document and the charges
they might be charged while a transmission is in progress via packet marks?
Anderson et al. (2006) consider this latter question in the static context with
a fixed number of flows.
Sharing bandwidth among a fixed set of long-lived flows in a propor-

tionally fair way with appropriate weights maximizes total system welfare if
throughput is the only QoS attribute of user utility. The proportionally fair
allocation is inherently flexible in that it takes into account user utility
functions and creates bandwidth-differentiated services with just the right
amount of granularity: one ‘‘class’’ for each user. Distributed adaptive flow
control and price algorithms are shown to attain this equilibrium band-
width allocation even when the network has no or incomplete information
on utility functions and users are unaware of capacity constraints and
congestion. The approach assumes that capacity is adequate and queueing
delays are small in comparison with propagation delays, a simple network
with a single packet class and differentiation provided by the algorithms
located at the end nodes. The approach is appealing since it offers users
flexibility, is based on local information and leverages simple and largely
existing flow control mechanisms. The uncertainty in prices and the com-
plexity of end node decisions due to the price fluctuations that are inherent
to this approach pose some challenges, but they may be overcome by de-
signing appropriate bandwidth markets and user agents.
further research.
Dynamic stochastic environments. The issues considered here have not
been studied in dynamic stochastic environments where the optimal band-
width allocations and the respective prices fluctuate as flows arrive and
depart. (See the discussion in Section 5.5.)
Pricing user congestion cost. In contrast to the models for best effort
services in Section 4, here a user’s utility function only depends on her own
throughput. The congestion cost, represented by the functions Cl, l A L, is a
known system cost for the network. Therefore the socially optimal prices
are immediate from the packet loss rate and the corresponding packet
marks. By contrast, if users are also sensitive to congestion via QoS metrics
such as packet loss or delay (see Section 5.4), then the congestion cost is not
known to the network and may vary across users. How to translate con-
gestion signals into prices in this case to attain a socially optimal network
allocation? A useful reference point are Gupta et al. (1997), Masuda and
Whang (1999) and Stidham (2003) who consider adaptive pricing of best
effort services for delay-sensitive customers, i.e., whose utility is an explicit
function of delay. Stidham (2003) introduces user heterogeneity with re-
spect to delay sensitivity in the model of Kelly et al. (1998). He shows that
this introduces a fundamental nonconvexity into the congestion-cost func-
tions. As a result, there are typically multiple stationary points, and the
rate-control algorithm may converge to a local rather than global maxi-
mum, depending on the starting point.
Flow level versus customer level decisions. The analysis discussed above
focuses on customers’ flow level payment and usage decisions. How do
these decisions relate to their overall subscription and usage decisions?
Profit maximization. The body of work discussed here focuses on social
welfare maximization whereas the Internet is operated by profit-maximizing
providers. How should a profit-maximizing provider who does not know
user preferences design prices and packet marking algorithms? What is the
impact of competition? How should interconnected networks share pay-
ments based on congestion marks and how can these marks be propagated
among networks with different technologies?
6 Discussion
After considering each contract type in isolation the survey now discusses
broad pricing and service design issues that concern all contract types: the
benefits and challenges of auctions versus posted prices; the debate on flat-
rate versus usage-based pricing; and the merits and challenges of alternative
QoS designs and their delivery machanisms.
6.1 Auctions versus posted prices
This survey has so far focused on posted price mechanisms. A fundamen-

tally different approach to selling data communication services is via auctions.
Customers submit bids and the provider determines prices and allocations
based on their bids as a function of previously announced rules. See Krishna
(2002) for a recent treatment of auction theory and Ockenfels, Reiley and
Sadrieh (2006) for a survey of online auctions. Vickrey’s (1961) second-price
sealed-bid auction inspired the ‘‘smart market’’ proposal of MacKie-Mason
and Varian (1995a) which received considerable attention in the data pricing
community. An important motivating factor for the smart market is the de-
mand-revealing property of the classic second-price auction that allows the
provider to attain the efficient allocation even if she has no information on
demand functions. This section considers the benefits and challenges of the
smart market proposal and other auction-based approaches that differ in the
auction format they study (first-price or second-price auction) and the objects
that users bid on (data packets, network resources, guaranteed services or best
effort priority services.) The discussion also considers under what conditions
the demand-revealing property of the second-price auction, which is only
known to hold in static settings, may be preserved in dynamic data commu-
nications settings. An exhaustive discussion of auction formats and issues is
beyond the scope of this survey. See Courcoubetis et al. (2001) for an example
of a descending bandwidth auction for a multi-link network. They propose a
version of a Dutch auction and discuss several implementation and economic
performance issues in comparison to other auction formats.
The second-price auction and the smart market proposal

In the classic second-price auction a single indivisible unit is sold to
bidders with independent private values and is awarded to the highest bid-
der at the second highest bid. Thus the winner pays the social cost or
surplus loss that her use inflicts on other bidders, measured based on the
reported valuations. This payment rule implies the celebrated demand-re-
vealing property of the second-price auction: it is a dominant strategy for
each bidder to bid her true preferences (regardless of others bidders’ ac-
tions) and so the equilibrium bids yield the efficient allocation. The Vickrey
auction and the Vickrey-Clarke-Groves mechanism extend the second-price
auction to multiple identical and nonidentical units, respectively, and share
its demand revelation and efficiency properties (Clarke, 1971; Groves, 1973;
see Loeb, 1977, for a comparison of these mechanisms.) The resulting prices
also reflect the marginal value of capacity which send the correct signal
about efficient capacity investments. MacKie-Mason and Varian (1995a)
propose to repeatedly execute at each network node packet-level Vickrey
auctions in short successive time intervals: in each time interval the band-
width available for transmitting K data packets is allocated to the K high-
est-bid packets present at that time. The smart market proposal has some
appeal, in part due the attractive theoretical properties of the underlying
second-price auction, and also because it allows prices to be determined
immediately based on user bids rather than gradually in response to user
demand, as in adaptive pricing approaches such as those studied by Low
and Varaiya (1993), Gupta et al. (1997), Kelly et al. (1998), Masuda and
Whang (1999) and Thomas et al. (2002). However, the nature of data
communications demand poses significant challenges for the implementa-
tion of the smart market or indeed any auction-based approach.
Resource and time aggregation

The basic smart market proposal envisions no aggregation: each auction
sells individual units of packet transmission capacity at a network node
over a short time horizon. However, users assign value to successful end-to-
end transmissions of complete data flows, e.g., email messages or database
queries, rather than to their component data packets whose values are
interdependent. Since complementary packets may be routed through
different nodes and arrive over multiple bidding periods, it is not feasible to
elicit valuations and run bandwidth auctions at the packet and node level.
MacKie-Mason (1997) and Lazar and Semret (1998) apply the smart mar-
ket approach at a higher aggregation level, allowing users to bid on service
units or resource bundles with more easily quantifiable values. In MacKie-
Mason (1997) users bid for rates of service units over longer demand periods,
each service unit for end-to-end delivery of an entire data stream over a
network with certain QoS guarantees. Users submit utility functions over
their allocation and the provider determines the allocation and payments
using the Vickrey–Clarke–Groves mechanism. In contrast to the packet-level
auctions, here users bid on resources only indirectly, via their bids for services
with QoS guarantees. In Lazar and Semret’s (1998) ‘‘progressive second price
auction’’ users with elastic demand and fixed routes directly bid for shares of
resource quantities along their routes. Their analysis applies to any infinitely
divisible resources where consumption is additive, but the notion that bidders
know their valuations of resource quantities may be most plausible if the
auction is for bulk bandwidth for aggregate flows like virtual paths. Lazar
and Semret (1998) propose and analyze an iterative variation of the second
price auction. They prove existence of a truthful and efficient A-Nash equi-
librium where all players bid at prices equal to their marginal resource val-
uations. This is weaker than the dominance result but it comes at the benefit
of a minimal message space: unlike in the standard format, users do not bid
their entire utility function at once, but rather a price-quantity pair in re-
sponse to the (observable) bids already in competition for the resource.
Convergence to the equilibrium is shown experimentally.
Dynamic effects
The properties of the Vickrey auction and indeed most auction results are
developed for static one-shot environments: all bidders and units for sale
are present at the start of the auction and all units are allocated at once. In
contrast, data communications demand is dynamic: it occurs as a random
sequence of asynchronous requests whose resource requirements partially
overlap in time. The static auction results need not hold in this dynamic
environment due to several potentially significant intertemporal effects: the
arrival and departure of bidders over time; their cost of delay between
service request, allocation and delivery; complementarities between re-
source requirements over time; and spill-overs of unserved demand from
one auction to the next. As a case in point, Milgrom and Weber (2000)
show that in a sequential second-price auction where bidders strategize over
time and the selling prices are revealed after each auction, bidding one’s
true value is a dominant strategy only in the last auction, but not in earlier
stages—there is an option value of winning at a later auction and lower
price, and unlike in the one-shot case the equilibrium bids depend on the
entire value distribution.
Therefore, mechanisms such as the smart market that run repeated one-shot
Vickrey auctions over nonoverlapping time periods inherit its demand-reveal-
ing and efficiency properties only if intertemporal effects are insignificant or
assumed away. One may argue that the extent of resource and time aggre-
gation has an impact on how closely the static results approximate those for
repeated one-shot auctions: intertemporal effects are fundamental to bidding
strategies in packet-level second-by-second auctions but perhaps less impor-
tant, e.g., in auctions that sell contracts for daily usage rates. This reflects the
trade-off between tractability and optimality inherent in the aggregation level:
the more aggregation, the more restrictions on bidding strategies and the fewer
the opportunities to finetune bids and allocations.
Several papers consider auctions for priority service at a single resource

among dynamically arriving and departing delay-sensitive bidders. With the
main exception of Balachandran (1972) who characterizes the equilibrium
of state-dependent bidding strategies, these papers (discussed below) as-
sume that bidders cannot observe the queue length, which reduces the en-
vironment to a static auction with delay externalities.
Dynamic auctions have received increased attention in recent years.
Broad issues include what information is available to bidders on competing
bidders and their history, how often bids may be updated and based on
what events (e.g., a new arrival) and when and how to allocate units. For
examples, see Etzion et al. (2006), Caldentey and Vulcano (2006), Pai and
Vohra (2006) who consider settings outside the data communications con-
text where a fixed number of units are sold over a fixed time horizon to
dynamically arriving bidders.
Delay externalities
The smart market approach does not explicitly consider externalities: a
bidder’s valuations only depend on her own consumption, and if users bid
on services (as opposed to resources) the network guarantees quality.
However, if the auction is for services without quality guarantees and con-
gestion is significant, then externalities are relevant and each bidder’s utility
also depends on the consumption of other bidders. Several papers study
auction-based service allocations with delay externalities in the context of
single-resource single- or multi-class queueing models that offer ‘‘best
effort’’ service to delay-sensitive customers. Whang (1990) studies the
Clarke–Groves mechanism for FIFO service. Glazer and Hassin (1985), Lui
(1985), Hassin (1995), Stahl (2002), Afèche and Mendelson (2004) and
Kittsteiner and Moldovanu (2005) study first-price auctions for priority
service and Stahl (2002) also considers a second-price priority auction.
The first-price priority auction is not demand-revealing and requires from
bidders common knowledge about all demand and supply factors. Further-
more, in the presence of externalities the resulting allocation is only efficient
with an appropriately chosen reserve price which in turn depends on the
demand distribution (Stahl, 2002; Afèche and Mendelson, 2004). An excep-
tion is the case of preemptive priorities where the optimal reserve price is zero
(see Hassin, 1995; Afèche and Mendelson, 2004.) Thus, the first-price auction
may only have an information advantage over a posted price mechanism if
bidders are better informed about each other’s demand than the provider.
Whether a second-price priority auction is demand-revealing likely de-
pends on its specific design. In general the demand-revealing property of the
static Vickrey–Clarke–Groves mechanism extends to settings with exter-
nalities, provided that bidders submit preferences over all users’ consumpt-
ions (MacKie-Mason and Varian, 1994; Jehiel and Moldovanu, 2005).
Extending this static result to a queueing setting raises two questions. First,
how to handle the fact that a queueing system is inherently dynamic? The
following common assumptions reduce the analysis to that of a static auc-

tion: A1. bidders do not observe the queue length (except in Balachandran,
1972) and do not choose their arrival time; A2. each bidder submits a single
bid and cannot renege, i.e., withdraw her bid; A3. the Poisson arrival rate,
value and delay cost distributions are common knowledge; A4. bidders
form rational expectations about the delay distribution of the system that is
assumed to be in steady-state. Assumptions A1–A4 allow bidders to com-
pute their delay and payoff distribution as a function of all bid strategies,
like in the static case. Second, under what conditions is a second-price
auction satisfying assumptions A1 and A2 demand-revealing and efficient?
Stahl (2002) presents examples where a second-price priority auction does
not have a pure-strategy bid equilibrium, unlike in the classic static case.
However, it is important to emphasize that Stahl (2002) has bidders submit
preferences only over their own transmission, rather than the overall usage
rates, as required for demand-revelation in the presence of externalities
(MacKie-Mason and Varian, 1994; Jehiel and Moldovanu, 2005). His find-
ing therefore is not inconsistent with static auction results and the apparent
discrepancy follows from restricting the bid strategy space in the presence of
externalities. Nevertheless, his conclusion that ‘‘efficiency arguments [from
static auctions] cannot be used to justify the implementation of simple y
second price auctions in dynamic stochastic environments’’ is certainly
valid. First, truly dynamic auctions, as opposed to static ones fitted to
dynamic resource allocation problems, indeed need not inherit known
properties from static auctions. Second, submitting preferences over all
consumptions in the presence of delay externalities requires bidders to un-
derstand how their delay distribution depends on overall usage rates, ca-
pacities and the priority mechanism—hardly a realistic requirement. What
may work instead, at least in principle, is for users and the provider to
contribute information. Each user declares her utility as a function of only
her own quantity and delay, regardless of how the delay relates to overall
usage, and the provider determines the allocation using her knowledge of
the relationship between usage and delay performance.
Revenue considerations
The demand-revealing and efficiency properties of second price auctions do
not address the issue of how to maximize provider revenues. The key is that
these auctions need not generally be the revenue-maximizing format in mul-
tiunit settings. Choosing the revenue-maximizing auction format and reser-
vation price typically does require knowledge of the demand distribution.
Computational and accounting overhead

Theoretical questions of demand-revelation, efficiency and revenue max-
imization aside, auctions may impose an unreasonably high overhead for
computation, communication and accounting. Bidders’ utility functions
may be complex and their strategies sophisticated, particularly if they bid
on multiple services or resources in a dynamic setting. The inherent price

uncertainty also makes it harder to plan usage compared to a quote-based
scheme. The provider’s tasks of tracking bids and determining the optimal
allocation and payments is computationally more intense than under a
posted price scheme, particularly in a second price auction where each
winner’s payment depends on all other bids.
Conclusions
In assessing the value of an auction design compared to a posted price
mechanism for data communication services one should trade off its po-
tential for demand revelation and efficiency gains with the complexity of the
bidding mechanism. The potential efficiency gains of an auction may be
large enough to offset the overhead in rapidly changing nonrepetitive de-
mand environments, since this implies a large information gap between
users and the provider, and for units with significant transaction values
where inefficiencies are costly. These requirements suggest that auctions are
only viable if they sell units with a minimum level of resource and time
aggregation, e.g., not individual packet transmissions over seconds, but
larger bandwidth chunks over days or weeks, probably sold to larger retail
or business customers. Such aggregation leads to less frequent auctions,
each for fewer but more bundled units; this sacrifices some optimality by
restricting the bid strategies but also increases the transaction values and
reduces the overall overhead. The following are to be considered in eval-
uating the demand revelation potential of a particular auction design. First,
the ideal outcome of ‘‘complete’’ demand revelation of users with no in-
formation on each others’ demand is only known to hold in the static
Vickrey auction—it extends neither to other static formats such as the first
price auction nor to dynamic auctions. Second, the performance of the
static Vickrey auction may be a good approximation of the outcome in
repeated one-shot auctions with negligible intertemporal effects, as may be
the case if each auction exhibits significant time aggregation. Each demand
period in a repeated auction should be long enough to ensure significant
transaction values but short enough so the allocation can adapt to changing
demand. Third, while any ‘‘truly dynamic’’ auction mechanism is likely to
lack the complete demand revelation property of the static Vickrey auction,
an ‘‘appropriately’’ designed dynamic auction may still have a partial de-
mand revelation advantage over a fixed price scheme in environments where
the provider cannot keep up with demand changes (cf. Whang, 1990). In-
deed, even if the optimal ‘‘truly dynamic’’ auction mechanism could be
identified, it would likely be very complicated, involving state and time-
dependent bids (Gupta et al., 2006), and require from users an unrealistic
amount of information on each others’ preferences. Fourth, whether users
have enough information on their own preferences, a necessary condition
for demand revelation even in the static Vickrey auction, depends in part on
the type of resources or services auctioned.
6.2 Flat-rate versus usage-based pricing
Like much of the research activity on data pricing this survey has focused
on usage-based pricing. This activity has been stimulated by the prevalence
of Internet congestion and by the development of many service technologies
that can implement sophisticated resource allocation and QoS differenti-
ation mechanisms and that raise challenging research questions.
In practice, usage-based pricing is far from universal. Traditional packet-
switched networks (e.g., based on the X.25 protocol) have typically charged
based on data volume and some Internet Service Providers have been
offering usage-based pricing for high bandwidth access links in recent years
(cf. Songhurst, 1999, p. 8). However, flat-rate pricing has been and still is
the prevalent tariff structure in the Internet retail market, whereby a user
pays a flat periodic subscription fee that only depends on the bandwidth of
her access link. Thus, her charge is related to her maximum sending rate but
independent of actual usage, quality or time of service.
Flat-rate versus usage-based pricing

The co-existence of these two polar tariff structures raises a simple ques-
tion: what are the relative benefits and challenges of flat-rate versus usage-
based pricing? Flat-rate pricing has many attractive features. It is cheap to
implement and operate, leads to simple and predictable user charges and
provider revenues, and is quite popular with customers, which has stim-
ulated the growth of the Internet. On the downside, flat-rate pricing typ-
ically results in a socially inefficient allocation and market segmentation
with light users subsidizing heavy users. Usage-based pricing, if appropri-
ately designed, is superior to flat-rate pricing in terms of economic per-
formance, giving the provider a more powerful tool to control congestion,
segment the market and increase revenues. This was also confirmed em-
pirically by Internet Demand Experiment, a market and technology trial
conducted at the University of California at Berkeley (cf. Edell and
Varaiya, 1999). In general, the potential economic benefit of usage-based
pricing must be weighed against the potential drawbacks, including user
adoption problems, a higher cost and complexity of implementation, and
the requirement for more information about user preferences and network
performance.
There are also arguments in favor of two-part tariffs that combine a flat
and a usage-sensitive component. For settings with congestion costs, Mac-
Kie-Mason and Varian (1995a) argue for an efficient two-part tariff with
subscription fees at levels that cover the infrastructure cost and usage-based
charges that depend on congestion. MacKie-Mason and Varian (1995b)
show for the monopoly case that a two-part tariff mitigates the problem of
inefficient use and capacity investment that arises under usage-based pricing
only. They also compare competitive market equilibria under two-part tariffs
and connection only fees. They argue that the absence of a usage-sensitive
tariff component may result in an equilibrium with lower usage but higher
congestion than when usage is also priced. Essegaier et al. (2002) study for
access services without congestion costs the optimal choice among flat-rate,
usage-based and two-part tariffs as a function of the capacity level and user
heterogeneity. They show that two-part tariffs are not always optimal.
Optimality versus implementability

The debate on flat-rate versus usage-based pricing points to a funda-
mental trade-off between optimality and implementability criteria that
should be considered in the design and evaluation of any pricing mecha-
nism. Optimality loosely refers to the economic performance of a pricing
mechanism in terms of total surplus and its distribution among providers
and customers. Implementability depends on several criteria on the side of
the network provider: the availability, cost-effectiveness, simplicity and
scalability of the technical infrastructure required to implement network
control, usage measurement and billing mechanisms; and the availability of
the necessary demand and network congestion information to determine
tariffs and charges. The implementability of a price mechanism also de-
pends on the adoption by customers who typically want tariffs and usage
decision processes to be simple and charges predictable, based on metrics
they can understand and control.
This trade-off highlights the importance of striking a good balance be-
tween optimality and implementability in designing a pricing scheme in
practice. For example, a mechanism that price discriminates at the level of
each individual packet may be theoretically optimal, but hardly implement-
able due to prohibitive implementation constraints and costs. Based on
these insights Shenker et al. (1996) call for a shift away from the ‘‘optimality
paradigm’’ and focus instead on structural and architectural pricing issues,
including edge pricing at the network access points instead of in the core
network, multicast and receiver charging. They also assert that computing
the true congestion costs and obtaining detailed knowledge of actual con-
gestion conditions along the entire path of a packet is not feasible and
propose approximations to compute the price based on the expected con-
gestion cost. This proposal laid the foundation for Clark’s (1997) expected
capacity model which was later expanded into the DiffServ model discussed
below.
Conclusions
The benefits and challenges of flat-rate pricing or any usage-based pricing
scheme must be evaluated based on optimality and implementability cri-
teria. Implementability criteria may not be as ‘‘quantifiable’’ as economic
performance measures—at least some criteria have hardly been explicitly
modeled so far, but they may be the decisive factor in choosing a pricing
scheme. Implementability issues also imply that usage-based pricing and
service differentiation schemes should not only be considered with the sole
objective of economic optimality; a significant improvement over flat-rate

pricing for uniform service may already be sufficient to justify their de-
ployment, especially if their implementation is relatively simple.
6.3 Providing QoS: overprovisioning versus control
The optimality and implementability criteria discussed in the previous

section imply that not only differentiated price schemes but also the QoS
delivery mechanisms that support them should be relatively simple. In this
sense the issue of implementability versus optimality is related to the debate
on overprovisioning versus control in providing QoS. Generally QoS per-
formance depends on demand characteristics and on ‘‘service production’’
which is a function of capacity and network control mechanisms. The two
extreme approaches for providing good QoS are overprovisioning (a lot of
excess capacity in the presence of very little control, e.g., a simple best-effort
network) and control (sophisticated control mechanisms in the face of little
excess capacity.)
Overprovisioning
The proponents of overprovisioning argue that the cost of excess capacity
has already been incurred for the network backbone, making it possible to
quickly and cheaply expand the capacity of this important network seg-
ment. By some measures networks are so lightly utilized that complex QoS
technologies seem to bring no value. Data networks are overall lightly
utilized on average: The Internet backbones exhibit average utilizations of
10–15% and peak utilizations of 25%. Private line networks have average
utilizations of 3–5% and peak utilizations of 15–25%. For comparison,
U.S. long distance voice networks have 33% and 70% average and peak
utilizations, respectively, (Odlyzko, 1999a).
However, parts of the Internet are highly congested, in particular: public
peering points, network access points, access links that aggregate traffic into
the backbone, and transatlantic links between the US and the rest of the
world. Global overprovisioning is considered an economically prohibitive
luxury for the foreseeable future (Gevros et al., 2001). The fact that capacity
does not come cheaply throughout the network infrastructure casts doubt
on the universal validity of the overprovisioning argument. Bandwidth
bottlenecks may well persist in certain network segments and may also
occur intermittently in transition periods between successive capacity up-
grades. Even if there were excess capacity throughout the network, it is
questionable whether a competitive equilibrium with excess industry ca-
pacity is sustainable in the long run.
Control: guaranteed services

At the opposite end of the spectrum, QoS can be delivered by guaranteed
services (as discussed in Section 3), exemplified by the IntServ architecture
proposal of the IETF and by the ATM, Resource Reservation Protocol and
Frame Relay standards. These approaches rely on sophisticated control
mechanisms including admission control, reservation and enforcement
based on per-flow state information within network routers. The challenge
is scalability since the network core must maintain state information for
each flow, which becomes a problem when the number of flows is high such
as in the Internet. A possibility is to deploy these approaches at the network
access level where the number of connections may be relatively small and
bandwidth may be tight, which is when the value of service guarantees is
large.
Moderate solution: differentiated services

This approach can be viewed as a moderate solution between over-
provisioning and control. In contrast to guaranteed services the DiffServ
architecture proposal of the IETF does not try to distinguish among in-
dividual data flows. Instead it classifies packets into one of a few categories.
All packets within a class are then handled (e.g., scheduled and routed) in
the same way, with the same QoS parameters. There is no per flow infor-
mation at routers. As such the DiffServ model can be viewed as augmenting
the basic elements of a multi-class best effort network (as discussed in
Section 4) with demand regulation mechanisms that enforce certain traffic
inflow constraints and help achieve statistical class-level QoS targets. Thus,
complexity is reduced at the expense of control. A key advantage of such an
approach is that it is more scalable than guaranteed services since the net-
work only needs to distinguish among aggregate flows that belong to the
different service classes, not among individual flows. It creates differenti-
ated services with the least amount of network control. This approach
should probably be viewed as increasing social welfare compared to no
service differentiation, without attaining economic optimality. A downside
of this approach is that it does not offer absolute QoS guarantees at the
individual flow level, but rather average class-level guarantees. It is also not
entirely clear exactly how distinctive levels of quality of service are to be
achieved. DiffServ does not by itself say anything about the assignment of
priorities to and the pricing of different traffic classes. Odlyzko’s (1999b)
PMP proposal for Internet pricing is to partition the Internet into several
logically separate ‘‘channels’’ for different service classes with higher prices
for using the better provisioned and therefore (allegedly) less congested
classes. One implication of these approaches is that the provider must
monitor and adjust the capacity allocation of individual service classes if
she is to maintain the desired average quality levels.
Conclusion
Extreme solutions such as guaranteed QoS require excessively complex
micromanagement that needs a lot of overhead, whereas overprovisioning
may not be feasible at all times and throughout the network. Simpler and
more effective solutions to QoS provisioning should draw on the benefits of

network control mechanisms and capacity expansion in a way that miti-
gates their respective drawbacks. For example, a broad service segmenta-
tion into 3–4 classes likely provides much of the benefits of price and service
differentiation without the prohibitive cost and complexity of tailoring a
service class to each individual connection. Such service differentiation and
the requisite control mechanisms should play a larger role when and where
network bandwidth is tight and expensive.
7 Conclusions
This chapter has surveyed research on data transport pricing decisions for
a single provider who pursues one of two economic performance objectives:
welfare maximization or profit maximization. It has largely focused on
studies that consider these issues in the context of mathematical models of
data transport service demand and production. The discussion is organized
according to an unifying framework that classifies data transport service
contracts based on the types of their QoS promises into guaranteed, best
effort and flexible bandwidth-sharing services. The survey highlights and
compares the key features of each service type, provides a broad roadmap
for important questions, modeling approaches and results, and outlines
open questions. It discusses pricing decisions for each service type along
common dimensions: basic pricing and allocation principles; incomplete
information and adaptive pricing; more differentiation; optimal dynamic
pricing; and conclusions and directions. The survey then examines benefits
and challenges of auctions versus posted prices; the debate on flat-rate
versus usage-based pricing; and the merits and challenges of alternative
QoS designs and their delivery mechanisms.
7.1 Research directions
Sections 3–5 suggest research directions for each service type. The sug-
gestions listed here refer to issues that appear relevant for all service types.
Revenue & profit maximization. The majority of studies consider socially
optimal pricing decisions; profit-maximization has received less attention so
far, but would benefit from more attention in the future. Clearly in com-
mercial data communications markets providers are primarily interested in
their own profits, not in the surplus generated for the entire system.
Service design. Most studies consider pricing for a given service design;
the problem of jointly designing services and prices is important and in-
teresting. This includes decisions on the number of service classes and the
types of QoS guarantees to offer, and on the supply allocation mechanisms
to deliver them.
Service plans and bundles. In a similar vein, most analyses focus on

determining the price for an individual connection or transmission, and
consider tariffs that are linear in the number or rate of transmissions. This
ignores potentially interesting bundling and nonlinear pricing options for
users that have demand for larger service quantities over longer horizons.
Similarly, it might be interesting to study the bundling of data transport
services with complementary services such as information content, data
hosting and computing applications (See Geng, Stinchcombe and Whins-
ton, 2006, for a survey of product bundling.).
Integrated services. The vast majority of studies consider networks that
offer a single type of QoS guarantee. Only a handful of studies consider a
mix of service types, e.g., best effort and strict QoS guarantees. There ap-
pears to be potential for more work along these lines.
Integration of empirical demand models into pricing studies. A number of
papers report on empirical analyses of demand for Internet-based data
transport services, including Edell and Varaiya (1999), Altmann et al.
(2000), Varian (2002) and Beckert (2005). However, these empirically de-
rived demand models appear not to be used in other pricing studies. Ideally,
more studies would build on such empirical models.
7.2 Further topics and readings
A number of issues have not been considered in this survey. Songhurst

(1999) discusses experimental and implementation issues. Courcoubetis and
Weber (2003) provide a discussion of interconnection and regulatory issues,
traffic splitting and multicasting; they also offer more discussion of some
issues considered here.
Gupta et al. (2006) offer another survey of data transport pricing. Hassin
and Haviv (2003) survey the economics of queueing systems, an area which
is closely related to the pricing of best effort services.
References
Adiri, L., U. Yechiali (1974). Optimal priority-purchasing and price decisions in non-monopoly and
monopoly queues. Operations Research 22, 1051–1066.
Afèche, P. (2004). Incentive-compatible revenue management in queueing systems: optimal strategic
delay and other delay tactics. Manufacturing & Service Operations Management, to appear.
Afèche, P., B. Ata (2005). Revenue management in queueing systems with unknown demand charac-
teristics. Working Paper, University of Chicago, Chicago, IL.
Afèche, P., H. Mendelson (2004). Pricing and priority auctions in queueing systems with a generalized
delay cost structure. Management Science 50, 869–882.
Alperstein, H. (1988). Optimal pricing policy for the service facilities offering a set of priority prices.
Altman, E., N. Shimkin (1998). Individual equilibrium and learning in processor sharing systems.
Operations Research 46, 776–784.
Altmann, J., B. Rupp, P. Varaiya (2000). Quality matters: some remarks on internet service provisioning
and tariff design. Telektronikk 96, 20–25.
Alvarez, J., B. Hajek (2002). On the use of packet classes in communication networks to enhance
congestion pricing based on marks. IEEE Transactions Automatic on Control 47, 1020–1026.
Anderson, E., F. Kelly, R. Steinberg (2006). A contract and balancing mechanism for sharing capacity
in a communication network. Management Science 52, 39–53.
Bajaj, S., L. Breslau, S. Shenker (1998). Is service priority useful in networks?. SIGMETRICS Per-
formance Evaluation Review 26, 66–77.
Balachandran, K. (1972). Purchasing priorities in queues. Management Science 18, 319–326.
Beckert, W. (2005). Estimation of stochastic preferences: an empirical analysis of demand for internet
services. Review of Economics and Statistics 87, 495–502.
Bell, C., S. Stidham (1983). Individual versus social optimization in the allocation of customers to
alternative servers. Management Science 29, 831–839.
Berger, A., W. Whitt (1998a). Effective bandwidths with priorities. IEEE/ACM Transactions on Net-
working 6, 447–460.
Berger, A., W. Whitt (1998b). Extending the effective-bandwidth concept to networks with priority
classes. IEEE Communications Magazine 36, 78–84.
Bergstrom, T., J. MacKie-Mason (1991). Some simple analytics of peak-load pricing. Rand Journal of
Economics 22, 241–249.
Bohn, R., H. Braun, K. Claffy, S. Wolff (1994). Mitigating the coming internet crunch: multiple service
levels via precedence. Technical Report. San Diego Supercomputer Center, University of California,
San Diego.
Bonald, T., L. Massoulié (2001). Impact of fairness on internet performance. Proceedings of the ACM
SIGMETRICS pp. 82–91.
Bradford, R.M. (1996). Pricing, routing, and incentive compatibility in multiserver queues. European
Journal of Operational Research 89, 226–236.
Braess, D. (1968). Über ein Paradoxon aus der Verkehrsplanung. Unternehmensforschung 12, 258–268.
Caldentey, R., G. Vulcano (2006). Online auction and list price revenue management. Working Paper,
New York University, New York.
Chao, H.-P., R. Wilson (1987). Priority service: pricing, investment, and market organization. The
American Economic Review 77, 899–916.
Chen, H., M. Frank (2001). State dependent pricing with a queue. IIE Transactions 33, 847–860.
Clark, D. (1997). A model for cost allocation and pricing in the internet, in: L. McKnight, J. Bailey
(eds.), Internet Economics, MIT Press, Cambridge, MA, pp. 215–252.
Clarke, E. (1971). Multipart pricing of public goods. Public Choice 11, 17–33.
Cocchi, R., S. Shenker, D. Estrin, L. Zhang (1993). Pricing in computer networks: motivation, for-
mulation, and example. IEEE/ACM Transactions on Networking 1, 614–627.
Courcoubetis, C., M. Reiman (1999). Pricing in a large single link loss system, in: P. Key, D. Smith
(eds.), Teletraffic Engineering in a Competitive World, Elsevier, pp. 737–746.
Courcoubetis, C., F. Kelly, R. Weber (2000a). Measurement-based usage charges in communication
networks. Operations Research 48, 535–548.
Courcoubetis, C., F. Kelly, V. Siris, R. Weber (2000b). A study of simple usage-based charging schemes
for broadband networks. Telecommunication Systems 15, 323–343.
Courcoubetis, C., M. Dramitinos, G.D. Stamoulis (2001). An auction mechanism for bandwidth al-
location over paths. Proceedings of the 17th International Teletraffic Congress.
Courcoubetis, C., R. Weber (2003). Pricing Communication Networks: Economics, Technology and
Modelling, Wiley, West Sussex, England.
Crowcroft, J., P. Oechslin (1998). Differentiated end-to-end internet services using a weighted propor-
tionally fair sharing TCP. ACM Communications Review 28, 53–67.
Dewan, S., H. Mendelson (1990). User delay costs and internal pricing for a service facility. Management
Science 36, 1502–1517.
Dolan, R. (1978). Incentive mechanisms for priority queuing problems. Bell Journal of Economics 9,
421–436.
Edell, R., N. McKeown, P. Varaiya (1995). Billing users and pricing for TCP. IEEE Journal of Selected
Areas of Communication 13, 1162–1175.
Edell, R., P. Varaiya (1999). Providing internet access: what we learn from INDEX. IEEE Network 13,
18–25.
Edelson, N.M., D.K. Hildebrand (1975). Congestion tolls for poisson queuing processes. Econometrica
43, 81–92.
Elwalid, A., D. Mitra (1992). Fluid models for the analysis and design of statistical multiplexing with
loss priorities on multiple classes of brusty traffic. Proceedings of IEEE INFOCOM, 415–425.
Elwalid, A., D. Mitra (1995). Analysis, approximations and admission control of a multiservice mul-
tiplexing system with priorities. Proceedings of IEEE INFOCOM, 463–472.
Essegaier, S., S. Gupta, Z.J. Zhang (2002). Pricing access service. Marketing Science 21, 139–159.
Etzion, H., E. Pinker, A. Seidmann (2006). Analyzing the simultaneous use of auctions and posted prices
for on-line selling. Manufacturing and Service Operations Management 8, 68–91.
Floyd, S. (1994). TCP and explicit congestion notification. ACM Computer Communication Review 24,
10–23.
Floyd, S., V. Jacobson (1993). Random early detection gateways for congestion avoidance. IEEE/ACM
Transactions on Networking 1, 397–413.
Ganesh, A., K. Laevans, R. Steinberg (2006). Congestion pricing and non-cooperative games in com-
munication networks. Operations Reasearch, to appear.
Geng, X., M. Stinchcombe, A. Whinston (2006). Product bundling, in: T. Hendershott (ed.), Handwork
on Economics and Information Systems, Elsevier, North Holland.
Gevros, P., J. Crowcroft, P. Kirstein, S. Bhatti (2001). Congestion control mechanisms and the best
effort service model. IEEE Network 15, 6–26.
Ghanem, S. (1975). Computing central optimization by a priority pricing policy. IBM Systems J 14,
272–292.
Gibbens, R., F. Kelly (1999a). Resource pricing and the evolution of congestion control. Automatica 35,
1969–1985.
Gibbens, R., F. Kelly (1999b). Distributed connection acceptance control for a connectionless network.
Proceedings of the 16th International Teletraffic Congress.
Gibbens, R., F. Kelly (2002). On packet marking at priority queues. IEEE Transactions on Automatic
Control 47, 1016–1020.
Gibbens, R., R. Mason, R. Steinberg (2000). Internet service classes under competition. IEEE Journal on
Selected Areas in Communications 18, 2490–2498.
Giridharan, P., H. Mendelson (1994). Free-access policy for internal networks. Information Systems
Research 5, 1–21.
Glazer, A., R. Hassin (1985). Stable priority purchasing in queues. Operations Research Letters 4, 285–288.
Golestani, S., S. Bhattacharyya (1998). A class of end-to-end congestion control algorithms for the
Internet. Proceedings 16th International Conference on Network Protocols.
Groves, T. (1973). Incentives in teams. Econometrica 41, 617–631.
Gupta, A., B. Jukic, D. Stahl, A. Whinston (2000). Extracting consumers’ private information for
implementing incentive compatible Internet traffic pricing. Journal of Management Information Sys-
tems 17, 9–29.
Gupta, A., D. Stahl, A. Whinston (1997). A stochastic equilibrium model of internet pricing. Journal of
Economic Dynamics and Control 21, 697–722.
Gupta, A., D. Stahl, A. Whinston (1999). The economics of network management. Communications of
the ACM 42, 57–63.
Gupta, A., D. Stahl, A. Whinston (2006). Pricing traffic on interconnected networks: issues, approaches,
and solutions, in: S. Majumdar, I. Vogelsang, M. Cave (eds.), Handbook of Telecommunications
Economics, Volume 2: Technology Evolution and the Internet, Elsevier, Amsterdam.
Ha, A. (1998). Incentive compatible pricing for a service facility with joint production and congestion
externality. Management Science 44, 1623–1636.
Ha, A. (2001). Optimal pricing that coordinates queues with customer-chosen service requirements.
Hassin, R. (1985). On the optimality of first come last served queues. Econometrica 53, 201–202.
Hassin, R. (1986). Consumer information in markets with random product quality: the case of queues
with balking. Econometrica 54, 1185–1196.
Hassin, R. (1995). Decentralized regulation of a queue. Management Science 41, 163–173.
Hassin, R., M. Haviv (2003). To Queue or not to Queue. Kluwer, Boston.
Haviv, M., J. van der Wal (1997). Equilibrium strategies for processor sharing and queues with relative
priorities. Probability in the Engineering and Informational Sciences 11, 403–412.
Hui, J.Y. (1988). Resource allocation for broadband networks. IEEE J. Selected Areas of Communi-
cation 6, 1598–1608.
Jacobson, V. (1988). Congestion avoidance and control. Proceedings ACM SIGCOMM, 314–329.
Jehiel, P., B. Moldovanu (2005). Allocative and informational externalities in auctions and related
mechanisms. Working paper, University of Bonn, Bonn, Germany.
Johari, R., D. Tan (2001). End-to-end congestion control for the Internet: delays and stability. IEEE/
ACM Transactions on Networking 9, 818–832.
Kelly, F. (1991). Loss networks. Annals of Applied Probability 1, 319–378.
Kelly, F. (1994). On tariffs, policing and admission control for multiservice networks. Operations Re-
search Letters 15, 1–9.
Kelly, F. (1996). Notes on effective bandwidths, in: F. Kelly, S. Zachary, I. Ziedins (eds.), Stochastic
Networks: Theory and Applications, Oxford University Press, pp. 141–168.
Kelly, F. (1997). Charging and rate control for elastic traffic. European Transactions on Telecommu-
nications 8, 33–37.
Kelly, F. (2001). Mathematical modelling of the internet, in: B. Engquist, W. Schmid (eds.), Math-
ematics Unlimited – 2001 and Beyond, Springer, Berlin, pp. 685–702.
Kelly, F. (2003). Fairness and stability of end-to-end congestion control. European Journal of Control 9,
159–176.
Kelly, F., A. Maulloo, D. Tan (1998). Rate control in communication networks: shadow prices, pro-
portional fairness and stability. Journal of the Operational Research Society 49, 237–252.
Kenyon, C., G. Cheliotis (2001). Stochastic models for telecom commodity prices. Computer Networks
36, 533–555.
Keppo, J. (2005). The pricing of point-to-point bandwidth contracts. Mathematical Methods of Op-
erations Research, to appear 6, 191–218.
Key, P., D. McAuley (1999). Differential QoS and pricing in networks: where flow-control meets game
theory. IEE Proceedings Software 146, 39–43.
Key, P., D. McAuley, P. Barham, L. Laevens (1999). Congestion pricing for congestion avoidance.
Microsoft Technical Report.
Key, P., L. Massoulié (1999). User policies in a network implementing congestion Pricing. Workshop on
Internet Service Quality Economics, Cambridge, MA.
Kittsteiner, T., B. Moldovanu (2005). Priority auctions and queue disciplines that depend on processing
time. Management Science 51, 236–248.
Kleinrock, L. (1964). Communication Nets. Dover, New York.
Kleinrock, L. (1967). Optimum bribing for queue position. Operations Research 15, 304–318.
Knudsen, N. (1972). Individual and social optimization in a multiserver queue with a general cost-
benefit structure. Econometrica 40, 515–528.
Korillis, Y., A. Lazar, A. Orda (1995). Architecting noncooperative networks. IEEE Journal on Selected
Areas in Communications 13, 1241–1251.
Krishna, V. (2002). Auction Theory. Academic Press, San Diego, CA.
Kulkarni, V., N. Gautam (1997). Admission control of multi-class traffic with service priorities in high-
speed networks. Queuing Systems 27, 79–97.
Kulkarni, V., L. Gun, P. Chimento (1995). Effective bandwidth vectors for multiclass traffic multiplexed
in a partitioned buffer. IEEE Journal Selected Areas in Communications 13, 1039–1047.
Kumar, A., D. Manjunath, J. Kuri (2004). Communication Networking: An Analytical Approach. Mor-
gan Kaufmann, San Francisco, CA.
Lapsley, D., S. Low (1999). Random Early Marking: an optimisation approach to Internet congestion
control. Proceedings of IEEE ICON.
Lazar, A., N. Semret (1998). Design, analysis and simulation of the progressive second price auction for
network bandwidth sharing. Technical Report, Columbia University.
Lippman, S., S. Stidham (1977). Individual versus social optimization in exponential congestion systems.
Low, D. (1974). Optimal dynamic pricing policies for an M/M/s queue. Operations Research 20, 545–561.
Low, S., D. Lapsley (1999). Optimization flow control, I: basic algorithm and convergence. IEEE/ACM
Low, S., P. Varaiya (1993). A new approach to service provisioning in ATM networks. IEEE/ACM
Loeb, M. (1997). Alternative versions of the demand-revealing process. Public Choice 29, 15–26.
Lui, F. (1985). An equilibrium queuing model of bribery. Journal of Polish Economy 93, 760–781.
MacKie-Mason, J., H. Varian (1994). Generalized vickrey auctions. Working Paper, University of
Michigan, Ann Arbor, MI.
MacKie-Mason, J., H. Varian (1995a). Pricing the internet, in: B. Kahin, J. Keller (eds.), Public Access
to the Internet, MIT Press, Cambridge, MA, pp. 269–314.
MacKie-Mason, J., H. Varian (1995b). Pricing congestible network resources. IEEE Journal of Selected
Areas in Communications 13, 1141–1149.
MacKie-Mason, J. (1997). A smart market for resource reservation in a multiple quality of service
information network. Working Paper. University of Michigan, Ann Arbor, MI.
Maglaras, C., A. Zeevi (2003). Pricing and capacity sizing for systems with shared resources: approx-
imate solutions and scaling relations. Management Science 49, 1018–1038.
Maglaras, C., A. Zeevi (2005). Pricing and design of differentiated services: approximate analysis and
structural insights. Operations Research 53, 242–262.
Marbach, P. (2004). Analysis of a static pricing scheme for priority services. IEEE/ACM Transactions
on Networking 12, 312–325.
Marchand, M.G. (1974). Priority pricing. Management Science 20, 1131–1140.
Massoulié, L. (2002). Stability of distributed congestion control with heterogeneous feedback delays.
IEEE Transactions on Automatic Control 47, 895–902.
Massoulié, L., J. Roberts (2000). Bandwidth sharing and admission control for elastic traffic. Tele-
communication System 15, 185–201.
Massoulié, L., J. Roberts (2002). Bandwidth sharing: objectives and algorithms. IEEE/ACM Trans-
actions on Networking 10, 320–328.
Masuda, Y., S. Whang (1999). Dynamic pricing for network service: equilibrium and stability. Man-
agement Science 45, 857–869.
Masuda, Y., S. Whang (2002). Capacity management in decentralized networks. Management Science
48, 1628–1634.
Masuda, Y., S. Whang (2006). On the optimality of fixed-up-to tariff for telecommunications service.
Mendelson, H. (1985). Pricing computer services: queueing effects. Communications of the ACM 28,
312–321.
Mendelson, H., S. Whang (1990). Optimal incentive-compatible priority pricing for the M/M/1 queue.
Mendelson, H., S. Whang (1994). Pricing for communication network service. Working Paper, Stanford
University, Stanford, CA.
Milgrom, P., R. Weber (2000). A theory of auctions and competitive bidding, II, in: P. Klemperer (ed.),
The Economic Theory of Auctions, Edward Elgar, Cheltenham, UK.
Mitchell, B., I. Vogelsang (1991). Telecommunications Pricing: Theory and Practice. Cambridge Uni-
versity Press, New York, NY.
Mo, J., J. Walrand (2000). Fair end-to-end window-based congestion control. IEEE/ACM Transactions
on Networking 8, 556–567.
Naor, P. (1969). On the regulation of queue size by levying tolls. Econometrica 37, 15–24.
Ockenfels, A., D. Reiley, K. Sadrieh (2006). ‘‘Online Auctions.’’ in: T. Hendershott (ed.), Handbook on
Economics and Information Systems. Elsevier, North Holland.
Odlyzko, A., P. Fishburn (1998). Dynamic behavior of differential pricing and quality of service options
for the internet. Proceedings First Intern. Confence on Information and Computation Economies (ICE-
98) 128–139.
Odlyzko, A. (1999a). The current state and likely evolution of the internet. Proceedings in Globecom’ 99,
1869–1875.
Odlyzko, A. (1999b). Paris metro pricing for the internet. Proceedings ACM Conference on Electronic
Commerce 140–147.
Odlyzko, A. (2003). Internet traffic growth: sources and implications, in: B. B. Dingel, W. Weiershausen,
A. K. Dutta, and K.-I. Sato (eds.), Optical Transmission Systems and Equipment for WDM Net-
working II, Proc. SPIE, Vol. 5247, pp. 1–5.
Oi, W. (1971). A disneyland dilemma: two-part tariffs for a mickey mouse monopoly. Quarterly Journal
of Economics 85, 77–96.
Oren, S., S. Smith (1981). Critical mass and tariff structure in electronic communications markets. The
Bell Journal of Economics 12, 467–487.
Pai, M., R. Vohra (2006). Optimal dynamic auctions. Working Paper, Northwestern University,
Evanston, IL.
Paschalidis, I.Ch., Y. Liu (2002). Pricing in multiservice loss networks: static pricing, asymptotic op-
timality, and demand substitution effects. IEEE/ACM Transactions on Networking 10, 425–438.
Paschalidis, I.Ch., J.N. Tsitsiklis (2000). Congestion-dependent pricing of network services. IEEE/ACM
Transactions On Networking 8, 171–184.
Rao, S., E.R. Petersen (1998). Optimal pricing of priority services. Operations Research 46, 46–56.
Roberts, J. (2004). A survey on statistical bandwidth sharing. Computer Networks: The International
Journal of Computer and Telecommunications Networking 45, 319–332.
Rohlfs, J. (1974). A theory of interdependent demand for a communications service. The Bell Journal of
Economics and Management Science 5, 16–37.
Ross, K. (1995). Multiservice Loss Models for Broadband Telecommunications Networks. Springer,
London.
Rump, C., S. Stidham (1998). Stability and chaos in input pricing for a service facility with adaptive
customer response to congestion. Management Science 44, 246–261.
Semret, N., A. Lazar. (1999). Spot and derivative markets in admission control. Proceedings of the 16th
International Teletraffic Congress.
Shenker, S. (1995). Service models and pricing policies for an integrated services Internet, in: B. Kahin,
J. Keller (eds.), Public Access to the Internet, Prentice Hall, Englewood Cliffs, NJ.
Shenker, S., D. Clark, D. Estrin, S. Herzog (1996). Pricing in computer networks: reshaping the research
agenda. ACM SIGCOMM Computer Communication Review 26, 19–43.
Shy, O. (2001). Dynamic peak-load pricing. Working Paper, University of Haifa.
Songhurst, D. (1999). Charging Communication Networks: From Theory to Practice. Elsevier, Amster-
dam.
Stahl, D. (2002). The inefficiency of first and second price auctions in dynamic stochastic environments.
Netnomics 4, 1–18.
Stahl, D., A. Whinston (1994). A general economic equilibrium model of distributed computing, in:
W. Cooper and A. Whinston (eds.), New Directions in Computational Economics, pp. 175–189.
Stidham, S. (1992). Pricing and capacity decisions for a service facility: stability and multiple local
optima. Management Science 38, 1121–1139.
Stidham, S. (2003). Pricing and congestion management in a network with heterogeneous users. Work-
ing Paper, University of North Carolina, Chapel Hill, NC.
Tanenbaum, A. (2003). Computer Networks 4th ed. Prentice-Hall, Upper Saddle River, NJ.
Thomas, P., D. Teneketzis, J. MacKie-Mason (2002). A market-based approach to optimal resource
allocation in integrated-services connection-oriented networks. Operations Research 50, 603–616.
Upton, D. (2002). Modelling the market for bandwidth. Ph.D. Dissertation, University of Cambridge,
Cambridge, UK.
Van Mieghem, J.A. (2000). Price and service discrimination in queuing systems: incentive compatibility
of Gcm scheduling. Management Science 46, 1249–1267.
Van Mieghem, J.A., P. Van Mieghem (2002). Price-coupled scheduling for differentiated services: GcM
vs. GPS. International Journal of Communication Systems 15, 429–452.
Varian, H. (2002). The demand for bandwidth: evidence from the INDEX project, in: R. Crandall, J.
Alleman (eds.), Broadband: Should We Regulate High-Speed Internet Access?, AEI-Brookings Joint
Center for Regulatory Studies, Washington, DC, pp. 39–56.
de Veciana, G., T. Lee, T. Konstantopoulos (2001). Stability and performance analysis of networks
supporting elastic services. IEEE/ACM Transactions on Networking 9, 2–14.
de Veciana, G., J. Walrand (1995). Effective bandwidths: call admission traffic policing and filtering for
ATM networks. Queueing Systems 20, 37–59.
Vickrey, W. (1961). Counterspeculation, auctions, and competitive sealed tenders. Journal of Finance 16,
8–37.
Vinnicombe, G. (2002). On the stability of networks operating TCP-like congestion control. Proceedings
15th IFAC World Congress on Automatic Control.
Walrand, J., P. Varaiya (2000). High-Performance Communications Networks 2nd ed. Morgan Kauf-
mann, San Francisco, CA.
Wang, Q., J. Peha, M. Sirbu (1997). Optimal pricing for integrated-services networks, in: L. McKnight,
J. Bailey (eds.), Internet Economics, MIT Press, Cambridge, MA, pp. 353–376.
Westland, J. (1992). Congestion and network externalities in the short run pricing of information system
services. Management Science 38, 992–1009.
Whang, S. (1990). Alternative mechanisms of allocating computer resources under queueing delays.
Wilson, R. (1993). Nonlinear Pricing. Oxford University Press, New York, NY.
Wilson, R. (1989). Efficient and competitive rationing. Econometrica 57, 1–40.
Wischik, D. (2001). How to mark fairly. Statistical Laboratory Research Report. University of Cam-
bridge, Cambridge, UK.
Wu, X., P. Marbach (2003). Pricing and rate control in a non-cooperative environment. IEEE Con-
ference on Decision and Control.
Yahalom, T., J.M. Harrison, S. Kumar (2006). Designing and pricing incentive compatible grades of
service in queueing systems. Working Paper, Stanford University, Stanford, CA.
Yechiali, U. (1971). On optimal balking rules and toll charges in the GI/M/1 Queuing Process. Op-
erations Research 19, 348–370.
Yechiali, U. (1972). Customers’ optimal joining rules for the GI/M/s queue. Management Science 18,
434–443.
Zhang, J. (1993). Performance study of Markov modulated fluid flow models with priority traffic.
Proceedings IEEE INFOCOM, pp. 10–17.
Zhang, Z. (1997). Large deviations and the general processor sharing scheduling for a two-queue system.
Queuing Systems 26, 229–254.
Zhang, Z., D.Dey, Y. Tan (2006). Pricing in communication services with delay guarantee. INFORMS
Journal on Computing, forthcoming.
Chapter 3
Firms and Networks in Two-Sided Markets
Daniel F. Spulber
Kellogg School of Management, Northwestern University, Evanston, IL 60208, USA
Abstract
The chapter presents a theory of two-sided markets. Firms create and operate
centralized allocation mechanisms both by matching buyers and sellers and
by market making. Buyers and sellers have the option of decentralized search,
matching, and bargaining. The chapter applies network theory to examine
the transaction costs of alternative centralized and decentralized allocation
mechanisms. It examines the efficiency of two-sided markets both for ho-
mogenous products and for differentiated products. Firms employ informa-
tion systems to improve communication between buyers and sellers and to
improve computation through centralized market mechanisms. Centralized
allocation mechanisms offered by firms can increase economic efficiency and
reduce transaction costs relative to decentralized exchange. Firms play a
critical economic role in establishing the microstructure of markets.
1 Introduction
All markets are two sided—they have a buyer side and a seller side. The
term two-sided market refers to a collection of individual buyers and indi-
vidual sellers such that the buyers on one side of the market can transact
only with the sellers on the other side of the market. Two-sided markets have
two main modes of organization: decentralized and centralized. In a decen-
tralized market, buyers and sellers match with each other and determine
transaction prices. In a centralized market, firms act as intermediaries be-
tween buyers and sellers, principally by matchmaking and market making.
In this article, I derive and present a number of new results on interme-
diation by firms in two-sided markets. I compare the relative efficiency of
137
138 D.F. Spulber
decentralized and centralized two-sided markets. I also review a number of

key developments in the literature on two-sided markets. The analysis sheds
light on the theory of the firm and on market mechanisms.
Two-sided markets present economic actors with a critical problem.
Buyers and sellers must find each other for the purpose of economic trans-
actions. In a decentralized market, buyers and sellers must search for each
other. In a centralized market, firms can provide matchmaking services that
improve information exchange between buyers and sellers or more ad-
vanced matchmaking services that implement assignments of buyers and
sellers.1
Another critical problem that buyers and sellers face in two-sided markets
is to determine their terms of trade. In a decentralized market, buyers and
sellers must handle their own negotiations through bilateral or multilateral
transactions. In a centralized market, firms provide market making services
that establish prices and balance supply and demand (see Spulber, 1999).
Firms employ information systems for communication and computation
functions that are necessary for matchmaking and market making services.
Firms use information systems to improve the effectiveness of information
exchange between buyers and sellers, to determine efficient assignments of
buyers and sellers, and to implement transaction mechanisms such as auc-
tions and posted prices. The result can be greater economic efficiency rel-
ative to decentralized exchange.
I apply network theory to represent decentralized and centralized two-
sided markets. I review some of the basics of network theory, which provides
highly useful tools for representing buyer–seller interactions and for mode-
ling transaction costs. The study of networks using graph theory covers
almost three centuries beginning in 1736 with Leonhard Euler’s celebrated
problem of the seven bridges of Königsberg (Euler, 1736). Network theory
sheds light on decentralized allocation mechanisms in which consumers
transact directly. Network theory also helps to explain centralized allocation
mechanisms operated by one or more firms acting as intermediaries.
The main contributions of the article are as follows. First, I explore the
implications of costly communication for the design of market mechanisms.
I extend existing results on allocation in buyer–seller networks (Kranton
and Minehart, 2000, 2001). I examine markets with homogenous products
in which not only buyers have differences in willingness to pay but also
sellers have cost differences. In addition, I extend the model of buyer–seller
networks to markets in which there are differentiated products. Using net-
work theory, I provide sufficient conditions under which allocations on
incomplete networks can be efficient both for markets with homogenous
products and markets with differentiated products.
1
An assignment is defined as a set of buyer–seller pairs that matches each buyer to a distinct seller or
each seller to a distinct buyer.
Ch. 3. Firms and Networks in Two-Sided Markets 139
Second, I examine the Gale–Shapley stable assignment in a decentralized

market and modify the model by adding money transfers between buyers
and sellers. I show that the stable assignment with differentiated products
and money transfers is unique. The stable assignment with homogenous
products is shown to be an efficient assignment. I show that with costly
communication due to an incomplete network, the stable assignment need
not be efficient. I consider the stable assignment with differentiated prod-
ucts and I give sufficient conditions for the stable assignment to be an
efficient assignment.
Third, I introduce a firm into the assignment problem. The firm can
charge subscribers for admission to a communications network. Once they
have subscribed to the network, buyers and sellers experience costless
communication so subscribers can attain a stable assignment. I examine
how a profit-maximizing firm chooses subscription prices for buyers and
sellers. I show the generality of Böhm-Bawerk’s (1891) method of marginal
pairs by extending it to a market with differentiated products. Additionally,
I show how the method of marginal pairs corresponds to a stable assign-
ment in markets for homogenous products and for differentiated products.
Fouth, I present a two-sided Vickrey auction by a monopoly firm in a
market with differentiated products. The method is related closely to dou-
ble Vickrey auctions in markets for homogenous and for differentiated
products. Finally, I consider the connection between search models and
random networks in two-sided markets. I compare the efficiency of buyer–
seller search with centralized assignments.
Economists have studied two-sided markets for well over a century.
Eugen von Böhm-Bawerk’s (1891) famous market for horses originates the
important idea of a two-sided market. Böhm-Bawerk presents what is
essentially a double auction 70 years before William Vickrey’s (1961) land-
mark work. Böhm-Bawerk introduces the term ‘‘two-sided competition,’’
stating that ‘‘The case of two-sided competition is the most common in
economic life, as it is the most important in the development of the Law of
Price. It demands, therefore, our most careful attention.’’ Rather than
characterizing supply and demand as aggregates, he begins with individual
buyers and sellers and their subjective valuations. Böhm-Bawerk (1891,
p. 213) observes that
If all are to exchange at one market price, the price must be such as to suit all exchanging parties;
and since, naturally, the price which suits the least capable contracting party suits, in a higher
degree, all the more capable, it follows quite naturally, that the relations of the last pair whom the
price must suit, or, as the case may be, the first pair whom it cannot suit, afford the standard for
the height of the price.
The market clearing price and quantity are determined by the marginal pair
who trade, that is, the buyer–seller pair who have the smallest positive
difference between the buyer’s value and that of the seller, or by the mar-
ginal pair who are excluded from trade.
140 D.F. Spulber
The two-sided market model is developed further by John von Neumann

and Oskar Morgenstern (1944). In their classic book Games and Economic
Behavior, they apply Böhm-Bawerk’s framework as their main, indeed their
only, economic interpretation of general non-zero-sum games.2 Lloyd
Shapley (1961) introduces a related multiplayer market game. David Gale
and Lloyd Shapley (1962) examine two-sided matching markets including
students and universities, marriage partners, and renters and housing.
Shapley and Shubik (1972, p. 111) employ the term ‘‘two-sided market’’
and observe that
Two-sided market models are important, as Cournot, Edgeworth, Böhm-Bawerk, and others
have observed, not only for the insights they may give into more general economic situations with
many types of traders, consumers, and producers, but also for the simple reason that in real life
many markets and most actual transactions are in fact bilateral—i.e., bring together a buyer and
a seller of a single commodity.
Shapley and Shubik (1972) analyze two-sided markets both for homoge-
nous products and for differentiated products.3
The traditional literature on two-sided markets in game theory focused
on decentralized exchange and thus has not given sufficient attention to the
role of the firm. The literature on market microstructure in economics and
finance has highlighted the role of the firm in intermediating exchange (see
Spulber, 1996a, b, 1998, 1999, 2002b). The industrial organization literature
has begun to extend analyses of product market competition to the con-
sideration of competition between intermediaries (see Stahl, 1988; Gehrig,
1993; Spulber, 1996b, 1999; Fingleton, 1997; Caillaud and Jullien, 2003;
Rust and Hall, 2003; Loertscher, 2004, 2005; Shevchenko, 2004; Weill,
2005; Hendershott and Zhang, 2006).
2 Firms in two-sided markets
The concept of a two-sided market is fundamental to models of inter-

mediation (see Spulber, 1996a, b, 1998, 1999, 2002a, b, 2003). Firms act as
matchmakers by bringing together buyers and sellers. Firms also act as
market makers by establishing and operating the institutions of exchange.
2
von Neumann and Morgenstern (1944, p. 562) observe that ‘‘The size of the transaction, i.e. the
number t0 of units transferred, is determined in accord with Böhm-Bawerk’s criterion of the ‘marginal
pairs.’’’
3
A substantial literature examines matching in two-sided markets. Much of this literature extends the
work of Gale and Shapley (1962) and Shapley and Shubik (1972). Hylland and Zeckhauser (1979)
examine efficient assignments by a social choice mechanism in which individuals report their preference
and the mechanisms assigns probabilities to position. Demange (1982), Leonard (1983), and Demange
and Gale (1985) examine auction-based allocation mechanisms. For an overview of this literature see
Roth and Sotomayor (1990). Two-sided matching markets include workers and employers, particularly
at the entry level in such specialized professional markets as law, medicine, and business (see Roth and
Xing, 1994). Becker (1973) presents a model of a marriage market in which types are multiplicative and
therefore act as complements.
Buyers Sellers
1 1
2 2
3 3
A decentralized market
Buyers Sellers
1 1
Firm
2 2
3 3
A centralized market
Fig. 1. In a decentralized market, buyers and sellers communicate directly and engage in
bilateral transactions. In a centralized market, a firm establishes a hub-and-spoke network
to provide matchmaking and market making services.
Spulber (1999) examines in a general setting how firms establish markets

and act as intermediaries between buyers and sellers.
Firms offer potential advantages over decentralized exchange in solving
problems associated with both matchmaking and market making. Firms
address the problems of costly communication and costly computation by
establishing hub-and-spoke networks of transactions. Fig. 1 represents de-
centralized and centralized transactions. In a hub-and-spoke network of
transactions, transaction costs are potentially lower because buyers and
sellers transact through the intermediary firm rather than forming bilateral
or multilateral transactions.4
In a decentralized market, individual buyers and sellers handle all com-
munication and computation tasks. Allocations in decentralized markets
are characterized by constraints on communication and computation.
Costly communication is likely to lead to random search and inefficient
matching of buyers and sellers. Costly computation is likely to involve
asymmetric information and inefficient allocation mechanisms. Individuals
thus encounter transaction costs that limit the efficiency of decentralized
exchange.
4
Such hub-and-spoke economies are widely recognized in communications, electric power transmis-
sion, and transportation. Hub-and-spoke economies also are well known in marketing, Alderson (1954)
(see also Townsend, 1978).
142 D.F. Spulber
In a centralized market, firms provide matchmaking and market making

services to buyers and sellers. Firms establish and operate information sys-
tems that supply buyers and sellers with some of the means to communicate
and process information. Firms engage in communication with buyers and
sellers to gather information about their characteristics and to provide
information about terms of exchange, such as prices and product features.
Firms also engage in computation to improve the efficiency of matchmak-
ing and market-making activities.
2.1 Firms in two-sided markets: matchmaking
One of the most important problems faced by buyers and sellers is to find
the best possible trading partner. By acting as matchmakers, firms help
buyers and sellers find each other. Matchmaking firms offer potential
advantages over random search by buyers and sellers.
Firms address the problem of costly communication by establishing hub-
and-spoke communications networks (see Fig. 1). Such networks effectively
connect all buyers to all sellers while offering advantages over point-
to-point networks that link buyers and sellers in decentralized exchange.
Point-to-point networks that are complete can be costly to establish because
they have so many links while incomplete point-to-point networks lead to
imperfect communication. The firm that serves many buyers and sellers
creates a network that gives its customers access to many suppliers and its
suppliers access to many customers (see Fig. 1).
Firms provide various types of centralized communications systems that
contribute to buyer and seller search activities. Postal systems involve cen-
tralized mechanisms for sorting and delivering mail (see Sidak and Spulber,
1996). The hub-and-spoke network also corresponds to a telecommunica-
tions network, including a traditional voice telephone system or computer
data transmission system. In a hub-and-spoke network, each buyer is linked
to every seller and each seller is linked to every buyer. The hub-and-spoke
communications network thus provides the same services as a more costly
complete point-to-point network.
Communications systems generally involve a physical infrastructure and
a system for making connections. A postal system not only involves facil-
ities for sorting and delivering mail but also depends on an established
system of addresses. Traditional telecommunications systems transmit voice
and data through a system of lines and switches. Traditional telecommu-
nications systems also give each user an address in the form of a telephone
number and provide guidance through telephone directories. The Internet
provides a transmission mechanism as well as a system of addresses for
electronic mail and for web sites. The Uniform Resource Locator (URL)
specifies both the protocol needed to retrieve information on a web site and
the domain name where the information is located. For basic postal,
telecommunications, and data transmission systems, subscribers choose
what other subscribers to contact without the help of the firm that supplies
communications services.
Some firms provide a very basic form of matchmaking in the form of
guides to subscribers of communications services. For example, suppliers of
telephone directories improve the information available to subscribers of a
telecommunications network. Yellow pages in particular allow buyers to
identify sellers of goods and services and obtain their telephone number.
Internet users require both the communications capability of the Internet to
link to web sites and the computational ability of search engines to locate
web sites. Internet auction sites, such as eBay offer both communication
between buyers and sellers and computation in the form of automated
auctions.
Firms offer more complex matchmaking services as well. Some match-
making firms operate in markets with homogenous products, introducing
buyers and sellers. Many other matchmaking firms operate in markets with
differentiated products, taking into account not just buyer and seller char-
acteristics but also the features of the products or services to be exchanged.
For example, real estate agents are matchmakers in the housing market
where every house is a differentiated product.
Firms that are brokers match buyers and sellers without involving them-
selves as a party in the economic transactions between buyers and sellers.
Table 1, while not exhaustive, gives some idea of the complexity and range
of available matchmaking services.
2.2 Firms in two-sided markets: market making
Market-making firms aggregate demand and aggregate supply, offering

potential advantages over bilateral transactions. Market-making firms also
perform critical price adjustment functions, applying computation to cal-
culate posted prices or to operate auctions. As market makers, firms per-
form dealer functions by buying from sellers and reselling to buyers. For a
partial listing of firms that provide market-making services see Table 1.
Firms play a critical role in coordinating transactions between buyers and
sellers. Buyers and sellers decide whether or not to trade on organized ex-
changes for financial assets depending on their expectations of market par-
ticipation. If there are sufficient numbers of buyers and sellers, the market for
financial assets is said to offer liquidity, while in product markets, the cor-
responding idea is immediacy.5 Buyers participate in markets based on their
expectations of seller participation and sellers participate in markets based on
their expectations of buyer participation. Market makers provide coordina-
tion services by standing ready to buy and sell, assuring buyers and sellers of
liquidity if there is not a sufficient number of counter parties available.
5
See Clower and Leijonhufvud (1975).
144 D.F. Spulber
Table 1
Partial listing of types of firms that provide matchmaking and market-making services
Firms are intermediaries in two-sided markets

Matchmakers Market makers
Communications intermediaries Posted price market makers

Telecommunications firms Retailers
Data networks Wholesalers
Postal systems Securities dealers, market makers, and
Basic information providers(infomediaries) specialist firms
Search engines and online portals (Yahoo, Mutual funds
Google) Banks and insurance companies
Directories and yellow pages Commodities firms (metals, energy, grains)
Publishers of classified advertisements Auction firms
Comparison shopping web sites Auctioneers (real estate, art, antiques,
Enhanced service matchmakers commodities, agricultural products,
Dating and marriage matchmakers industrial equipment)
Interest group matchmakers Online auctions (eBay)
Residential and commercial real estate brokers Organized exchanges for securities, currencies,
Employment and recruiting agencies financial assets, commodities
Staffing and temporary agencies Business-to-business market places
Agents and brokers
Representative agents (literary, talent, sports)
Travel agents and freight transportation
brokers
Media (book publishers, journals,
broadcasters, movie studios)
Payments systems (credit cards, check clearing,
money transfers)
Securities brokers
Business brokers (for buyers and sellers of
businesses)
Technology and intellectual property brokers
Commodity and metals brokers
Ship brokers
Art brokers
Consignment, resale, and antique stores
Rubinstein and Wolinsky (1987) consider a random matching model with

buyers, sellers, and intermediaries that act as dealers. In a related setting,
Yavas (1994a) examines the role of intermediaries that act as matchmakers
in two-sided markets where both buyers and sellers search, considering such
intermediaries as employment agencies and real estate brokers. In Yavas
(1994a), the presence of the matchmaker affects the endogenously chosen
search intensity of buyers and sellers.
Gehrig (1993) addresses the coordination problem when there is compe-
tition between a market maker and a decentralized search market. Buyers
and sellers choose between dealing with the market-making intermediary
and direct exchange in the search market based on the price spread offered
by the intermediary and expectations about the participation in the search
market. Yavas (1996a) considers competition between a market maker and
a search market when buyer and seller search intensities are endogenous;
Yavas (1992) lets the intermediary choose between being a market maker
and a matchmaker.
Spulber (2002a) considers competition between a market maker and an
intermediary when buyers and sellers make investments that increase the
returns from exchange. Extending Spulber’s (1996b) model of competition
between intermediaries in a search market, Rust and Hall (2003) consider
competition between multiple dealers and a central market maker. Buyers
and sellers choose between searching across dealers for the best price and
transacting with the market maker at observable posted prices. Hendershott
and Zhang (2006), in a model related to Spulber (1996b), examine com-
petition between upstream firms who can sell directly to consumers or
through intermediaries. Ellison, Fudenberg, and Mobius (2003) consider
how two auctions can compete without one driving out the other from the
market.
Competition between market makers is a subject that draws increasing
interest. Stahl (1988) considers competition between intermediaries engaged
in Bertrand competition for inputs and subsequent Bertrand competition
for outputs. Spulber (1999, Chapter 3) examines competition between in-
termediaries in various settings. Spulber (1999, Chapter 3) introduces a
model of a Hotelling-type differentiated duopoly in a two-sided market.
Because the intermediary firms compete in a differentiated duopoly, the
equilibrium bid–ask spread is positive and depends on the relative transport
costs on the buyer side and the seller side of the market. Spulber (1999,
Chapter 3) also introduces a model of a two-sided market in which both
buyers and sellers face switching costs in changing intermediaries and the
equilibrium bid–ask spread is also positive.
Gehrig (1993) allows Bertrand competition between intermediaries
who also compete with the decentralized search market and shows that
competition eliminates the bid–ask spread. Loertscher (2004) introduces
capacity constraints on intermediaries leading to rationing and positive
bid–ask spreads. Spulber (1996b) obtains price dispersion in both bid and
ask prices when intermediaries compete in a search setting. Fingleton (1997)
examines a model of competition similar to Stahl (1988) that allows direct
trade between buyers and sellers. Shevchenko (2004) examines competi-
tion between intermediaries in a search setting when market makers can
hold inventories of varying size. Weill (2005) examines how compet-
ing market makers provide liquidity in financial markets during financial
disruptions.
The firm as an intermediary can be observed in a wide variety of industry
applications. Yavas (1996b) models the matching of buyers and sellers by
real estate brokers and compares alternative commission structures (see also
146 D.F. Spulber
Yavas, 1994b). Baye and Morgan (2001) study Internet intermediaries that
act as information gatekeepers. Baye et al. (2004) consider Internet com-
parison sites that perform match-making functions in two-sided markets
and their effects on price dispersion across online retailers. Lucking-Reiley
and Spulber (2001) examine business-to-business intermediaries that act as
dealers and operate online marketplaces. Some other examples of networks
and two-sided markets include information services (Caillaud and Jullien,
2003) and yellow pages (Rysman, 2004). Ju et al. (2004) consider oligopo-
listic market makers in the natural gas industry.
In product markets, retailers perform various types of market-making
functions. Retailers include online merchants, automobile dealers, super-
markets, discount stores, department stores, general merchandise stores,
specialty apparel stores, warehouse clubs, drug stores, convenience stores,
and variety stores (United States Census Bureau, 2000). Retailers provide a
wide variety of intermediation services including posting prices, market
clearing, marketing, inventory holding, selection of suppliers, quality cer-
tification, and management of transactions. Retailers employ information
systems in intermediation including web sites for communication with cus-
tomers and suppliers, and data collection and processing through bar cod-
ing of merchandise, point-of-sale scanners, and computerized inventory
tracking and reordering.
Wholesalers also provide market-making functions. Wholesalers prima-
rily include merchant wholesalers that purchase and resell durable and
nondurable goods including distributors, jobbers, drop shippers, import/
export merchants, grain elevators, and farm product assemblers. They also
include agents such as brokers, commission merchants, import/export
agents and brokers, auction companies, and manufacturers’ agents (United
States Census Bureau, 2000). Wholesalers also include direct manufacturer–
retailer transactions (retail chain stores, warehouse clubs, discount stores,
and home center stores), mail order, catalog sales, manufacturer–industrial
user transactions, and retail sales to industrial users. In addition, there are
also manufacturer’s sales branches, agents, brokers, and commission mer-
chants. Wholesalers also provide a variety of intermediation services as
intermediaries between businesses; they distribute goods, manage invento-
ries, communicate price and product information, certify quality, and pro-
vide credit. Wholesalers employ information systems in communicating
with retailers and manufacturers and many types of information gathering
devices for bar coding, electronic data interchange, product tracking, in-
ventory controls, and distribution.
Financial firms that provide intermediation services include banks, secu-
rities brokerages, mutual funds and insurance companies (United States
Census Bureau, 2000). Financial intermediaries provide many types of
services including pricing of some financial assets, providing liquidity, risk
allocation, allocation of financial assets over time, combining assets to re-
duce the transaction costs of diversification, supplying information, and
managing transactions. Depository institutions intermediate between bor-

rowers and lenders, setting rates of interest for loans and deposits, screening
borrowers for credit worthiness, and monitoring their repayment perform-
ance. Securities and commodity brokers provide a range of intermediation
services managing complex financial transactions, carrying out trades on
the organized exchanges, and supplying investors with information. Insur-
ance companies manage transactions, allocate risk, and intermediate be-
tween investors and buyers of insurance contracts.
Organized exchanges for securities, commodities, currencies, and other as-
sets provide market-making services. Some organized exchanges are operated
by firms. Although many organized exchanges are associations of members,
the incorporation of some exchanges such as the New York Stock Exchange,
demonstrates that these markets also are owned and operated by firms.
2.3 Information systems and platforms
Firms are intermediaries that create and operate markets. Information

systems are part of these firms’ production technology. Firms use infor-
mation systems to communicate with buyers and sellers and to provide
computation services needed to carry out transactions. Although informa-
tion systems are a critical component of transaction technology, informa-
tion systems are not the same thing as a firm—they are tools used by firms.
Information systems contain two essential components: communication,
which refers to the exchange of information between individuals, and com-
putation, which refers to the processing of data that is being exchanged.
Information systems generally comprise physical networks involving tele-
communications and connected computers.6 Information systems must in-
teract with human intelligence and facilitate interaction within economic
and social networks.
Information technology is part of the firm’s production function for cre-
ating transactions. The costs of communication and computation are part of
the transaction costs both in decentralized and centralized markets. When
communication is costly, buyers and sellers deal with incomplete networks.
When computation is costly, buyers and sellers engage in bilateral trans-
actions rather than more complex multilateral transactions such as those
represented by the Core. There is a role for firms to provide communication
and computation in the form of matchmaking and market–making services.
6
An industry definition of information system states ‘‘1. A system, whether automated or manual, that
comprises people, machines, and/or methods organized to collect, process, transmit, and disseminate
data that represent user information. 2. Any telecommunications and/or computer related equipment or
interconnected system or subsystems of equipment that is used in the acquisition, storage, manipulation,
management, movement, control, display, switching, interchange, transmission, or reception of voice
and/or data, and includes software, firmware, and hardware. y 3. The entire infrastructure, organ-
ization, personnel, and components for the collection, processing, storage, transmission, display, dis-
semination, and disposition of information.’’ (see Committee T1A1, 2000).
148 D.F. Spulber
Table 2 presents a classification of market mechanisms that is based on the

costs of communication and computation. This classification is useful in
understanding the role of the firm in two-sided markets.
Markets consist of transactions between buyers, sellers, and firms. Mar-
kets require the services of information systems to function effectively. In-
dividuals must communicate to find each other, negotiate the terms of
transactions, and monitor performance. Individuals must perform compu-
tations to choose between trading partners and to evaluate alternative terms
of exchange. Market outcomes are affected by costs of communication and
costs of computation. Transaction costs and the capabilities of information
systems affect the structure and performance of market mechanisms.
The classification scheme in Table 2 connects transaction costs and the
properties of information systems to market mechanisms. If communica-
tion is costless, buyers and sellers can find each other with greatly reduced
search costs. Buyers and sellers may still face problems of asymmetric in-
formation and the time costs of search. When computation is costless,
buyers and sellers face greatly reduced costs of determining the best as-
signment of buyers and sellers, market-clearing prices or the allocation of
goods. If both communication and computation are costless, then in prin-
ciple decentralized exchange can achieve efficient outcomes by implement-
ing such market mechanisms as auctions or solutions to cooperative games.
When communication or computation is costly, there may be advantages to
centralized markets.
Firms offer customers various types of matchmaking and market-making
services through the use of information systems. Firms that provide tele-
communications and data transmission are information intermediaries.
Additional, firms extensively use communications systems internally to
communicate with their customers and suppliers. Firms employ commu-
nications systems to automate transactions with their customers and sup-
pliers as well as automating the myriad internal transactions associated
production operations, back office record keeping, and organizational
management.
Electronic commerce refers to the automation of economic transac-
tions through the use of information systems. Electronic commerce lowers
transaction costs by enhancing communications and computation in ex-
change between consumers, between consumers and firms, and between
firms. Electronic commerce substitutes capital for labor services in the
production of transactions, potentially displaces costly labor services ap-
plied to routine commercial tasks including the time that employees spend
communicating with customers and suppliers regarding prices, product
availability, ordering, billing, and shipping. Moreover, electronic commerce
enhances the productivity of labor services in commercial activities such as
sales, distribution, and procurement. Firms improve efficiency by linking
external transaction systems with their internal computer systems, thus
increasing the frequency, rapidity, and accuracy of communication and
Table 2
Classification of allocation mechanisms depending on the costs of communication and
computation
Computation Costless computation Costly computation
Communication
Costless communication Buyers and sellers in a Buyers and sellers in a

decentralized market can decentralized market
communicate costlessly and communicate over
can costlessly implement complete networks (due
efficient mechanisms for to costless
exchange: Core allocations communication) and
and Walrasian equilibria engage in bilateral
exchange (due to costly
computation)
There is no need for firms to Firms offer a centralized
provide communication or market alternative to
computation services buyers and sellers. Firms
provide computation
services by offering
advanced matchmaking
services to assign buyers
to sellers by offering
market making services
to specify terms of trade
and clear markets
Costly communication Buyers and sellers in a Buyers and sellers in a
decentralized market can decentralized market
implement efficient engage in bilateral
mechanisms (auctions and exchange due to costly
cooperative game solutions) computation and
due to costless computation transact on incomplete
but they transact on networks or on random
incomplete networks due to networks due to costly
costly communication communication
Firms offer a centralized Firms offer a centralized
market alternative to buyers market alternative to
and sellers. Firms provide buyers and sellers. Firms
communication services provide separate or
(gathering and distributing bundled communication
information, facilitating and computation
information exchange) and services to buyers and to
basic matchmaking services sellers through
with buyers and sellers matchmaking and
handling their own market making services
computation activities.
150 D.F. Spulber
allowing links to production and inventory management systems within

each organization (see Lucking-Reiley and Spulber, 2001).
A platform is a collection of related technological standards. The term
‘‘platform’’ sometimes is used in the economics literature as shorthand to
designate firms that act as intermediaries in electronic commerce. Although
firms use information systems that obey common technical standards, a plat-
form is not a firm. The firm’s technology and the applicable technology
standards are distinct from the firm itself. The use of technological standards
or information system platforms to implement the linkage of buyers and
sellers through networks is related to the earlier discussions of compatibility of
products in a network (see Farrell and Saloner, 1985, 1986; Katz and Shapiro,
1985, 1986, 1994; Liebowitz and Margolis, 1994, 1999; Liebowitz, 2002).
Platforms play an important role in information systems, both in com-
munications and in computing. In computers, a platform is a ‘‘reconfigur-
able base of compatible components on which users build applications’’ and
is identified with ‘‘engineering specifications for compatible hardware and
software,’’ Bresnahan and Greenstein (1996) (see also Bresnahan and
Greenstein, 1999; Greenstein, 1998). For example, IBM devised standards
for the personal computer that were adapted by manufacturers of software
designers, internal components, such as memory and microprocessors, and
peripheral devices, such as printers and monitors. In turn, Microsoft’s
standards for the personal computer operating systems are used by design-
ers of software applications that are compatible with operating systems.
In communications networks, platforms permit compatible transmission
of information in communications and interconnection of transmission
equipment (see Spulber and Yoo, 2005). Platforms in telecommunications
include computer hardware and software standards for computer-based
switching and transmission systems. Platforms in communications include
computer software standards such as the Transmission Control Protocol/
Internet Protocol (TCP/IP) used for Internet communications between
computers. A network is said to be modular or to exhibit an open architecture
if most suppliers of complementary services can gain access to the network.
Collections of technical standards exist in many industries where inde-
pendent producers supply substitute products that are interchangeable and
complementary products that must work together. Thus, cameras and film
share technological standards that allow the products to be used together,
and there are multiple providers of cameras and of film that follow the
technical standards. These standards exist in many high-tech industries such
as audio systems, video systems, and mobile phones. Platforms exist in
many other types of industries in which compatible components are needed
including automobiles, aircraft, and industrial machinery.
Platforms exist in electronic commerce, in the form of technical standards
for the electronic exchange of data between companies. Innovations in
communications and computation as applied to business documents avoids
the need to translate computer files into paper documents, thereby increasing
the speed and accuracy of transactions. There is a wide-ranging set of

standards for electronic data interchange (EDI) on private networks that
predates the Internet. Extensible markup language (XML) provides stand-
ards for documents and data transmission over the Internet developed by the
World Wide Web Consortium. The advantage of document standardization
is ease of communication and computation between businesses, including
retailers, wholesalers, manufacturers, and parts suppliers.
Electronic commerce further enhances communication of transaction in-
formation by allowing buyers and sellers to transact with the firm at remote
locations and at different times. Thus, the buyers and sellers in an auction
on eBay need not be present at the same location and can participate in the
auction at different times. This reduces the transaction costs by avoiding the
costs of travel and the costs of holding meetings whether those costs would
be borne by the firm or its customers and suppliers. Thus, technological
change in information processing and communications result in innovations
in transaction methods and changes in the organization of firms. Increasing
standardization of economic transactions provides additional economic
benefits from centralization of markets.
3 Networks in two-sided markets
Graphs are sets of points and lines that connect some of those points to
each other. The points are referred to as nodes and the lines are referred to
as links. Networks are graphs in which numerical values are assigned to the
links.7 In economic networks studied in this article, nodes in the graph
represent economic actors. The links that connect the nodes in a graph
represent some important aspect of the relationship between those eco-
nomic actors. The architecture or configuration of the set of nodes and links
provides a representation of the market mechanism. The pattern of links
can be used to represent decentralized exchange in which buyers and sellers
search for each other or centralized exchange with firms as intermediaries,
as shown previously in Fig. 1. This section presents some of the basics of
graph theory that will be useful in the discussion.
3.1 Transactions on networks in two-sided markets
The theory of networks provides a highly useful framework for under-

standing how markets work. The theory of networks uses the mathematics
7
For an introduction to graph theory see Bollobás (1998), Aldous and Wilson (2000), Tutte (2001),
Gross and Yellen (1999, 2004), and Diestel (2000). For a more popular treatment see for example Watts
(2003). Networks can represent facilities used by firms to provide communication of information,
transmission of energy, and transportation of goods and services. Networks also can represent indi-
viduals in society and their relationships. There are also many scientific applications using networks in
physics, chemistry, biology, and other areas.
152 D.F. Spulber
of graph theory to represent markets as networks. Important developments

in the field of graph theory offer a rich set of useful mathematical tools.
These tools can be incorporated in economic models to show the interaction
of buyers and sellers. These concepts help our understanding of markets
and the role of the firm. They also illuminate the critical function of in-
formation systems in economic transactions.
I apply network theory to understand the performance of markets with
and without intermediation by firms. Networks can represent decentralized
allocation mechanisms used by buyers and sellers in decentralized markets.
Networks also can represent centralized allocation mechanisms established
by firms. Techniques adapted from graph theory can be used to study how
firms design efficient allocation mechanisms. The results shed light on the
rules, organization and institutions of exchange, referred to as market
microstructure, for all types of markets including those for products, fi-
nancial assets, and labor services.
Network theory contributes to the theory of the firm by allowing com-
parisons between decentralized exchange and markets mediated by firms.
First, networks can be used to find those connections between buyers and
sellers that yield the greatest gains from trade. Second, networks can be
used to find the connections between buyers and sellers that have the lowest
transaction costs. Third, networks can be used to determine what connec-
tions between buyers and sellers are the most efficient. The most efficient
connections are those that generate the greatest benefits from transactions
net of the costs of transactions. Firms exist when they can provide the most
efficient networks in comparison to decentralized networks.
Network theory helps to illustrate the complexity of assignments of buy-
ers and sellers. The discussion thus far showed that buyers and sellers in
decentralized exchange face two difficult problems. Buyers and sellers must
establish the network communication links that permit the selection of an
efficient assignment. This may be costly since carrying out an efficient as-
signment can require establishing a complete network. Second, buyers and
sellers must make the complex computations that are required to select an
assignment among the many possible permutations.
There is an extensive literature on the strategic formation of networks by
buyers and sellers (see Dutta and Jackson, 2003) and the references therein.8
Kranton and Minehart (2001) present a theory of buyer–seller networks
that has the following features. Products are homogenous and sellers have
identical costs equal to zero. Buyers each purchase a single unit and sellers
8
Myerson (1977) introduces networks to cooperative game theory and obtains a constrained Shapley
value that depends on available links between players. Jackson and Wolinsky (1996) consider coop-
erative games in which the value function depends on the structure of the network. They define an
equilibrium in the endogenous formation of networks as occurring when no pair of individuals wants to
create a link that is absent or to dissolve a link that is present. They contrast efficiency and stability of
networks, although it should be pointed out that their equilibrium notion of stability in link formation
differs from the notion of stability in matching models (see also Dutta and Mutuswami, 1997).
each offer a single unit. Buyers differ in their willingness to pay for a unit of
the good. Buyers can only transact with sellers with whom they are con-
nected in a network. For those buyers and sellers that are connected, sellers
hold ascending-bid auctions with the same going price across all sellers. The
price rises until there is a subset of sellers for whom demand equals supply.
The auctions of these sellers then clear at the current price. If there are
remaining sellers, the price rises until all sellers have sold their goods.
Kranton and Minehart (2001) show that the equilibrium allocation of
goods is efficient since the highest-value buyers receive the goods, subject to
constraints imposed by the pattern of links. They further show that the
allocation of goods is pair-wise stable in the sense that the surplus that any
buyer and seller could obtain by exchanging their goods does not exceed
their joint payoffs. They further demonstrate that efficient link patterns are
equilibrium outcomes of the game played by buyers and sellers.
The present discussion differs from Kranton and Minehart (2000, 2001)
in several critical areas. First, they focus on decentralized markets in which
buyers and sellers handle their own communication and computation, so
that buyers and sellers can establish their own links to communicate and
transact with each other. Kranton and Minehart (2001) refer to their proc-
ess as a ‘‘frictionless model of network competition’’ because they allow
buyers and sellers to hold auctions without incurring transaction costs.
Since I am interested in transaction costs and the role of the firm, I assume
that buyers and sellers face constraints in communication and computation.
In my framework, when buyers and sellers are in a decentralized market,
they are limited to bilateral exchange. I require the presence of a firm to
establish centralized allocation mechanisms such as auctions. Second, while
Kranton and Minehart (2000, 2001) focus on homogenous products, I
consider both homogenous and differentiated products which have impor-
tant consequences for efficiency. Third, while Kranton and Minehart (2000,
2001) assume that sellers have equal costs, I allow sellers to have different
costs, which creates some difficulties for allocation mechanisms. I do not
assume that buyers and sellers can hold auctions on their own. In my
framework, centralized matching and operation of auctions require inter-
mediation by firms.
Graph theory has many economic applications. Graphs can be applied to
determine how buyers and sellers are matched in two-sided markets. The
links represent transactions and the properties of those transactions. The
values assigned to the links between two economic actors can represent
transaction benefits. The value assigned to the link can show the potential
gains from trade of a transaction. Those transactions with nonnegative
gains from trade are economically feasible. Thus, the value of the links
shows what transactions are feasible.
The links between economic actors also can represent transaction costs.
One can assign numerical values to the links equal to the potential cost of
transacting between the two individuals. The presence or absence of links
154 D.F. Spulber
provides a highly useful if somewhat more dramatic type of transaction

costs. If a link is present, transaction costs are zero and if a link is absent,
transaction costs are infinite. This gives a picture of feasible transactions.
Such links represent communication that allows economic interaction bet-
ween buyers and sellers.9 Such preconditions for trade also include whether
or not the two parties are already acquainted, perhaps they are connected
through personal, business, or social ties. They may be part of the same
social group or business association.10
The links between buyers and sellers can be established endogenously.
Suppose that a link between a buyer i and a seller j can be established at a
cost kij then, kij represents the transaction cost required before trade can
take place. This is the cost of linking buyers and sellers. If links must
precede trade, then kij is the ex ante transaction cost. After links are es-
tablished, then the ex post transaction costs again are either zero or in-
finite.11 If k is the cost of a link, the total costs of a point-to-point network
kmn are replaced by the lower costs of a hub-and-spoke network, k(n+m).
3.2 Basics of network theory
A graph is a pair G ¼ (J, L) consisting of a set of nodes J and a set of

links L that connect the nodes.12 For example, in Fig. 2, the set of nodes is
J ¼ {1, 2, 3, 4} and the set of links is L ¼ {(1, 2), (3, 4), (1, 4)}. A network is
a graph with numerical values assigned to the links. The numerical values of
the links can represent such things as gains from trade, transaction costs,
and transportation capacity. For example, in Fig. 2 the gains from trade
associated with the links (i, j) are represented as aij.
The interaction of buyers and sellers can be represented using bipartite
graphs. In a bipartite graph, the set of nodes is partitioned into two sets, and
the set of links has nodes that are in different sets. Let J ¼ (B, S) represent a
partition of the set of nodes into those in the set of buyers and those in the
9
The link between two economic agents can represent a physical communication channel, particularly
one provided by telecommunications and Internet systems for transmitting voice and data. Commu-
nications networks are an important type of information system and they allow buyers and sellers to
establish economic transactions. Links also represent pathways for travel and transportation, travel so
that buyers and sellers can visit each other and transportation so that trading partners can send goods
and services to each other.
10
Granovettor (1985) emphasizes the importance of strong ties in connecting individuals within social
groups and the importance of weak ties in connecting social groups with each other. Weak ties can be
more important than strong ties for adding value to economic transactions because they help establish
critical connections between many individuals.
11
One might suppose that with full information, buyers and sellers could establish the minimum
number of links necessary for efficiency, n*. But, if we require information to be transmitted over links,
then full information would require a complete set of links for consumers to communicate that in-
formation. Thus, if links are channels needed for communication, full information is not possible
without a complete set of links.
12
In the mathematics literature, nodes also are referred to as vertices and links also are referred to as
edges.
1 2
a12 = 5
a14 = 7
a34 = 2
4 3
Fig. 2. A network consisting of a graph G ¼ (J, L) with nodes J ¼ {1, 2, 3, 4}, links
L ¼ {(1, 2), (3, 4), (1, 4)}, and numerical values assigned to the links.
set of sellers, B[S ¼ J and B\S ¼ null set. Let L be a set of links such that
the nodes for each link are not adjacent, that is, one link is in the buyer
partition class and the other link is in the seller partition class. For con-
venience, we write a link (b, s) to indicate that the first node is in the set of
buyers and the second node is in the set of sellers. The set of buyers is
B ¼ {1, y, n} and the set of sellers is S ¼ {1, y, m}.
Define a neighbor of a node as the other node of a link that is connected
to that initial node. Let N(v) be the set of all neighbors of the node v. Thus,
N(b) is the set of sellers connected to the buyer b. Also, N(B) is the set of
neighbors of all buyers in B. If two nodes are neighbors, they are said to be
adjacent. Two links are said to be adjacent if they share a node. Two links
are said to be independent if they are not adjacent, that is, if the two links do
not share a node.
A bipartite graph is complete if all nodes in one partition class are adjacent
with all nodes in the other partition class. Thus, if a bipartite graph is com-
plete, each buyer is connected with every seller. This would require nm links if
there are n buyers and m sellers. Let LC represent the complete set of links for
the set of nodes J ¼ (B, S). Then, the complete graph is GC ¼ ((B, S), LC).
In a bipartite graph, an assignment of the set of buyers B associates each
buyer with a distinct seller in S. Notice that an assignment in a bipartite
graph is defined with respect to one of the partition classes. Such an as-
signment would involve n links and only exists if there are enough sellers,
nrm. An assignment is a set of independent links contained in a graph. A
necessary and sufficient condition for a bipartite graph to contain such an
assignment is Hall’s (1935) Marriage Theorem (see also Diestel, 2000, p. 31).
Marriage Theorem (Hall). Let N(A) be the neighbors of A in the set S for
ADB. The bipartite graph G ¼ ((B, S), L) contains an assignment of B if
and only if |N(A)|Z|A| for all ADB.
The theorem states that the graph contains an assignment if and only if every
subset of buyers has enough neighbors in the set of sellers. A corresponding
necessary and sufficient condition applies to the set of sellers. Clearly, there
156 D.F. Spulber
Buyers Sellers
1 1
2 2
3 3
4 4
Fig. 3. A bipartite graph that does not contain an assignment.
is a one-to-one matching of buyers and sellers when|B| ¼ |S|, that is, n ¼ m.

Fig. 3 shows a bipartite graph that does not contain an assignment.
Hall’s Marriage Theorem gives conditions under which, if each buyer is
connected to a subset of sellers, each buyer will be able to be matched with a
seller with whom that buyer is connected. The condition also can be stated
in terms of sellers being connected to buyers. Informally, the theorem con-
siders the number of connections needed for transactions to take place.13
The theorem is useful for assigning buyers to sellers or sellers to buyers.
Hall’s Marriage Theorem addresses the case in which all buyers are
identical and all sellers are identical. The only issue is the pattern of links
that connect them. Kranton and Minehart (2000, 2001) examine an im-
portant application of Hall’s Marriage Theorem. In their setting, buyers are
no longer identical but instead have different valuations of a good. All of
the sellers are identical because they have the same cost of supplying the
good to a buyer. Buyers then are assigned to the set of sellers. Their analysis
focuses on making sure that members of the set of sellers are allocated to
the highest-value buyers. In the next section, I consider a more general case
that allows there to be both diverse buyers and diverse sellers.
4 Assignments of buyers and sellers in a network: costly communication
Network theory provides insights into how costly communication affects

economic efficiency. Costly communication takes the form of incomplete
networks. This section assumes that there is costless computation so that
the allocation mechanism performs efficiently. The outcome is inefficient
due to communication constraints in the form of missing links between
buyers and sellers. When there are missing links, it may not be possible to
achieve efficient assignments of buyers to sellers.
Consider a market with consumers divided into two distinct groups, a set
of buyers B ¼ {1, y, n} and a set of sellers S ¼ {1, y, m}. Let i denote an
13
The traditional story of the Marriage Theorem discusses marriage to someone with whom the
woman (or man) is acquainted, depending on the direction of the assignment.
element of the set of buyers and let j denote an element of the set of sellers.
To give some idea of the complexity of the problem, consider the number of
possible assignments. When these are n buyers and n sellers, there are n!
possible assignments. Thus, with 10 buyers and 10 sellers there are over 3.6
million solutions. Moreover, the efficient assignment need not be unique.
Choosing an efficient assignment depends on both gains from trade and
the costs of trade. In an economy without transaction costs, buyers and
sellers will choose only the best transactions, that is, those transactions that
maximize gains from trade. When there are transaction costs, the set of
efficient transactions changes. If a buyer–seller pair (i, j) is in the set of links
L, then buyer i and seller j can transact. Thus, if (i, j) is in the set of links L
then the transaction costs for buyer i and seller j equal zero. Conversely, if a
buyer–seller pair (i, j) is not in the set of links L, the buyer i and seller j
cannot transact. This means that if (i, j) is not in the set of links L, then the
transaction costs for buyer i and seller j are infinite. This section applies
network theory to determine the set of efficient transactions when there are
both transaction benefits and transaction costs.
4.1 Assignments with homogenous products
Every buyer wishes to purchase a unit of a homogenous good and every

seller can supply a unit of the good. Buyers have valuations {v1, y, vn}, that
are arranged in decreasing order, v1>v2>y>vn. Sellers have costs {c1, y,
cm} that are arranged in increasing order, c1oc2oyocm. For any link
between a node i in the set of buyers and a node j in the set of sellers, we can
assign a value equal to
vij ¼ max f0; vj cj g. (1)
This represents the total of gains from trade that buyer i and seller j would
obtain from a transaction. This is also referred to as the value of a match.
See Fig. 4 for a network in which all gains from trade are positive.
An assignment of buyers to sellers matches every buyer i in B to a distinct
seller j in S. There is no loss of generality in framing the discussion in terms
of assigning buyers to sellers. If the number of buyers exceeds the number
Buyers Sellers
v1 – c1
1 1
v1 – c2
v2 – c1
v2 – c2
2 2
Fig. 4. A network with values assigned to the links representing the potential gains from
trade that would result from a transaction between pairs of buyers and sellers.
158 D.F. Spulber
of sellers, then one can create dummy sellers with high costs such that there
would be no trade between a buyer and such a seller. The benefits of an
assignment equal the sum of the gains from trade for any buyer–seller pair
in the assignment that has positive gains from trade vi–cj. For any bipartite
graph with a set of buyers and a set of sellers, there can be many possible
assignments of buyers to sellers. An efficient assignment of B to S is an
assignment that maximizes total net benefits over the set of assignments of
B to S. Thus, efficiency is defined strictly in terms of gains from trade.
When transaction costs are introduced, we will define a constrained effi-
cient assignment as one that maximizes gains from trade subject to feasibility.
A feasible assignment of B to S for a graph G is an assignment such that each
buyer–seller pair (i, j) in the assignment is in the graph. Thus, a feasible
assignment is a subset of the set of links in the graph G such that not all
buyers need to be assigned to a seller but each buyer that is assigned is
matched to a distinct seller. A constrained efficient assignment is defined to be
a feasible assignment that maximizes total net benefits over the set of feasible
assignments from B to S.
Consider the set of efficient assignments. Let n be the largest integer i
such that
vi cj
where i ¼ j, iAB and jAS. The number n is the market-clearing quantity.
Let B ¼ {1, y, n} and S ¼ {1, y, n}. The sets B and S describe the
buyers and the sellers that are on the left-hand side of the supply and
demand diagram. The excluded buyers i>n and the excluded sellers j>n
are those on the right-hand side of the supply and demand diagram
(see Fig. 5).14 This outcome corresponds to Böhm-Bawerk’s (1891) method
of marginal pairs.
Let E be an assignment of B to S. In other words, E is a set of distinct
pairs that matches every buyer in B with a different seller in S. The sets B
and S have the same number of elements by construction. Any such as-
signment E is an efficient assignment of B to S. The proof of this result is
immediate. For any such E, number the distinct pairs k ¼ 1, y, n. Then,
X n
X n
X
ðvk ck Þ ¼ vi cj . (2)
k2E i¼1 j¼1
The right-hand side equals the maximum net benefits that can be obtained
from matching buyers in B and sellers in S.
This brings us to our first result. In a market with homogenous products,
given any set of buyers B and sellers S, and the subset of buyers B and the
14
We restrict attention to bilateral trades between a buyer and a seller. We thus rule out side payments
between more than two parties.
Supply
v, c
B*
Demand
S*
n* m n i, j
Fig. 5. The set of efficient transactions that maximizes gains from trade.
subset of sellers S as defined previously, an assignment of buyers B to

sellers S is efficient if and only if it contains an assignment from B to S.
The result holds for the following reason. If an assignment from B to S
contains an assignment from B to S then those pairs of buyers and sellers
outside of B and S will not trade. In an efficient assignment, the buyer–
seller pairs that trade are those on the left-hand side of the supply-
and-demand diagram while the buyers and sellers that do not trade are
those on the right-hand side of the supply-and-demand diagram.
Can buyers and sellers attain an efficient outcome through direct ex-
change? To answer this question, it is necessary to make explicit what types
of transactions are involved in direct exchange. Suppose that there is full
information so that the identity and values of buyers and the identity and
costs of sellers are common knowledge. Suppose further that any trans-
action must be a bilateral exchange, which rules out coalitions of more than
two players. Suppose further that any pair of players resolves their nego-
tiation by the same bargaining procedure, such as the Nash Bargaining
Solution, so that a buyer and a seller split the gains from trade. Given this
framework, the highest-value buyer will trade with the lowest-cost seller.
There will be at least one trade if v1Zc1. This pairing continues until viocj
is reached at which point trade ceases. This approach yields an efficient
outcome since it matches the buyers in B with the sellers in S and all other
buyers and sellers are inactive.
This allows a characterization of the efficiency of any graph that connects
the set of buyers and sellers. Hall’s Marriage Theorem can be applied to this
task. Let G ¼ ((B, S), L) be any bipartite graph. Define N(i) as the neigh-
bors of iAB that are contained in the restricted set of sellers S. Then, the
160 D.F. Spulber
Marriage Theorem implies that G contains an assignment of B with S if

and only if|N(A)|Z|A|for all ADB.
I now introduce a useful benchmark to determine how many links are
needed to assure that there will be an efficient assignment. All pairs are said
to be viable with homogenous products if the lowest-value buyer has a value
greater than or equal to the cost of the highest-cost seller, vnZcm.
Complete Viability Condition. All pairs are viable and the number of buy-
ers equals the number of sellers.
The condition states that all buyer–seller matches are viable, vnZcm and
there is no need for rationing, n ¼ m. Later the condition will also be
applied to markets with differentiated products.
Recall that a graph G ¼ ((B, S), L) contains an efficient assignment if and
only if G contains an assignment of B to S. The condition guarantees that
the demand curve is everywhere above the supply curve. If the condition
holds, then B ¼ B and S ¼ S. The following result shows the connection
between the viability condition and efficiency.
Proposition 1. Let products be homogenous. Given the Complete Viability
Condition, any assignment of B to S is efficient. If the Complete Viability
Condition does not hold, then the conditions of Hall’s Marriage Theorem on
the graph G are not sufficient to guarantee that there exists an efficient
assignment of buyers to sellers.
Given the Complete Viability Condition, all assignments contained by the
graph G ¼ ((B, S), L) will be efficient. However, if this condition does not
hold, then assignments need not be efficient.
To illustrate this result, suppose first that the number of buyers and
sellers is equal but that demand and supply cross. For example, let there be
three buyers and three sellers, with v1>v2>v3 and c1oc2oc3. Assume that
v2>c2 and v3oc3. The bipartite graph in Fig. 6 does not contain an efficient
assignment. Buyer 2 in the set B ¼ {1, 2} does not have a neighbor in the
set S ¼ {1, 2}.
Next, consider a graph G ¼ ((B, S), L) in which the number of sellers is
greater than the number of buyers, m>n. Suppose that vnZcn. This case
Buyers Sellers
1 1
2 2
3 3
Fig. 6. Given v1>v2>v3 and c1oc2oc3, and assuming that v2>c2 and v3oc3, the graph
contains an assignment but does not contain an efficient assignment.
Buyers Sellers
1 1
2 2
3 3
4
Fig. 7. The graph contains assignments but does not contain an efficient assignment.
corresponds to the situation in which the market demand function is greater

than the market supply function. Then, B ¼ B and SCS so that not every
assignment in the graph G is an efficient assignment. Suppose for example
that there are three buyers and four sellers. Then, B ¼ {1, 2, 3} and
S ¼ {1, 2, 3}. The bipartite graph in Fig. 7 contains assignments but does
not contain an efficient assignment. The reason is that buyer 3 is not
matched with any seller in the set of efficient sellers S but only with in-
efficient seller 4.
Therefore, a network can fail to be efficient for two reasons. The network
may contain assignments, but unless all assignments are efficient, the net-
work may not contain the efficient ones. Secondly, even if all possible
assignments are efficient the network may not contain any assignments,
that is, the conditions of Hall’s Marriage Theorem may not be satisfied.
Without knowing whether or not the Complete Viability Condition is sat-
isfied, guaranteeing efficiency requires the network to have all possible
links.
Proposition 2. Let products be homogenous. Only a complete bipartite
graph, that is one containing all possible links between buyers and sellers, is
sufficient to guarantee a priori the existence of an efficient assignment.
The proposition imposes a very strong condition. It suggests that informa-
tion networks must offer a complete set of connections between individuals
to guarantee that the set of efficient transactions is attainable. Of course, it
may be costly to establish and operate an information network that offers a
complete set of connections, from anyone to anyone. There is a tradeoff
between the benefits and costs of a complete set of connections.
4.2 Assignments with differentiated products
It is even more difficult to assign buyers to sellers (or sellers to buyers)

when products are differentiated. With differentiated products, such as a
housing market or a more realistic marriage market, buyers differ from
each other and sellers differ from each other. The number of efficient
162 D.F. Spulber
assignments is likely to narrow, possibly to only one assignment if such an

assignment exists. This means that the need for complete networks is even
greater with differentiated products than it is with homogenous products.
Consider a market with a set of buyers B ¼ {1, y, n} and a set of sellers
S ¼ {1, y, m}. The sellers each offer an indivisible unit of a good or
service. The products offered by the sellers are differentiated goods. Also,
buyers have different preferences. Thus, the ith buyer derives value from the
jth good equal to aij, i ¼ 1, y, n, j ¼ 1, y, m. Sellers have costs cj, j ¼ 1,
y, m. The total of gains from trade that buyer i and seller j would obtain
from a transaction equals
vij ¼ maxf0; aij cj g. (3)
As before, vij represents the value of a match. All information is contained
in the value of the match because products are differentiated. The assign-
ment problem with differentiated products is discussed in Shapley and
Shubik (1972) and Shubik (1984, Chapter 8).
An efficient assignment is a set of matches that maximizes total surplus. A
buyer or a seller can be matched with at most one counterpart. Buyers and
sellers also can be viewed as partners in a productive enterprise. Buyers and
sellers also can represent firms and workers in a market where each firm is
matched with a worker, and where firms have different technologies and
workers have different skills.
Consider the general assignment problem for either homogenous prod-
ucts or differentiated products, where vij is defined either by (1) for ho-
mogenous products of by (3) for differentiated products. Let the graph G be
complete. An efficient assignment considers the following linear program-
ming problem,
!
XX
VðB; SÞ ¼ maxxij vij xij (4)
i2B j2S
X
subject to xij 1; 8 j2S
i2B
X
xij 1; 8 i 2 B and
j2S
xij 0; 8 i 2 B; j 2 S.
This problem is presented in Shapley and Shubik (1972) and Shubik (1984,
Chapter 8) (see also Roth and Sotomayor, 1990). There exists a solution to
the maximization problem xij that involves only P values of xij equal to either

one or zero,
P see Dantzig (1963, p. 318). Thus, ij vij xij is greater than or
equal to ij vij xij for all xij satisfying the constraints in problem (4). The
P
solution allows for someP buyers to be unassigned, j xij ¼ 0; or for some
sellers to be unassigned, i xij ¼ 0:
Only a mixed group of buyers and sellers can create value, V(A) ¼ 0 if
ADB or ADS. The value of any mixed coalition is an assignment of buyers
and sellers that maximizes the benefits of the coalition,
XX
VðAÞ ¼ vij xij ; A B [ S, (5)
i j
where xij solves the linear programming maximization problem (4) for i
and j restricted to the set ACB[S. Clearly, the value of the grand coalition
consisting of all buyers and sellers equals the value of the set of efficient
matches that solve the linear programming problem (4). Note that
V ðfi; jgÞ ¼ vij for i in B and j in S.
Buyers and sellers are complements in determining the value of a coa-
lition,
VðA þ i þ jÞ VðA þ iÞ V ðA þ jÞ V ðAÞ;
where iAB and jAS, ACB[S, and ieA, jeA (see Shapley, 1962). The value
of a coalition is non-decreasing when adding a buyer or a seller,
V(A+i)ZV(A) and V(A+j)ZV(A). The incremental value of a buyer or
a seller is non-decreasing in the addition of a player of the opposite type.
The additional buyer and seller pair adds at least vijZ0 and possibly more
because the entire set of matches for the coalition can be reshuffled and
improved.
To illustrate the assignment problem with differentiated products, con-
sider the classic example of a housing market due to Shapley and Shubik
(1972). There are three sellers each with a house on the market with sellers’
costs are equal to c1 ¼ 18, c2 ¼ 15, and c3 ¼ 19. There are three buyers
whose valuations aij are shown in Table 3, where rows refer to buyers
and columns refer to sellers. Buyer values net of seller costs are shown in
Table 4. The efficient assignment is shown in bold and is given by (1, 3),
(2, 1), and (3, 2), which yields a total value of 16.
To illustrate the effects of incomplete networks, I establish a benchmark
for markets with differentiated products. Suppose that each buyer i has a
productivity parameter zi and each seller j has a productivity parameter yj.
Suppose that the productivity parameters are in decreasing order for buyers
Table 3
Buyer valuations in the Shapley and Shubik housing market
Seller 1 Seller 2 Seller 3
Buyer 1 23 22 21
Buyer 2 26 24 22
Buyer 3 20 21 17
164 D.F. Spulber
Table 4
Buyer valuations net of seller costs in the Shapley and Shubik housing market
Buyer 1 5 7 2
Buyer 2 8 9 3
Buyer 3 2 6 0
and for sellers, z1>z2>?>zn and y1>y2>?>yn. The benefits from a

match are given by an additive value function,
aij ¼ aðzi ; yj Þ ¼ zi þ yj .
Given this additive value function, all pairs are viable with differentiated
products if any pair that contains the lowest-value buyer covers the seller’s
cost, zn+yjZcj for all j ¼ 1, y, n.
Proposition 3. Let products be differentiated and suppose that the value
function is additive. If the Complete Viability Condition holds, then any
assignment of B to S is efficient. If the Complete Viability Condition does
not hold, then the conditions of Hall’s Marriage Theorem on the graph G
are not sufficient to guarantee that there exists an efficient assignment of
buyers to sellers.
Any assignment E is efficient because additivity implies that
X X X
n Xn
vij ¼ ðzi þ yj cj Þ ¼ zi þ ðyj cj Þ (6)
ði;jÞ2E ði;jÞ2E i¼1 j¼1
The right-hand side equals the maximum net benefits that can be obtained
from matching buyers in B and sellers in S. Given the Complete Viability
Condition, all assignments contained by the graph G ¼ ((B, S), L) will be
efficient. The statement that without the Complete Viability Condition,
Hall’s Marriage Theorem is not sufficient to guarantee that there exists an
efficient assignment of buyers to sellers follows from similar arguments
made for homogenous products.
Without knowing whether or not the Complete Viability Condition is
satisfied, guaranteeing efficiency requires the network to have all possible
links.
Proposition 4. Let products be differentiated. Only a complete bipartite
graph, that is one containing all possible links between buyers and sellers, is
sufficient to guarantee a priori the existence of an efficient assignment.
The marriage market is another example of the assignment problem with
differentiated products (see Shapley and Shubik, 1972). A useful special case
of a marriage market is given by Gary Becker (1973) in which each buyer i
has a productivity parameter zi and each seller j has a productivity param-

eter yj. Again, suppose that the productivity parameters are in decreasing
order for buyers and for sellers, z1>z2>?>zn and y1>y2>?>yn. Also,
suppose that all sellers have zero costs. The value of a match is given by a
multiplicative value function,
aij ¼ aðzi ; yi Þ ¼ zi yi .
The value function exhibits complementarity in productivity parameters z
and y, azy(z, y)>0. The marginal product of a player’s type increases in the
partner’s type.
As Becker (1973) observes, an efficient assignment requires matched
partners to be identical. Therefore, in Becker’s marriage market, the effi-
cient assignment exists and is unique, and is given by (1, 1), (2, 2), y, (n, n).
The aggregate value is the sum of the values generated by matching iden-
tical types. A graph G does not contain an efficient assignment unless it
contains all of the diagonal pairs (1, 1), (2, 2), y, (n, n). Recall that this is
only one assignment out of n! possible assignments. When assignments are
constrained, it is necessary to reconsider all of the matches in the assign-
ment. We now examine how to choose a constrained efficient assignment.
4.3 Second-best assignments
To represent costly communication, consider how the choice of efficient

assignments is constrained by the possible absence of links between
some buyer–seller pairs. Define a second-best assignment as an assignment
that maximizes benefits subject to the constraint that the network is
not complete. Given a bipartite graph G ¼ ((B, S), L), let dij ¼ 1 if the link
(i, j) is in G and dij ¼ 0 otherwise. Then, define the weighted value of a
match hij by
hij ¼ maxf0; dij vij g. (7)
This applies whether products are homogenous and vij ¼ max{0, vicj}, and
it applies when products are differentiated and vij ¼ max{0, aij–cj}. Thus,
hij>0 if buyer i and seller j have positive gains from trade and if there is a
link between buyer i and seller j.
There exists a solution to the maximization problem xij ¼ xP
ij (G) that
involves only values of xijPequal to either one or zero. Thus, ij hij xij is
greater than or equal to ij hij xij for all xij satisfying the constraints in
problem
P (4). Again, the solution allows for some buyers
P to be unassigned,
j xij ¼ 0; or for some sellers to be unassigned, i xij ¼ 0: The solution
xij ¼ xij(G) is efficient if and only if G contains an efficient assignment.
The value function
XX
V L ðB; SÞ ¼ xij ðGÞ (8)
i2B j2S
166 D.F. Spulber
gives total benefits given that L is the set of links. Since the assignment is
second best, the following result holds. The total benefit is non-decreasing
in the set of links, VL(B, S) rVK(B, S) for LDK.
5 Networks and the Core in a two-sided market
The Core in a two-sided market is the ideal case, because the solution
concept implicitly assumes that there are no costs of communication or
computation. Players in a cooperative game can engage in the communi-
cation and computation that are necessary to examine the value of all
possible coalitions and to implement an allocation of value in the Core.
Costless communication in a cooperative game also means that all coali-
tions of buyers and sellers are feasible. Myerson (1977), in contrast, exam-
ines the value in cooperative games on networks, where some coalitions are
ruled out by the absence of links.
5.1 The Core with homogenous products
The Core of the market game with homogenous products corresponds to

a market allocation with a uniform price (see Shubik, 1984, p. 228) and
Shapley and Shubik (1972). Even if the efficient assignment were unique,
there are at least min{n, m} degrees of freedom in the Core, although the
dimensionality of the Core cannot exceed min{n, m} since the buyers’ pay-
offs determine those of the sellers (see Shubik, 1984, p. 199).
The Core allocation yields a surplus vi–p for active buyers, a surplus of p–
ci for active sellers, and zero for inactive buyers and sellers. The market
price lies between the value of the marginal buyer and the cost of the
marginal seller. The market price is constrained above by the cost of the
highest-cost seller who does not trade if cn+1ovn. The market price is
constrained below by the value of the highest-value buyer who does not
trade if vn+1>cn.
Thus, the market price is in the interval
maxfcn ; vnþ1 g p minfvn ; cnþ1 g.
This range of prices determines the set of core allocations. Each allocation
in the Core corresponds to an allocation determined by one of the prices in
this range.15 Because the Core corresponds to an allocation based on a
common price, it effectively requires a central market mechanism. The
equivalence of market outcomes and outcomes in the Core illustrates the
relationship between allocation mechanisms in cooperative games and
market allocation mechanisms.
15
For a proof of core equivalence in the assignment game, see Shapley and Shubik (1972).
5.2 The Core with differentiated products
Consider next the cooperative assignment game with differentiated prod-

ucts. The discussion in this section is from Shapley and Shubik (1972) and
Shubik (1982). Define the characteristic function of the cooperative game,
v(A) ¼ V(A), where the value function is given by the linear programming
problem (4). The Core of the assignment game with differentiated products
consists of a set of payoffs for buyers uiZ0, i ¼ 1, y, n and a set of payoffs
for sellers rjZ0, j ¼ 1, y, m such that the following two conditions hold:
X X
ui þ rj ¼ vðB [ SÞ. (9)
i2B j2S
X X
ui þ rj vðAÞ. (10)
i2A\B j2A\S
Imputations in the Core correspond exactly to the (nonempty) set of

solutions to the dual of the linear programming problem for the assignment
games (see Shapley and Shubik, 1972). To find the dual of the linear pro-
gramming problem (4), find the pair of payoff vectors (u, r) that solve the
following problem:
X X
minu;r ui þ rj (11)
i2B j2S
subject to ui 0; rj 0; and
ui þ rj vij ; for all i 2 B and j 2 S.

The dual problem has a solution because the primal problem has a so-
lution and the objective functions of the two problems reach the same value,
Dantzig (1963, p. 129). Thus, we obtain a solution (u, r) that satisfies the
following condition:
X X XX
ui þ rj ¼ vij xij ¼ vðB [ SÞ. (12)
i2B i2S i j
Note that from the constraint of the dual problem,

ui þ rj vij ,
so that no buyer–seller coalition can improve on the imputation, which
establishes that it is a core allocation.
Any core allocation (u, r) is associated with a vector of prices. Consider
a buyer–seller pair for a given core allocation. Recall the definition
vij ¼ max{0, aij–cj}. If aijZcj, then buyer i and seller j transact at price pij
where the price is determined as follows,
pij ¼ cj þ rj ¼ aij ui ; i ¼ 1; . . . ; n j ¼ 1; . . . ; m. (13)
168 D.F. Spulber
Otherwise, the buyer–seller pair receives zero payoffs and no transaction

occurs between them. Product differentiation implies that prices may differ
for each transaction. This contrasts with the matching of buyers and sellers
when products are homogenous, which can be achieved with a uniform
market-clearing price. Notice that all payments are made within buyer–seller
pairs without the need to make side payments to other buyers and sellers. Just
as in the assignment problem, the simplicity of the dual linear programming
problem hides the complexity of determining the set of core allocations.
The complexity of the assignment problem strongly suggests that an effi-
cient decentralized solution is not available for buyers and sellers. First, the
assignment problem presents many challenges that are likely to entail sub-
stantial transaction costs in practice. Solving the assignment problem re-
quires information about buyer values and seller costs. Then, it is necessary
to assign buyers to sellers by choosing from the set of n! possible assign-
ments or equivalently solving the linear programming problem. Second,
buyers and sellers must determine transfer payments that reflect the value of
coalitions and alternative assignments.
6 Stable assignments in a decentralized two-sided market: costly

computation
Consider a decentralized two-sided market. To represent the situation in

which buyers and sellers face costly computation, consider a game in which
buyers and sellers are restricted to bilateral exchange with no money trans-
fers. The equilibrium solution for such a game is a stable assignment defined
by Gale and Shapley (1962). The Gale–Shapley assignment game corre-
sponds to a decentralized market since the stable assignment does not re-
quire any central coordination, it can be reached with only bilateral offers
and acceptances or rejections.
Gale and Shapley (1962) consider a two-sided market in which equal
numbers of buyers and of sellers have ordinal preferences over matches. An
assignment of buyers to sellers is stable if it does not contain two buyers i1
and i2 who are assigned to sellers j1 and j2, respectively, such that buyer i1
prefers seller j2 to seller j1 and buyer i2 prefers seller j1 to seller j2.
Gale and Shapley (1962) show that a stable assignment always exists and
that it can be found with a simple algorithm. Sellers can break off a match if
they get a better offer. Any buyer that does not have a match, proposes a
match to a seller that ranks highest among all sellers that have not yet
rejected him. If the seller is not matched, the seller accepts the offer. If the
seller is matched, the seller can either accept or reject the buyer’s offer
depending on whether or not it is preferred to the seller’s existing match.
The buyers can be chosen in any order. The algorithm continues until all
buyers are matched with a seller.
The Gale–Shapley assignment game contrasts with that of Shapley and

Shubik (1972), who allow for money transfers and who focus on core al-
locations. As already noted the Core allocation corresponds to a game in
which there is costless communication and computation. Recall that in the
Core, there need only be money transfers within buyer–seller pairs.
In this section, I obtain some interesting insights by combining elements
of these two fundamental studies. I allow for the possibility of money
transfers. I then consider stable assignments in the two-sided market game.
The stable assignment with homogenous products is shown to be an effi-
cient assignment. The stable assignment for a market with differentiated
products is analogous to Böhm-Bawerk’s method of marginal pairs for a
market with homogenous goods. The outcome is not necessarily in the
Core, although I give sufficient conditions under which a stable assignment
is in the Core.
6.1 Stable assignments with homogenous products
In contrast to Gale and Shapley (1962), consider the assignment game

with the possibility of money transfers within any buyer–seller pair. Once
buyers and sellers make a match they lose the outside option of other
matches. Suppose that buyers and sellers engage in generalized Nash bar-
gaining. Let a represent the buyer’s share of the surplus, so that 1–a is the
seller’s share of the surplus, where 0oao1.
The generalized Nash bargaining solution specifies a consistent propor-
tional division of surplus across buyer–seller pairs. A match yields mon-
etary payoffs for the two parties i and j equal to avij for buyer i and (1–a)vij
for seller j. Thus, buyers and sellers have cardinal as well as ordinal pref-
erences rankings across matches. The values of matches in the assignment
game with money transfers treat buyers and sellers symmetrically as
Shapley and Shubik (1972) observed.
Proposition 5. Given money transfers within each buyer–seller pair and
generalized Nash bargaining, there is a unique stable assignment.
We know from Gale and Shapley (1962) that with ordinal preferences a
stable assignment always exists. With money transfers and generalized
Nash bargaining, there can be only one stable assignment.16
Consider the market with homogenous products. As before, let buyer
valuations be arranged in decreasing order, v1>v2>y>vn and let seller
16
A B-stable assignment is defined as an assignment that is optimal for the buyer side of the market,
that is, all members of the buyer side of the market are as well off as they would be in any other stable
assignment. With ordinal preferences, if there is a B-stable assignment, then the B-stable assignment is
unique. An S-stable assignment is defined as an assignment that is optimal for the seller side of the
market, that is, all members of the seller side of the market are as well off as they would be in any other
stable assignment. With ordinal preferences, if there is a S-stable assignment, then the S-stable assign-
ment is unique (see Bollobás, 1998).
170 D.F. Spulber
costs be arranged in increasing order, c1oc2oyocm. The best buyer and

the best seller wish to trade with each other to obtain the greatest gains
from trade. The highest-value buyer will match with the lowest-cost seller,
the next-highest-value buyer will match with the next-lowest cost seller, and
so on, until there are no more subscriber pairs with non-negative surplus.
Therefore, a stable matching takes the form of Böhm-Bawerk’s method of
marginal pairs. This implies the following result.
Proposition 6. In the market with homogenous products, the unique stable

assignment of buyers and sellers is an efficient assignment.
Total net benefits of buyers and sellers therefore are maximized by decen-
tralized exchange between buyers and sellers under full information.
Even though the stable assignment is efficient, the bilateral allocation of
surplus between buyers and sellers does not correspond to a core allocation.
In a market mechanism as described by the Walrasian auctioneer or by
Böhm-Bawerk’s method of marginal pairs in which the marginal trade sets
the price, the law of one price holds.17 However, when buyers and sellers
match themselves, every buyer–seller pair (i, j) that completes a trade di-
vides the surplus proportionally and thus trades at a potentially distinct
price. Since buyers and sellers split the surplus according to the generalized
Nash Bargaining Solution, the exchange price equals
pij ¼ ð1 aÞvi þ acj . (14)
With decentralized exchange, the law of one price does not hold and the
allocation of surplus does not correspond to a Walrasian allocation. Al-
though the stable assignment with decentralized matching corresponds ex-
actly to the set of pairs in the Core, the allocation with decentralized
matching generally is not a Core allocation because prices will vary across
matched pairs.
The efficiency of the stable assignment with homogenous products de-
pends on costless communication. Suppose instead that communication is
costly, so that the bipartite graph linking buyers and sellers is incomplete.
Then, there exist stable assignments on the incomplete graph. Such stable
17
The theory of multiperson cooperative games sometimes portrays equilibrium solutions as decen-
tralized mechanisms. For example, Shubik (1982, p. 127) considers the solutions to games as ‘‘the end
results of rational, motivated activity by the players.’’ However, solutions to multiperson cooperative
games, such as the Core, also resemble centralized allocation mechanisms. Much coordination goes into
finding core allocations. Determining the solution to a game, even under full information, requires an
evaluation of the value of all possible coalitions and a calculation of the set of feasible payoff vectors
that cannot be improved on by any coalition. The imputations then must be provided to individual
players. The complexity of these procedures represents centralized mechanisms rather than multilateral
negotiations. The equivalence of core allocations and market outcomes suggests that there exists a
fundamental connection between mechanisms needed to discover and implement a market equilibrium
and a solution to a cooperative game. It should be emphasized that both the neoclassical market
equilibria and the solutions to cooperative games are attained without transaction costs.
Buyers Sellers
v1 c1
1 1
v1 c2
v2 c1
2 2
Fig. 8. The stable assignment on the incomplete network is (1, 1).
assignments need not be efficient even subject to the constraint imposed by

the available links.18
Suppose for example that there are two buyers and two sellers. Suppose
that buyer 1 is linked to both sellers and suppose that buyer 2 is linked only
to seller 1 (see Fig. 8). Then, since v1>v2 and c1oc2, the only stable as-
signment is for buyer 1 to trade with seller 1. Given generalized Nash
bargaining, buyer 1 prefers seller 1 to seller 2 and seller 1 prefers buyer 1 to
buyer 2. To follow the Gale–Shapley algorithm, buyer 1 makes an offer to
seller 1 that is accepted. Buyer 2 makes an offer to seller 1 that is declined,
and the process is complete. The assignment yields value equal to v1–c1. The
three other feasible assignments yield v1–c2, v2–c1, and (v1–c2)+(v2–c1). The
stable assignment (1, 1) dominates the assignment (1, 2) and the assignment
(2, 1). It also dominates the assignment composed of both matches (1, 2)
and (2, 1) if v2oc2. However, if v2>c2, the stable assignment is not efficient
since v1–c1o(v1–c2)+(v2–c1).
This situation corresponds to the Complete Viability Condition. Thus,
when the Complete Viability Condition holds, a stable assignment on an
incomplete network need not be efficient on that network. It is important to
observe that the simple bipartite graph in Fig. 8 satisfies Hall’s Marriage
Theorem and therefore contains an assignment.
Proposition 7. Consider a market with homogenous products and suppose

that the Complete Viability Condition holds. Let the graph G satisfy the
conditions of Hall’s Marriage Theorem. Then, the stable assignment on the
graph G need not be efficient on that network.
Any assignment on the graph is efficient by the Complete Viability Con-

dition, which guarantees that all assignments are efficient. The incomplete-
ness of the graph means that the stable assignment need not be an
assignment since it does not necessarily map all buyers into the set of all
18
This differs from Kranton and Minehart’s (2001) result in their auction framework since they have
identical sellers.
172 D.F. Spulber
sellers, as in the previous example. Accordingly, the stable assignment need

not be efficient, which establishes the result in Proposition 7.
6.2 Stable assignments with differentiated products
Consider now the market in which the products offered by the sellers are
differentiated goods and buyers have different preferences. Again, suppose
that buyers and sellers find their own matches and engage in bilateral ex-
change. Also, once they are matched buyers and sellers lose the outside
option of other matches. As before, the ith buyer derives value from the jth
good equal to aij, i ¼ 1, y, n, j ¼ 1, y, m. Sellers have costs cj, j ¼ 1, y,
m. The value of a match is vij ¼ max{0, aij –cj}. Assume that buyers and
sellers engage in generalized Nash bargaining so that the buyer obtains
surplus avij and the seller obtains surplus (1–a)vij.
The main insight of the discussion is that a stable assignment with
differentiated products can be obtained in a manner that is analogous to
Böhm-Bawerk’s method of marginal pairs for the homogenous products
case. This corresponds to a sequential process with the buyer–seller pair
that has the highest value being matched first, the buyer–seller pair with the
next-highest value matched next and so on. However, in contrast to the case
of homogenous products, matching by buyers and sellers need not yield an
efficient assignment even on a complete network.
Proposition 8. In a market with differentiated products, a stable assignment
need not be efficient.
To see why, consider again the housing market example due to Shapley
and Shubik (1972). Recall that the efficient assignment equals (1, 3), (2, 1),
and (3, 2), which yields a total value of 16. This is shown in bold in Table 5.
The stable assignment, which results from decentralized matching by buyers
and sellers, consists of the two matches (2, 2) and (1, 1), with buyer 3 and
seller 3 inactive. This outcome is indicated with asterisks in Table 5. Thus,
the stable assignment yields a total value of 14 and only two transactions.
The stable assignment is not efficient.
Becker’s (1973) marriage market provides conditions under which a sta-
ble assignment is efficient with differentiated products. The production
Table 5
The efficient assignment (in bold) and the stable assignment (with asterisks) in the housing
market
Buyer 1 5* 7 2
Buyer 2 8 9* 3
Buyer 3 2 6 0*
function is a(z, y) ¼ zy. Let the zi and yj be arranged in decreasing order,

z1>z2>y>zn and y1>y2>y>yn, and let zi ¼ yj for i ¼ j. Let seller costs
equal zero. There is only one efficient assignment, which is to match each
buyer with a seller of the same type. With decentralized exchange, buyers
and sellers will self-select with the highest-value buyer matching with the
highest-value seller and so on.
Proposition 9. Suppose that buyers and sellers make their own bilateral
matches in a market with differentiated products. Then, in Becker’s marriage
market, the unique stable assignment of buyers and sellers is also the unique
efficient assignment.
Consider the market for differentiated products in which the technology
is additive a(z, y) ¼ z+y. Then, given the Complete Viability Condition, all
assignments are efficient. Since a unique stable assignment exists, it is effi-
cient as well.
Proposition 10. Suppose that buyers and sellers make their own matches in
a market with differentiated products. Then, given an additive production
function and the Complete Viability Condition, the unique stable assign-
ment is efficient.
The stable assignment need not be efficient if the network is not complete,
as was seen with homogenous products. The same problem arises with
differentiated products. Given the conditions of Hall’s Marriage Theorem,
the graph contains an assignment. If the production function is additive and
the Complete Viability Condition applies, that assignment must be efficient.
The stable assignment may not be a complete assignment as was shown in
Fig. 8. Seller 1 prefers to deal with buyer 1 instead of buyer 2 since z1>z2
implies that z1+y1–c1>z2+y1–c1. Buyer 1 prefers to deal with seller 1
instead of seller 2 if z1+y1–c1>z1+y2–c2, in which case the stable assign-
ment is (1, 1). However, the assignment (1, 2), (2, 1), which is feasible, yields
greater value than (1, 1) since it is an efficient assignment on the complete
network by the Complete Viability Condition. This implies the following
result.
Proposition 11. Consider a market with differentiated products. Suppose
that the production function is additive and the Complete Viability Con-
dition holds. The graph G is not complete but satisfies the conditions of
Hall’s Marriage Theorem. Then the stable assignment on the graph G need
not be efficient on the network.
7 Firms and stable assignments in a centralized two-sided market
Consider a market in which a monopoly firm provides centralized com-

munication using information systems. Buyers and sellers who are the firm’s
174 D.F. Spulber
subscribers find each other and form their own bilateral matches, so that the
matching process itself is decentralized. The firm incurs costs in creating
and operating the centralized market. Once they become members of the
network, buyers and sellers have zero costs of communication. The firm
provides communication in the form of a centralized hub-and-spoke com-
munications network that allows every buyer to be linked to every seller.
This provides the links described by a complete bipartite graph that con-
nects every buyer with every seller. Once buyers and sellers are members of
the network their costs of communication are zero. In addition, the firm
provides every buyer and every seller with full information about all other
buyers and sellers. The cost savings of establishing a centralized hub-and-
spoke network reinforces potential efficiency advantages of full informa-
tion.
The firm charges buyers a subscription fee P to participate in the network
and charges sellers a subscription fee W to participate in the network. There
are two alternative scenarios. In the first scenario, the firm charges the fees
only to buyers and sellers after they complete a match based on transac-
tions. In the second scenario, the firm charges the fees to buyers and sellers
before they complete a match based on subscription to the network. These
two scenarios affect the nature of the equilibrium.19
7.1 Firms and stable assignments with homogenous products
The firm reduces but does not eliminate coordination problems for buyers
and sellers. Because buyers and sellers have complete information about
each other after they become subscribers, they coordinate easily after join-
ing the firm’s network. However, they still face a coordination problem
before becoming subscribers because each side must anticipate which buy-
ers and sellers will choose to participate in the network. Buyers and sellers
play a coordination game in participation decisions.
To illustrate the coordination problem, consider a simple game with a
single buyer and a single seller. The buyer has willingness to pay v and the
seller has cost c. Assume that they have potential gains from trade, v>c.
Suppose that the firm charges a fee after a match is made. Then, the buyer
and the seller allocate the surplus from exchange net of the fees charged by
the firm. They will become members of the network if the total gains from
trade are non-negative, (v–c–P–W)Z0.
Suppose now that the firm charges a fee before a match is made, so that
the subscriber fees are sunk costs when the transaction occurs. Let buyers
and sellers engage in Nash bargaining (a ¼ 1/2) and thus evenly divide the
quasirents from exchange so that exchange between the buyer and the seller
19
Caillaud and Jullien (2003) refer to ex ante fees as connection or registration fees and they refer to ex
post fees as transaction fees. Their framework differs from the present one since in their model, all
buyers are identical and all sellers are identical but there is only one acceptable match for each one.
Table 6
A participation game where payoffs are (buyer, seller)
Seller Subscribe Do not subscribe
Buyer
Subscribe (v–c)/2–P, (v–c)/2–W P, 0

Do not subscribe 0, W 0, 0
gives (v – c)/2. Then, the buyer and seller face a coordination game resem-
bling the battle of the sexes (see Table 6). Then, if the buyer anticipates
trading with the seller, the buyer will become a member if (v–c)/2ZP. Also,
if the seller anticipates trading with the buyer, the seller will become a
member if (v–c)/2ZW. If both these conditions are satisfied, then both
subscribing is a Nash equilibrium, although neither subscribing also is a
Nash equilibrium.
Consider the market with consumers divided into two distinct groups, a
set of buyers B ¼ {1, y, n} and a set of sellers S ¼ {1, y, m}. As before,
buyers have valuations that are arranged in decreasing order,
v1>v2>y>vn, and sellers have costs that are arranged in increasing or-
der, c1oc2oyocm. Only buyers and sellers subscribing to the firm’s serv-
ice can communicate with each other and form matches. Communication is
costless since buyers and sellers are linked by a complete bipartite graph
and buyers and sellers have full information about each others’ types.
Buyers and sellers form bilateral matches and split the gains from trade in
some way.
Consider the first scenario in which buyers and sellers make their par-
ticipation decision and make a match before paying the fee. The modified
value of a match between buyer i and seller j is
hij ¼ maxf0; vi cj P W g. (15)
Let n be the largest integer i such that vi P cj þ W ; where i ¼ j, iAB
and jAS. The number n is the market-clearing quantity with fees P and
W. Let B ¼ {1, y, n} and S ¼ {1, y, n}. The sets B and S
describe the buyers and the sellers that are on the left-hand side of the
supply-and-demand diagram such that if they are paired their gains from
trade exceeds the sum of the two subscription fees.
In defining the two sets B and S, buyers and sellers are paired in
order, just as in Böhm-Bawerk’s (1891) method of marginal pairs. It is
possible that an excluded buyer i>n, if paired with a seller in S, could
have gains from trade that exceed the sum of subscriber fees. The same
applies to excluded sellers j>n. However, the two sets reflect the best
matches. Because buyers and sellers have full information on the firm’s
network, buyers and sellers will choose to participate only in the best
176 D.F. Spulber
possible match. Thus, the highest-value buyer trades with the lowest-cost
seller, and so on with the remaining buyers and sellers. As in Böhm-Bawerk
(1891), buyers and sellers form pairs until the marginal pair is reached.
However, in contrast to Böhm-Bawerk, the terms of trade are determined
by bilateral bargaining rather than with two-sided competition managed by
a central auctioneer.
Let R ¼ P+W be the sum of the subscription fees.
Proposition 12. Let subscription fees be paid after a match is made. Then,
in a market with homogenous products, only buyers in B and sellers in S
will choose to become subscribers on the firm’s network. Only the sum of
subscription fees matters for participation, not the amounts assigned to
buyers and to sellers.
Those buyers and sellers that choose to join the network as subscribers
correctly expect to find a match. Other buyers and sellers do not participate
since they would not find a match.
Only the sum of the subscriber fees matters since buyers and sellers an-
ticipate the assignment process on the network, see Fig. 9. The allocation of
the fees between buyers and sellers does not affect the participation decision
since the total of the two fees is subtracted from the value of the bilateral
transaction. A buyer and seller engaged in bilateral exchange bargain over
the allocation of rents from exchange, which equals the value to the buyer
minus the cost to the seller and minus total fees. The allocation of fees
between buyers and sellers also does not affect the firm’s profit. Suppose
that the firm has a cost T per subscriber pair. Since n is the volume of
Supply
v, c
B**
R{
Demand
S**
n** m n i, j
Fig. 9. Subscribers to a network when the sum of the buyers’ and sellers’ fee equals R.
transactions, the firm’s profit equals

P ¼ n ðP þ W TÞ ¼ n ðR TÞ. (16)
Consider the second scenario in which buyers and sellers must pay a fee
before making a match. The fees become sunk costs that do not affect the
quasirents from exchange. After fees are paid and buyers and sellers are
members of the firm’s network, the value of a match between buyer i and
seller j is the same as in decentralized exchange hij ¼ max{0,vi–cj}. Because
buyers and sellers have full information on the firm’s network, buyers and
sellers will choose to participate only in the best possible match. Thus, the
highest-value buyer trades with the lowest-cost seller, and so on with the
remaining buyers and sellers, regardless of the subscription fees.
The allocation of the subscription fees does affect the participation de-
cision of buyers and of sellers in the second scenario since buyers and sellers
make decisions about participation before paying the fees. The outcome of
bilateral exchange affects the participation decision as well.
Suppose for example, that buyers and sellers engage in Nash bargaining
(a ¼ 1/2) and thus evenly divide the quasirents from exchange. Even di-
vision of quasirents implies that exchange between buyer i and seller j gives
each party (vi–cj)/2. Then, if buyer i anticipates trading with seller j, buyer i
will become a member if and only if (vi–cj)/2ZP. Also, if seller j anticipates
trading with buyer i, seller j will become a member if and only if (vi–cj)/
2ZW. If P is greater than W, then fewer buyers than sellers will find it
worthwhile to subscribe based on the division of surplus. Since buyers and
sellers are paired in order, the number of sellers that choose to subscribe in
equilibrium equals the number of buyers that choose to participate. There-
fore, participation is determined by P when P is greater than W. The con-
verse is true when W is greater than P.
Therefore, when the buyer and seller evenly split the quasirents from
exchange, the higher of the two fees effectively determines participation of
both buyers and sellers in equilibrium.
Proposition 13. Let subscription fees be paid before a match is made. Then,
in a market with homogenous products, if buyers and sellers engage in Nash
bargaining, the greater of the buyer and seller fees determines participation
in the network.
Since the higher of the two fees determines participation with Nash bar-
gaining between buyers and sellers, the firm will choose to set equal sub-
scriber fees for buyers and for sellers, P ¼ W ¼ R/2. Since the higher of the
two fees determines participation, the profit-maximizing firm will raise the
lower of the two fees to equal the higher one since that increases revenue
without affecting participation.
From Propositions 12 and 13, it follows that the same level of partic-
ipation will be observed under the two scenarios for the same total of the
fees R. This is because the allocation of subscriber fees is neutral in the first
178 D.F. Spulber
scenario and because the firm chooses equal subscriber fees P ¼ W ¼ R/2 in
the second scenario.
I now derive the monopoly firm’s choice of buyer and seller fees. For ease
of presentation, consider the limiting case in which there is a continuum of
buyers and a continuum of sellers both of which are uniformly distributed
on the unit interval with unit density.
Recall in the first scenario that if fees are paid after a match is made, only
the total of the fees matters. In the second scenario, with fees paid before a
match is made, the firm will set equal buyer and seller fees when buyers and
sellers engage in Nash bargaining. Thus, in either case, a monopoly firm
chooses only the total of buyer and seller fees, R.
For any R less than one, the marginal buyer–seller pair is such that the
market clears,
1 v ¼ c. (17)
The marginal buyer–seller pair obtains gains from trade equal to the total
buyer and seller fee R,
v c ¼ R. (18)
Thus, solving (17) and (18), the marginal pair is given by
v ¼ ð1 þ RÞ=2; c ¼ ð1 RÞ=2.
The volume of transactions is a function of the total of the buyer and seller
fee, Q(R) ¼ (1R)/2.
A firm with transaction cost T per subscriber pair has profit equal to
PðRÞ ¼ ðR TÞð1 RÞ=2. (19)
Maximizing profit in (19) gives the firm’s profit-maximizing fee, which
equals R ¼ (1+T)/2. The firm’s profit is P ¼ (1T)2/8, and the market-
clearing volume of transactions is Q ¼ (1T)/4.
For any given fee R, consumers’ surplus equals buyers’ benefits minus
sellers’ costs and minus the payments to the firm,
Z QðRÞ Z QðRÞ
D
CSðRÞ ¼ P ðQÞdQ PS ðQÞdQ RQðRÞ
0 0
¼ ð1=4Þð1 RÞ2 . ð20Þ
The firm’s fee R is a transaction cost that has two effects on gains from
trade for buyers and sellers. There is a direct cost RQ(R) and there is a
deadweight loss due to those efficient transactions that cannot be com-
pleted. The deadweight loss to buyers and sellers is the transactions fore-
gone in comparison with a situation in which the transaction cost R equals
zero. These two effects are shown in Fig. 10.
Total surplus including the firm’s profit equals
VðRÞ ¼ CSðRÞ þ PðRÞ ¼ ð1=4Þð1 R2 2T þ 2TRÞ. (21)
S(p)
R I II
D(p)
Q(R) Q
Fig. 10. Transaction costs to buyers and sellers due to a positive subscriber fee R (in
comparison to the case of R ¼ 0) result in a resource cost (I) and a deadweight loss (II).
Evaluated at the profit-maximizing fee R ¼ ð1 þ TÞ=2; total surplus

equals
V ¼ ð3=16Þð1 TÞ2 . (22)
7.2 Firms and stable assignments with differentiated products
Consider again a market with differentiated products as represented by

Becker’s (1973) marriage market in which each buyer i has a productivity
parameter zi and each seller j has a productivity parameter yj. Again, sup-
pose that the productivity parameters are in decreasing order for buyers
and for sellers, z1>z2>y>zn and y1>y2>y>yn, and let zi ¼ yj for i ¼ j.
Also, suppose that all sellers have zero costs. The value of a match is given
by a multiplicative value function, aij ¼ a(zi, yj) ¼ ziyj.
Consider the first scenario in which the firm charges buyers and sellers
subscription fees after a match is made. The value of a match between buyer
i and seller j with differentiated products is
hij ¼ maxf0; aij P W g. (23)
The stable assignment in the marriage market consists of buyer–seller pairs
with identical types. Define the sets B and S as consisting of those
buyers and sellers for whom a match yields surplus that covers the total of
180 D.F. Spulber
the subscribers’ fees. Let n be the largest n such that

aðzi ; yj Þ ¼ R for zi ¼ yj . (24)
The sets B and S consist of buyers and sellers 1, 2, y , n.
Only those buyers and sellers that expect to be matched will subscribe to
the network. Buyers and sellers will only choose to participate in the net-
work if they can complete matches that generate sufficient gains from trade
to cover the fee R.
Proposition 14. Let subscribers’ fees be paid after a match is made. In
Becker’s marriage market, only buyers in B and sellers in S will choose
to become subscribers on the firm’s network. The sum of subscription fees
matters for participation decisions, not the amounts assigned to buyers and
to sellers.
As with homogenous products, when subscription fees are charged after a
match is made, the allocation of subscriber fees between buyers and sellers
in the marriage market does not affect participation decisions.
Consider now the second scenario in which subscriber fees are charged
before buyers match with sellers. The outcome of bilateral exchange affects
the participation decision. Suppose that buyers and sellers engage in Nash
bargaining and thus evenly divide the quasirents from exchange so that
exchange between buyer i and seller j gives each party aij/2. Then, if buyer i
anticipates trading with seller j, buyer i will become a member if and only if
aij =2 P: Also, if seller j anticipates trading with buyer i, seller j will be-
come a member if and only if aij/2ZW.
The effects of unequal fees are more complex with differentiated products
than with homogenous products. Changes in the participation decisions of
buyers or sellers can fundamentally change the configuration of efficient
assignments and thus will affect buyer and seller expectations about par-
ticipation on the other side of the market However, in Becker’s marriage
market, the effects of unequal fees are similar to those in the homogenous
products market. Suppose that P is greater than W. Then, since buyers and
sellers are matched in decreasing order, only those buyers participate if aij/
2ZP and the corresponding number of high-value sellers will participate.
Thus, the greater of the two subscriber fees determines participation.
Proposition 15. Let subscription fees be paid before a match is made. Then,
in Becker’s Marriage Market with Nash Bargaining, the greater of the
buyer and seller fees determines participation in the network.
As in the case of homogenous products, the firm that charges subscriber
fees before a match is made will set equal fees with differentiated products,
P ¼ W ¼ R/2. The firm raises the lower fee to the level of the higher fee to
increase revenue without reducing participation.
In either scenario, with differentiated products, the profit-maximizing
firm is only concerned with the choice of the total fee R. This is again
because the sum of subscription fees is neutral if they are charged after a
match is made, or if they are charged before a match is made, the firm will
choose to make them symmetric, P ¼ W ¼ R/2, To examine the firm’s
choice of a total fee, consider the marriage market with a continuum of
buyers and a continuum of sellers, both of which are uniformly distributed
on the unit interval. Since z2ZR for completed transactions, the firm’s
volume of transactions equals
QðRÞ ¼ 1 R1=2 . (25)
Note that the marginal buyer and the marginal seller equal z ¼ y ¼ 1–Q(R).
The firm’s profit is therefore
PðRÞ ¼ ðR TÞð1 R1=2 Þ. (26)
Arranging terms in the first-order condition, the firm’s profit-maximizing

fee R solves
3R 2ðR Þ1=2 T ¼ 0. (27)
Given the firm’s fee R, consumers’ surplus is the sum of the value of
buyer–seller matches,
Z 1
CSðRÞ ¼ ðz2 RÞdz ¼ ð2=3ÞR3=2 R þ 1=3. (28)
1=2
R
Buyers and sellers pay a transaction cost in the form of the sum of the buyer
and seller subscriber fees, R. This fee has both a resource cost effect on total
buyer and seller benefits RQ(R) and a deadweight loss effect since it pre-
vents efficient matches at the margin. As in the homogenous products case,
the direct cost is a transfer to the firm and is viewed as a transaction cost to
buyers and sellers. The deadweight loss to buyers and sellers is the trans-
actions foregone in comparison with a situation in which the transaction
cost R equals zero. This is shown in Fig. 11. Total surplus including the
firm’s profit equals
VðRÞ ¼ CSðRÞ þ PðRÞ ¼ 1=3 ð1=3ÞR3=2 þ TR1=2 T. (29)
8 Matchmaking and market making by a firm using double auctions
Firms offer a combination of communication and computation that takes

advantage of both centralized networks and centralized market-making
mechanisms. For buyers and sellers that are members of the firm’s network,
any buyer can be matched with any seller. The firm handles communication
and computation by matching buyers and sellers through double auctions.
182 D.F. Spulber
z2
z2
II I
1 Q 1 z
Fig. 11. Transaction costs to buyers and sellers due to a positive subscriber fee R (in
comparison to the case of R ¼ 0) result in a resource cost (I) and a deadweight loss (II).
8.1 Market making by a firm using double auctions for homogenous

products
Consider a market in which the firm assigns buyers to sellers using a

double auction. The auction is a modified Vickrey auction. The firm charges
a fixed commission R for each completed transaction. The firm has a
transaction cost T for each completed trade. This includes the costs of
communicating with buyers and sellers, operating the auction, and com-
pleting purchases and sales. The firm is profitable if and only if RZT.
The firm does not know the types of buyers and sellers in advance, just as
buyers and sellers themselves do not know each others’ types. The firm can
apply a double-auction mechanism to match buyers with sellers efficiently.
Using the double auction mechanism, the firm takes bids from buyers and
sellers and uses the bids to clear the market. The firm sells goods to buyers
and purchases goods from sellers.
There are n potential buyers who choose bids bi, i ¼ 1, y, n. There are m
potential sellers who choose bids sj, j ¼ 1, y, m. Buyers and sellers are risk
neutral and symmetric. Their types are independently distributed with the
uniform distribution on the unit interval. As before, buyers demand at most
one unit of the good and sellers supply at most one unit. Suppose that bids
received by the firm are numbered in decreasing order for buyers,
b1>b2>y>bn, and in increasing order for sellers, s1os2oyosn.
Assume that the firm chooses a uniform market-clearing price p. Buyers
pay p to the firm and the firm pays p–R to sellers. The firm follows a process
for choosing the market-clearing price based on the uniform-price auction.
First, the firm finds the highest i such that

bi si þ R.
Denote that i by i. If bi þ1 si þ R; then let p ¼ bi+1 and sell to the first
i buyers and buy from the first i sellers based on their bids. If that is not
feasible but si þ1 bi and si þ1 si þ R; then let p ¼ si þ1 and sell to the
first i buyers and buy from the first i sellers. Otherwise, let p ¼ bi and sell
to the first i–1 buyers and buy from the first i–1 sellers. For convenience,
let Q represent the market-clearing quantity that results from the auction,
where Q ¼ i if p ¼ bi þ1 or p ¼ si þ1 ; and Q ¼ i–1 if p ¼ bi :
Given the modified uniform-price auction, it can be shown that it is a
weakly dominant strategy for buyers to bid their willingness to pay and for
sellers to bid their cost, bi(vi) ¼ vi and sj(cj) ¼ cj. When arranged in de-
creasing order, buyer values are reversed order statistics. When arranged in
increasing order, seller costs are standard order statistics.
To see why bidding truthfully is a weakly dominant strategy, consider a
buyer’s decision. A buyer never will bid above the true willingness to pay
because the buyer would make a loss if a unit of the good were received. A
buyer never would bid strictly below the true willingness to pay. If the buyer
were not to receive a unit of bidding at willingness to pay, then bidding less
would not have any effect. If the buyer were to receive a unit bidding at
willingness to pay there are two possibilities. If the buyer is marginal then
lowering the bid slightly below value will have no effect on the price paid or
the likelihood of receiving a unit, while lowering the bid farther below value
will result in the buyer not obtaining a unit. An inframarginal buyer does
not have any effect on price from small reductions in the bid and does not
benefit from becoming the marginal buyer by making large reductions in
the bid. Therefore, each buyer bids their true value as a weakly dominant
strategy, bi ¼ vi for all buyers.
Similar arguments apply to seller bids. Each seller bids his true cost as a
weakly dominant strategy, sj ¼ cj for all sellers. Sellers who supply a unit
receive the market-clearing price net of the firm’s commission. The firm uses
the double auction to match buyers and sellers efficiently by including high-
value buyers and low-cost sellers and by excluding low-value buyers and
high-cost sellers. The firm extracts rents from buyers and from sellers, so
that the marginal buyer and the marginal seller have positive surplus. Be-
cause the firm earns rents, it excludes some buyers and some sellers on the
left-hand side of the supply-and-demand diagram.
8.2 Matchmaking and market making by a firm using double auctions for
differentiated products
The firm in a two-sided market for differentiated products can use double
auctions to match buyers with sellers of differentiated products. In addition
184 D.F. Spulber
to being a matchmaker, the firm is also a market maker because the auction
mechanism serves to determine the prices at which each buyer–seller pair
will transact.
In this section, I present a double auction for the two-sided market with
differentiated products. The discussion uses the dominant-strategy mech-
anism due to Vickrey (1961), Clarke (1971), and Groves (1973), and the
generalized Vickrey–Clarke–Groves mechanism of Krishna and Perry
(1998). An efficient Vickrey–Clarke–Groves mechanism that is incentive
compatible and individually rational needs to satisfy a balanced-budget
constraint, see Green and Laffont (1977).
It has been shown that under some conditions, there may exist an efficient
Bayes–Nash mechanism that is incentive compatible and individually ra-
tional that satisfies a balanced-budget constraint, d’Aspremont and Gérard-
Varet (1979). Spulber (1988) shows that an efficient Bayes–Nash mecha-
nism will satisfy budget balancing if the gains from trade between buyers
and sellers are sufficiently large that they cover information rents used to
induce buyers and sellers to tell the truth.
I now consider the problem of a firm that is a monopoly intermediary in a
two-sided market. Since the firm is maximizing profits there is not the
problem of operating with a surplus, but the firm must be profitable to
operate. The firm provides buyers and sellers with incentives to participate
in the auction and to reveal their true types. The firm is profitable if the
rents from exchange between buyers and sellers is sufficient to cover the
incentives for truth-telling. The firm extracts rents from buyers and sellers,
so that the firm necessarily excludes some matches that would be in the
Core.
Demange (1982) and Leonard (1983) design a related one-sided Vickrey
auction for the two-sided market with differentiated products. They assume
that all sellers have identical costs that equal zero, so that in their frame-
work the auction serves to allocate the differentiated products across buy-
ers. Demange (1982) and Leonard (1983) use buyer bids to determine
imputations in the Core of the cooperative game, which yields payoffs to
both buyers and sellers (see also Roth and Sotomayor, 1990). In their
framework, the auction satisfies a break-even constraint since buyer payoffs
plus seller payoffs equal total value. In contrast, I allow sellers to have
different costs so that sellers must also make bids that reveal their cost
information.
In the general setting of the two-sided market, sellers have different costs,
cj>0, j ¼ 1, y, m. Also, the double auction determines ask prices for
buyers and bid prices for sellers. Accordingly, the firm designing the auction
earns a profit equal to the difference between payments received from buy-
ers and payments made to sellers.
There are n buyers, i ¼ 1, y, n. Buyer types are represented by a vector
of benefits that the buyer would obtain from consuming a unit of the
various goods j ¼ 1, y, m, ai ¼ (ai1, ai2, y, aim). Seller types are simply
scalar costs cj for sellers j ¼ 1, y, m. Let Hi be the set of possible types for a
buyer where Hi is a compact, convex subset of Rm, and let Cj be the set of
possible cost types for a seller, where Cj is a closed interval on the positive
real line. Buyers report their vector of preference types ai in Hi and sellers
report their cost types cj in Cj. Let H ¼ in¼ 1Hi and let C ¼ jm¼ 1Cj.
Let the n m matrix of values x represent an allocation of goods. The
values xij in the matrix
P x equal either zero or one. A buyer receives at most a
unit
P of one good, ixijr1. A seller provides at most a unit of one good
x
i ijr1. A good j can be transferred from seller j to buyer i only if aij>cj.
Define a benefit function for buyers bi(x) such that
bi ðxÞ ¼ aij
if for some j in S, xij ¼ 1, and let bi(x) ¼ 0 otherwise. Define a benefit
function for sellers bj(x) such that
bj ðxÞ ¼ cj
if for some i in B, xij ¼ 1 and let bj(x) ¼ 0 otherwise. Let X be a dis-
crete finite set containing all possible allocations of the differentiated
products.
The firm designs a direct mechanism consisting of an allocation rule Q:
H C-X and a payment rule c : H C-Rn+m. The firm obtains the
allocation rule Q by calculating the valuations vij ¼ max{0, aij–cj} and
choosing x to solve the linear programming problem (4). The firm’s pay-
ment rule c consists of a vector of ask prices for buyers p1 , p2 , y, pn and a
vector of bid prices for sellers w1 , w2 , y, wm. A buyer i reports his type,
which consists of a preference vector ai1, ai2, y, aim and a seller j reports his
cost type cj. Then, given the allocation rule Q and the payment rule c, a
buyer i receives bi(x)–pi, and a seller j receives wj+bj(x).
By the definition of the individual benefit functions, it follows that the
total benefit equals the sum of the individual benefit functions
X X
VðJÞ ¼ bi ðx Þ þ bj ðx Þ. (30)
i2B j2S
Also, define V(J–i) as the total benefits obtained from optimization problem
(4) with the set of buyers except for buyer i, B–i and the set of sellers S.
Define V(J–j) as the total benefits obtained from optimization problem (4)
with the set of buyers B and the set of sellers except for seller j, S–j. Define
V(J–i–j) as the total benefits obtained from optimization problem (4)
with the set of buyers B and the set of sellers S, excluding the buyer–seller
pair (i, j).
Given the generalized Vickrey–Clarke–Groves mechanism, the payment
for buyer i equals
X X
pi ¼ VðJ iÞ bk ðx Þ bj ðx Þ þ R=2. (31)
k2B; kai j2s
186 D.F. Spulber
The Vickrey–Clarke–Groves payment to seller j equals

X X
wj ¼ V ðJ jÞ þ bi ðx Þ þ bk ðx Þ R=2. (32)
i2B k2s; kaj
Therefore, buyer i receives a benefit from the auction equal to

bi ðx Þ pi ¼ V ðJÞ V ðJ iÞ R=2. (33)
Seller j receives a benefit from the auction equal to
wj þ bj ðx Þ ¼ V ðJÞ V ðJ jÞ R=2. (34)
The buyer and seller payment mechanism is thus
c ¼ ðp1 ; . . . ; pn ; w1 ; . . . ; wm Þ, (35)
where prices depend on the reported types, that is pi(a, c), wj(a, c) with (a, c)
in H C.
The firm assigns buyer i to seller j to maximize the sum of net benefits. Let
E be the firm’s profit-maximizing assignment and let n be the number of
matches in the assignment. Assume that the firm chooses R to maximize
profit after observing buyer and seller bids. Buyers and sellers choose to
participate if they have positive net benefits, so that the mechanism satisfies
individual rationality.
The mechanism is incentive compatible. Buyer i’s payoff is V(J)–V(J–i)–
R/2 so that only the first term depends on the buyer’s reported type
ai ¼ (ai1, y, aim). The buyer maximizes the first term by correctly reporting
his type. Seller j’s payoff has the same property. Since truth-telling is a
weakly dominant strategy with the Vickrey–Clarke–Groves mechanism, the
mechanism is incentive compatible. Thus, in equilibrium every buyer and
every seller receives their marginal contribution to total benefits (see
Krishna and Perry, 1998).
A mechanism P if the allocation rule Q: H C-X maximizes
P is efficient
total benefits ibi(x)+ jbj(x). Krishna and Perry (1998) demonstrate that
the Vickrey–Clarke–Groves mechanism maximizes the expected payments
of each agent among all mechanisms that are efficient, incentive compatible,
and individually rational. The Vickrey–Clarke–Groves mechanism is effi-
cient since it selects an allocation x that yields the maximum benefits V(J)
for the set of buyers and sellers J ¼ (B, S).
Except for the payment R/2, the payoff to a buyer in the Vickrey double
auction, V(J)–V(J–i)–R, exactly equals the maximum payoff to a buyer in
the Core since it equals each buyer’s contribution to the grand coalition’s
value. This is the same payoff observed in the one-sided Vickrey auction
considered by Demange (1982) and Leonard (1983). In the double auction,
the payoff to a seller also equals the maximum payoff since it equals their
contribution to the grand coalition’s value V(J)–V(J–j), again net of R/2.
Given the Vickrey–Clarke–Groves mechanism, the firm’s profit is

X X
P¼ pi wj . (36)
i2B j2S
If buyer i is matched with seller j, the price spread equals

pi wj ¼ bi ðx Þ bj ðx Þ ðV ðJÞ V ðJ iÞÞ
ðVðJÞ V ðJ jÞÞ þ R. ð37Þ
By the definition of the benefit functions, note that bi(x)bj(x) ¼ aijcj.
The incremental value contributed by a matched pair (i, j) equals
aij cj ¼ V ðJÞ V ðJ i jÞ. (38)
Substituting into the price spread equation gives
pi wj ¼ ðV ðJÞ þ V ðJ i jÞ V ðJ iÞ V ðJ jÞÞ þ R. (39)
The term in parenthesis is positive since buyers and sellers are complements.
Thus, the price spread is less than or equal to R for each buyer and seller
pair, piwjrR.
Summing over buyer–seller pairs in the assignment chosen by the firm,
the firm’s profit equals
X
P¼ ðV ðJÞ þ V ðJ i jÞ V ðJ iÞ VðJ jÞÞ þ n R.
ði;jÞ2E
(40)
To illustrate the firm’s optimization problem, consider an example with
two buyers and two sellers. Assume that a11–c1oa22–c2.
a11 c1 40; a12 c2 o0; a21 c1 o0; a22 c2 40.
The optimal assignment is (1, 1) and (2, 2). Then, the total benefit equals
V(J) ¼ a11c1+a22c2. The benefit without buyer 1 or without seller 1
equals V(Ji) ¼ V(Jj) ¼ V(Jij) ¼ a22c2 for i ¼ j ¼ 1. The benefit
without buyer 2 or without seller 2 equals
V(Ji) ¼ V(Jj) ¼ V(Jij) ¼ a11c1 for i ¼ j ¼ 2. The firm’s profit is
then
P ¼ ða11 c1 þ a22 c2 Þ þ 2R.
The firm sets the fee R such that
R=2 ¼ a11 c1 .
Substituting for the fee R, the firm’s profit equals
P ¼ 3ða11 c1 Þ ða22 c2 Þ.
The firm is profitable if 3(a11c1)>(a22c2).
188 D.F. Spulber
9 Two-sided markets in random networks
In this section, I consider decentralized exchange between buyers and

sellers with both costly communication and costly computation. I examine
both costly search and transactions on random networks. The inefficiencies
of the random matching process are the manifestation of transaction costs
for buyers and for sellers.
When communication is costly, buyers and sellers may incur search costs
in trying to find trading partners. Before searching, consumers may not
know the characteristics of other consumers. The costs of search to con-
sumers can include the costs of travel and the time it takes to search. Also,
the search market represents a market with costly computation since the
allocation mechanism is decentralized with transactions taking the form of
independent, bilateral transactions.20
The search process is inherently uncertain, consumers do not know
whether or not they will find a trading partner and they do not know the
characteristics of prospective trading partners. The risks involved in the
search process represent yet another cost of search. Because search takes
time, consumers experience delays in consumption. The loss of future ben-
efits is another cost of search if consumers discount future benefits. In
addition to the resources expended in the search process and the costs of
risk and time, costly search results in deadweight losses either from ineffi-
cient matches or from the failure to make efficient matches. Consumers may
not find the best trading partner, settling instead for a less-desirable match
that is available so as to avoid further search. If search is sufficiently costly,
consumers may choose not to seek any trading partners. Thus, the costs of
search include lost gains from trade.
9.1 Search and random assignments
A random selection from the set of assignments corresponds to random

matching of buyers and sellers. Random matching is one way of representing
the transaction costs of consumer search. If there are n buyers and n sellers,
then a complete assignment consists of n matchings and there are n! equally
likely assignments. Buyers and sellers do not know each others’ identities
before they meet to trade. Buyers and sellers are matched only once. After a
buyer and seller meet, they observe each others’ type. After they are matched,
trade takes place between buyer i and seller j if and only if gains from trade
are nonnegative. In the market with homogenous products this requires
viZcj, and in the market for differentiated products this requires aijZcj.
This implies the following result. Consider either (i) the market with
homogenous products or (ii) the market for differentiated products with
20
See Shimer and Smith (2000) for a model of assortative matching with costly search.
additive production function and different costs across sellers. Then, ran-
dom matching is efficient if and only if the Complete Viability Condition
holds.
Consider first the market for homogenous products. Random matching is
efficient if and only if there is no right-hand side of the supply-and-demand
diagram. The demand curve must be everywhere above the supply curve.
Otherwise, buyers and sellers on the right-hand side of the supply-and-
demand diagram are active in the search market. These low-value buyers
and high-cost sellers may be matched with players who would otherwise
trade under an efficient assignment, thus displacing efficient matches. The
displacement of efficient matches can be viewed as an externality that re-
sults from search inefficiencies. Buyers and sellers cannot choose efficient
matches due to imperfect information about the types of potential trading
partners. Costly search prevents buyers and sellers from visiting multiple
trading partners as a way of improving matches.
Sufficiency of the Complete Viability Condition is evident since random
matching will result in a selection from the set of efficient assignments. To
establish necessity, suppose that all pairs are viable vnZcm, but the set of
sellers has one more member than the set of buyers. Since one of the buyers
can be matched with the highest-cost seller, a better pairing is displaced,
which is inefficient. Suppose that there is the same number of buyers and
sellers and viZci, i ¼ 1, y, n–1 but vnocn. Then, if buyer 1 is matched with
seller n, the match displaces an efficient match. Therefore, the Complete
Viability Condition is necessary for random matching to be efficient. So, the
Complete Viability Condition is both necessary and sufficient for random
matching to be efficient.
When the Complete Viability Condition does not hold therefore, ran-
dom matching is always inefficient in expectation. The demand and supply
functions cross, due to the highest-value buyer having a lower value than
the high-cost seller’s cost or due to the number of buyers and sellers
not being equal. Thus, inefficient matches can occur with positive prob-
ability.
If the Complete Viability Condition does not hold, random matching can
generate too much trade. Consider, for example, a market with four buyers
and four sellers. Assume that v1>c4>v2>c3>v3>c2>v4>c1 (see Fig. 12). Sup-
ply and demand cross so the Complete Viability Condition does not hold.
The efficient number of trades equals 2. At an efficient outcome, buyers v1
and v2 would be paired with sellers c1 and c2 in either combination. Buyers
v3 and v4 and sellers c3 and c4 would be inactive. However, there are 4! ¼ 24
possible assignments, so only 4 out of 24 are efficient. With random
matching, efficiency occurs with probability 1/6. Otherwise, successful
matches with inefficient buyers or sellers are possible. Out of 24 assignments
there is only 1 assignment with 4 trades. There are 10 assignments with 3
trades, 12 assignments with 2 trades, and 1 assignment with 1 trade. Thus,
the expected number of trades with random matching equals 4(1/24)+3(10/24)
190 D.F. Spulber
v, c
v1
Supply
c4
v2
c3
v3
c2
v4
c1
Demand
Q
Fig. 12. A market with four buyers and four sellers.
+2(12/24)+1(1/24) ¼ 59/24. This exceeds 2, which is the number of effi-

cient trades.
If buyer types and seller types are chosen randomly from overlapping
intervals of buyer values and seller costs, the Complete Viability Condition
is not likely to hold. To illustrate the inefficiency of random matching,
suppose that there are many buyers and sellers. Let buyer values v be
uniformly distributed on the unit interval so that market demand is
DðpÞ ¼ Prfv4pg ¼ 1 p. (41)
Let seller costs c be uniformly distributed on the unit interval so that mar-
ket supply is
SðpÞ ¼ Prfc pg ¼ p. (42)
D S
Let P (Q) ¼ 1Q and P (Q) ¼ Q be the inverse demand and the inverse
supply. The efficient volume of trade is Q ¼ 1/2 and the total benefits with
an efficient assignment of buyers to sellers are
Z 1 Z 1=2

V ¼ vdv cdc ¼ 1=4. (43)
1=2 0
The total expected value with random matching and a continuum of

buyers and sellers equals
Z 1Z v
0
V ¼ ðv cÞdcdv ¼ 1=6, (44)
0 0
so that random matching is not efficient. The total expected volume of trade
with random matching and a continuum of buyers and seller is Q0 ¼ 1/2,
which is the same as the efficient volume. Now compare decentralized
search by buyers and sellers with centralized exchange with a network op-
erated by monopoly firm. The firm’s network is more efficient than decen-
tralized search, VZV0 ¼ 1/6 if the firm’s costs are not too large,
Tr1(1/3)O8.
Consider now the market for differentiated products. Any costly search
process under imperfect information introduces randomness in the assign-
ment process. It is straightforward to show that random matching will be
inefficient
P for the assignment
P game. Let bij be a set of probabilities such that
bijZ0, ibij ¼ 1, and jbij ¼ 1. Then, for any arbitrary set of probabilities,
it follows by optimization that the efficient solution must yield a greater
total value than random matching,
XX XX
vij xij vij bij .
i j i j
The preceding inequality is likely to be strict. The maximum value in the

linear program is attained for xij taking values of zero or one. Suppose for
example that there is a unique efficient matching where xij ¼ 1 for i ¼ j,
xij ¼ 0 for i6¼j and n ¼ m ¼ 2. Thus, v11+v22>v12+v21. Then, the expected
value of any random matching with 0obijo1 for some ij pair is strictly less
than the maximum total value,
v11 þ v22 4X 11 ðv11 þ v22 Þ þ ð1 X 11 Þðv12 þ v21 Þ,
XX
¼ vij bij ;
i j
since b11 ¼ 1–b12 ¼ 1–b21 ¼ b22.

The inefficiency of random matching applies to any set of probabilities
including a random process that matches buyers and sellers with equal
probability of any match. If the numbers of buyers and sellers are not equal,
then any player on the long side of the market has the same probability of
not being matched. Let vij ¼ 0 represent the value of matches between
players on the long side of the market with ‘‘dummy’’ players that are
added to the short side of the market to equalize the number of players on
each side. Then, the inefficiency of random matching also applies when the
number of buyers differs from the number of sellers.
Consider Becker’s marriage market with a continuum of buyers and sell-
ers who have productivity parameters uniformly distributed on the unit
192 D.F. Spulber
interval. When z ¼ y, a(z, y) ¼ z2. Thus, the value of the grand coalition of
buyers and sellers is
Z 1

V ¼ vðB [ SÞ ¼ aðz; zÞdz ¼ 1=3. (45)
0
Buyers and sellers can attain this outcome if there is full information and
costless matching. To attain the outcome requires a network with a com-
plete set of links between all buyers and all sellers. If establishing links is
costly, then it will be very costly to establish a complete network so as to
attain the efficient assignment.
Compare the value generated by efficient matching with that generated by
random matching. With random matching all buyer–seller pairs are feasible
since they generate nonnegative value. The total gains from trade with
random matching equals
Z 1Z 1
0
V ¼ zydydz ¼ 1=4: (46)
0 0
The difference between V ¼ 1/3 and V0 ¼ 1/4 represents the inefficiency of

random matching in a large market with differentiated products. In a
Becker marriage market, all matches generate positive benefits so that ran-
dom matching performs relatively well. More generally, with seller costs
and general differentiated products, there can be matches that do not result
in trade, so that random matching need not perform well.
To compare random matching with intermediation by the firm, let the
firm costs T equal zero. Then, the monopoly commission rate is R ¼ 4/9
and total surplus is
VðR Þ ¼ ð1=3Þð19=27ÞoV 0 ¼ 1=4.
Since any positive value of costs T will lower total surplus, it follows that
centralized matching by the monopoly firm yields lower total gains from
trade in comparison with random matching. Recall that random search in the
marriage market performs well because all matches yield positive benefits.
The monopoly firm also performs less well than random search because
of the efficiency losses from monopoly pricing. Centralized matching by the
firm offers advantages in comparison with decentralized matching that can
be realized by lowering the fee. Suppose that the firm competes with de-
centralized exchange and suppose that buyers and sellers must decide
whether to enter the search market or deal with the firm before observing
their type. The firm will be constrained in its pricing by the search market
alternative. The highest commission rate that the firm can charge and still
attract potential buyers and sellers must yields the consumer surplus from
decentralized exchange,
CSðRMAX Þ ¼ V 0 ¼ 1=4, (47)
so RMAX solves 2R3/2–3R+1/4 ¼ 0 and RMAX is approximately 1/10. For

the market-making firm to be profitable and to improve economic effi-
ciency, its transaction cost T must be less than RMAX. Then, the market-
making firm can attract all potential buyers and sellers in a market with
differentiated products.
Given consumers’ surplus evaluated at the maximum commission
CS(RMAX) ¼ V0, the firm’s profit at the maximum commission rate equals
PðRMAX Þ ¼ V ðRMAX Þ V 0 . (48)
MAX
If transaction costs are not too high, ToR , the firm earns a profit by
improving allocative efficiency relative to the search market. The firm offers
the advantage of efficient assignments of buyers and sellers that are its
subscribers but at the cost of excluding some buyers and sellers. The search
market yields inefficient assignments but all buyers and sellers participate
and receive positive value from any match. The firm adds value if the direct
costs of making assignments and the deadweight losses from excluding
some buyers and sellers are not too high.
Hoppe, Moldovanu and Sela (2005) consider a Becker-type marriage
market. They compare the total gains from random matching with that
from assortative matching under incomplete information when players can
engage in costly signaling. For a class of distributions with increasing fail-
ure rate, they find that random matching performs better than assortative
matching, although for distributions with a decreasing failure rate, the
efficiency gains from assortative matching can exceed those under random
matching.
9.2 Markets and random networks
Mathematicians, scientists, and sociologists have studied extensively the

random formation of networks.21 Random networks are of great value in
understanding the structure of networks. First, random networks provide a
contrast with the design of efficient networks by firms that establish net-
works. Second, random networks are highly interesting because they sug-
gest patterns of usage that might be observed within an existing network.
For example, telephone calls on a telecommunications network within a
given period of time activate various nodes and links to constitute a tem-
porary network that reflects traffic patterns on the underlying physical
network. Network reliability depends on the supply of network capacity
and random demand for that capacity.
Even managed networks evolve over time as new facilities are added and
old facilities are abandoned. Companies alter their network facilities in
response to variations in customer demand, production costs, technology,
21
On random networks, see Solomonoff and Rapoport (1951), Erdös and Rényi (1960, 1961), Janson
et al. (2000), and Bollobás, (2001). For popular introductions, see Barabási (2002) and Watts (2003).
194 D.F. Spulber
competitor behavior, and regulation. Thus, demand shocks or cost shocks

alter the plans of network designers and introduce randomness in the ev-
olution of the network. Random networks offer some insights into the
evolution of designed networks that are subject to random shocks.
Networks that offer open access experience randomness in terms of the
types of complementary services that are provided by independent suppli-
ers. Innovation by suppliers of complementary services introduces uncer-
tainty about traffic and usage patterns on a network. The rapid growth of
electronic commerce applications on the Internet created demand patterns
and usage of the network exhibited considerable randomness. For example,
consider the impact of search engines on the development of the Internet.
Consider a simple random network with three preexisting nodes. Suppose
that in a given period of time, there is a probability equal to b that a link
forms to connect any pair of nodes where 0obo1. Then, by the end of the
period, there are four possibilities. There might be no links, one link, two
links, or the full set of three links. The likelihood of these events is as
follows: no links (1–b)3, one link 3b(1–b)2, two links 3b2(1–b), and three
links b3. These possibilities are shown in Fig. 13. This assumes that there is
no distinction between any of the nodes. The probability distribution over
the number of links follows the basic Bernoulli distribution.
The notion of random networks can be extended to examine the evo-
lution of networks over time. Suppose for example that at most one link can
form per unit of time, again with likelihood b. Then, the four configurations
shown in Fig. 13 also would depict the growth of the network over time.
The network would have no links after the first period with likelihood 1–b
and one link with likelihood b. After two periods, the network would have
no links with likelihood (1–b)2, one link with likelihood 2b(1–b), and two
links with likelihood b2. The likelihood of a link forming in a period thus
will determine the speed of network formation.
Another extension of this approach is to make the likelihood of links
forming depend in some way on the characteristics of the nodes. For ex-
ample, the nodes might be separated by geographic distance. A link might
be more likely to form between two nodes that are closer to each other than
No links One link Two links Three links

1 1 1 1
3 2 3 2 3 2 3 2
(1 – b)3 3b(1 – b)2 3b2(1 – b) b3
Fig. 13. The likelihood of various configurations of a random network with three links and
likelihood b of a link forming between any pair of nodes.
between two nodes that are farther apart. This might describe the formation
of business relationships when individuals are separated geographically.
The discussion of random networks assumed that the nodes of the net-
work were already established but that the formation of links was random.
Instead, one can suppose that the number of characteristics of the nodes are
also random. For example, the existence of a node can represent the de-
cision of a consumer to enter or exit a given market. The number of nodes
at any given time would be the net result of the entry and exit of individuals
up to that time, in a similar way to demographic analysis of the evaluation
of a population. Then, networks would reflect both the random number of
nodes and the random formation of links between them.
Random graphs have many interesting properties. Consider a set of
nodes N ¼ {1, y, n}. Suppose that in each time period starting with the
first time period, one new link is added on, so that at time t the graph has
exactly t links. Suppose that any new link is equally likely. This is known as
a random graph process. Such a process provides a description of how a
graph might grow over time if the links were to be added randomly.
As links are added randomly to a graph, the largest component of the
graph has an important feature. Suppose that there is an initial set of n
nodes. Then, after t periods of time, where t is not much greater than n/2, a
giant component or cluster forms that swallows up other large components.
The critical time n/2 is the phase transition time for the random graph
process. For example, a graph process beginning with eight unconnected
nodes has formed a giant component shortly after only four periods. After
more time passes, the giant component swallows up all other components
and the graph becomes connected! This means that by adding links ran-
domly to n nodes, a network is very likely to be connected shortly after n/2
time periods.22
Assignments on bivariate graphs can be extended to incorporate uncer-
tainty in the creation of links. Suppose that the numbers of buyers and of
sellers are equal, n ¼ m. Let G(n, p) denote the set of all bipartite graphs
with n nodes in each of the two parts of the graph. Let p be the probability
that any pair of nodes, one from each set, is connected by a link. The
probability that a graph G0 in G has y links is given by

y Ny n
PðG0 Þ ¼ p ð1 pÞ ; where N ¼ .
2
The greater is the likelihood of link formation p, the greater is the expected
number of links.
Since a random network is likely to be incomplete, assignments on the
network are likely to be inefficient. The expected efficiency of the network
will therefore be increasing in the probability of link formation. Proceeding
22
This discussion on random graphs draws from Béla Bollobás (2001) and Janson et al. (2000).
196 D.F. Spulber
informally, let V(p) represent the expected gains from trade generated by a
random network with probability of link formation p. As p approaches one,
the likelihood of a complete network approaches one and decentralized
exchange will be more efficient than a market with a firm that centralizes
assignments.
Suppose in contrast that a firm establishes a network with links to all
buyers and sellers. The firm matches buyers and sellers efficiently as long as
the surplus from a match exceeds the fee charged by the firm R. Then, the
expected total benefits generated by the firm equals V(R), the firm offers an
advantage over decentralized exchange if p is sufficiently small.23
10 Conclusion
Firms play an essential role in the economy if they can coordinate trans-
actions between buyers and sellers more efficiently than decentralized
search and bargaining. Firms employ information systems to improve
communication between buyers and sellers and to improve computation
through centralized market mechanisms. Whether goods are homogenous
or differentiated, firms can employ dominant-strategy auction mechanisms
to match diverse buyers with diverse sellers. Such centralized matching
mechanisms can offer efficiencies that reduce or avoid some of the resource
costs and deadweight losses from search.
When communication between consumers requires establishing costly
networks, firms can offer some efficiencies from centralized networks. Hub-
and-spoke networks with the firm at the center offer potential benefits
relative to costly point-to-point networks that link consumers. There is a
trade off between benefits of establishing more links in a decentralized
network, which improves the performance of the network, and the costs of
the links. By establishing and coordinating buyer networks and seller net-
works, firms offer potential advantages relative to decentralized exchange
over a network of complete links between consumers. Centralized networks
created by firms can also offer benefits relative to consumer networks that
form randomly.
23
We can also represent random networks as a selection from a set of graphs. Let G(n, Y) be a set of
bipartite graphs with n nodes in each set and Y links connecting nodes in one set with nodes in the other.
n
graphs are equally likely. Then, since G(n, Y) has Y elements,
Suppose that all graphs in the set of such
each graph occurs with probability 1= Yn : Then, G(n, Y) approaches a complete graph as Y goes to n2.
As y increases, G(n, Y) yields a bipartite graph that is more likely to permit an efficient assignment. Let
V(Y) be the expected efficiency of decentralized exchange with Y randomly chosen links. Compare this
with a firm that establishes a centralized network at a fixed cost of T and matches buyers and sellers
efficiently. Then, compare V(R) with g(Y). As Y increases toward n2, the expected benefits of decen-
tralized approach full efficiency. The firm offers an advantage over decentralized assignments of buyers
and sellers if Y is sufficiently small. As Y increases, any randomly chosen network will permit more
efficient assignments. Thus, if Y is sufficiently large, decentralized exchange is likely to be more efficient
than a market with a firm that centralizes assignments.
Economic analysis of network theory suggests that the design of infor-

mation systems can be enhanced by adding economic efficiency criteria to
technical performance criteria. Engineers commonly use network theory to
design communications and information systems to satisfy performance
criteria such as reliability and transmission capacity. The interaction be-
tween information networks and markets is most clearly observed in elec-
tronic commerce such as Internet retail transactions and business-to-
business wholesale transactions. Electronic trading systems in financial
markets are another key example of the interaction between information
networks and markets. Designers of information systems can benefit from a
better understanding of the connections between networks and markets.
Such an understanding should yield innovations in information systems
that are designed to enhance the efficiency of economic transactions.
Additional research is needed to explore the role of the firm as an in-
termediary. Competition between firms as matchmakers and as market
makers should continue to provide a rich area of study. Network theory
provides a powerful tool for understanding economic interactions between
buyers and sellers that can be applied in a variety of settings. Network
theory provides critical insights on how firms help to establish the micro-
structure of markets.
Acknowledgments
Daniel F. Spulber is the Elinor Hobbs Distinguished Professor of Inter-

national Business and Professor of Management Strategy at the Kellogg
School of Management, Northwestern University. The support of a re-
search grant from the Searle Fund is gratefully acknowledged. I thank
Michael Baye, George Deltas, Yuk-fai Fong, Terry Hendershott, Heidrun
Hoppe, Simon Loertscher, and John Morgan for helpful comments.
References
Alderson, W. (1954). Factors governing the development of marketing channels, in: R.M. Clewett (ed.),
Marketing Channels for Manufactured Products, Irwin, Homewood, IL.
Aldous, J.M., R.J. Wilson (2000). Graphs and Applications: An Introductory Approach. Springer, New
York.
Barabási, A.-L. (2002). Linked: The New Science of Networks. Perseus Publishing, New York.
Baye, M., J. Morgan (2001). Information gatekeepers on the internet and the competitiveness of ho-
mogenous product markets. American Economic Review 91, 454–474.
Baye, M., J. Morgan, P. Scholten (2004). Price dispersion in the large and in the small: evidence from an
internet price comparison site. Journal of Industrial Economics 52(December), 463–496.
Becker, G.S. (1973). A theory of marriage: part I. The Journal of Political Economy 81(4), 813–846.
Böhm-Bawerk, E. von (1891). Positive Theory of Capital, Translated from German by W.E. Smart.
Macmillan, London.
Bollobás, B. (1998). Modern Graph Theory. Springer, New York.
198 D.F. Spulber
Bollobás, B. (2001). Random Graphs 2nd Edition. Cambridge University Press, Cambridge.
Bresnahan, T., S.M. Greenstein (1996). Technical progress and co-invention on computing and in the
uses of computers. Brookings Papers on Economic Activity: Microeconomics 1996, 1–78.
Bresnahan, T., S.M. Greenstein (1999). Technological competition and the structure of the computer
industry. Journal of Industrial Economics 47(March), 1–40.
Caillaud, B., B. Jullien (2003). Chicken & egg: competition among intermediation service providers.
Rand Journal of Economics 34, 309–328.
Clarke, E. (1971). Multipart pricing of public goods. Public Choice 2, 19–33.
Clower, R., A. Leijonhufvud (1975). The coordination of economic activities: a Keynesian perspective.
Committee T1A1 [renamed Network Performance, Reliability and Quality of Service Committee
(PRQC)] (2000). ATIS Telecom Glossary 2000. Alliance for Telecommunications Industry Solutions,
Washington, DC.
d’Aspremont, C., L.-A. Gérard-Varet (1979). Incentives and incomplete information. Journal of Public
Economics 11(February), 25–45.
Dantzig, G. (1963). Linear Programming and Extensions. Princeton University Press, Princeton, NJ.
Demange, G. (1982). Strategyproofness in the assignment market game, Laboratoire d’Economie de
L’Ecole Polytechnique, Paris, Working Paper.
Demange, G., D. Gale (1985). The strategy structure of two-sided matching markets. Econometrica 53,
873–888.
Diestel, R. (2000). Graph Theory 2nd Edition. Springer, New York.
Dutta, B. and M.O. Jackson (eds.) (2003). Networks and Groups: Models of Strategic Formation.
Springer, Berlin.
Dutta, B., S. Mutuswami (1997). Stable networks. Journal of Economic Theory 76, 322–344.
Ellison, G., D. Fudenberg M. Mobius (2003). Competing auctions, Department of Economics, MIT,
Working Paper, March.
Euler, L. (1736). Solutio problematis ad geometriam situs pertinentis [The solution of a problem relating
to the geometry of position], 8 Commentarii Academiae Scientiarum Imperialis Petropolitanae
128,128—140, reprinted in Biggs, N.L., E.K. Lloyd and R.J. Wilson (1976). Graph Theory 1736—
1936. London, Oxford University Press, 3–11.
Erdös, P., A. Rényi (1960). On the evolution of random graphs. Publications of the Mathematical
Institute of the Hungarian Academy of Sciences 5, 17–61.
Erdös, P., A. Rényi (1961). On the strength of connectedness of random graphs. Acta Mathematica
Hungarian Academy of Sciences 12, 261–267.
Farrell, J., G. Saloner (1985). Standardization, compatibility, and innovation. Rand Journal of Eco-
nomics 16, 70–83.
Farrell, J., G. Saloner (1986). Installed base and compatibility: innovation, product preannouncements,
and predation. American Economic Review 76, 940–955.
Fingleton, J. (1997). Competition among middlemen when buyers and sellers can trade directly. The
Journal of Industrial Economics, XLV(4), 405–427.
Gale, D., L. Shapley (1962). College admissions and the stability of marriage. The American Math-
ematical Monthly 69(1), 9–15.
Gehrig, T. (1993). Intermediation in search markets. Journal of Economics and Management Strategy 2,
97–120.
Granovettor, M. (1985). Economic action and social structure: the problem of embeddedness. American
Journal of Sociology 91, 451–510.
Green, J., J.-J. Laffont (1977). Characterization of satisfactory mechanisms for the revelation of pref-
erences for public goods. Econometrica 45, 427–438.
Greenstein, S.M. (1998). Industrial economics and strategy: computing platforms. IEEE Micro 18(May–
June), 43–53.
Gross, J., J. Yellen (1999). Graph Theory and Its Applications. CRC Press, Boca Raton.
Gross, J. and J. Yellen (eds.) (2004). Handbook of Graph Theory. CRC Press, Boca Raton.
Groves, T. (1973). Incentives in teams. Econometrica 41, 617–631.
Hall, P. (1935). On representatives of subsets. Journal of the London Mathematical Society 10, 26–30.
Hendershott, T., J. Zhang (2006). A model of direct and intermediated sales. Journal of Management &
Economics Strategy 15, 279–316.
Hoppe, H., B. Moldovanu, A. Sela (2005). The theory of assortative matching based on costly signals.
Department of Economics, University of Bonn, Working Paper, November.
Hylland, A., R. Zeckhauser (1979). The efficient allocation of individuals to positions. The Journal of
Political Economy 87(2), 293–314.
Jackson, M.O., A. Wolinsky (1996). A strategic model of social and economic networks. Journal of
Economic Theory 71, 44–74.
Janson, S., T. Luczak, A. Rucinski (2000). Random Graphs. Wiley, New York.
Ju, J., S.C. Linn, Z. Zhu (2004). Price dispersion in a model with middlemen and oligopolistic market
makers: a theory and an application to the North American natural gas market. University of
Oklahoma, Working Paper.
Katz, M.L., C. Shapiro (1985). Network externalities, competition and compatibility. American Eco-
nomic Review 75, 424–440.
Katz, M.L., C. Shapiro (1986). Technology adoption in the presence of network externalities. Journal of
Political Economy 94, 822–841.
Katz, M.L., C. Shapiro (1994). Systems competition and network effects. Journal of Economic Per-
spectives 8(Spring), 93–115.
Kranton, R.E., D.F. Minehart (2000). Competition for goods in buyer–seller networks. Review of
Economic Design 5, 301–331.
Kranton, R.E., D.F. Minehart (2001). A theory of buyer–seller networks. American Economic Review
91(June), 485–508.
Krishna, V., M. Perry (1998). Efficient mechanism design. Pennsylvania State University, PA, Working
Paper, April.
Leonard, H.B. (1983). Elicitation of honest preferences for the assignment of individuals to positions.
Journal of Political Economy 19(3), 461–479.
Liebowitz, S.J. (2002). Rethinking the Network Economy. American Management Association, New
York.
Liebowitz, S.J., S.E. Margolis (1994). Network externality: an uncommon tragedy. Journal of Economic
Perspectives 8, 133–150.
Liebowitz, S.J., S.E. Margolis (1999). Winners Losers and Microsoft: Competition and Antitrust in High
Technology. The Independent Institute, Oakland, CA.
Loertscher, S. (2004). Market-making oligopoly. University of Bern, Working Paper.
Loertscher, S. (2005). Horizontally differentiated market makers. University of Bern, Working Paper.
Lucking-Reiley, D., D.F. Spulber (2001). Business-to-business electronic commerce. Journal of Eco-
nomic Perspectives 15(Winter), 55–68.
Myerson, R. (1977). Graphs and cooperation in games. Mathematics of Operations Research 2, 225–229.
Roth, A.E., M.A.O. Sotomayor (1990). Two-Sided Matching: A Study in Game Theoretic Modeling and
Analysis. Cambridge University Press, Cambridge.
Roth, A.E., X. Xing (1994). Jumping the gun: imperfections and institutions related to the timing of
market transactions. American Economic Review 84(4), 992–1044.
Rubinstein, A., A. Wolinsky (1987). Middlemen. Quarterly Journal of Economics 102, 581–593.
Rust, J., G. Hall (2003). Middlemen versus market makers: a theory of competitive exchange. Journal of
Rysman, M. (2004). Competition between networks: a study of the market for yellow pages. Review of
Economic Studies 71(April), 483–512.
Shapley, L.S. (1961). Values of Large Games V: An 18-Person Market Game. The Rand Corporation,
RM-2860, Santa Monica, CA, November.
Shapley, L.S. (1962). Complements and substitutes in the optimal assignment problem. Naval Research
Logistics Quarterly 9, 45–48.
Shapley, L.S., M. Shubik (1972). The assignment game I: the Core. International Journal of Game Theory
1(2), 111–130.
200 D.F. Spulber
Shevchenko, A. (2004). Middlemen. International Economic Review 45(1), 1–24.

Shimer, R., L. Smith (2000). Assortative matching and search. Econometrica 68(2), 343–369.
Shubik, M.S. (1982). Game Theory in the Social Sciences, Concepts and Solutions. MIT Press, Cam-
bridge, MA.
Shubik, M.S. (1984). A Game-Theory Approach to Political Economy. MIT Press, Cambridge, MA.
Sidak, J.G., D.F. Spulber (1996). Protecting Competition from the Postal Monopoly. American Enter-
prise Institute, Washington, DC.
Solomonoff, R., A. Rapoport (1951). Connectivity of random nets. Bulletin of Mathematical Biophysics
13, 107–117.
Spulber, D.F. (1988). Bargaining and regulation with asymmetric information about demand and sup-
ply. Journal of Economic Theory 44(April), 251–268.
Spulber, D.F. (1996a). Market microstructure and intermediation. Journal of Economic Perspectives
10(Summer), 135–152.
Spulber, D.F. (1996b). Market making by price-setting firms. Review of Economic Studies 63, 559–580.
Spulber, D.F. (1998). The Market Makers: How Leading Companies Create and Win Markets. McGraw
Hill/Business Week Books, New York.
Spulber, D.F. (1999). Market Microstructure: Intermediaries and the Theory of the Firm. Cambridge
University Press, New York.
Spulber, D.F. (2002a). Market microstructure and incentives to invest. Journal of Political Economy
110(April), 352–381.
Spulber, D.F. (2002b). Transaction innovation and the role of the firm, In: M.R. Baye (ed.), The
Economics of the Internet and E-commerce, Advances in Applied Micro-Economics, Vol. 11, JAI
Press/Elsevier Science, 159–190.
Spulber, D.F. (2003). The intermediation theory of the firm: integrating economic and management
approaches to strategy. Managerial and Decision Economics 24, 253–266.
Spulber, D.F., C.S. Yoo (2005). Network regulation: the many faces of access. Northwestern University,
Working Paper.
Stahl, D.O. (1988). Bertrand competition for inputs and Walrasian outcomes. American Economic
Review 78, 189–201.
Townsend, R.M. (1978). Intermediation with costly bilateral exchange. Review of Economic Studies 55,
417–425.
Tutte, W.T. (2001). Graph Theory. Cambridge University Press, Cambridge.
United States Census Bureau (2000). 1997 Economic Census, Wholesale Trade, Geographic Area Series.
U.S. Department of Commerce, Washington, DC, March.
Vickrey, W. (1961). Counterspeculation, auctions and competitive sealed tenders. Journal of Finance 16,
8–37.
von Neumann, J., O. Morgenstern (1944). Theory of Games and Economic Behavior. Princeton Uni-
versity Press, Princeton, NJ.
Watts, D.J. (2003). Six Degrees: The Science of a Connected Age. Norton, New York.
Weill, P.-O. (2005). Leaning against the wind. Department of Finance, University Stern School of
Business, New York, Working Paper.
Yavas, A. (1992). Marketmakers versus matchmakers. Journal of Financial Intermediation 2(March), 33–
58.
Yavas, A. (1994a). Middlemen in bilateral search markets. Journal of Labor Economics 12(July), 406–
429.
Yavas, A. (1994b). Economics of brokerage: an overview. Journal of Real Estate Literature 2(July), 169–
195.
Yavas, A. (1996a). Search and trading in intermediated markets. Journal of Economics and Management
Strategy 5, 195–216.
Yavas, A. (1996b). Matching of buyers and sellers by brokers: a comparison of alternative commission
structures. Real Estate Economics 24, 97–112.
Chapter 4
Organization Structure
Thomas Marschak
Haas School of Business, University of California, Berkeley
Abstract
This chapter concerns formal models of organizations that regularly acquire

information about a changing environment in order to find actions which are
appropriate to the new environment. Some or all members of the organi-
zation are specialists. Each of them privately learns something about a par-
ticular aspect of the new environment. The organization operates a
mechanism, which assembles relevant pieces of the specialists’ private obser-
vations and uses the assembled information to obtain the desired new action.
The mechanism has various informational costs and they are measured in a
precise way. The research seeks to characterize mechanisms that strike an
appropriate balance between informational cost and the performance of the
mechanism’s final actions. As costs drop, due to improved information tech-
nology, the properties of good mechanisms, and hence the structure of the
organizations that adopt them, may change. The chapter starts by examining
research in which the organization’s members reliably follow the mecha-
nism’s rules and so incentives are not an issue. It then turns to research in
which each member is self-interested and needs an inducement in order to
make the informational efforts that the mechanism requires. A number of
unmet research challenges are identified.
1 Introduction
This chapter concerns organizations that acquire information about a

changing environment in order to take appropriate actions.
The term ‘‘organization’’ is used in many disciplines and is applied to
many groups of persons. The term covers government agencies, markets,
entire economies, firms, nonprofit institutions, the users of the Internet, the
201
202 T. Marschak
firms in a supply chain, and so on. We shall be concerned with models of

organizations where some or all members are specialists: each of them
privately learns something about a particular aspect of the organization’s
newly changed environment. The organization then adjusts its actions. It
seeks new actions that are appropriate to the new environment. It could find
them by ‘‘direct revelation’’ (DR)—collecting in one central place all the
information about the changed environment that the members have pri-
vately collected, and using it to obtain a new action. But that would be
wasteful: much of the transmitted information would not be needed. In-
stead, the organization balances the benefit of appropriate actions against
the costs of learning about the current environment, transmitting some of
what has been learned, and using the transmitted information to choose
new actions. Advances in information technology (IT) may reduce those
costs, and that may change the structure of successful organizations. In all
the models discussed in this chapter, there is some precise measure of one or
more informational costs.
The members of our modeled organization might be totally self-inter-
ested, or they might behave like loyal and selfless team members (or well-
programmed robots), choosing their informational efforts and their actions
so as to contribute in the best way to a common goal. In either case, we
consider some appropriate measure of the organization’s ‘‘gross’’ perform-
ance, before informational costs are subtracted. For the self-interested
traders in a market, gross performance might be closeness to Pareto op-
timality or perhaps the sum of individual utilities. For a multidivisional firm
whose divisions are loyal and selfless, gross performance might be total
profit. That might also be our gross performance measure when each di-
vision pursues its own interests. We are concerned with the structure of
organizations whose gross performance is high given the informational
costs they incur.
The research discussed here approaches the problem in the style that is
usual in economic theory: a formal model, with well-defined concepts, is
studied, in order to find the conditions under which certain interesting
conjectures can be shown to hold. Typically, the conjectures arise from
certain empirical observations, which suggest loose claims. When it comes
to organizations and information, empirical work is hard and scarce, the
issues are complex, and the gap between a general empirical claim and a
tractable formal model might be very large. Consider, for example, the
following loose claim which appears to date back almost 50 years, to the
time when dramatic advances in IT first loomed on the horizon: As IT
advances, firms (and perhaps other organizations) will become ‘‘flatter’’:
middle management will fade away, and decisions will be made centrally, at
the top of a flat hierarchy. (That paraphrases the conjectures in Leavitt
and Whistler, 1958.) Or consider a much more recent and quite different
conjecture: As IT advances, firms will find it advantageous to adopt
decentralized modes of organization, in which unit managers are given wider
Ch. 4. Organization Structure 203
authority.1 Finally, consider a classic claim in economics, very much older

than the preceding two claims: For organizations whose task is to allocate
resources efficiently, the only informationally cheap allocation mechanisms
are those that use prices, since the ‘‘amount’’ of information transmitted in
such mechanisms is exactly what is needed to find an efficient allocation; in
any mechanism that achieves the same thing without prices, the information
transmitted exceeds what is needed.
The formal modeler who wants to explore these claims faces major chal-
lenges. Terms have to be defined precisely, while still capturing informal
usage. (How, for example, should we define ‘‘decentralized’’?) The informal
claim has to be stated in a precise manner. A way of showing that appro-
priate (and plausible) assumptions imply the formalized claim has to be
developed. When the modeler is finished, his demonstrated propositions
may be far from those that existing empirical work can support or refute.
Even so, the future empirical study of organizations might be usefully
guided by some of the modeler’s results.
In the formal models which we consider, the organization uses a mech-
anism to obtain appropriate new actions once its members have learned
about the organization’s newly changed environment. The mechanism uses
messages and it has informational costs, which are measured in a precise
way. We are particularly interested in informationally cheap mechanisms,
and we often take the viewpoint of a ‘‘designer’’ whose task is to construct
an informationally cheap mechanism that generates actions meeting the
organization’s goals. There is a vast literature on mechanisms, but only a
small part of it makes informational costs explicit. That is the part with
which we shall be concerned. We interpret the term ‘‘organization struc-
ture’’ as referring to the mechanism the organization uses, or to certain
properties of the mechanism. Some examples of structural questions: Are
the mechanism’s message flows arranged in a hierarchical pattern? Are the
actions it generates chosen by one central person or are the action choosers
dispersed? Is each external environment variable observed by a single per-
son, who is the sole specialist in that variable, or is each variable observed
by several persons?2
We discuss past research as well as new paths.3 A number of unmet
Research Challenges are identified. In Section 2, we consider organizations
whose members reliably follow the rules of the mechanism which the
1
That conjecture is consistent with the following statement in Bresnahan et al. (2002), which performs
a careful analysis of IT adoptions and organizational characteristics in a variety of industries: ‘‘IT use is
more likely to be effective in organizations with a higher quality of service output mix, decentralized
decision-making, and more skilled workers.’’ See also Bresnahan et al. (2000).
2
There is a literature which refers to the former case as the ‘‘U-form’’ (unitary form) and the latter
case as the ‘‘M-form’’ (multidivisional). See, for example, Harris and Raviv (2002) and Creti (2001).
3
Some new (unpublished) results are reported in Section 2.7 (on ‘‘speak-once-only mechanisms’’) and
in Section 3.3 (on networks of self-interested decision-makers, who bear the network’s informational
costs).
204 T. Marschak
designer has constructed, whatever those rules may be. The designer is not
concerned with incentive issues. In Section 3, the members become self-
interested and the designer has to take incentives into account. In Section 4,
we look very briefly at some formal models in which the primitive terms are
no longer ‘‘external environment’’ and ‘‘organizational action.’’ Section
5 offers a quick retrospective impression.
2 Goals, mechanisms, and informational costs: the ‘‘incentive-free’’ case,

where individuals obey the designer’s rules without inducement
Consider an organization composed of n persons. The organization con-

fronts a changing environment e ¼ (e1, y ,ek). The set of possible values of
each ej is denoted Ej. The set of possible values of e is denoted E and is the
Cartesian product E1 y Ek.4 In many settings we will have k ¼ n, and
ej will describe person j’s local environment, which he privately observes.
At any moment of time, the organization has in force an action a. The
action is a vector (a1, y ,am). The set of possible values of each a‘ is denoted
A‘. The set of possible values of a ¼ (a1, y ,am) is denoted A and is the
Cartesian product A1 y Am.5 In many settings, m will equal n and a‘
will be the action variable for which person ‘ has responsibility.
2.1 Two general frameworks for judging the organization’s actions
2.1.1 First framework: there is a performance function on A E

Suppose there is a real-valued function F on A E which measures the
organization’s performance, when the current environment is e and the ac-
tion in force is a. The organization, or its designer, wants F (a, e) to take
high values. If e never changes and is known to everyone, then the problem
reduces to that of finding, once and for all, an a in A which maximizes F for
the perfectly known and unchanging e. We may study that optimization
problem, and it may be challenging. But if we do so, we strip away the
organizational-design issue, and the question of how information about
e might be obtained and who might obtain it. Accordingly, we shall suppose
that e varies. We are then in a position to define the organization’s task. Its
task is to learn something about the current e, so that it can find an ap-
propriate action a. It seeks to strike a good balance between the costs
of learning about e and the value of F (e, a) that it achieves when it chooses
an action a that is best, given the information about e that it has collected.
4
There are settings in which one has to relax the requirement that E be the Cartesian product. For
example, each ej may be a parameter which identifies a production technology in a location j, i.e., it
identifies the input/output bundles that are feasible in that location. It may be that because of exter-
nalities between the locations, some k-tuples belonging to the Cartesian product E1 y Ek cannot
occur.
5
Again, in some settings one would want to relax the requirement that A be the Cartesian product.
(Here F has to have properties guaranteeing the existence of a best action.)

A mechanism is used by the organization to fulfill its task. We defer, until
Section 2.2, formal definitions of ‘‘mechanism,’’ and of the informational costs
associated with a mechanism.
Note that this framework covers the case where each component of e is a
signal, privately observed by some person, about a random variable x,
whose probability distribution is common knowledge. The true value of x,
which does not become known until after the action has been taken, de-
termines a payoff W (a, x) for every action a. Then we may let F (a, ē) be the
conditional expectation E (W(a, x)|e ¼ ē).
2.1.2 Second framework: for each environment e in E, there is a satisfactory

(goal-fulfilling) set of actions
Suppose now that for every e in E, there is a set of satisfactory actions (a
subset of A), denoted G (e), where G may be called the goal correspondence.
Any action in G (e) meets the organization’s goal for the environment e.
The organization, or its designer, asks: Is there a mechanism which gen-
erates a satisfactory action for every environment e in E, and what are the
informational requirements of such a mechanism?
In some settings, any action in G (e) has to meet certain constraints,
which are defined by e. In other settings, one starts by defining the per-
formance function F of the first framework. Then an action a belongs to
G(e) if and only if F(a, e) is within some given distance, say d, of its maximal
value. Thus for any fixed ē in E, we have G (ē) ¼ {a AA:|F (a, ē)F
(a, ē)|rd for all aAA}.
2.1.3 Three examples

First example: an organization that provides health services. Suppose each
person iA{1, y , n} is a diagnostician. Each week he examines a group of Ti
patients (who are examined by no one else). For the tth patient, he correctly
determines a diagnostic category, say dti. Thus the week’s local environment
for i is ei ¼ (di1, y , d iT i ). The organization’s action for the week is an
assigment of a treatment modality rti to each patient t inspected by diag-
nostician i. (We do not specify who chooses the modality.) The possible
modalities comprise the finite collection Q. So the organizational action
a belongs to the set
A ¼ fððr11 ; . . . ; r1T 1 Þ; . . . ; ðrn1 ; . . . ; rnT n ÞÞ : rit 2 Q for all ði; tÞg.
Let pe (a) denote the proportion of patients who, X weeks later, are found
to be ‘‘significantly improved’’ given that the observed diagnostic vector is e
and the chosen treatment vector is a. For every (e, a) in E A, the pro-
portion pe (a) is a random variable with a known probability distribution. In
our first framework we may have a performance function F, where F (a, e) is
206 T. Marschak
a weighted sum of (i) the total cost, denoted C (a), of the treatment vector a
and (ii) the expected value of pe (a).
In the second framework, we have a goal correspondence G, which might,
for example, assign to every e the set of actions for which cost does not
exceed an upper bound C, and at the same time the probability that the
proportion of significantly improved patients exceeds a specified lower
bound p is at least P (with 0oPo1). Thus,

GðeÞ ¼ a : CðaÞ C ; P probfpe ðaÞ p g .
Second example: a multidivisional firm. Now person i is the Manager of a
production facility which produces product i. The environment component
ei is the current week’s cost function for that product: it costs ei (q) to
produce the quantity q. The week’s cost function ei is known with certainty
as soon as the week starts. There are n action variables a1, y , an, where ai is
the current week’s quantity of product i. (We suppose that any positive
quantity is feasible.) The products are related, and the price at which the
quantity ai of product i is sold depends on the entire vector of quantities
a ¼ (a1, y , an). The price is denoted ri (a). For the function F, measuring
the week’s performance (profit), we have
Xn
Fða; eÞ ¼ ½ri ðaÞ ai ei ðai Þ.
i¼1
It may be informationally costly to gather the information about the

current e that is needed in order to find an F-maximizing a. If so, it may be
appropriate to consider the second framework. The goal correspondence G
might specify, for each e, the set of actions for which profit exceeds a lower
bound F. Thus
GðeÞ ¼ fa : F ða; eÞ F g.
Third example: an exchange economy. This organization’s n members are
consumers in an L-commodity economy with trade but no production.
Person i has a (privately observed) local environment ei ¼ (Ui, wi), where Ui
is a utility function defined on all vectors with L real nonnegative com-
ponents and wi ¼ (wi1, y , wLi) is the vector of person i’s (nonnegative)
initial endowments of the L commodities. An action a specifies trades. It is
an nL-tuple a ¼ ((a11, y ,aL1),y,(an1, y ,aLn)), where ai‘ may be positive (an
addition to i’s endowment of commodity ‘), negative (a deduction), or zero.
Let ai denote (ai1, y , aiL). An action a is feasible
P n for ei ¼ (e1, y , en) if
ai‘Zwi‘ for all ‘ and the trades balance, i.e., i ¼ 1 a ¼ 0. The set of
actions that are feasible for e is denoted Ae.
A possible performance function F is as follows:
X
n
FðeÞ is a value of a which maximizes U i ðwi þ ai Þ on Ae :
i¼1
That performance function would be appropriate if the designer of re-

source-allocating schemes is willing to compare utilities across individuals
and to take their sum as an appropriate standard.
On the other hand, it is more conventional to take Pareto optimality and
individual rationality as the goal. A trade vector a is individually rational for
e if
U i ðwi þ ai Þ U i ðwi Þ for all i. (1)
The trade vector aAAe is Pareto optimal for e if
8
< for all i and all ā in Ae ; the following holds :
>
‘‘U i ðāi þ wi Þ4U i ðai þ wi Þ’’ implies that for some j we have
>
:
‘‘U j ðāj þ wj ÞoU j ðaj þ wj Þ’’: ð2Þ
Note that for a given environment e ¼ ((U1, w1), y , (Un, wn)), conditions
(1) and (2) are restrictions on a alone.
To study the informational requirements of schemes that achieve this
goal, the second framework is needed since, for a given e, there may be
more than one trade a that is feasible (belongs to Ae), individually rational,
and Pareto optimal. (We would have to restrict the Ui in order to guarantee
uniqueness.) Formally, the goal correspondence is:

Gða; eÞ ¼ a 2 Ae : a satisfies ð1Þ and ð2Þ given e .
2.2 How the organization finds its current action when incentives are not an
issue
We now consider mechanisms which the organization may repeatedly use

in order to find its current action. A mechanism requires the transmission of
messages. In our first framework, the organization (or its designer) seeks a
mechanism which strikes a good balance between the performance measure
and the mechanism’s informational costs. In the second framework, it seeks
a mechanism which always yields goal-fulfilling actions and is, at the same
time, informationally cheap. We start with the assumption that once the
mechanism has been chosen, all members of the organization reliably follow
its rules. They may be thought of as robots. So we need not worry about
designing the mechanism so that each member will want to follow the
mechanism’s rules. Incentives are introduced in Section 3.
2.2.1 Decentralized many-step broadcast mechanisms which obtain the

organization’s action at the final step
Until we reach Section 2.7, where the designer decides who shall observe a
given environment variable, we will assume that our n-person organization’s
current environment e is a vector (e1, y , en), and the local environment ei is
208 T. Marschak
privately observed by person i. The possible values of ei comprise the set Ei,
and the possible values of e comprise the Cartesian product E ¼ E1 y
En. There is a set A of possible organizational actions. In our first
characterization of a mechanism, we shall suppose that it proceeds in a
sequence of steps. At each step each person i broadcasts or announces an
individual message to everyone. The vector of n individual messages is simply
called a message. Person i’s announcement at a given step (his part of the
broadcast message) is a function, denoted fi, of the preceding broadcast
message and of ei. But the variable ej, for any j6¼i, does not directly enter the
function fi. The privacy of every person is preserved. Others can only learn
about the current value of ei indirectly, through the broadcast message.
‘‘Informational decentralization’’ is an alternative term for privacy preser-
vation.6 Suppose that, for a given e, the message m̄ has the property that
once it is broadcast, the next broadcast message is again m̄. Then m̄ is called
an equilibrium message for e or a stationary message for e. When an equi-
librium message, say m, has been reached, the sequence stops and the
organization takes the action h (m). The function h: M-A is called the
outcome function.
Formally, let Mi be i’s individual message space, i.e., the set of individual
messages that person i is able to announce. Then the message space (the
set of possible messages) is M ¼ M1 y Mn. At step t, person i
broadcasts the message mit ¼ fi(mt1, ei), where mt denotes (mt1, y , mnt).
There is an initial message m0(e) ¼ (m01(e1), y ,mn0(en))AM. The
message m ¼ (m1 , y , mn) is an equilibrium message for e if, for all i,
we have
mni ¼ f i ¼ ðmn ; ei Þ. (3)
The quadruple /(M1, y ,Mn), (m01, y ,mn0), (f1, y ,fn),hS is an n-person

privacy-preserving (decentralized) broadcast mechanism on E, with action
set A, written in dynamic form. The term ‘‘broadcast’’ will often be omitted
but understood, until we reach Section 2.2.9, where individually addressed
messages are introduced.
In many studies one ignores the message-forming functions fi. Moreover,
one does not require that a message have n components, one for each person.
Instead it suffices, for the purposes of the study, to define, for each e and
each i, the set of messages mi (ei) (a subset of M) for which the equilibrium
condition (3) is satisfied. That set can be specified without taking the trouble
to specify the individual message spaces and the functions fi.
Then the individual-message-correspondence form of a decentralized (pri-
vacy-preserving) mechanism on the environment set E with action set A is a
6
In Section 3.2, the term ‘‘decentralized’’ is given another meaning, related to incentives: in a ‘‘de-
centralized’’ organization, each person is free to pursue her own self-interest. In particular, she makes a
self-interested choice as to whether or not to obey a proposed mechanism’s rules.
triple /M,(m1, y , mn),hS. Its elements are as follows.

M is the message space.
mi is a correspondence from Ei to M (i.e., mi is a function from Ei to the
subsets of M); the mi define a correspondence m, from E to M, namely
m(e) ¼ m1(e1)\ y \mn(en); a message m in m(e) is called an equilibrium
message for e.
m has the coverage property7, i.e., for all e in E, the set m(e) is not empty
(for every e there is at least one equilibrium message).
h (the outcome function) is a function from M to A.
One can give the broadcast mechanism /M,(m1, y , mn),hS an interpre-

tation that is sometimes called the verification scenario. In this scenario, we
imagine a central agent who broadcasts a sequence of trial message an-
nouncements. When a message m is announced, each person i responds by
saying YES if he finds, using his private knowledge of ei, that m belongs to
the set mi (ei), and he says NO otherwise. The announcements stop when
and only when the announcer has announced an m for which all n persons
say YES. The organization then takes the action h (m). The message m
lies in the set m(e) ¼ m1(e1)\ y \mn(en).
In a still further condensed formalism, one does not trouble to identify
the individual mi, but merely specifies the correspondence m. Then a mech-
anism is a triple /M,m,hS. The term ‘‘decentralized’’ or ‘‘privacy-preserv-
ing’’ is a restriction on m. It means that there exist correspondences
m1, y ,mn such that m(e) ¼ m1(e1)\ y \mn(en), even though we do not iden-
tify them. Thus it is understood, without being made explicit, that mAm(e)
means that person i has determined that m belongs to mi(ei), using his own
private knowledge of the current ei to do so.
Now suppose the mechanism designer is given a goal correspondence G
from E to A, as in our second framework. Then we say that the mechanism
/M, m, hS realizes G if
for every e in E; ‘‘m 2 mðeÞ’’ implies ‘‘hðmÞ 2 GðeÞ’’.

An important observation is that any goal correspondence can be realized by
a Direct Revelation (DR) mechanism. In a DR mechanism each person i
reveals his current ei, i.e., his announced message mi belongs to Ei. An
action in G (e) is taken once a complete description of the current e is
assembled. Formally, we have
M¼E
mi(ei) ¼ {ei}
h(m)AG(e).
7
In an alternative terminology, introduced by Hurwicz (1960) in an early and fundamental discussion
of mechanisms, the mechanism is called decisive on E if it has the coverage property.
210 T. Marschak
Fig. 1. A three-message two-person broadcast mechanism.
The only equilibrium message for e is e itself. A DR mechanism is, of

course, informationally costly. On the other hand, it has the merit that it
can reach an equilibrium message in just one step.8 Moreover DR mech-
anisms play a central role in the incentive literature. That literature often
confines itself to DR mechanisms, on the ground that any mechanism can
be rewritten as a DR mechanism (the ‘‘revelation principle’’). That may be a
correct claim, but it is not useful if one seeks mechanisms that are infor-
mationally cheap.
2.2.2 A three-message example wherein messages may be visualized as

rectangles that cover E
The organization has two persons. For person i, the set of local envi-
ronments is the real interval Ei ¼ [0, 1]. Consider the three-message mech-
anism portrayed in Fig. 1.
8
Define the initial message to be m0(e) ¼ (m01(e1), y ,mn0(en)) ¼ (e1, y ,en), and let person i’s message-
forming rule have the property that
f i ðm; ei Þ ¼ mi if and only if mi ¼ ei .

Then the initial message is already an equilibrium message. The message formed at Step 1 just repeats it.
At Step 1, the action is taken, and it belongs to G (e).
The message space is M ¼ {m1, m2, m3}, where mj identifies the rectangle
labeled mj. Then
8
> fm ; m g for 0 e1 o12;
< 1 3
1
m1 ðe1 Þ ¼ fm1 ; m2 ; m3 g for e1 ¼ 2;
>
: fm ; m g
2 for 1oe 1:
3 2 1
8
> fm ; m g for 0 e2 o34;
< 1 2
m2 ðe2 Þ ¼ fm1 ; m2 ; m3 g for e2 ¼ 34;
>
: fm g
3 for 34oe2 1:
Let the outcome function h be the following:
5 9 11
hðm1 Þ ¼ ; hðm2 Þ ¼ ; hðm3 Þ ¼ .
8 8 8
It is easy to verify that the mechanism /M, (m1, m2),hS so defined realizes
the following goal correspondence:

5
GðeÞ ¼ a : ja2ðe1 þ e2 Þj .
8
That is the case because h (mj) is the value of e1+e2 at the center of the
rectangle mj. Call the center (ej1, ej2). The largest value taken by the distance
jðej1 þ ej2 Þ ðe1 þ e2 Þj; over all e in the rectangle mj, occurs at the ‘‘north-
east’’ and ‘‘southwest’’ corners of the rectangle. At those corners the dis-
tance equals one quarter of the rectangle’s perimeter. All three rectangles
have the perimeter 5/2.
The above goal correspondence belongs to a class of goal correspond-
ences Gd, where d>0 and Gd (e) ¼ {a:|a(e1+e2)|rd}. To interpret this
class, go back to our first framework. Suppose that â(e) is the organiza-
tional action which uniquely maximizes a performance function F, defined
on A E, where the set A is the positive reals and (as in our example)
E ¼ [0, 1] [0, 1]. Suppose that it is very costly for the action taker to learn
the exact current value of e (as he would in a DR mechanism). Instead the
action taker only learns the rectangle in which e lies. Having learned that
the rectangle is mj, he takes the action h (mj). As a result there is an ‘‘er-
ror’’|â(e)h(mj)|at each e in the rectangle mj. It is straightforward to verify
that for any correct-action function â, no outcome function h achieves a
lower maximum error (with respect to â), on any given rectangle, than the
function which assigns to that rectangle an action that is midway between
the minimum of â on the rectangle and the maximum.9 So a mechanism
which minimizes error on each rectangle will use that outcome function,
and such a mechanism minimizes error on all of E.
9
That statement holds as well if the rectangle is a ‘‘generalized’’ one, i.e., it is the Cartesian product of
its e1-projection and its e2-projection and may consist of disjoint pieces.
212 T. Marschak
In the Fig. 1 example, â equals e1+e2 and our ‘‘midpoint’’ action is the
value of e1+e2 at the rectangle’s center. The maximum error of our mech-
anism (relative to the true e1+e2) is 5/8. It is natural to ask: Is our value of
d, namely d ¼ 5/8, the smallest d such that Gd can be realized by a three-
rectangle mechanism? Is there some other three-message mechanism in
which the maximum error (relative to the true value of e1+e2) over all e in
E is less than 5/8? The answer turns out to be NO. The argument which
establishes that fact has not yet been generalized to the case of k messages.
We do not know, in general, the smallest maximum error (relative to the
true e1+e2) that is achievable by a k-message (k-rectangle) mechanism.10
2.2.3 The ‘‘rectangle’’ definition of a broadcast mechanism

The preceding example suggests that we can define a decentralized (pri-
vacy-preserving) mechanism by specifying a covering of E ¼ E1 ? En,
provided that the sets in the covering are generalized rectangles, i.e., each is
the Cartesian product of its n projections. Let S be such a covering of E. Its
typical element, denoted sm, is a generalized rectangle, i.e., it is a Cartesian
product, sm ¼ sm1 ? smn, where smi is a subset of Ei. The collection of
possible values of the index m is denoted M.
To complete our rectangle definition of mechanism, we only need an
outcome function h from M to A. Then a broadcast mechanism is defined
by the triple /M, S, hS, with S ¼ {sm}mAM. We obtain the /M,(m1, y ,
mn),hS specification from the /M, S, hS specification, by letting mi(e) be the
set {mAM: eiAsmi}. We obtain the /M, S, hS specification from the /
M,(m1, y , mn),hS specification by letting smi be the inverse of mi, i.e.,
smi ¼ {eiAEi : mAmi(ei)}.
The verification scenario provides one way to interpret a broadcast
mechanism that is specified in the /M, S, hS form. Imagine a central
announcer who displays successive rectangles sm to all n persons. Each
responds with YES if he finds that his privately observed ei lies in the
projection smi and NO otherwise. When the announcer has found a rec-
tangle sm to which all say YES, then he takes the action h(m).
10
For the general case, one first has to establish that nothing is lost by confining attention to rectangles
(such as the three in Fig. 1), which are ‘‘proper,’’ rather than being generalized rectangles (each the
Cartesian product of its two projections) consisting of disjoint pieces. We next have to argue that we lose
nothing by further confining our attention to proper rectangles of equal perimeter. (Recall that the
maximum error on a rectangle, relative to the true e1+e2, equals its quarter-perimeter.) That can be
shown for our three-rectangle case and certain other cases, but a general argument, for arbitrarily many
proper rectangles, is not available. Finally, we have to calculate the smallest obtainable perimeter when
we cover E (the unit square) with k proper rectangles of equal perimeter. For our case (k ¼ 3) that can
indeed be shown to be 5/2, as in Fig. 1. For general k, there is no known formula giving the smallest
obtainable common perimeter. There is a conjectured formula, and bounds on the distance between that
formula and the unknown true one have been obtained. See Alon and Kleitman (1986).
2.2.4 A broadcast mechanism in ‘‘agreement function’’ form

In the verification scenario, as we have described it so far, person i re-
sponds to an announced message mAM by inspecting his privately observed
ei and saying YES if he finds that m lies in the set mi(e). In a number of
settings it is useful to be more explicit about person i’s procedure, by spec-
ifying a function which he computes. Let gi be a function whose domain is
M Ei and whose range is a finite dimensional Euclidean space. Let person
i say YES to the message m if he finds that gi(m, ei) ¼ 0. We may call gi
person i’s agreement function. When gi(m, ei) ¼ 0, we may think of person
i’s YES announcement as ‘‘agreement with’’ the message m. A mechanism
in agreement-function form is a triple /M,(g1, y , gn),hS. We obtain the
/M,(m1, y , mn),hS form from the /M,(g1, y , gn),hS form by specifying
that mi (ei) is the set {mAM : gi(m, ei) ¼ 0}. We obtain the /M,(g1, y ,
gn),hS form from the /M,(m1, y , mn),hS form by choosing, for each i, any
function gi which takes the value zero if and only if mAmi(ei).
Now suppose that M is the Cartesian product of n individual message
spaces Mi, where each Mi is a linear vector space, so that subtracting one
value of mi from another is well defined. Suppose we have written the
mechanism in dynamic form, i.e., we have specified a message-forming rule
fi for each person i. Suppose we are interested in the action generated by the
mechanism when m is an equilibrium message for e. Then we can rewrite the
mechanism in agreement-function form. Let gi express i’s equilibrium con-
dition for the rule fi. That is to say, we define
gi ðm; eÞ ¼ f i ðm; ei Þ mi .
When (and only when) gi(m, ei) ¼ 0, person i’s response to the announced
message m ¼ (m1, y , mn) (in the dynamic version of the mechanism) is to
repeat his piece of that announcement. Thus, message m is an equilibrium
message for e (in the dynamic version) when and only when all n persons i
find that gi (m, e) ¼ 0.
2.2.5 A summary
We have identified several different ways of specifying a decentralized
(privacy-preserving) n-person broadcast mechanism on the environment set
E ¼ E1 ? En, which the n-person organization may use to obtain
actions in reponse to a new environment e in E. The alternative specifi-
cations are
A mechanism with individual messages in dynamic form. This is a triple
/(M1, y , Mn), ((m01, y , mn0)), (f1 y , fn),hS. Here mi0 is a function
from Ei to Mi and mi0 (ei) is i’s initial individual message when i’s local
environment is ei. In each sequence of announced messages, person i
forms his next individual message by using the message-forming rule fi.
A mechanism with individual message correspondences. This is a triple
/M,(m1, y , mn),hS, where mi is a correspondence from Ei to M.
214 T. Marschak
A mechanism in which only an equilibrium message correspondence is

identified. This is a triple /M, m, hS where m is a correspondence from
E to M, and it is understood that there are (unspecified) individual
message correspondences m1, y , mn such that for all e we have
m(e) ¼ m1(e1)\ y \mn(en).
A mechanism in rectangle form. This is a triple /M,{sm}mAM,hS
where each sm is a generalized rectangle in E.
A mechanism in ‘‘agreement function’’ form. This is a triple
/M,(g1, y , gn),hS, where gi is a function from M Ei to a (finite-
dimensional) Euclidean space.
2.2.6 An example: a price mechanism for an exchange economy

Return now to the n-person L-commodity exchange economy discussed
in Section 2.1.3. In the classic (Walrasian) mechanism for obtaining indi-
vidually rational and Pareto optimal allocations, the typical message,
broadcast to all n persons, consists of
a nonnegative price vector p ¼ (p2, y , pL), with the price of com-
modity 1 (the numeraire) being one, and
a proposed trade vector ((a11, y , aL1),y,(an1, y , aLn)), whose compo-
nents may be positive, negative, or zero.
PThe proposed trades specified in any message m have the property that
n i
i¼1 a ¼ 0 (where ai ¼ (ai1, y , aLi)), or equivalently, for every i and every
commodity ‘:
X j
ai‘ ¼ a‘ . (y)
jai
The prices and proposed trades in m have the further property that each
person i’s budget balances,11 i.e.,
X
L
p‘ ai‘ ¼ ai1 . (yy)
‘¼2
Let us write the mechanism in agreement-function form. For person i, the

local environment is ei ¼ (Ui, wi). Assume Ui to be differentiable and in-
creasing in each of its arguments. The agreement function gi has L1
components, denoted gi2, y , giL. Each corresponds to one of the com-
modities 2, y , L. All of person i’s functions gi‘ equal an arbitrary nonzero
number if i’s proposed trade vector is infeasible with regard to some com-
¯ i.e., if ai¯ o wi¯ (where, as (yy) specifies, ai1 ¼ S‘L¼ 2 p‘ a‘i). For
modity ‘; ‘ ‘
the feasible case, where ai‘Zwi‘ for all ‘, consider the bundle that person
11
That is to say, the amount that i spends on positive additions to his endowment must equal the value
of the quantities that he subtracts from his endowment and makes available to others.
i holds after the proposed trades have taken place, and consider the ratio of
person i’s marginal utility for commodity ‘ (at that post-trade bundle) to his
marginal utility for commodity 1. The function gi‘ equals that ratio minus
the price of ‘. For an equilibrium message, each function gi‘ equals zero.
That implies that for the prices p, person i’s bundle (wi1+ai1, y ,wli+aLi)
satisfies the first-order condition for utility maximization P L subject ito the
constraint that the bundle’s value not exceed ‘ ¼ 1 p‘ w‘ . Let
U‘i(x1, y ,xL;ei) denote i’s marginal utility for commodity ‘ (the partial
derivative of Ui with respect to x‘) when he consumes the bundle (x1, y ,
xL) if his utility function is the function Ui specified in ei.
Using (y), (yy), we have the following for a message m ¼ (a, p) such that
a‘iZw‘i for all ‘: for ‘ ¼ 2, y , L,
P
U i‘ wi1 2 L‘¼2 p‘ ai‘ ; wi2 þ ai2 ; . . . ; wiL þ aiL ; ei
gi‘ ðða; pÞ; ei Þ ¼ P p‘ .
U i1 wi1 2 L‘¼2 p‘ ai‘ ; wi2 þ ai2 ; . . . ; wiL þ aiL ; ei
In view of condition (y), we can reduce the size of m, by deleting the

proposed-trade vector of one person, say person n. In person n’s agreement
rule we replace each a‘n (where ‘ ¼ 2, y , L), with Sn1 i
j¼1 a‘ : We replace the
commodity-1 term S‘¼2 p‘ a‘ ; with the term S‘¼2 p‘ ½Sn1
L n L i
j¼1 a‘ : Then the
message m is a vector of n(L1) real message variables, namely (n1)(L1)
trade variables plus (L1) prices.
To complete our definition of the mechanism, we have to provide the
outcome function h. We let that function be a simple projection operator,
i.e., h(a, p) ¼ a. If we now assume that each utility function Ui is strictly
concave, then the mechanism has the coverage property: for every eAE,
there exists an equilibrium message (a, p). Moreover the allocation a is
feasible, Pareto optimal, and individually rational.
Now consider any other mechanism whose equilbrium actions (trades) are
also individually rational and Pareto optimal. Under what further restric-
tions on the rival mechanism can we claim that its message space cannot be
‘‘smaller’’ than that of the mechanism we have just constructed? In partic-
ular, if the rival mechanism’s messages are again real vectors, when can we
claim that those vectors cannot have fewer than (n) (L1) components?
That is a well-studied question. We shall return to it in Section 2.2.8.
2.2.7 Another example: a price mechanism for a firm with managers and a
resource allocator12
In this organization, persons 1, y , n1 are Managers and person n is an
allocator. Manager j is in charge of nj activities. An activity uses resources and
it generates profit. There are L resources and the quantity e‘n of
12
This example is discussed in Ishikida and Marschak (1996).
216 T. Marschak
resource ‘ is available to the firm. Manager j operates each of his nj activities,

say activity k, at some level xki (a nonnegative real number). We assume nj>L,
j ¼ 1, y , n1. When Manager j’s activity-level vector is xj ¼ (xj1, y , xnjj ),
he contributes ej0(xj) to the organization’s profit and he uses the quantity
e‘j(xj) of resource ‘, for ‘ ¼ 1, y , L, provided that the Allocator gives him
those resource quantities. In each period there are new resource availabilities
e‘n, and these become known to the Allocator. For each Manager j, there are
also new profit functions ej0 and new resource-requirement functions ej‘, and
those become known to Manager j.
So the Allocator’s local environment is an L-tuple of resource availabilities,
namely en ¼ (en1, y , eLn), and Manager j’s local environment is an (L+1)-
tuple of functions, namely ej ¼ (ej0,ej1, y , eLj). The organization’s action is an
allocation vector y ¼ ((y1, y , yL1),y,(yn1, y , yLn1)), where yj‘ is the
quantity of resource ‘ allocated to Manager j. The action y meets the or-
ganization’s goal if it permits the managers to choose activity-level vectors
that maximize the firm’s total profit subject to the current resource avail-
abilities. So the goal correspondence is G, defined by

GðeÞ ¼ y : some nonnegative ðx1 ; ; xn1 Þ
satisfies ej‘ ðxj Þ ¼ yj‘ ; j ¼ 1; . . . ; n 1; ‘ ¼ 1; . . . ; L,
X
n1
and maximizes ej0 ðxj Þ
j¼1
)
X
n1
subject to ej‘ ðxj Þ en‘ ; ‘ ¼ 1; . . . ; L .
j¼1
Now suppose that the sets of possible local environments are as follows. For
each Manager j
E j ¼ fej : ej0 is strictly concave and differentiable;
ej‘ is convex and differentiable; ‘ ¼ 1; . . . ; Lg;
þ
while for the Allocator we have E n ¼ RL : Assume that for every e in E ¼ E1
? En, the set G(e) is nonempty. Then we can construct a mechanism /
M, m, hS which uses prices and realizes G on E. To do so, consider a vector p
of nonnegative resource prices, (p1, y , pL), and, for each Manager j consider
the following local problem:
X
L
j
find x so as to maximize ej0 ðxj Þ p‘ ej‘ ðxj Þ subject to xj 0:
‘¼1
Let Sj (ej, p) denote the set of solutions to that problem.

We use the individual-message-correspondence form þ
/M,(m1, y , mn),hS
to define our mechanism. The message space M is RnL : The typical message,
broadcast to everyone, is a pair m ¼ (p, y), where p is a price vector and y is

a proposed allocation vector. For Manager j ¼ 1, y , n1, define
mj ðej Þ ¼ fðp; yÞ 2 M : for some xj in Sj ðej ; pÞ and for all ‘ ¼ 1; . . . ; L,
we have ð1Þ ej‘ ðxj Þ yj‘ and ð2Þ p‘ ðyj‘ ej‘ Þ ¼ 0g.
For the Allocator, define
(
X
n21
mn ðen Þ ¼ ðp; yÞ 2 M : yj‘ en‘
j¼1
! )
X
n21
and p‘ en‘ yj‘ ¼ 0; ‘ ¼ 1; . . . ; L .
j¼1
Finally, the outcome function is a projection operator (just as it was in the

exchange-economy price mechanism): h (p, y) ¼ y. It is quickly verified that
if (p, y) is an equilibrium message for e (i.e., (p, y)Am1(e1)\ y \mn(e1)), then
y, together with some activity-level vector (x1, y , xn1), satisfies the first
requirement in our definition of G. Moreover that activity-level vector sat-
isfies the Kuhn–Tucker conditions associated with the maximization de-
scribed in the second requirement of our definition of G. The equilibrium
vector p is the vector of Lagrange multipliers in those Kuhn–Tucker con-
ditions. So we have yAG(e). Under our assumptions on E, a Kuhn–Tucker
solution exists for every e in E. That means that for every e in E, there exists
an equilibrium (p, y). So our mechanism has the coverage property.
We conclude that our price mechanism indeed realizes our goal corre-
spondence G on the environment set E. Its message space has dimension nL.
It is natural to ask: Is there another mechanism (with appopriate regularity
properties) which also realizes G on E and does so with a message space of
dimension less than nL? The next section deals with questions of that sort.
2.2.8 Using the ‘‘uniqueness property,’’ or ‘‘fooling sets,’’ to obtain a useful

lower bound to the message-space size of a ‘‘smooth’’ goal-realizing broadcast
mechanism
The best-known cost measure for a broadcast mechanism is the size of its
message space M. Suppose each broadcast message m is a vector of QZ1
real numbers m1, y , mQ, and M is the cube {m: AqrmqrBq, q ¼ 1, y ,
Q}, where AqoBq for all q. Then the natural size measure for M is its
dimension, namely Q. More generally, the dimension of M is a suitable cost
measure as long as M is any subset of RQ for which ‘‘dimension’’ is defined.
(For example, M may be a differentiable manifold.) Even more general
classes of message space have been studied, and message-space size meas-
ures have been defined for those classes as well.
Whatever the definition, we typically seek to identify the broadcast
mechanisms that realize a given goal correspondence G while using the
218 T. Marschak
smallest possible message space. When we do so, however, smuggling of

many numbers into a single number is a basic difficulty. If, for example, we
start with a mechanism in which each message is a vector of Q>1 real
numbers, then we may define a new mechanism, which realizes the same
goal correspondence as the original mechanism, but has messages com-
prised of just a single real number. One way to do this is to let the single real
number be a decimal, which encodes the qth of the Q original numbers as a
string composed of our decimal’s qth digit, its (Q+q)th digit, its (2Q+q)th
digit, and so on. That particular smuggling trick is ruled out if, when
writing the mechanism in the message-correspondence form, we require that
every set mi(ei) contain an element ti(ei), where ti is a continuous function. A
more elaborate smuggling trick uses the Peano ‘‘space-filling-curve’’ map-
ping. (See, for example, Apostol, 1957, pp. 396–398). That mapping allows
us to recover the Q numbers from a single number in a continuous manner.
A requirement stronger than continuity (e.g., differentiability) is therefore
needed in order to exclude it. If we want to give a nontrivial meaning to the
message-space minimality of a particular mechanism within some interest-
ing class of mechanisms, then smoothness requirements on the candidate
mechanisms are unavoidable.
Now consider the case of a two-person organization. We are given a goal
correspondence G from E ¼ E1 E2 to an action set A. Suppose we suspect
that no suitably smooth mechanism can realize G with a message space of
dimension less than D. Suppose further that E is a subset of a Euclidean
space, and that we have found Ē, a subset of E—called a test class of
environments—which has dimension D. Suppose further that G has the
uniqueness property on Ē. That means that there is no action which is goal-
fulfilling for all four corners of the ‘‘cube’’ defined by a pair of distinct
environments e, e in Ē, i.e., the four environments (e1 , e2 ), (e
1 , e2 ), (e1 ,

e2 ) and (e1 , e2 ). Thus for any e , e in Ē, the following holds:
If Gðen1 ; en2 Þ \ Gðenn nn n nn nn n n nn

1 ; e2 Þ \ Gðe1 ; e2 Þ \ Gðe1 ; e2 Þa+; then e ¼ e .
Now note that if a mechanism realizes G on all of E, then, in particular, it

realizes G on the test class Ē Next, recall that a mechanism is privacy
preserving: in determing whether or not to agree to a broadcast message m
(or determining whether or not m lies in mi(ei)), each person i looks only at
his own local environment ei. Thus if a message m̄ is an equilibrium message
for both (e1 , e2 ) and (e
1 , e2 ), then m̄ is also an equilibrium message for

(e1 , e2 ) and (e1 , e2 ). So if the mechanism’s action h(m̄) is indeed going to
lie in the sets G(e) and G(e), as realization of G requires, then h(m̄) lies in
G(e1 , e
2 ) and G(e1 , e2 ) as well. Since G has the uniqueness property on Ē,
it follows that e ¼ e. So a pair of distinct environments in the D-dimen-

sional test class Ē cannot share the same equilibrium message. Since M must
contain at least one equilibrium message for all e in Ē (the coverage re-
quirement), we can claim, informally, that M must be at least as large as Ē.
More precisely, consider any mechanism which realizes G on E and hence on

Ē. Consider the restriction of m to Ē and call that restriction m: ~ So m~ is a
correspondence from Ē to a subset of M, namely the set mð ~ ĒÞ ¼ mðĒÞ: Now
since Ē has the uniqueness property for G, the inverse of the correspondence m~
is a function, which we may call t : mð ~ ĒÞ ! Ē: For example t might be the
Peano mapping, which assigns a member of Ē to every message in mð ~ ĒÞ; even
though (Ē) has higher dimension than mð ~ ĒÞ: A message m in mð ~ ĒÞ is an
equilibrium message for the environment t(m)AĒ and for no other environ-
ment in Ē. Since every environment in Ē must have some equilibrium message
in m(Ē) (by the coverage property), our function t is onto (it is a surjection).
Now suppose our smoothness requirement on the candidate mechanisms is
that the function t, the inverse of m;~ be differentiable. If M, and hence mðĒÞ ¼
~ ĒÞ; had a smaller dimension than Ē, then t would be a one-to-one function
mð
from one set onto a second set having higher dimension. That cannot be the
case if t is differentiable. (For example, the Peano mapping, while continuous,
is not differentiable, so a mechanism in which t is the Peano mapping from
~ ĒÞ onto the higher-dimensional set Ē violates our smoothness requirement.)
mð
So we have confirmed our suspicion that D is indeed a lower bound for smooth
mechnisms which realize G on all of E. We obtain the same conclusion for
other smoothness requirements. Some of them are weaker than our require-
ment that the function t (the inverse of the message correspondence, restricted
to the test class) be differentiable.13 Moreover, there is another way to force M
to have a dimension at least as large as Ē. We can impose requirements directly
on the correspondence m~ rather than on its inverse. Suppose we require m~ to be
locally threaded. That means that for any neighborhood N in Ē, we can find a
continuous function u: N-M, which is a selection from m; ~ i.e., vðeÞ 2 mðeÞ
~ for
all e in N. The uniqueness property of Ē tells us that for any two distinct
environments ē, ē¯ in N, we have vðēÞavðē¯ Þ: It can be shown that this fact rules
out a continuous u if M indeed has a smaller dimension than Ē.14
13
In particular, we may use the weaker requirement that t be ‘‘Lipschitz-continuous’’, i.e., there exists
K>0 such that for all m0 , m00 in m̄ðEÞ; we have||t(m0 )t(m00 )||rK ||m0 m00 ||. (Here the symbol||x||, for
x ¼ (x1, y , xl), denotes max {|xj|: jA{1, y ,l}}.) The Peano mapping is not Lipschitz-continuous.
14
We may want to consider mechanisms whose message space M is not Euclidean but consists, for
example, of certain infinite sequences, or of integer k-tuples, or of preference orderings (as when an
environment specifies agents’ individual preference orderings and a message identifies a set of possible
environments). That has motivated the study of mechanisms with message spaces that are general
topological spaces. Then, instead of comparing dimensions, we use a general topological definition of
the statement that one message space is ‘‘at least as large as’’ another. If M is Euclidean, then ‘‘at least as
large as’’ reduces to ‘‘having a dimension no smaller than.’’ For example, one may define M to be at
least as large as M if and only if there is a subspace of M which is homeomorphic to M. For each
such topological definition, we seek an associated smoothness requirement on the message correspond-
ence m used by a G-realizing mechanism, so that a smooth mechanism’s message space is at least as large
as a test class Ē having the uniqueness property for G. One such requirement is ‘‘spot-threadedness’’ of m
on the test class. That is a weaker requirement than local threadedness. It means that there is an open set
WDĒ and a function q : W-M such that q(e)Am(e) for all e in W. If a G-realizing mechanism obeys
that condition, while its message space and the test class Ē are both Hausdorff and Ē is locally compact,
then the message space must be at least as large as the test class, where ‘‘at least as large as’’ has the
meaning just given. The details are carefully developed in Section 2.3 of Hurwicz (1986), which concerns
220 T. Marschak
The technique extends to n-person mechanisms. Let the symbol e/ei

denote the vector obtained from e(e
1 , y ,en ) when we replace ei

with

ei . Then the statement ‘‘the correspondence G from E ¼ E1 ? En to
an action set A has the uniqueness property on the test class Ē
E’’ means
that

if en ; enn 2 Ē and Gðen Þ \ \ni¼1 Gðenn =eni Þ a+; then en ¼ enn .
It is interesting to note that the idea just sketched was developed

independently by computer scientists and economists.15 In the computer
science field known as ‘‘communication complexity,’’16 one studies dia-
logues between n persons that end with one person having enough infor-
mation to compute a function F of n numbers, each of them privately
known by one of the n persons. The dialogue is a sequence of binary strings.
The dialogue changes when the privately known numbers change. One
wants the worst-case dialogue to be as short as possible. If the function
possesses a ‘‘fooling set,’’ then the size of the fooling set provides a lower
bound to the length of the worst-case dialogue. In the terminology we
have just developed, a fooling set is a set on which F has the uniqueness
property.
An illustration of the uniqueness technique: resource allocating mechanisms
for a class of exchange economies. Let us return to the n-person L-com-
modity exchange economies discussed in Sections 2.1.3 and 2.2.6. Recall
that person i’s local environment is a pair ei ¼ (Ui, wi), where Ui is a utility
function and wi ¼ (wi1, y , wLi) is an endowment vector. As before, there is
a set Ei of possible local environments ei. Recall also that the action
the economy chooses is a trade nL-tuple a ¼ (a1, y , an), where
ai ¼ (ai1, y , aLi). Call the action a interior for e if ai‘>wi‘ for all (i, ‘).
(footnote continued)
‘‘a strategy for obtaining minimality results.’’ See also (among others) Mount and Reiter (1974), Walker
(1977), and Hurwicz and Marschak (1985).
15
Economic applications of the technique, so as to establish lower bounds to the message-space size
required to achieve various resource-allocation goals, include the following papers, in each of which the
lower bound is shown to be attainable by a particular mechanism that uses prices: Hurwicz (1997),
Mount and Reiter (1977), Osana (1978), Sato (1981), Jordan (1982) (which shows that the only mech-
anisms that have minimal message space while realizing Pareto-optimality are versions of the compet-
itive mechanism); Chander (1983), Aizpura and Manresa (1995), Calsamiglia and Kirman (1998), Tian
(2004), Stoenescu (2004), and Osana (2005). On the other hand, the following papers find that real-
ization of the resource-allocation goal requires a message space of infinite dimension: Calsamiglia (1977)
(which permits increasing returns in production), Hurwicz and Weinberger (1990), Manresa (1993),
Kaganovitch (2000) (which consider efficient intertemporal resource allocation), and Jordan and Xu
(1999) on expected profit maximation by the managers in a firm.
16
See, for example, Lovász (1990), Karchmer (1989), and Kushilevitz and Nisan (1997).
Slightly modify the goal correspondence defined in Section 2.1.3 so that it is

now defined by
Xn
GðeÞ ¼ a: i¼1
ai ¼ 0; for the economy defined by e; a is interior,

Pareto-optimal and individually rational .
The price mechanism introduced in Section 2.2.6 has a message space of

dimension n (L – 1) and realizes G on certain environment sets E ¼ E1
? En. Can we show n(L1) to be a lower bound for any suitably
smooth mechanism which realizes G on E? Yes we can, using the uniqueness
technique, provided E contains a suitable test class Ē whose dimension is
n(L1). In the test class most used in the literature,17 utility functions have the
Cobb–Douglas form and endowments are fixed. Specifically, if e is in Ē, then
person i’s endowment
Q L vector wi is (1, 1, y , 1), and his utility for the bundle
ai
(X1, y , XL) is ‘ ¼ 1X‘ , where the ai are positive numbers such that
a1+ ? +aL ¼ 1. Thus each ei is uniquely determined by L1 real numbers
and so the set Ēi has dimension L1. The dimension of Ē is n(L1), which
equals the dimension of our price mechanism’s message space.
The uniqueness property of G on the set Ē is readily shown, using the first-
order conditions that characterize the unique n interior trade vectors ai that
maximize each person i’s Cobb–Douglas utility subject to the balancing con-
straint Sai ¼ 0. One then has a choice of several smoothness conditions to be
imposed on the candidate G-realizing mechanisms. If the set of messages that
are equilibrium messages for the environments in Ē has dimension less than
n(L1), then each of these smoothness conditions rules out a one-to-one
mapping from that set of equilibrium messages onto the n(L1)-dimensional
environment set Ē. But if the mechanism indeed realizes G on Ē, then such a
mapping must exist, by the uniqueness property of G on Ē. So if the entire set E
contains our test class Ē, we can rule out18 a smooth broadcast mechanism that
realizes G on all of E and has a message space of dimension less than n(L1).19
17
See, among others, Mount and Reiter (1977) and Jordan (1982).
18
For the organization considered in Section 2.2.7 (N1 Managers and an Allocator), a similar
argument (given in Ishikida and Marschak, 1996) establishes that no suitably smooth mechanism can
realize the goal function G (defined in Section 2.2.7) with a message-space dimension less than nL (the
message-space dimension of the G-realizing price mechanism that we constructed). In our nL-dimen-
ffiffiffiffiffiactivity for Manager j (who has n oL activities) uses only the resource k and
j
kth
sional test class, the q
earns a profit of 2ajk xjk when it is operated at level xkj Each member of the test class is defined by an
nL-tuple of positive numbers (the numbers akj), so the test class has dimension nL. It is straightforward
to show that G has the uniqueness property on that test class.
19
Another interesting setting for dimensionally minimal broadcast mechanisms is the allocation of one
or more objects among n persons, each of whom has a private valuation for each object. Consider the
case of a single object. Let ei be i’s valuation for the object and let Ei be the interval [0, H], where H>0.
Let the action set be 1, y , n, where ‘‘action i’’ is the allocation of the object to person i. Let the goal be
allocation to a maximal-value person, i.e., the goal correspondence G is defined by G (e) ¼ {j: ejZei} for
all i. Then one G-realizing broadcast mechanism uses messages m ¼ (t, J), where t is a real number and
J is an integer in 1, y , n. Person i agrees to m if and only if: (1) i6¼J and eirt or (2) i ¼ J and ei ¼ t. The
222 T. Marschak
2.2.9 Mechanisms in which messages are not broadcast but are individually
addressed, and responsibility for each action variable is assigned to a selected
person
We have assumed thus far that all parts of a given message are broadcast.
They are ‘‘seen’’ or ‘‘heard’’ by all of the organization’s n members. It may,
of course, be true that while an agreement function gi has the entire broad-
cast message m as an argument, the function gi is sensitive to only a portion
of m, namely the portion that i hears. Even though m is broadcast to
everyone, we can interpret ‘‘hearing only a portion of the broadcast mes-
sage m’’ as ignoring all of m except that portion. Formally, for every person
i, we may be able to write every message m as a pair (mi, mi), where mi is
the portion of m that i hears, and to write the agreement function gi(m, ei) as
gi(mi, ei). Similarly, we may be able to write the set mi(ei) as {m ¼ mi,
mi)AM : miAmi(ei)}, where mi is a correspondence from Ei to Mi and
Mi is the set of possible values of mi.
But if we want to permit messages to be individually addressed, and if we
want to study the cost born by person i as he hears and processes the messages
he receives, and responds to them by sending further messages to certain other
persons, then it is more convenient to extend our previous formalism by
introducing network mechanisms.20 That will also have another advantage: it
will allow us to be explicit about who is responsible for a given action variable.
Our mechanism concept thus far has been silent on this matter.
In defining a network mechanism we may again use the agreement-func-
tion form, but an agreement function’s domain and range are now different.
We start by letting M denote an n-by-n matrix of sets Mij, where Mij is the set
of possible messages that i may send to j. The set Mii on the diagonal of M
may be empty, or, if it is not, we may interpret the self-addressed messages in
Mii as stored information. Moreover, Mik may be empty for some k6¼i. That
means that i never sends a message to k. Next, let Mi denote the Cartesian
product of the sets in the ith row of M, i.e., Mi ¼ Mi1 ? Min. Let Mi
denote the Cartesian product of the sets in the ith column of M, i.e.,
Mi ¼ M1i ? Mni. Let Pi (M) denote the Cartesian product of the sets
(footnote continued)
outcome function h is a projection: h (t, J) ¼ J. This mechanism is one way to model a Dutch
(descending) auction. Consider a subclass of E ¼ E1 ? En, namely the ‘‘diagonal’’ class
Ē ¼ {eAE:e1 ¼ e2 ¼ y ¼ en}. It is easily seen that G has the uniqueness property on Ē. But Ē has
dimension one, so it does not provide a useful lower bound for mechanisms whose messages are real
vectors. The ‘‘auction’’ mechanism, however, uses both real numbers and integers. One has to be careful
in choosing a cost measure, and smoothness requirements, for mechanisms of that sort. Much more
challenging is the case of several objects, when each person has a valuation for each subset of the set of
objects, and each person may be allocated a subset. One seeks a mechanism which finds (at equilibrium)
an allocation that maximizes the sum of the valuations. Lower bounds for such mechanisms have been
developed by Nisan and Segal (2005). The mechanisms considered again use both real numbers and
integers in their messages. The uniqueness technique, using a counterpart of the ‘‘diagonal’’ test class,
plays a central role in that study. The ideas are extended to a much larger class of allocation problems in
Segal (2004).
20
They are studied in Marschak and Reichelstein (1995, 1998).
that are in the ith row or the jth column of M. Thus Pi(M) ¼ Mi {Mi/Mii}.
We shall say that a message muuAMuu is heard by i if it is received by
i (so that u6¼i, u ¼ i), sent by i (so that u ¼ i, u6¼i), or stored by i (so
that u ¼ u ¼ i). Then Pi (M) is the set of the possible message vectors that
i can hear.
We shall speak of a message array mAM. Its typical component is an
individually addressed message mijAMij, where mij is a vector of sij real
numbers; sij may be zero. The symbol Pi(m) will denote the portion of m
that i hears; Pi(m) is an element of the set Pi(M). Let the domain of person
i’s agreement function gi be the Cartesian product of Ei with the set n
Pi(M)
of possible message vectors that i can hear, and let its range be RSj¼1 sij : The
statement ‘‘gi (Pi(m), ei) ¼ 0,’’ means that person i finds the message array
m to be acceptable: given his current local environment ei, and given that he
has received the messages m1i, y , mni, he finds it appropriate to send
the messages mi1, y , min. The message array m is an equilibrium message
array for the environment e ¼ (e1, y , en) if all persons find it acceptable,
i.e., gi (Pi (m), ei) ¼ 0 for all persons i.21
To complete the definition of a network mechanism, we have to specify
how the action variables are chosen once an equilibrium message array is
found. Let the organization have k action variables, z1, y , zk; let Zj be the
set of possible values of zj, j ¼ 1, y , k; and let Z ¼ Z1 ? Zk be
the set of possible organizational action k-tuples z ¼ (z1, y , zk). Partition
the index set {1, y , k} into n sets (some of them may be empty), namely
J1, y , Jn. The (possibly empty) set Ji identifies the action variables for
which i is responsible. Those are the variables zr, where rAJi; they comprise
a vector zji belonging to the set ZJ i ¼ t2J i Z t : Person i chooses the value of
the action variables that are in his charge as a function of what he has
heard. So he uses an outcome function hi : Pi ðMÞ ! Z ji :
As before, we want the agreement functions to have the coverage prop-
erty: for every eAE, there exists a message array m which is an equilibrium
array for e. If coverage is satisfied, then a triple /M, (g1, y ,gn),
(h1, y ,hn)S, whose elements we have just defined, is a (privacy-preserving)
n-person network mechanism on the environment set E ¼ E1 ? En with
action space Z ¼ Z1 ? Zk. As before, we may be given a goal cor-
respondence G: E-Z, where the set G(e) consists of the organizational
actions z which meet a certain goal when the environment is e. As before, we
shall say that a given network mechanism realizes G if, for every e in E,
the organizational action (h1(Pi (m)), y , hn (Pi(m)) lies in the set G(e)
whenever m is an equilibrium message array for e. Note that every network
mechanism has a communication graph. Its nodes are the n persons, and
there is an edge between i and j if and only if at least one of the sets Mij, Mji
is nonempty.
21
We can also write a network mechanism in dynamic form, in message-correspondence form, or in
rectangle form, just as we can for broadcast mechanisms.
224 T. Marschak
An example: a network ‘‘price’’ mechanism for a three-plant four-person

firm. Consider a four-person firm. Person 4 markets two products. He ob-
tains revenue from the quantities Q1 and Q2 of the products, which are
produced, respectively, by Person 1 and Person 2. Person 4’s privately
observed local environment is a function, namely the revenue function
e4(Q1, Q2). Person 1’s local environment is the cost function e1(Q1). For
Person 2, cost depends not only on product quantity but also on the
quantity I of an intermediate material used in production (the material is
supplied by Person 3). So Person 2’s local environment is the cost function
e2(Q2, I). Person 3 produces the intermediate material; his local environ-
ment is the cost function e3(I). For Persons 1 and 3, the local-environment
set Ei (where i ¼ 1 or 3) is the set of all continuous convex functions from a
closed interval [Ai, Bi] (with 0rAioBi) to the positive real numbers. For
Person 2, the local-environment set E2 is the set of all continuous convex
functions of two variables, from a set [A2, B2] [C, D] (with 0rA2oB2,
0rCoD) to the positive real numbers. For Person 4, E4 is the set of all
continuous concave functions from a closed interval [A4, B4] (with
0rA4oB4) to the positive real numbers. The numbers Ai, Bi, C, D stay
the same for all the environments e ¼ (e1, y , en).
Now consider a network mechanism in which Person 4 sends prices u1
and u2 to Persons 1 and 2, respectively, and 1 and 2 reply with quantities
Q1,Q2 that they are willing to supply to 4 at those prices. Similarly, Person 2
sends an intermediate-material price u to 3, who replies with an interme-
diate-material quantity I that he is willing to supply to 2 at that price. So
there are six message variables: m14 ¼ Q1, m23 ¼ u, m24 ¼ Q2, m32 ¼ I,
m41 ¼ u1, m42 ¼ u2. Let each of the sets Mij of possible message-variable
values be the nonnegative real numbers. For all other i, j combinations let
the set Mij be empty. Ignore the empty sets and note that for the typical
message array, say m ¼ (Q̄1 ; Q̄2 ; Ī; ū1 ; ū2 ; ū), we have P4(m) ¼ (u1, u2, Q1,
Q2), P1(m) ¼ (u1,Q1), P2(m) ¼ (u2, Q2, u, I), and P3(m) ¼ (I, u).
Fig. 2 portrays the message flows in this six-message-variable mechanism.
The agreement rules of our price mechanism will express the usual con-
ditions for divisional profit maximization. Consider the typical message
array m̄ ¼ ðQ̄1 ; Q̄2 ; Ī; ū1 ; ū2 ; ūÞ: Person 4 agrees with m̄ (he finds that
g4(p1(m̄),e4) ¼ 0) if and only if his divisional profit e4(Q1, Q2)u1Q1u2Q2
is maximized by (Q̄1 ;Q̄2 ). He can determine whether or not that is so without
knowing two components of the array m̄, namely I¯ and ū: Person 1 agrees if
and only if the profit u1Q1e2(Q1) is maximized by Q̄1 ; and does not need to
know Ū2,Q̄2 ; I¯, or ū: Person 2 agrees if and only if u2e2(Q2, I) – uI is
maximized by (Q̄2 ; I¯), and does not need to know u% 1 or Q̄1 : Finally Person 3
agrees if and only if uIe3(I) is maximized by I¯, and does not need to know
ū1 ū2 Q̄1 , or Q̄2 :
The organization’s action variables are Q1, Q2, and I. (To minimize no-
tation, we use the same symbol for the action variable as for the message
variable associated with it.) For each action variable, we let the set of
Fig. 2. A four-person network mechanism with six individually addressed message vari-
ables.
possible values (one of the sets Zk in our general definition) be the non-
negative reals. We have many choices in designing the outcome function.
We may, for example, give Person 4 responsibility for the action variables
Q1, Q2, while Person 3 has responsibility for I, the remaining action var-
iable. Then we write
h4 ðP4 ðmÞÞ ¼ h4 ðQ1 ; Q2 ; u1 ; u2 Þ ¼ ðQ1 ; Q2 Þ; h3 ðP3 ðmÞÞ ¼ h3 ðI; uÞ ¼ I.
(The outcome function is simply a projection operator, just as it was in the

exchange-economy price mechanism that we considered before introducing
network mechanisms.)
Under our assumptions on the Ei, our agreement functions have the
coverage property. Moreover, if Q̄1 ; Q̄2 ; and I¯ are the actions of an equi-
librium message array, then they maximize the firm’s profit. That is to say,
our network mechanism realizes the following goal correspondence:
GðeÞ ¼ fQ̄1 ; Q̄2 ; ĪÞ : ðQ̄1 ; Q̄2 ; ĪÞ maximizes

e4 ðQ1 ; Q2 Þ e1 ðQ1 Þ e2 ðQ2 ; IÞ e3 ðIÞg:
Note that we may, if we wish, reverse the directions of the flows depicted
in Fig. 2. We may, for example, let Person 1 send a price to Person 4, who
replies with a quantity. Let the agreement rules stay as they were. Then the
set of equilibrium messages for any e does not change and hence the set of
226 T. Marschak
actions obtained at equilibrium for any given e does not change. The revised
mechanism again realizes G.
The costs of a network mechanism. One cost measure is simply the di-
mension of the entire message space M, the set of possible message Parrays.
Since each mij is a vector of sijZ0 real variables, we have dim M ¼ i,j sij. If
there are no self-addressed messages (i.e., the sets Mii are empty), and if we
think of each of the sij real message variables as requiring a ‘‘pipeline’’
between i and j, then the dimension of M is the total number of pipelines. In
the preceding example, there are no self-addressed messages and there are
six pipelines.
But it is also of considerable interest to study the n individual commu-
P burdens. Person i’s burden is the number of variables he hears,
nication
i.e, fj:M ji a0g sij : That is also the dimension of the set Pi (M), so we may
use the symbol dim Pi (M) for i’s burden. Note that if there are no self-
addressed messages, then dim M equals half the sum of the individual
burdens, since that sum counts each pipeline twice. In the example, sij is
either one or zero and the vector of individual burdens, for 1,2,3, and 4,
respectively, is (2, 4, 2, 4). It is natural to ask: Is there another network
mechanism which also realizes G, but does so with fewer than six message
variables, and with an individual-burden vector that dominates (2,4,2,4)—
i.e., one person’s burden is less than in our price mechanism and no per-
son’s burden is higher? Once again a smoothness requirement has to be
imposed on the candidate mechanisms, to avoid the smuggling of many
numbers into one.
We shall define one such smoothness requirement in a general way, for a
class of n-person network mechanisms where (as in our three-plant exam-
ple) (i) each person i’s environment is a real-valued valuation function ei
whose arguments are certain action variables and (ii) the mechanism re-
alizes a goal correpondence, in which the goal-fulfilling action variables
maximize the sum of the valuation functions. (That is the case in our three-
plant example if we define the valuation function for 1,2, and 3 to be the
negative of the cost function, while 4’s evaluation function is his revenue
function; then the firm’s profit is indeed the sum of the four valuation
functions.)
We start by considering the k action variables z1, y , zk. Let each set Zr—
the set of possible values of the action variable zr—be a closed real interval,
so that Z ¼ Z1 ? Zk is a closed cube in Rk : Next we shall say that the
action variable zr is one of person i’s concerns if it enters his function ei. A
given action variable may be the concern of several persons. (Thus, in our
example, Person 4’s concerns are the action variables Q1 and Q2; Person 2’s
concerns are Q2 and I.) Let Ai be the index set that identifies i’s concerns,
i.e., Ai ¼ {r A {1, y , k}: zr enters the function ei}. Then a given vector zAi
specifies a value for each of i’s concerns, and ZAi (a closed cube or closed
interval) denotes the set of possible values of zAi : We now fix the concern
sets Ai and we consider mechanisms and environment sets for which the
following is true:
8
> The local environment set for person i is E i ¼
<
Ai
fei : ei is a concave function from Z~ to Rþ g; where Z ~ Ai is an
>
:
open convex set that includes Z Ai :
(y)
Now consider the goal correspondence P defined by
X
n
PðeÞ ¼ fz 2 Z : z max imizes ei ðzAi Þ on Zg:
i¼1
Consider a subclass of the environment set E, namely Ē ¼ Ē1 ? Ēn,

where each concave valuation function eiAĒi takes a separable quadratic
form. That is to say, if eiAĒi, then
X 1 a a2

Ai a a
ei ðz Þ ¼ xi z yi ðz Þ ,
a2A
2
i
a
where xia ,yi are numbers privately observed by i. So we may identify each
local environment ei in Ēi by a point in R2jAi j : (For a finite set H, we let |H|
denote the number of elements it contains.) Moreover if eAĒ, then the goal-
fulfilling action
P set P(e) has a single element, since there is a unique max-
imizer of ni¼1 ei ðzAi Þ on the cube Z. We now let P(E) denote that unique
maximizer. Call an element e of Ē interior if that unique maximizer is an
interior point of Z. Let t (a) denote the set of persons concerned with the
action variable za, i.e., t (a) ¼ {i: aAAi}. Assume (to avoid trivial cases) that
each set t (a) has at least one member. Write the action k-tuple P(e) as
(P1(e), ? , Pk(e)). It is quickly checked that for an interior environment
ēAĒ, we have, for every aA{1, y , k}
P a
a i2tðaÞ xi
P ðeÞ ¼ P a.
i2tðaÞ yi
We are now ready to define our smoothness requirement.

Consider a network mechanism /M,(g1, y ,gn), (h1, y ,hn)S on the
environment set defined in (y). Let each message mij in a nonempty set Mij
be a vector of real numbers. The mechanism is smooth on the separable
quadratic subset Ē if for some interior environment ēAĒ, there exists a
neighborhood U(ē) and a continuously differentiable function r: U(ē)-M,
such that
for all e 2 UðēÞ we have g1 ðrðeÞ; e1 Þ ¼ 0; . . . ; gn ðrðeÞ; en Þ ¼ 0.
Thus the function r identifies an equilibrium message for each e in the
neighborhood U (ē), and that message varies in a continuously differen-
tiable fashion as we move away from ē. Using a variant of the uniqueness
228 T. Marschak
argument sketched in Section 2.2.8, we obtain a lower bound on each

person’s communication burden. The following can be shown:
Proposition A. Suppose /M,(g1, y ,gn), (h1, y ,hn)S is a network mech-
anism on the environment set E defined in (y), and each message mij in
every nonempty set Mij is a vector of real numbers. If the mechanism is
smooth on the separable quadratic subset Ē and realizes the goal corre-
spondence P on E, then the number of real message variables that each
person hears is at least twice the number of his concerns, i.e., for each
person i we have dim Pi(M)Z2|Ai|.
If we now return to our four-person three-plant example, we see that
the concern-set sizes are (1, 2, 1, 2) for Persons 1,2,3,4, respectively. But (as
already noted) the vector of message variables heard is (2, 4, 2, 4).
So Proposition A tells us that no smooth mechanism whose equilibrium
messages yield actions that maximize the firm’s profit can improve on the
Fig. 2 price mechanism with regard to any person’s communication burden.
That is a strong result in favor of price mechanisms. Is it confined
to situations that do not depart significantly from our example? The answer
is not known. In particular, suppose we consider P-realizing mechanisms
in which each individual burden need not be minimal but instead a
weaker efficiency requirement is met. Call a mechanism efficient in a
class if no other mechanism in the class has a lower burden for some
person and not higher burdens for the others. The following challenge
is unmet.
Research challenge # 1. Consider the class of all P-realizing mechanisms

on the above environment set E which are smooth on the separable
quadratic subclass Ē. If such a mechanism is efficient in that class, is it
always possible to write it so that it becomes a price mechanism, where
each message variable can be interpreted as a price or a quantity, and the
agreement functions express divisional profit maximization?
Other fundamental results on network mechanisms concern the size of
the overall message space M rather than the individual burdens. These
results require a stronger condition than the smoothness we have defined.
Call the stronger condition regularity on the separable quadratic subclass
Ē.22 For regular mechanisms, there is a useful lower bound to the size of M.
For any P-realizing mechanism which is regular on Ē we can show that
X
dim M 2 ðjtðaÞj 1Þ.
a2f1; ... ;kg
22
For regularity on Ē, we have to add the requirement that the matrix of second partial derivatives of
the gi has full rank at (m̄,ē), where ēAĒ is the interior environment in our previous smoothness condition
and m̄ ¼ r(ē); moreover, the rank does not change when we vary e in a neighborhood of ē while keeping
the message array constant at m̄.
A mechanism is dimensionally minimal in a certain class if no other mech-

anism in the class has a smaller value of dim M. It is of particular interest to
know when the communication graph of a dimensionally minimal regular
P-realizing mechanism is hierarchical, i.e., the graph is a tree. That may
have advantages that are related to incentives (e.g., it may facilitate ‘‘con-
trol’’). But trees may have communication costs that are higher than
needed. Using the above lower bound we can characterize the situations
where trees turn out to be dimensionally minimal. Before doing so, note
that the communication graph of a network mechanism defines a subgraph
for every subset of persons. In particular, there is a subgraph for t (a), the
set of persons who are concerned with the action variable za. If the com-
munication graph of the mechanism is a tree, then that subgraph may or
may not be a tree as well. The following has been shown.
Proposition B. There exists a P-realizing mechanism, that is (i) regular on
the separable quadratic subset Ē, (ii) dimensionally minimal among all such
mechanisms, and (iii) hierarchical, if and only if there is an n-node tree with
the property that for every action variable za, the tree’s subgraph for the
persons concerned with that variable is also a tree.
Propositions A, B, and further propositions that we do not summarize

here, all deal with environments that are valuation functions and with a
goal correspondence that requires maximization of their sum. The valua-
tion functions have a ‘‘public good’’ property, since a given action variable
may enter several of them. Are there analogues of these propositions for
other classes of environment sets and goal correspondences, where there are
no public goods? In particular, suppose that person i’s privately observed
local environment is a ‘‘revenue’’ function ei whose argument is a ‘‘private’’
action vector zi that does not enter any other ej, where j6¼i. Suppose, how-
ever, that all the action variables have to meet a common resource con-
straint. We require z A C, where C is some subset of a finite-dimensional
Euclidean space. Suppose, in particular, that the set C can be written in the
form fðz1 ; . . . ; zn Þ : r1 ðzB1 Þ 0; . . . ; rk ðzBn Þ 0g: Here the Bi are index sets
and each is a subset of {1, y , n}. The functions r1, y , rk are fixed and
known to all persons. Now we may view person i as being ‘‘concerned with’’
those constraint functions rt in which his own action vector enters , i.e.,
those for which iABt. Consider the goal correspondence P defined by
X n
P ðeÞ ¼ fz : z maximizes ei ðzÞ on the set Cg.
i¼1
Research challenge # 2. Are there propositions characterizing those P-

realizing network mechanisms which are regular (in a suitable sense) and
are efficient with regard to individual burdens, or are dimensionally
minimal? In particular, are there propositions which describe the situa-
tions where efficient or minimal mechanisms are hierarchical?
230 T. Marschak
2.3 Finite approximations of mechanisms whose message spaces are

continua
In most mechanisms studied in the literature, the message space M is a

continuum. That is not surprising, since typically the mechanism’s final
action maximizes some function, in which actions and environments are the
arguments. Maximization is most easily studied with the tools of calculus
and those tools deal with continua, not with finite sets. Thus in a classic
price mechanism, a message specifies prices and proposed quantities. The
first-order conditions for the required maximization can be expressed in a
statement about prices and quantities, and that statement holds at the
mechanism’s equilibrium messages. The message space is a continuum
composed of possible prices and possible quantities.
But continua are not realistic. In practice, one cannot send all the mes-
sages in a continuum (e.g., all the points in a real interval). Moreover it may
take infinite time to find that message in a continuum which satisfies the
required equilibrium conditions. If an organization wants to use a contin-
uum mechanism in a practical way, it has no choice but to approximate the
continuum mechanism with an appropriate finite mechanism, whose mes-
sage space is a finite set. The penalty paid for such finite approximation
may be an error: the actions generated (at equilibrium) by the finite ap-
proximation may differ from the goal-fulfilling actions which the original
continuum mechanism generates.
If we take the issue seriously, then the following question immediately
comes to mind: Will the informational advantages of the original contin-
uum mechanism be reflected in its finite approximation? In particular, if we
have found (using the tools of smooth mathematics) a continuum mech-
anism that realizes a given goal correspondence and does so with minimal
message space dimension, are finite approximations to that mechanism
superior (in an appropriate sense) to finite approximations of a continuum
mechanism which also realizes the goal correspondence but has a higher-
than-minimal message-space dimension? Does dimension still matter when
we turn from continuum mechanisms to their finite approximations?
In fact, for broadcast mechanisms, a theory of finite approximations has
been begun23 and several ‘‘dimension still matters’’ propositions have been
established. In these propositions, the environment set is a continuum as
well as the message space of the mechanism we are approximating. The
view taken is that ‘‘nature’’ is able to choose the organization’s current
external environment from some continuum of possible environments, but
the continuum message space is ‘‘man-made’’ (for analytic convenience)
and we are free to replace it with a finite message space.
The first step is to define a style of approximation. Suppose we are
given an n-person broadcast mechanism L on an environment set E ¼ E1
23
In Marschak (1987) and Hurwicz and Marschak (2003a, b, 2004).
? En, with action set A, where both M and each Ei are continua.
In particular (as in our introductory discussion of mechanisms in
D1
Section 2.2.1), let M be the Cartesian product M M Dn : Let
Di Di
each M be a closed Di-dimensional cube in R ; while each Ei is a
closed Ji-dimensional cube in RJ i : Write the continuum mechanism L in
agreement-function form: L ¼ /M,(g1,y,gn),hS. Recall that each agrem-
ent function gi has M Ei as its domain and M Di as its range. Thus each
gi has P Di real-valued components, say gi1, y , gik, y , giDi. Let D
denote in¼ 1Di.
In a finite approximation to L, our finite message space, denoted MA is
the intersection of the D-dimensional continuum message space M with a
mesh-A lattice of points, which are spaced 2A apart (A>0). That lattice,
denoted SD A, is the D-fold Cartesian product of the set
S 2 ¼ f. . . ; 2ð‘ þ 1Þ 2; 2‘ 2; . . . ; 4 2; 2 2; 0; 2 2 ,
4 2; . . . ; 2‘ 2; 2ð‘ þ 1Þ 2; . . .g.
Next we replace each agreement function gi with a new function gZ i ¼

ðgZ Z Z2
i1 ; . . . ; giDi Þ: Each gik is the following two-valued function:
(
0 if jgik ðm; ei Þj Z;
for every m 2 M 2 ; gn2
ik ðm; ei Þ ¼
1 otherwise;
where Z>0 is called the tolerance. Finally, we have to specify the outcome
function of our finite approximation. In the simplest approach, we let the
outcome function be the original one, i.e., it is the restriction of h, the
outcome function in L, to the new finite message space MA (which is a
subset of M). Denoting the new outcome function h0, we have
h0(m) ¼ h(m).
Suppose that our new agreement functions satisfy the coverage require-
ment: i.e., for every e A E, there exists m A MA such that gZ2 i ðm; ei Þ ¼ 0; all
i. Then the finite mechanism LZ2 ¼ hM 2 ; ðgZ21 ; . . . ; g Z2
n Þ; h0
i is called the finite
exact-outcome approximation of L with message mesh A and tolerance Z. To
obtain it we have, in effect, rounded off the original functions gi to a
specified accuracy. The accuracy is determined by the the mesh A and the
tolerance Z. In an alternative approximation of L, we do not require the
outcome for the message m to be exactly what it was in L. Rather we place a
mesh-n lattice on the action set A, so that our finite mechanism’s action set
becomes A\Sna We then choose the outcome to be a lattice point that is
closest to the action chosen in L. Suppose there are a real-valued action
variables and that A is contained in a closed cube in Ra : Then in the finite
rounded-outcome approximation of L with message mesh A, action mesh n
and tolerance Z, all elements except the outcome function are the same as
232 T. Marschak
those just defined. The outcome function hn : M 2 ! A \ S an is defined as

follows:
8
> the element a ¼ ða1 ; . . . ; aa Þ of A \ Sav
>
>
>
< which is closest to hðmÞ; where distance
v
h ðmÞ ¼ is measured by max ðjhðmÞ ar jÞ
>
>
>
> r2f1;...;ag
:
and ties are broken downward:
For both versions, we shall say that our finite mesh-A approximation has
the minimal tolerance property if Z has the smallest value that permits
coverage. (Such a smallest value can be shown to exist.)
If the mechanism L is a finite approximation of the continuum mech-
anism L, then for any e A E, we define the error at e of L to be the worst-
case distance between the (equilibrium) value of an action variable in the
continuum mechanism and its value in the approximation. Let h# denote
the finite approximation’s outcome function. (This is either the exact
outcome of the continuum mechanism or it is a rounded-outcome approx-
imation.) Since there are a real-action variables, the function h# has a real-
valued components ha. The error at e of L is
supfjh#j ðm̄Þ2hj ðmÞj : m 2 M; m̄ 2 M 2 ;
gi ðm; ei Þ ¼ 0; g2Z
i ðm̄; ei Þ ¼ 0; i ¼ 1; . . . ; n; j 2 f1; . . . ; agg.
The overall error of L is supeAE (error at e of L).
We define the cost of the approximation L to be the number of messages
in its finte message space. We impose some regularity conditions24 and
obtain the following ‘‘dimension still matters’’ proposition.
Proposition C. Consider two regular continuum mechanisms L ¼
/M,(g1, y , gn),hS and L ¼ /M,(g1 , y ,gn, hS. Each is a mecha-
nism on the same environment set, namely the compact set E ¼ E1
? En, and each has an action set which is a subset of Ra : The two
message spaces are distinct: M is a D-dimensional subset of RD , while M is
a D-dimensional subset of RD , where D>D. Let L be a mesh-A
rounded-outcome minimal-tolerance approximation of L, and let L be a
mesh-2̄ rounded-outcome minimal-tolerance approximation of L. Suppose
that L costs no more than L. Then if A is sufficently small, the overall
error of L exceeds the overall error of L:
Note that the proposition does not require us to specify goal corre-
spondences realized by the two continuum mechanisms. But it can certainly
be applied to two continuum mechanisms which realize the same goal
24
In a regular mechanism each function gik is continuously differentiable. Moreover there exists a
number d>0 such that for all dikA[d,d] and for all e in E, there is a unique message m satisfying all the
equations gik(m,ei) ¼ dik.
correspondence. In that case it tells us, informally speaking, that if we want

to come close to achieving our goal, then the low-dimensional mechanism
L is a better candidate for finite approximation than the high-dimensional
mechanism L. We achieve lower overall error, for a given ‘‘budget’’ if we
approximate L than if we approximate L. So dimension indeed continues
to matter.25
It remains open whether or not there are similar propositions for network
mechanisms.
Research challenge #3. If two continuum network mechanisms realize the
same goal correspondence and have suitable regularity properties, but the
second has a higher number of individually addressed message variables,
is it better to approximate the first? (Do we achieve a smaller overall
error, for a given ‘‘budget,’’ when we do so?)
Returning to the case of broadcast mechanisms, a very difficult question
remains unaddressed. In constructing finite broadcast mechanisms, we have
confined our attention to finite mechanisms which approximate regular
continuum broadcast mechanisms. Do we ignore certain efficient finite
mechanisms when we do so?
Research challenge # 4. Given a goal correspondence G, can there be a
finite broadcast mechanism which is NOT an approximation of any reg-
ular G-realizing continuum broadcast mechanism but makes better use of
a given ‘‘budget’’ (achieves lower overall error with respect to G while not
using more messages) than any such approximation?
2.4 The dynamics of a mechanism
Return now to the dynamic form of a broadcast mechanism, which

started our discussion in Section 2.2.1. The mechanism is a quadruple
/(M1, y ,Mn), (m01, y ,mn0),(f1, y ,fn),hS. It defines a difference-equation
system, namely:
mt ¼ f ððmt1 t1
1 ; . . . ; mn Þ; ei Þ; i ¼ 1; . . . ; n;
with an initial message m0(e) ¼ (m01(e1), y ,mn0(en)). We have been inter-
ested thus far in the achievements of the mechanism once it has reached an
equilibrium message. A difficult missing piece in the story has to do with the
stability properties of the difference-equation system. We would like the
action taken once an equilibrium message is reached to meet a specified
goal, but we would also like the difference-equation system to display some
25
Another proposition lets us be ‘‘kind’’ to the high-dimension mechanism by permitting its approx-
imation to have the original exact outcomes, while we are ‘‘harsh’’ to the low-dimension mechanism by
requiring its outcomes to be rounded off. Even so, it is better to approximate the low-dimension
mechanism. This proposition, however, requires the high-dimension mechanism to have a ‘‘projection’’
property: each message is a pair (a, q) and the outcome function is h (a, q) ¼ a.
234 T. Marschak
sort of convergence to the system’s equilibria. Do the mechanism’s infor-

mational requirements (e.g., its message-space size) grow if we require sta-
bility as well as goal realization? One can construct examples where the
answer is Yes.26
Some progress has been made on this question when the difference
equations are replaced by differential equations. In particular, Jordan
(1995) developed the following new mechanism concept.27 For each person
i there is a real-valued ‘‘control message’’ ci, whose possible values comprise
a set Ci. There is also a broadcast ‘‘state message’’ m ¼ (m1, y , mq), with q
real components, which is continually adjusted as a function of both
c ¼ (c1, y , cn) and m. It is not required that q ¼ n. We have the differ-
ential-equation system
dmj
¼ aj ðc; mÞ; j ¼ 1; . . . ; q;
ci ¼ f i ðm; ei Þ; i ¼ 1; . . . ; n;
dt
where t is a time point. Note that the function aj does not have ei as an
argument. The interpretation is that each person i continually adjusts his
own control variable ci in a privacy-preserving manner. We do not specify
who adjusts a given component of m itself, but that adjustment does not
require direct knowledge of the privately observed eis. In the case where
q ¼ n and i has responsibility for the message variable mi, we have a com-
plete privacy-preserving scenario: person i observes the entire broadcast
message m, adjusts his own part of m, and chooses his control variable as a
function of the broadcast message and his local environment.
The general question is then as follows. Suppose we are given a parti-
cular message correspondence m from E ¼ E1 ? En to the state-
message space M (the set of possible values of m). We are interested in
this correspondence because it realizes some goal (i.e., if m A m(e), then
there is an action h (m) which lies in a set G (e) of goal-fulfilling actions),
but the goal itself is not part of the research question. Instead we ask:
How large do the sets Ci have to be if the equilibria of the differential-
equation system always lie in the set m(e) and the system has a local stability
26
A two-person example, due to Reiter and discussed in Hurwicz (1986) is as follows. Person i’s local
environment is a real-number pair ei ¼ (ei1,ei2). The action set is R: For each e the goal specifies a unique
action for all e such that e116¼e12, namely F(e) ¼ [e11e22e21e12]/[e11e21]. If we do not require stability, we can
realize the goal with a two-message-variable mechanism. The typical message is a pair m ¼ (m1, m2). The
mechanism’s difference-equation system is mt1 ¼ 2mt1 1 t1 2 t t1 t1
1 e1, m2 e1,m2 ¼ m1 +m2 e2m2 e2. The
1 t1 2
outcome function is a projection: h (m) ¼ m1.

But this system fails to satisfy the following local stability requirement: for m0 sufficiently close to an
equilibrium value of m, the system should converge to that equilibrium. Moreover, if we seek any two-
message-variable difference equation system which realizes F at equilibrium and uses the projection
outcome function, we find that if the functions f1, f2 have continuous partial derivatives, then the system
is not locally stable. On the other hand, a locally stable F-realizing mechanism with four message
variables can be constructed.
27
See also Jordan (1987), Mount and Reiter (1987), and Saari and Simon (1978).
property for every e? How large, in other words, is the extra informa-
tional cost of stability? Several general results characterize the required size
of the Ci.
One would like applications of these results to classic questions, notably
the informational merits of price mechanisms. Consider, once again, the
n-person L-commodity exchange economy and the Pareto-optimal mes-
sage correspondence. If we construct a privacy-preserving dynamic mecha-
nism which uses prices, has suitable regularity properties as well as local
stability, and achieves Pareto-optimality at the equilibrium message, then is
its total message space (i.e., M C) dimensionally minimal among all dy-
namic mechanisms with those properties? Much remains to be learned about
this question, but for certain reasonable classes of economies, and certain
versions of the dynamic price mechanism the answer is Yes. For one
such class, Jordan (1995) studies dynamic mechanisms in which the message
m specifies current trades, the control variable ci is a vector of i’s ‘‘demand
prices’’ (marginal utilities), and the adjustment rules for m (i,.e., the func-
tions aj) adjust the trades so that they are Pareto-improving. It is shown that
if we delete the stability requirement for such mechanisms, then a lower
bound on the dimension of each C is n (L–1). It is then shown that stability
can be achieved without increasing the dimension of C beyond n (L–1).
In other types of dynamic mechanism, the control variables are trades as
well as prices. It turns out that if such a mechanism is formulated so that
privacy is preserved and local stability is achieved, then C has to be very large
and the stabilized price mechanism may no longer be minimal among all such
mechanisms.
Note that for a finite broadcast mechanism we have an upper bound on
the time required to achieve equilibrium—namely the time required to an-
nounce all the messages in the finite set M, in some predetermined sequence.
If M is large, that upper bound is of little interest. We may then want to
choose the sequence with care, perhaps by approximating (in some suitable
way) the rules of a locally stable mechanism in which the sets M and C are
continua. Such approximation remains to be studied.
Note also that if we truncate a difference-equation broadcast mecha-
nism /(M1, y ,Mn), (m01, y ,mn0),(f1, y ,fn),hS after T steps, then we
have defined a new privacy-preserving broadcast mechanism in which
every possible broadcast message describes a possible T-step ‘‘conversa-
tion,’’ namely a proposed sequence of announcements m0, m1, y ,
mT. Person i agrees to the proposed conversation if he finds—given his
ei, given the proposed sequence of announcements by others, and given
his function fi—that the proposed sequence mi0, mi1, y ,miT is exactly
what he would announce. So, using the uniqueness technique discussed
in Section 2.2.8, one could study the goal correspondence realized by the
T-step truncation and could ask whether there are broadcast mecha-
nisms which realize the same goal correspondence using a smaller mes-
sage space.
236 T. Marschak
2.5 Constructing an informationally efficient mechanism
The informationally efficient (or cheap) mechanisms that have app-

eared in the literature are, in a sense, historical accidents. They are mech-
anisms which allocate resources so as to satisfy a goal correspondence
that expresses Pareto-optimality, or perhaps profit maximization. Typically
each message proposes prices and quantities, and agreement to the mes-
sage by all persons means that its prices and quantities correspond to
the required first-order maximization conditions. With some exagge-
ration one might say that the literature started by asking: ‘‘What is
the precise information-related question to which ‘price mechanisms’
(or ‘competitive mechanisms’) is the answer?’’ That was a natural chall-
enge, given the long history of sweeping but never rigorously defen-
ded claims about the price (or competitive) mechanism’s informational
merits.
But what if prices had not yet been discovered? Imagine looking for low-
cost mechanisms among all those that realize a goal correpondence. If the
goal required Pareto optimality, then a search for such mechanisms would
eventually discover mechanisms that use prices. How might such a search
proceed?
Two new books, one by Hurwicz and Reiter (2006), and the other by
Williams (2007), deal with this fundamental puzzle. It would be futile to
attempt any kind of summary here. But we can roughly visualize one of the
main issues in the Hurwicz/Reiter agenda by going back to the ‘‘rectangle’’
definition of a mechanism in Section 2.2.3 and the two-person, three-mes-
sage example in Fig. 1 of Section 2.2.2. Suppose that our environment sets
are Ei ¼ [0, 1], i ¼ 1, 2 and that we have an action variable a that takes
three values, namely u, v, w. Suppose we have not yet constructed a mech-
anism which yields (at equilibrium) a value of the action for every e A
E ¼ E1 E2. Relabel the three rectangles m1, m2, m3 in Fig. 1 as U, V, W,
respectively. Let those rectangles (which overlap at boundary points) define
the goal correspondence G that we want to realize. Thus
GðeÞ ¼ fa : a ¼ u if e 2 U; a ¼ u if e 2 V ; a ¼ w if e 2 W g.
We may call U the level set of the correspondence G for the action u, and
similarly for V and W. Formally the level set corresponding to the action a
is G1(a) ¼ {e A E: a A G (e)}. Consider the mechanisms which realize G
and suppose that we write all of them in rectangle form, so that there is a
generalized rectangle sm for every message m, and m is an equilibrium
message for all the environments in sm. (Recall that a generalized rect-
angle is a set of environments e that is the Cartesian product of its
E1-projection and its E2-projection.) There are many such mechanisms, but
they all use more generalized rectangles (messages) than we need except the
three-message mechanism defined in Section 2.2.2. In that mechanism there

are just three rectangles sm, namely the same three rectangles U, V, W that
define the goal correspondence. An inefficient G-realizing mechanism
might, for example, needlessly add a fourth message, by dividing the rec-
tangle sm3 ¼ W into two further generalized rectangles.
The three-message mechanism defined in Section 2.2.2 is efficient: it cov-
ers each level set of our goal correspondence with rectangles sm in such a way
that the total number of rectangles sm is minimized. To illustrate further,
suppose we modify our example by allowing just the two actions u and w
and letting w be goal-fulfilling for all environments that lie in W or V. (The
action u remains goal-fulfilling for all e A U.) Now the level set for the
action w is no longer a (generalized) rectangle; it is now the union of W
and V. Nevertheless, an efficient goal-fulfilling mechanism has to cover
that level set (as well as the level set corresponding to u) with genera-
lized rectangles sm and it has to do so in a minimal way. The efficient
mechanism will again require three messages. It will cover the level set
corresponding to the action w with two generalized rectangles, namely the
rectangles W and V.
So the search for an efficient goal-realizing privacy-preserving mecha-
nism requires us to inspect the level sets, to find a way of covering each
of them with generalized rectangles, to find a way of indexing (labeling)
each of our generalized rectangles with an index m, and to do all this
while keeping the size of M (the set of values if the index m) as ‘‘small’’
as possible. The basic set-theoretic properties of such a mechanism-
designing algorithm are worked out in the Hurwicz/Reiter book. The
algorithm yields efficient mechanisms whatever the goal correspondence
may be, whether the action set and the environment sets are finite or are
continua.
The book of Williams also deals with a mechanism-designing algorithm,
but from a very different point of view. Smoothness requirements are im-
posed on the goal and on the candidate mechanisms. Tools of differen-
tial topology are used rather than purely set-theoretic tools. Some of the
results imply that the agreement functions of an efficient goal-realizing mech-
anism can be found by solving an appropriate system of partial differential
equations.
Once these two books are understood, they may open a massive research
agenda for the designers of practical computer-friendly algorithms that
construct mechanisms (protocols). It remains to be seen, for example,
whether the general results in these books will eventually allow a computer
to generate protocols that yield the minimal-length dialogues studied in the
computer-science communication-complexity literature (briefly discussed
in Section 2.2.8). At present that literature finds bounds on the length of
the dialogues but does not tell us how to construct the minimal-length
dialogues themselves.
238 T. Marschak
2.6 Finding a best action rule (outcome function) once a mechanism has
conveyed information about the environment to each person: the methods of
the Theory of Teams
The central problem studied in the Theory of Teams (Marschak

and Radner, 1972) is the choice of a rule that tells each member of an
organization—called a team—what action to choose, given certain infor-
mation about the organization’s randomly changing environment. The rule
has to maximize the expected value of a payoff function whose arguments
are the environment and the team action. Mechanisms, as we have defined
them, do not appear in the statement of the central problem studied in the
Theory of Teams, but they are part of the story which implicitly precedes
that problem.28
Here is one version of the central n-person team problem. The team
has to chose an action, namely a vector a ¼ (a1, y , an), where ai is the
responsibility of person i. Let Ai denote the set of possible values of ai and
assume that every n-tuple a in A1 ? An is a possible team action. The
team earns a (real-valued) payoff H (a, e), where e ¼ (e1, y , en) is a vector of
random local-environment variables. We study a given information structure,
specifying what each person knows about a given e. The set of possible values
of e is denoted E. Let Zi be a function from E to a signal set Yi. Let E be a
subset of a (finite-dimensional) Euclidean space, and similarly for each Yi
and each Ai. A probability distribution P on E is given. In the information
structure Z ¼ (Z1, y , Zn), person i observes the signal yi ¼ Zi (e) when the
environment is e.
The signal Zi (e) might be a vector, and ei might be one of its components.
In our discussion of privacy-preserving broadcast mechanisms, each ei was
automatically known to one person, namely person i. We may interpret
Zi (e) as the information about the current e that person i possesses once
the mechanism has terminated. Then Zi (e) indeed includes ei, but it also
describes the information about the other ej that is revealed to i by the
mechanism’s terminal message.29
For a given information structure Z, we consider the possible team action
rules a ¼ (a1, y , an), where ai is a function from Yi to Ai. Thus the team
action is (a1(y1), y , an (yn)) when the signal vector is y ¼ (y1, y , yn).30 An
action rule a^ is
team-best for the payoff function H and the information struc-
ture Z if EH ðâ1 ðZ1 ðeÞÞ; . . . ; a^ n ðZn ðeÞÞÞ; e EHðða1 ðZ1 ðeÞÞ; . . . ; an ðZn ðeÞÞÞ; eÞ
28
In Chapter 8 of The Economic Theory of Teams, there is a discussion of ‘‘networks’’, with a number
of examples. One may interpret the network concept developed in that chapter as a mechanism in our
sense.
29
Thus we can express an information structure as a message correspondence m, where m lies in m(e) if
and only if m ¼ (Z1(e), y ,Zn(e)).
30
Appropriate measurability assumptions have to be made when E is not finite. They guarantee that
(i) P implies a probability distribution on the set ai(Z1(E)) for every i and (ii) H((aiZ1(e)), y ,an(Zn(e))),e)
has a finite expected value.
for all action rules a, where e again denotes expectation. A necessary con-
dition for an action rule a^ to be team-best for H, Z is that it be person-by-
person-satisfactory (pbps) for H, Z. That means that for every i and every y,
the action a^ i (yi) maximizes the conditional expected value E H ðâ1 ðZ1 ðeÞÞ;
. . . ; a^ i21 ðZi21 ðeÞÞ; ai ; a^ iþ1 ðZiþ1 ðeÞÞ; . . . ; a^ n ðZn ðeÞÞÞ; eÞ Zi ðeÞ ¼ yi : on the set Ai.
If H is differentiable and strictly concave in a for each e, then the pbps
condition is sufficient as well as necessary.
Consider the case of a team with the linear-quadratic team payoff func-
tion W (a, e) ¼ 2a0 ea0 Qa. Here Ai is the real line. The random variables
e1, y , en also take values in the real line and they are independently dis-
tributed with finite second moments; Q ¼ ((qij)) is an n n symmetric pos-
itive definite matrix. The function W is differentiable and strictly concave in
a for each e. Accordingly the pbps condition is both necessary and suffi-
cient. A best team action rule is linear. Its coefficients can be found by
solving a system of linear equations.
That permits the explorations of information structures with interesting
organizational properties. (For some explorations, it is also helpful to
assume that each ei is normally distributed.) For example in a ‘‘manage-
ment-by-exception’’ information structure, each person i 6¼ n knows only his
own ei. But person n is a manager who learns the value of every ej whenever
that ej lies in a specified ‘‘exceptional’’ region. A best team action rule a will
take advantage of the manager’s exceptional information. In a variant of
this structure, an ‘‘emergency conference’’ of all n persons is called whenever
some person j observes an exceptional value of ej. When that happens, all
persons learn that exceptional value. We can vary the exceptional regions
and in each case we can compute the structure’s ‘‘gross’’ performance, i.e.,
expected team payoff when a best action rule is used. The gross performance
of one interesting structure can be compared to that of another. The cost of
each structure, however, needs to be measured in some consistent way if we
are to characterize structures whose net performance is high.
Unfortunately, it is difficult to obtain similar explicit results about gross
performance once we leave the linear-quadratic case. It is difficult even
though we remain in the class of payoff functions W that are strictly con-
cave in a for each e, so that the pbps condition is sufficient as well as
necessary. We are, after all, performing a search in function space, and that
is difficult unless the functions can be parametrized in some convenient
way. Nevertheless it would seem plausible that for some class of concave
payoff functions, algorithms could be constructed that come close to yield-
ing best team action rules.
Research challenge # 5. Construct an algorithm that yields best (or nearly

best) team action rules for a wide class of concave functions W (con-
tainimg the linear-quadratic function and others as well), and does so for
a wide class of information structures and probability distributions on the
environment variables
240 T. Marschak
2.7 Designing an organization ‘‘from scratch’’: choosing its members, what

each observes, and the speak-once-only mechanism that they use31
So far we have assumed (without comment) that the n members of our

organization are already in place, and that we have no choice as to the
external variables (the local-environment variables) which a given member
observes. That is natural if our organization is an economy and its members
are a given collection of consumers and producers. There is a natural pri-
vately observed external variable for each consumer; it describes her indi-
vidual preferences and perhaps her individual resource endowment. For a
producer, the natural external variable describes his current technology.
Similarly, when modeling a firm with several divisons, it is natural to let the
members of the organization be the division managers, who are already in
place. For each manager, the natural privately observed external variables
are those that characterize his production technology.
Once we leave such settings, we may want to enrich our modeling toolkit.
We may want to take the view that the organization does not yet exist but is
being designed. The designer has a clean slate. He is given external variables
e1, y , ek, and a set E which contains their possible values. He is given a
set A of possible organizational actions. He is given a goal, which identifies
at least one appropriate action in A for each external environment
e ¼ (e1, y , ek) in E. But he can choose the following:
The size and composition of the collection of persons who make up the
organization.
The identity of the person who will observe each external variable;
some external variables may be observed by more than one person;
some persons may not observe any external variable.
The speak-once-only mechanism, which the organization uses to find a
new action when the environment changes.
2.7.1 Speak-once-only mechanisms: introductory comments

In a speak-once-only mechanism, the newly changed external variables
are observed by their designated observers. Each observer sends forward
to others a message based on those observations and then stays silent.
The recipients of those messages send forward messages to others, based on
the messages they have received and on their external observations (if they
are designated external-variable observers), and then stay silent; the
recipients of those messages send forward still further messages and then
stay silent. And so on. When all sending has stopped, one member, called
the action taker, takes an action based on the messages he has received
31
Some of the ideas in this section grew out of conversations with Ilya Segal. Some of the results in
Section 2.7.6 are due to Jeff Phan. The discussion of delay in Section 2.7.10 as well as the result in
Section 2.7.7 are largely due to Dennis Courtney.
(and on his own external observations if he is a designated external-

variable observer). The mechanism realizes a given goal if the action taker’s
action is always goal-fulfilling for the environment e which initiates the
process.
Our speak-once-only requirement is certainly restrictive. For full
generality, we would let the designer choose a mechanism in which each
sender sends messages to others at each of a sequence of steps. But our
clean-slate assumption, where the designer chooses the size and composition
of the organization, the observing assigments, and the mechanism used
to generate new actions presents a formidable challenge. Confining attention
to speak-once-only mechanisms is a reasonable compromise, if one wants
to start learning something about the structure of the designer’s chosen
organization.
To help motivate a research agenda on speak-once-only mechanisms
consider the following question32:
What might be inefficient about a one person mechanism;
wherein a single person observes all the external variables
and then finds a goal fulfilling action?
A first step in making sense out of this question (and more complicated
ones) is to choose some cost measures for a speak-once-only mechanism, so
that ‘‘efficiency’’ (and inefficiency) can be defined. First we assume that
each external variable ek is real-valued, and that every message in the
mechanism is a vector of real numbers. For simplicity we assume that for
every external vector e, there is a unique goal-fulfilling action F (e) where F
(e) is a real number. Consider the following three cost measures for a speak-
once-only mechanism:
The number of persons.
Each person’s burden, defined as the number of real variables observed
or received.
The mechanism’s delay, i.e., the total elapsed time until the action-
taker has computed F, given that (i) no one sends message variables or
computes F until he has finished doing all the receiving and observing
that the mechanism requires of him, and (ii) it takes one time unit for a
person to observe or receive one real variable, but no extra time is
required for the sender of a message to compute and send it or for the
action taker to compute F.
32
One can ask a version of this question, even if one specifies that n, the number of persons, is greater
than one and cannot be changed by the designer, and that each external variable is observed by one and
only one person. A mechanism in which one person does essentially all the work, would be one in which,
say, Person 1 collects full information about e from all the others and thereupon finds the goal-fulfilling
action.
242 T. Marschak
2.7.2 The approaches of Radner/Van Zandt and Mount/Reiter

Papers by Radner (1993), Radner and Van Zandt (1995), and Van Zandt
(1997, 1998, 2003a, b) study a class of speak-once-only mechanisms. The
messages transmitted are composed of real numbers, and there is an action
taker who acquires enough information from others to compute a real-
valued goal function F of external real variables e1, y , ek, each of which is
observed by someone. The cost measures are number of persons and de-
lay.33 But the function F has to have the form F ¼ H (H1(e1) * H2(e2)
* ? * Hk (ek)), where * is an associative operation. Moreover, any person
i, who receives, say, the real numbers u1, y , ur, is only able to compute u1 *
u2 ? * ur, where * is the same associative operation. It takes one time unit
to perform the operation * and no time is required for communication, so the
Radner/Van Zandt work may be viewed as a model of organizational
computing. It can be shown that if F is differentiable in the k variables, then
no generality is lost if the operation * is constrained to be addition. That is
to say, any differentiable F having the above associative form can be re-
written so that it becomes a function of the sum of certain terms H~ i ðei Þ:
Nondifferentiable goal functions with the associative property include max
(e1, y , ek), where e1, y , ek are real numbers.
In the book by Mount and Reiter (2002),34 the goal function F (e1, y , ek)
can be any analytic function. The problem is to assemble a minimal-delay
F-computing network of processors (persons). Each of them receives inputs
from other processors; each of those inputs consists of at most d real num-
bers. In one time unit a processor computes a certain analytic function of the
numbers received and sends the result to one or more other processors. The
class of functions that the processors are able to compute is a primitive of
the model. If F itself has d real arguments, and is one of the available
functions, then the problem is trivial, since a single processor can then
compute F itself in one time unit. Instead, one seeks networks (directed
graphs) that compute F with minimal delay when the functions available for
each processor belong to an interesting class. The network is not required to
be a tree, so cycles are permitted. But it is shown (under weak assumptions)
that no generality is lost if one confines attention to trees. A tree has no
33
While Radner (1993) and Radner and Van Zandt (1995) consider a ‘‘one-shot’’ situation, where each
environment vector e ¼ (e1, y , ek) is processed before a new e appears, papers by Van Zandt (1999,
2003a, b) go on to study the much more challenging situation where a new e arrives before the previous
one has been completely processed, and the successive es follow a stochastic process. The current
computed action, which is a function of the previous e, is then somewhat obsolete. The penalty due to
obsolescence is studied. In particular, Van Zandt (2003b; sketched also in Van Zandt, 1998, Section 3.3)
studies the performance of an organization which repeatedly allocates resources and thereby earns, in
each time period, a payoff that is a quadratic function of the allocation and of that period’s environment
vector. But the information used by the allocator reflects an earlier period’s environment vector, since it
takes time for that information to reach him. Results are obtained by exploiting the fact that the mean of
k variables is a sum and hence it can be computed by a sequence of associative operations. Early
discussions of obsolescence as one of the costs of a mechanism appear in Marschak (1959, 1972).
34
An easily accessible summary of some of the book’s ideas is given in Mount and Reiter (1998).
cycles and so it is, in our terminology, an F-computing speak-once-only

mechanism.
While similar models of networks of processors (automota)
have been studied by others, the Mount/Reiter research has a major nov-
elty: the possible values taken by the input numbers, by F, and by each
processor’s function, can be continua rather than finite sets. The research
reported in the book does not explicitly seek efficient combinations of
number of processors, individual burdens, and delay. Instead it seeks to
characterize the networks (trees) that are capable of computing certain
functions F with a given delay, when the processor functions obey key
conditions like twice differentiability. Some propositions, for example,
concern the number of processors to which each processor sends. There
are also results about the relation between the delay for a continuum-
valued F and the delay for each of a sequence of finite-valued functions
that approximates F. It turns out that the former delay is the limit of the
latter sequence of delays.
The Mount–Reiter model is a highly innovative way to study the com-
plexity of a given goal function F. Since it does not treat number of proc-
essors (persons) and individual burdens as explicit costs, it does not easily
lend itself to the study of some of the efficient-organization questions that
we shall now consider. Note that if each processor’s function is required to
be addition, and if number of processors is a cost element, then the Mount/
Reiter model becomes the Van Zandt/Radner model. Note also that in both
models one finds, in many interesting cases, that there is an upper limit to
the useful number of persons (processors). Going above that number does
not further decrease delay.
2.7.3 A formal definition of a speak-once-only mechanism

To define a speak-once-only mechanism, a directed graph has to be
specified. It will be useful to modify the conventional terminology of di-
rected graphs to fit our context. A directed graph is defined by a set of
nodes and a set of ordered pairs whose elements belong to the set of nodes.
We shall say that the first node in the pair sends to the second and the
second receives from the first. There is an arc between them. Then node i0 s
received-from set, denoted Ri is the set of nodes from which i receives, while
i’s sent-to set, denoted Si is the set of nodes to which i sends. We shall call a
node j a leaf if it sends but does not receive. We call a node r a root if it
receives but does not send. (Usually, the terms leaf and root are reserved for
trees, but our graph need not be a tree.)
An n-person speak-once-only mechanism on the local-environment sets
E 1
RD1 ; . . . ; E k
RDk ; with action space A
Ra is a pair L ¼
hG; ð~rkþ1 ; . . . ; ~
rn Þi: Here G is a graph and ~
rkþ1 ; . . . ; ~
rn are vectors of send-
ing functions. If j A {1, y , k}, then ej is an environment variable; its
244 T. Marschak
possible values comprise the set E j

RDj : We specify that
G is a connected directed graph with nodes denoted 1, y , k, k+1, y ,
n, where nodes 1, y , k, with kon, are leaves (they correspond to the
environment variables), node n (the action taker) is the only root, there
are no directed cycles, and there is at least one directed path from every
node to the root.
for i ¼ k+1, y , n – 1, the vector ~ ri has one sending function, ri‘, for
every ‘ in the sent-to set Si. Person i sends Di‘ real variables to person ‘
in Si, so the range of ri‘ is RDi‘ : The domain of ri‘ is the set of possible
values of the variables i receives or observes, i.e., the domain is a subset
of RSt2Ri Dti ; where DtiDt if t A {1, y , k}.
The remaining vector of functions is ~ rn : It has a single component,
denoted simply rn, which yields the organization’s action. The range of
rn is A
Ra and the domain is a subset of RSt2Rn Dtn :
The function rn, and every function rij, where i A {k+1, y , n1},
jASi, satisfies a smoothness requirement, e.g., it is differentiable at all
points of its domain.
Thus any person i in {k+1, y , n} whose received-from set Ri contains
t A {1, y , k} is an observer of the environment variable et. In addition,
person i receives a message, say mji, from every j A Ri with j>k. The message
mji is a vector with Dji real components. Person i sends a message, say mij, to
every person j A Si; that message is determined by the function ri. The
action taker n receives message variables, say mjn, from certain persons j>k.
His choice of the action a in A is determined by the function rn. For person
i inP{k+1, y , n} we shall define i’s individual burden in the mechanism L to
be t2Ri Dti :
2.7.4 An illustration, which leads to three questions

To illustrate, consider four real and positive local-environment variables.
It will be notationally convenient to call them w, x, y, and z. Suppose we
want to realize the following goal function F:

1 1 1 1
F ðw; x; y; zÞ ¼ w þ þ x þ þ y þ þ z þ þ ðwxyzÞ.
w x y z
Here is a two-person mechanism that realizes F.
Each person’s burden is 3 and the delay35 is 5.
35
Suppose a new (w, x, y, z) occurs. Three time periods then pass. At that point, Person 1
has completed his observing (processing) of his three assigned variables w, x, y and Person 2 has
completed his observing of z. Now two more time periods pass. During the first of them, Person 2
receives (processes) the first of the two message variables which Person 1 sends, and during the
second, Person 2 receives (processes) the second message variable. At the end of the fifth period,
Person 2 has received (from Person 1) the messages he needs, and is able to take the action F
(w, x, y, z).
Fig. 3. A two-person mechanism that realizes F .
We can ask the following questions about the Fig. 3 mechanism and other
possible speak-once-only mechanisms which also realize F:
(1) If we reduce the number of persons to just one, must the burden and
the delay of an F-realizing mechanism rise? Our one-person mech-
anism would be a five-node tree, with four nodes correponding to w,
x, y, and z, and the sole person at the root. Clearly the burden of
the sole person will be four and the delay will be four. We actually
reduce delay when we move to the one-person mechanism, since
no time is used for the processing of messages received from other
persons.36
(2) Is there a two-person F-realizing mechanism in which neither person
has a burden more than two and delay is not more than four? The
mechanism would improve on the Fig. 3 mechanism with regard to
delay and it would improve on the one-person tree with regard to
burden. Such a mechanism would exist if it were possible to aggregate
three of the four external variables into one, i.e., if it were possible to
write F in the form G (H (w, x, y), z)), where G, H are real-valued
functions (with suitable smoothness properties). It is natural to call
the function H an aggregator. The single number H (w, x, y) contains
36
That illustrates a deficiency of our definition of delay, where the computing of the final action takes
no extra time, once the action-taker has collected the information needed to do so. On the other hand, if
we insisted on measuring computing time, then we would need a detailed model of computing, such as
those studied in the Van Zandt/Radner and Mount/Reiter work.
246 T. Marschak
all the information about w, x, y that is needed in order to compute

the goal function. Consider a mechanism in which Person 1 observes
w, x, y and then sends the number H (w, x, y) to Person 2, the action
taker, who observes z and is able to compute the action F(w, x, y, z)
once he receives the number H. After three periods, Person 1 is done,
and the action taker knows z. It takes just one more period for the
action taker to learn H. So delay is indeed four. The key issue is
whether the functions G, H exist.
(3) Is there a two-person F -realizing mechanism in which neither
person has a burden more than 3 and one person has a burden less
than 3?
Additional questions about F can be posed. In trying to answer them, it

is clear that we have to restrict the F-realizing mechanisms that we are
permitting. Once again the issue of smoothness is inescapable. If we per-
mitted the sort of smuggling of many numbers into one number that we
have already discussed (in Section 2.2.8), then no message need ever contain
more than one number, and so we would obtain trivial and uninteresting
answers to our questions. A workable definition of a smooth mechanism is
that its sending functions rij be differentiable on their domains. Weaker and
stronger definitions can be explored as well.
What is known about efficient speak-once-only mechanisms and what
might one hope to learn? To organize the remarks we now make, we shall
use the above illustration, and its accompanying three questions. For each
question, we consider the tools available to answer it, as well as the answer
itself and its possible generalization.
2.7.5 General remarks suggested by Question (1)

For a goal function F that is sensitive to all k external variabes, it is
obviously true that a one person F-realizing mechanism has a tree structure,
the sole person’s burden is k, and the delay is also k. A nontree F-realizing
mechanism may improve on the one-person mechanism with regard to de-
lay if the observing of the external variables is split among several persons
who do their observing simultaneously, provided that the reduction in
observing time exceeds the message-reading time.
2.7.6 General remarks suggested by Question (2)

Let us relabel the external variables as x1, y , xm, y1, y , yn. Given a real-
valued differentiable goal function F (x1, y , xm, y1, y , yn), the general
aggregation question is as follows:
Is there some neighborhood on which we can aggregate the m variables
x1, y , xm into rom variables, i.e., do there exist a neighborhood U
Rmþn ; and rom real-valued functions G, H1, y , Hr which are differentiable

on U, such that on U we have
ðþÞ F ðx1 ; . . . ; xm ; y1 ; . . . ; yn Þ
¼ GðH 1 ðx1 ; . . . ; xm Þ; . . . ; H r ðx1 ; . . . ; xm Þ; y1 ; . . . ; yn Þ?
An important contribution to answering this question is a theorem of

Abelson (1980). Let Fi denote the partial derivative of F with respect to xi.
Abelson’s theorem is as follows:
There exists a neighborhood U
Rmþn and differentiable functions G;
H 1 ; . . . ; H r ðwhere romÞ; which satisfy ðþÞ on U if and only if at
most r of the functions F i are linearly independent on U:
Checking the linear independence of the Fi, is not, in general, a straight-

forward matter. But a technique closely related to the Abelson theorem
provides answers to the aggregation question in certain cases. To introduce
the technique, let us use the following notation:
For an n-tuple of nonnegative integers a ¼ (a1, y , an), and for
y ¼ (y1, y , yn) 2 Rn ; let the symbol ya denote the array of symbols
ya11 yann : (This array will be used to identify partial derivatives of
varying orders.) The symbol|a|denotes the sum a1+ ? +an.
The symbol D (F, xt, ya) denotes the following partial derivative of
order 1+|a|:
@1þjaj F
.
@xt @ya11 . . . @yan
n
The symbol D (G, z) denotes the partial derivative @G=@z:

Now let us associate with the goal function F the Hessian
0 1
DðF ; x1 ; y1 Þ . . . DðF; x1 ; yn Þ
B .. .. .. C
HðF Þ B@ . . . C,
A
DðF; xm ; y1 Þ . . . DðF ; xm ; yn Þ
and the bordered Hessian
0 1
DðF; x1 Þ
B ..
BHðF Þ B . HðFÞ C
A
@

DðF ; xm Þ
248 T. Marschak
For easy reference, note that

The m rows of BH may be indexed by the m variables x1, y , xm that are
being aggregated. The first of the n+1 columns may be indexed by the function F
and the remaining columns by the nonaggregated variables y1, y , yn.
We now state two propositions about aggregation.37 We call them Prop-
osition (*) and (**).
Proposition (*)38 provides bordered-Hessian conditions that are neces-
sary for the existence of functions G, H1, y , Hr satisfying (+).
Proposition (*). Let F be a C3 function. If there exist C2 functions G,
H1, y , Hr such that (+) holds on some neighborhood U, then
rank BHðF Þ r at every point of U.

Now we turn to Proposition (**), which provides conditions that are suffi-
cient for the existence of functions G, H1, y ,Hr satisfying (+). In stating
Proposition (**), we consider open subsets of Rmþn of the form W V where
W
Rm ; V
Rn : We denote points in W V by (p, q), where pAW, qAV.
Proposition (**). Let F be a Ckþ1 function, kZ2. Suppose that rank
BHðF Þ r everywhere on W V
Rmþn and rank HðF Þ ¼ r everywhere
on W. Then there exists a neighborhood UDW V and Ck functions G,
H1, y Hr for which (+) holds.
We can successfully apply Proposition
(*) to the goal function
1 1 1 1
F ¼ w+x+y+z+wxyz+ w þ x þ y þ z considered in Question 2. Any
three variables, say w, x, y, can be aggregated into two real variables, using
37
They are found in Mount and Reiter (1996) and in Appendix B of the book by Mount and Reiter.
38
The proof of Proposition (*) is as follows: If (+) holds on U, then everywhere on U we have
X
r
DðF ; xi Þ ¼ DðG; H k ÞDðH k ; xi Þ
k¼1
and
X
r
DðF ; xi yj Þ ¼ DðG; H xyj ÞDðH k ; xi Þ;
k¼1
since Pthe Hk are independent Pr of the variables yi y ,yn We Pr can therefore write 1 BHðFÞ
0 r
k¼1 DðG; H k DðH k ; x1 Þ k¼1 DðG; H k ; y1 ÞDðH k ; x1 Þ ... k¼1 DðG; H k ; yn ÞDðH k ; x1 Þ
B .. .. .. .. C
asB
@ . . . . C: So
A
Pr Pr Pr
k¼1 DðG; H k ÞDðH k ; xm Þ k¼1 DðG; H k ; y1 ÞDðH k ; xm Þ . . . k¼1 DðG; H k ; yn ÞDðH k ; xm
everywhere
0 1 U, 0each column
on 1 of BHðF Þ is a linear combination of the r column vectors
DðH 1 ; x1 Þ DðH r ; x1 Þ
B .. C B .. C
B . C; . . . ; B . C: That implies that the rank of BHðF Þ is at most r at each point
@ A @ A
DðH 1 ; xm Þ DðH r ; xm Þ
of U. ’
the C2 functions H 1 ¼ x þ x1 þ y þ 1y þ z þ 1z; H 2 ¼ xyz; G ¼ H 1 þ w þ w1 þ

wH 2 : Can they be aggregated into just one real variable, i.e., do there exist
differentiable functions G, H such that on every neighborhood in Rþ we
have F ¼ G (H (w, x, y), z)? To check this, the relevant BH is
0 1
1 1=w2 þ xyz xy
B 1 1=x2 þ wyz wy C
@ A.
2
1 1=y þ wxz wx
That has rank two on any neighborhood in which (for example) all four
variables are neither zero nor one, and no two variables take the same
value. Proposition (*) tells us that if the rank exceeds one on a neighbor-
hood, then on that neighborhood we cannot aggregate three variables into
one. There is no two-person F-realizing mechanism with differentiable
sending functions in which neither person has a burden more than two and
delay is not more than four.
On the other hand, consider the following goal function F, identical to

F except that we now raise the term w1 þ x1 þ 1y þ 1z to the power 2. Thus,

1 1 1 1 2
F ðw; x; y; zÞ ¼ ðw þ x þ y þ zÞ þ ðwxyzÞ þ þ þ þ .
w x y z
Suppose we now ask whether we can aggregate three variables into two (as
we could for F), i.e., do there exist differentiable functions G, H1, H2 such
that on every neighborhood in R4 we have F ¼ G (H1(w, x, y), H2(w, x,
þ
y), z)? Again BH has three rows (indexed by the three variables being
aggregated) and two columns (indexed by F and by the remaining variable).
Proposition (*) says that if the proposed aggregation is possible then BH
has rank at most 2. But that is the case whether or not the proposed
aggregation is possible. So Proposition (*) cannot be used to rule out the
proposed aggregation. Proposition (**) says that the proposed aggregation
is indeed possible if the rank of BH is at most 2 and at some point the rank
of H is exactly 2. The latter condition cannot be satisfied, since H has just
one column. So Proposition (**) does not tell us that the proposed aggre-
gation is possible.
More generally39:
The necessay condition of Proposition ðnÞ has no force ðit is automatically
satisfiedÞ if mon þ 1
39
The limitations of Propositions (*), (**) (and the extended Proposition (*0 ) which follows) might
lead one to explore algebraic approaches. In particular, one might seek counterparts of the Abelson
Theorem in which linear independence is replaced by algebraic independence. T functions are algebra-
ically dependent if there is a polynomial, with the T functions as its arguments, which takes the value
zero at all points in the functions’ domains. The T functions are algebraically independent if there is no
such polynomial. Such a counterpart of the Abelson Theorem might hold for rational goal functions F,
i.e., F is the quotient of two polynomials.
250 T. Marschak
ðand hence ron þ 1Þ: The sufficient condition of Proposition ðnnÞ cannot be
satisfied if nor
We can, however, generalize Proposition (*). The following generaliza-

tion is more widely applicable than Proposition (*).
Proposition (*0 ). Let F be a Ck+1 function, kZ2. Fix an integer ‘Z0.
Using our definition of expressions of the form D (F, xj, ya), consider the
matrix
0 1
DðF ; x1 ; yað1Þ Þ . . . DðF ; x1 ; yað‘Þ Þ
B .. .. .. C
M¼B @ . . . C,
A
að1Þ að‘Þ
DðF; xm ; y . . . DðF; xm ; y Þ
where, for all tA{1, y ‘}, the symbol a(t) denotes a t-tuple of nonnegative
integers with|a(t)|ok. If there exist Ck functions G, H1, y ,Hr such that
on some neighborhood
F ¼ GðH 1 ðx1 ; . . . ; xm Þ; . . . ; H r ðx1 ; . . . ; xm Þ; y1 ; . . . ; yn Þ,
then for all ‘ the matrix
0
DðF ; x1 Þ 1
B .. C
M B . M A
@

DðF ; xm Þ
has rank at most r on that neighborhood.40
While Propositions (*) and (**) did not allow us to resolve the aggre-
gation question for the function F ¼ (w+x+y+z)+(wxyz)+
ðw1 þ x1 þ 1y þ 1zÞ2 ; we can now do so, using Proposition (*0 ). To fit the
notation of Proposition (*0 ), we first relabel the four variables as x1, x2, x3,
y. Our function is:

1 1 1 1 2
F ¼ ðx1 þ x2 þ x3 þ yÞ þ ðx1 x2 x3 yÞ þ þ þ þ .
x1 x2 x3 y
40
The proof is essentially the same as the proof of Proposition (*). The existence of functions
G,H1, y , Hr on some neighborhood U implies that each column of M is a linear combination of the
column vectors
0 1 0 1
DðH 1 ; x1 Þ DðH r ; x1 Þ
B .. C B .. C
B . C; . . . ; B . C.
@ A @ A
DðH 1 ; xm Þ DðH r ; xm Þ
That implies that the rank of M is at most r on U. ’
We ask whether there exist real-valued C2 functions G, H1, H2 such þthat

F ¼ G (H1(x1, x2, x3), H2(x1, x2, x3), y) on all neighborhoods in R4 :
To apply Proposition (*0 ), we first have to choose the partial derivatives
that will make up the matrix M. It is clear that if we are going to rule out
the existence of a given number of aggregators, then we want to make
M (and hence M) as large as possible without repeating columns or in-
troducing redundant columns like those consisting entirely of zeros. It
turns out that the following matrix M suffices to resolve the aggregation
question for our function F:
0 1
DðF ; x1 Þ DðF ; x1 ; y1 Þ DðF ; x1 ; y21 Þ
B DðF ; x2 ; y1 Þ DðF ; x2 ; y21 Þ C
M ¼ @ DðF ; x2 Þ A.
DðF ; x3 Þ DðF ; x3 ; y1 Þ DðF ; x3 ; y21 Þ
If we now choose a suitable neighborhood U and perform a Gauss–Jordan

reduction on M for any point in U, we obtain the identity matrix. So M
has rank 3 and Proposition (*0 ) tells us that the proposed aggregation
cannot occur on U. The neighborhood U must be chosen so that we do not
divide by zero during the reduction process. To do so, it suffices to pick a U
that does not intersect any of the zero sets of the numerators and denom-
inators of any of the entries of M at any stage of the Gauss–Jordan
process. Since those numerators and denominators are always polynomials
in x1, x2, x3, y1, it is possible to find such a U.41
While Propositions (*), (**), and (*0 ) are useful for a variety of goal
functions, a general difficulty is the absence of a usable necessary AND
sufficient condition for a proposed aggregation. The Abelson condition is
both necessary and sufficient, but for many goal functions it is very difficult
to verify.
2.7.7 A general puzzle suggested by Question (3): Can Question (3), and
similar questions, be answered using the Abelson condition?
Consider again the two-person mechanism in Fig. 3, which realizes the
goal function F ¼ ðw þ w1 þ x þ x1 þ y þ 1y þ z þ 1zÞ þ ðwxyzÞ: For a general
discussion of two-person F-realizing mechanisms it will be useful to start
by relabeling the two persons. Let us call the action taker A and the other
person B. In the Fig. 3 mechanism, Person B observes three environment
variables and has a burden of 3, while Person A, the action taker, also has a
burden of 3. (He observes one environment variable and he receives two
41
Precisely stated, our result is as follows:
There exists a finite number of polynomials f1, y , fk in the variables x1, x2, x3, y1 such that on
any open subset U of R4 not intersecting the zero sets of f1, y , fk, there do not exist C2 functions
G, H1, H2 such that F ðx1 ; x2 ; x3 ; yÞ ¼ GðH 1 ðx1 ; x2 ; x3 Þ; H 2 ðx1 ; x2 ; x3 Þ; yÞ at every point of U.
252 T. Marschak
numbers from B.) Turning to arbitrary two-person mechanisms, including

mechanisms wherein some environment variable is observed by more than
one person, we shall argue that
F cannot be realized by a two-person mechanism in which the
ðþÞ
maximum burden is 3 and one person has a burden less than 3:
Suppose a two-person mechanism realizes F and has maximum burden

less than or equal to 3. Since the values of all four variables are needed in
order to compute F, we can claim:
the mechanism’s graph must have a path from each
ðþþÞ
environment variable to A:
Now suppose that B observes exactly two variables, say x and y. Then,
by (++), A must observe the other two, z and w. A cannot observe both x
and y (in addition to z and w) since the maximum burden is less than 3. A
cannot observe just one of x, y (in addition to z, w), for if he did, then his
burden would exceed 3, since (by (++)) he must also receive at least one
number from B. If A’s burden is less than 3, then B must send just one number.
So F may be written F (x, y, z, w) ¼ A (B (x, y), z, w), where A and B (in a
slight abuse of notation) denote scalar-valued functions computed by persons
A and B. Consulting Abelson’s theorem (for the case r ¼ 1), we see that for
2
any fixed values of x, y, say x0, y2 0, the functions F x ðx0 ; y0 ; z; wÞ ¼ 1 1=x0 þ
y0 wz; F y x0 ; y0 ; z; w ¼ 1 1=y0 þ x0 wz must be linearly dependent, where
the symbol Fj again denotes the partial derivative of F with respect to the
variable j. But that is the case if and only if the vectors ð1 x12 ; y0 Þ; ð1 y12 ; x0 Þ
0 0
are linearly dependent in a neighborhood in R2 ; which is, in turn, the case if
and only if x0 x10 y0 þ y1 ¼ 0 in a neighborhood in R2 : That, however, is
0
not true. We conclude that F cannot be realized by a two-person mechanism
with a minimal burden of 3 in which B observes exactly two variables.
On the other hand, we can easily rule out the possibility that B observes
exactly one variable. If he does, then (by (++)) A must observe the other
three. But then A has reached the maximum burden of 3 and is not able to
recieve any variable from B.
So B must observe exactly three variables, say x, y, z. Then the only way
that (+) could be violated (i.e., the only way to improve on the mechanism
of Fig. 3) would be to give A a burden less than 3. By (++), A must
observe w and since A needs to know more than w in order to compute F, A
must have a burden of at least 2. Since A cannot compute F knowing only
two environment variables, he must receive one or more numbers from B.
To keep A’s burden below 3, he must receive only a single number from B.
If that were so, we could write
Fðx; y; z; wÞ ¼ AðBðx; y; zÞ; wÞ,

where both of the functions A, B are scalar-valued. But this is again ruled
out by the Abelson criterion (for r ¼ 1), since the three functions
1 1
F x ðx0 ; y0 ; z0 ; wÞ ¼ 1 2 þ y0 z0 w; F y x0 ; y0 ; z0 ; w ¼ 1 2
x0 y0
1
þ x0 z0 w; F z x0 ; y0 ; z0 ; w ¼ 1 2 þ x0 y0 w
z0
are linearly independent on every neighborhood in R: So we have estab-
lished (+).
Looking carefully at the preceding argument, one sees that it generalizes
so as to yield the following Proposition. The statement ‘‘s of the n variables
in F variables may be smoothly aggregated into ros’’ means that F may be
written as a continuously differentiable function whose arguments are r
continuously differentiable scalar-valued functions of those s variables,
together with the remaining n – s variables.
Proposition D. Let F: R4 ! R be continuously differentiable and sensitive
to all its variables.42 Suppose that
(I) three of F’s variables can be smoothly aggregated into two;
(II) no two of F’s variables may be smoothly aggregated into one; and
(III) no three of F’s variables may be smoothly aggregated into one.
Then:
(IV) there is a two-person F-realizing mechanism which has the graph
shown in Fig. 3, is smooth (its sending functions are continuously
differentiable), and has burden 3 for both persons, and
(V) there is no smooth two-person F-realizing mechanism in which each
person has burden at most 3 and one of them has burden less than 3.
2.7.8 Goal functions for which perfect aggregation is possible

It is now clear that one faces formidable difficulties in characterizing the
efficient speak-once-only mechanisms that realize a given goal function. Is
there some general class of goal functions for which the task becomes much
easier? One such class consists of the functions for which perfect aggregation
(PA) is possible. A goal function F : E 1 E k ! R has the PA property if it
behaves like a sum: for any partitioning of the external variables, the var-
iables in each set of the partitioning can be replaced by a single scalar-valued
aggregator. If F is a sum, then the aggregator for a given set is the sum of the
set’s variables. As before, we confine attention to the case where each en-
vironmental variable is a scalar, i.e., each Ej is a subset of the real line.
Formally,
The function F : E 1 E k ! R has the PA property if for
ANY partitioning of {1,y, k}, say into sets T1,y, Ts, srk, there
42
I.e., there is a neighborhood on which all partial derivatives are nonzero.
254 T. Marschak
exist differentiable real-valued functions G, A1,y, As such that for any e in

E ¼ E1 y Ek we have
FðeÞ ¼ GðA1 ðeT 1 Þ; . . . ; As ðeT s ÞÞ
Note that each of the functions A1, y , As must itself have the PA property.
Examples of a function with the PA property are:
e1+ y +ek
e1 e2 y ek–1 ek
be1 þ...þek :
If our goal function F has the PA property, we have a strong result about
the structure of efficient F-realizing mechanisms. It says that we can confine
attention to trees if we seek mechanisms that are efficient with respect to
number of persons and individual burdens.
Proposition E. Suppose F is real-valued and has the PA property. Suppose

that L ¼ hG; ð~ rkþ1 ; . . . ~
rn Þi is an F-realizing mechanism and that the graph
G, with received-from sets Rk+1, y , Rn, is not a tree. Then there exists an
F-realizing mechanism L0 ¼ hG 0 ; ðr~0 kþ1 ; . . . ~ r0n0 Þi, with received-from sets
0 0
Rkþ1 ; . . . ; Rn0 such that
(a) G0 is a tree
(b) n0 rn
(c) for every i A{k+1, y , n0 }, we have
#R0i #Ri ,
where the symbol # means ‘‘number of elements in’’.43
The Leontieff Theorem. It is remarkable that many years ago Leontieff
(1947), inspired by the search for useful production functions in economics,
and unaware of any implications for the theory of organization, proved
a theorem that anticipated the Abelson theorem that we considered above.
(Abelson’s theorem can be viewed as a generalization of Leontieff’s.) The
theorem says that any function with the PA property is essentially a sum.44
43
The proof has two steps.
Step 1 Since F has the PA property, it is clear that given any n-node-directed tree G which meets the
requirements of the graph of a speak-once-only mechanism on the k environment variables, one can
construct an F-realizing mechanism such that (i) its graph is the tree G, and (ii) exactly one real number
is sent over each arc. (In that mechanism some functions rij may be the identity function.)
Step 2 Without loss of generality we may suppose that on any arc of G a single real number is sent. (If
more than one number is sent, then they can be replaced by an aggregator.) Now convert the given
graph G into a tree as follows: Whenever there is more than one directed path between a node i6¼n and
the node n, delete all but one of those paths. The result will be a tree, say G0 , with fewer than n nodes.
The tree retains the original action-taking node. By Step 1, we can construct a new F-realizing mech-
anism with the tree G0 as its graph and with exactly one real number sent over each arc. So the new
mechanism satisfies conditions (a), (b), and (c). ’
44
Leontieff did not use our terminology, but his theorem can be restated in the form that follows.
Proposition F (Leontieff). Let F : E 1 E k ! R be a differentiable

function with the PA property. Then there exist differentiable functions f,
f1, y ,fk such that for all (e1, y , ek) in E1 ? Ek,
FðeÞ ¼ fðf1 ðe1 Þ þ . . . þ fk ðek ÞÞ.

So the PA property is not only a very strong one (yielding Proposition
E) but it is also more restrictive than one might have supposed. One
can develop further propositions about trees that realize PA goal func-
tions, by imposing further conditions on the costs associated with the
individual burdens. For example one might attach different costs to the
‘‘observation’’ burden (the number of variables that a node i>k receives
from nodes in {1, y , k}), than the other burdens. Then one simple prop-
osition says that if observation cost is linear in the number of observed
variables, while the cost associated with the receipt of nonenvironment
variables is increasing in the number of those variables, then the only effi-
cient goal-realizing tree—and hence (by Proposition E) the only efficient
goal-realizing design—has just one person, who observes all k environment
variables.
Replacing the PA property by a more general additive-structure property.
Once we drop the strong PA requirement, it is natural to begin by studying
goal functions that lack the PA property but have an additive structure
somewhat more general than the ‘‘generalized sum’’ f(f1(e1)+ ? +fk(ek))
of the Leontieff Theorem. Let us relabel the environment components
e1, y , ek so that they become the m real variables x1, y , xm. For 1rrom,
we shall say that F : Rm ! R is r – additive on the open set S ¼ S1
? Sm in Rm ; if there exist CN functions H, H1, y , Hr, V1, y , Vr
such that for all (x1, y , xm)AS, we have
F ðx1 ; . . . ; xm Þ ¼ HðH 1 ðV 1 ðx1 þ þ V 1 ðxm ÞÞ þ

þ H r ðV r ðx1 Þ þ þ V r ðxn ÞÞÞ. ð4Þ
We shall say that F is weakly r-additive on S if
F ¼ HðH 1 ðV 1 ðx1 Þ þ þ V 1 ðxm ÞÞ; ; H r ðV r ðx1 Þ þ þ V r ðxm ÞÞÞ.

(5)
We shall say that F is minimally r-additive (minimally weakly r-additive)

on S if for any positive integer r0 or, F is r-additive on S but not r0 -additive
(weakly r-additive on S but not weakly r0 -additive). An economic interpre-
tation might be as follows. The function F gives the organizational action
appropriate to the environment (x1, y , xm). That action depends on the
aspects of the environment given by the Vi. There are rom aspects. The ith
aspect influences the appropriate action through the function Hi. We may
256 T. Marschak
suppose that the goal function F has been parsimoniously written, so that
minimality is assured.45
An Example: Suppose there are four real environment variables: w, x, y, z
and consider once again our function

1 1 1 1 2
w x y z
(6)
This function is 3-additive on any open set in which all four variables are
nonzero, since it can be written
F ¼ ðw þ x þ y þ zÞ þ expðln w þ ln x þ ln y þ ln zÞ

1 1 1 1 2
þ þ þ þ .
w x y z
The results we obtained above imply that F is minimally 3-additive.
By contrast consider our function F, which we obtain by deleting the
exponent 2 in the third term of (6). The function F is 3-additive but it is not
minimally 3-additive, since it can also be written in the 2-additive form

1 1 1 1
wþ þxþ þyþ þzþ þ ðwxyzÞ.
w x y z
2.7.9 Conjectures about the aggregation properties of goal functions with an
additive structure
What can be said about efficient speak-once-only mechanisms which
realize a goal function that has an additive structure? If certain key
conjectures were established, we could develop a substantial theory about
such mechanisms. First, let us define F : Rm ! R to be t-aggregative on the
open set S
Rm ; if for any proper subset W of {1, y , m}, there exist C2
real-valued functions G, A1, y , At, with to#W, such that for all x ¼ (xW,
xW) in S we have46
Fðx1 ; . . . ; xm Þ ¼ GðA1 ðxW Þ; . . . ; At ðxW Þ; xW Þ.
45
A possible scenario: The organization operates a production facility in each of m locations. Each
location always produces a fixed proportion of the total quantity produced by the organization. That
total quantity is the action to be chosen. The appropriate total quantity depends on total production
cost, which depends, in turn, on the amounts of each of r centrally procured inputs required for each
unit of product in each location. For every location, say location i, the r input requirements are
determined by the environmental parameter xi, which describes the current technology in location i. The
r input requirements are V1(xi), y ,Vr (xi).
Once the case of a scalar-valued additive goal function F is understood, one could turn to the case of a
vector-valued additive goal function F. That would, of course, permit more flexible scenarios, including
scenarios wherein the organization’s action is a vector, specifying (for example) an output level for each
of the m locations.
46
Here, W denotes the complement of W.
Thus, on the set S, it is enough to know xW and the values of the t
aggregators A1, y , At, in order to compute F. However we partition the m
variables into the sets W and W, we need no more than t aggregators of
the variables in W in order to compute F.
The following key conjectures remain open.
Conjecture 1. If F is symmetric and r-additive on S and t-aggregative on
S, with torom, then F is also t-additive on S.
Conjecture 2. (stronger than Conjecture 1). If F is symmetric and
t-aggregative on S, with tom – 1, then F is t- additive on S.
Weak versions of these conjectures:
Put ‘‘weakly’’ in front of ‘‘r-additive’’ and ‘‘t-additive’’
Suppose F is symmetric, and minimally r-additive with ron1. Suppose
Conjecture 1 holds. Then we can claim that if we want to aggregate
x ¼ (xW, xW) in S, with ro#W, into as few variables as possible, while
retaining our ability to compute F, we cannot do better than to use
GðA1 ; . . . ; Ar ; xW Þ

P P (7)
¼ H H 1 ðA1 þ V 1 ðx‘ ÞÞ þ þ H r ðAr þ V r ðx‘ ÞÞ ;
‘2W ‘2W
where X
Ai V i ðx‘ Þ; i ¼ 1; . . . ; r: (8)
‘2W
If it were possible to aggregate the variables xW into fewer than r numbers—

i.e., if there existed G, A1, y , Aq, with qor, such that F ¼ G(A1(xW), y , Aq
(xW), xW)—then Conjecture 1 says that F is q-additive. But that contradicts
the assumed minimal r-additivity of F. Note that the converse of Conjecture
2 holds. If F is t-additive, then we can write F in a t-aggregative form, using
(7) and (8) (with r ¼ t). Note also that Conjecture 2 is correct for the case
m ¼ 3. To see this, call the three environment variables x, y, z. Since
m1 ¼ 2, Conjecture 2 takes the following form:
If F is symmetric and there exist C2 functions G, A such that
Fðx; y; zÞ ¼ GðAðx; yÞ; zÞ
then F is 1-additive.
But that is just the simplest (three-variable) version of the Leontieff
Theorem.
A modest proposition about mechanisms with a tree structure. Consider
again our function

1 1 1 1
F ¼ wþ þxþ þyþ þzþ þ ðwxyzÞ.
w x y z
258 T. Marschak
As we saw in Fig. 3, F can be realized by a two-person mechanism whose

graph is a six-node tree. Each person’s burden is 3. There is, however, no
F-realizing mechanism whose graph is a tree with all the individual bur-
dens being two or less. Our discussion of the application of Proposition (*)
to F showed that while F is 2-aggregative it is not 1-aggregative. (If we
choose the set W to contain three variables, then it is not the case that a
single aggregator provides the information about those variables that is
needed to compute F). Generalizing from the example of F and the Fig. 3
tree, we can obtain a rather modest proposition:
Proposition G. Suppose the function

F : Rk ! R,
where k>2, is 2-aggregative but not 1-aggregative. Then in any two-
person tree which realizes F, each person’s burden is at least 2.
Research challenge #6. Find interesting classes of goal functions F for
which Conjectures 1 and 2 hold.
2.7.10 Computing a speak-once-only mechanism’s delay

A mechanism’s delay is the total elapsed time until the action taker
has completed his computation of the organization’s final action. We
have considered delay informally, and we have illustrated delay, but we
have not yet discussed how one might compute it. To do so we first define
delay more carefully. We start by making the following simplifying
assumption.
(a) The environment variables change their values every G time units. For all the
speak-once-only mechanisms we consider, G is sufficiently large that the
mechanism’s delay is less than G.
We repeat, in addition, three assumptions already made in our
introductory discussion of delay in Section 2.7.1.
(b) It takes one time unit for a person to receive or observe one (scalar-valued)
variable.
(g) A person receives all the message variables that the mechanism specifies, and
does all the observing required of him, before he sends messages or computes
the final action.
(d) It takes zero time for a person to compute a message variable to be sent
onward, and for the action taker to compute and put into force the final
action.
Thus if a person observes X (scalar-valued) environment variables,
receives Y (scalar-valued) message variables from other persons, and
computes Z scalar-valued functions of the observed environment variables
and the received message variables (to be sent forward to other

persons), then X+Y time units are required for him to complete these
tasks. He is silent until the mechanism has operated for at least X+Y
time units.
Next, note that for a given speak-once-only mechanism, the sequence
of tasks may be subject to choice. Suppose, for example, that in a
given three-person mechanism Person 1 observes three environment
variables, Person 2 observes three other environment variables, and
each sends forward to Person 3 (the action taker) values of two scalar-
valued functions. In one possible sequence, Person 1 first observes all
his environment variables, then Person 2 observes all of his, and immedi-
ately after that (with no further time required), both persons compute
and send forward the values of their scalar-valued functions. That would
appear to be an inefficient sequence, since time would be saved if 1 and 2
did their observing simultaneously. Nevertheless it is a possible sequence.
In general, we may define a protocol p for a given mechanism as a
specification of exactly what tasks (observing, sending, receiving, com-
puting) each person performs at each time in a sequence of times (one
time unit apart), terminating with the action taker’s final-action compu-
tation. A protocol is required to respect the above assumption g. (A formal
definition of a protocol can be provided, but it requires an array of new
symbols.)
It will be useful to use the term entity for a node of the mechanism’s
directed graph, whether that node identifies an environment variable or a
person. For the protocol p, we may define a time-to-completion function fp
on the set of entities. For person p, the number fp (p) is the shortest time
after which person p ‘‘does nothing,’’ i.e., he does no further observing,
receiving, sending, or computing. If the entity j is an environment variable
(i.e., it is a node belonging to the set {1, y , k}), then we define fp (j) to be
zero. If the protocol p is used, then the total elapsed time until the mech-
anism terminates is fp (n) where (as before) n is the action taker. Our
defintion of delay is then as follows:
Definition. A speak-once-only mechanism’s delay is the minimum value,

over all protocols p, of fp (n). A protocol p for which fp (n) is minimal will
be called a minimal protocol for the mechanism.
A mechanism’s delay may be computed recursively. To simplify notation,

let T denote the function fp, where p is minimal, and write Tp instead of
T(p). (Thus Tp is the smallest time until p does nothing further.) We shall
use the term ‘‘listens to’’ to cover both observation of an environment
variable and receipt of a message variable. Suppose p listens to just one
entity, say 1, from whom he receives N1 variables. Then Tp ¼ T1+N1, since
260 T. Marschak
p has to wait T1 units for 1 to finish and then takes N1 units to listen to the
N1 variables sent to him by 1. At that point, p is done. Now suppose that p
listens to more than two entities. We make the following claim:
Claim. There is a minimal protocol in which p listens to each of the
persons from whom he receives in the order in which those persons are
done with their sending to him.
To argue this claim, start by supposing that p listens to two persons, say 1
and 2; 1 sends N1 variables to P and 2 sends N2 variables. Assume that
T1rT2. Then there is a minimal protocol in which
T p ¼ maxðT 1 þ N 1 ; T 2 Þ þ N 2 .
To see this, note first that p must wait T1 time units for 1 to finish. He
spends the next N1 time units listening to 1. If T2 has elapsed by then, (i.e.,
T2rT1+N1), then p spends the next N2 time units listening to 2. If T2 has
not yet elapsed (T2>T1+N1), then p must wait for T2 units to elapse and
then spends the next N2 units listening to 2. If p listens to three persons, say
1,2,3, with T1rT2rT3, then we obtain, for some minimal protocol:
T p ¼ maxðmaxðT 1 þ N 1 ; T 2 Þ þ N 2 ; T 3 Þ þ N 3 .
Recalling that Tp ¼ 0 if the node p corresponds to an environment variable,

we now see that there is some minimal protocol such that for each person p,
the following procedure computes the number Tp for that protocol.
(1) Label the entities that p listens to as 1, 2, y , m, where

T1rT2r ? rTm.
(2) Initialize values u ¼ T1+N1 and k ¼ 1. If k ¼ m, then set Tp ¼ u and
stop.
(3) Replace u with u ¼ max (u, Tk+1)+Nk+1 and replace k with k+1.
(4) If k+1om, repeat step (3). If k+1 ¼ m, then set Tp ¼ u and stop.
Note that the recursive scheme just described is indeed realized by some
protocol for a given mechanism, since the four steps can clearly be carried
out for a person p who observes environment variables but does not receive
messages. Since a speak-once-only mechanism has no directed cycles, it
follows that the four steps can also be carried out for a person p who
receives messages. Since we can, in particular, carry out the four steps for
p ¼ n (the action taker), the four steps provide a recursive way of com-
puting the delay; the delay is Tn ¼ fp(n) where p is a minimal protocol.
Roughly speaking, the steps insure that each person is ‘‘as busy as pos-
sible.’’ Each person begins the work of listening as soon as signals (envi-
ronment variables or message variables) begin to arrive, and no person
stays idle while there is work (listening) that he could be doing. (Assump-
tion (g) is crucial for this claim.)
Some preliminary computer experiments with the recursive procedure
suggest that typically the time required to compute Tn rises rapidly as n and
k grow. It seems likely that one can construct improved versions which
exploit the properties of narrow classes of goal functions.
Research challenge # 7. For interesting classes of goal functions, refine

the delay-computation algorithm so that it performs rapidly for mech-
anisms that realize a function in the class.
Examples of the computation of delay

(a) In a one-person mechanism, where the action taker observes all en-
vironment variables, delay is just the number of those variables.
(b) Suppose there are k ¼ 2M environment variables, where M is a pos-
itive integer not less than 2. Suppose the mechanism’s graph is a binary tree
with a person at each nonleaf node. Then delay is 2 log2 k ¼ 2 M. Similarly
if k ¼ tM, where t and M are integers greater than 2, and the mechanism’s
graph is a t-ary tree with a person at each nonleaf node, then delay is t logt
k ¼ tM. (Since each person listens to t variables, it takes t time units for all
the persons at a given tier of the tree to finish, and all persons in the tier
finish at the same time; there are M tiers, so delay is tM.) For such a
mechanism no protocol can achieve a delay less than tM. In contrast, a one-
person mechanism has a delay of k ¼ 2M as well as a burden of 2M.
(c) Consider once again the goal function

1 1 1 1 2
w x y z
Consider the following two nontree mechanisms that realize F. For each
of them delay can be checked visually or it can be found using our recursive
four-step procedure.
The six-person Fig. 4 mechanism has a property of ‘‘U-forms’’ in the
economic literature, briefly mentioned in Section 1. Each environment var-
iable is observed by just one person (a specialist). Persons 1 and 2 are the
observers. They report to a ‘‘sum’’ specialist (Person 3), a ‘‘product’’ spe-
cialist (Person 4), and a ‘‘reciprocal’’ specialist (Person 5). Those three
specialists report, in turn, to the action taker (Person 6), who then has all
the information needed to compute F.
The five-person Fig. 5 mechanism has a property of ‘‘M-forms’’ in the
economic literature: each environment variable is separately observed by
several persons. Persons 1,2, and 3 all observe w, x, y. For those variables,
Person 1 is a ‘‘sum’’ specialist, Person 2 is a ‘‘product’’ specialist, and
Person 3 is a ‘‘reciprocal’’ specialist. Person 1 reports directly to Person 5
262
T. Marschak
Fig. 4. This is a six-person nontree F realizing mechanism with delay 7. It has the ‘‘U-form’’ property: each environment variable is
observed by a single person.
Ch. 4. Organization Structure
Fig. 5. This is a five-person nontree F realizing mechanism with delay 6. It has the ‘‘M-form’’ property: each environment variable is
observed by more than one person.
263
264 T. Marschak
(the action taker), but Persons 2 and 3 report to an intermediary (Person 4),
who also observes z directly, as does Person 5. We find that
the Fig. 4 mechanism uses 6 persons. Five of them have a burden of 2
and one of them (the action taker) has a burden of 3. Delay is 7.
the Fig. 5 mechanism uses 5 persons. Each has a burden of 3. Delay47 is 6.
The comparison illustrates a general conjecture:
Conjecture y. Consider a goal function F which is a sum of three scalar-
valued functions of four environment variables and cannot be expressed
as a function of fewer than three scalar-valued functions. There does not
exist a nontree F-realizing mechanism which dominates all the other
nontree F-realizing mechanisms with regard to number of persons, in-
dividual burdens, and delay.
Research challenge # 8. Apply the recursive delay-computation algorithm
to establish Conjecture y and similar conjectures.
3 Models in which the designer is concerned with incentives as well as

informational costs
A gigantic literature considers the designing of schemes that induce the self-
interested members of an organization to make choices that meet a given
organizational goal. One branch of this literature studies contracts. Another
studies ‘‘implementation’’; it considers schemes in which each self-interested
member chooses an individual message, knowing that an organizational action
will be chosen once all the messages are sent, knowing the outcome function
that will be used to assign an action to the individual messages, and knowing
the impact of each action on his own personal welfare. Such a scheme defines a
message-choosing game. If the scheme is properly designed then it ‘‘imple-
ments’’ a goal correspondence that has been given to the designer: at an
equilibrium of the game, the messages chosen are those for which the assigned
action (prescribed by the outcome function) is goal-fulfilling.48 Contracting
schemes and goal-implementing mechanisms have informational costs, and it
would be of great interest to be able to find the cheap ones. Informational costs
are often discussed informally. Indeed some of the classic goal-implementing
mechanisms (e.g., those that lead the organization’s self-interested members to
make the correct choice as to the quantity of a public good) are striking in their
47
At the end of the first three periods, 1,2, and 3 have each learned w, x, y, while 4 and 5 have
learned z. No messages have yet been sent. In the fourth period, 5 receives w+x+y from 1, while 4
receives wxy from 2. In the fifth period, 4 receives 1/w+1/x+1/y from 3. In the sixth period, 5 receives
2
wxyz þ w1 þ x1 þ 1y þ 1z from 4 and is now able to compute F.
48
Two excellent introductory surveys of the implementation literature are: Serrano (2004) and Jackson
(2001).
informational simplicity.49 But it is very rare for a paper to present a goal-

implementing mechanism and then to argue its informational minimality in a
precise way. The managerial accounting literature studies mechanisms or con-
tracting schemes to be used by a profit-maximizing principal and the self-
seeking ‘‘responsibility centers’’ (cost or profit centers) of a firm. The infor-
mation transmitted by the centers (perhaps by using transfer prices) is far more
modest than, for example, that transmitted in a DR mechanism which achieves
the same result. But again, informational minimality is not formally shown.50
Occasionally, however, papers in this literature explicitly measure an infor-
mational cost, e.g., the cost of reducing the variance in the noise that accom-
panies a signal sent from the agents to the principal.51
We now turn to some research in which initial steps toward incorporating
informational costs as well as incentives are taken.
3.1 The message-space size required for implementation of a goal
Suppose, as before, that each person i has a privately observed local

environment ei whose possible values comprise the set Ei. In most imple-
mentation discussions, ei determines i’s preferences over the possible
organizational actions (and may contain other information about the or-
ganization’s environment as well). Then an n-person implemen-
tation scheme, often called a game form, has two elements: an n-tuple
S ¼ (S1, y ,Sn) of individual strategy spaces, and an outcome function h:
S-A, where A is the set of possible organizational actions. In particular, a
strategy siASi may be a rule that tells i how to behave at each stage of a T-
step message-announcement process, where the organization’s action is a
function of the step-T announcements. It tells him what message mi, in a set
Mi of possible messages, to announce at step t, given what he knows about
the announcements thus far and given his current ei. The organizational
action is therefore a function of the current e ¼ (e1, y , en). Given his
current ei, and given any strategy choices by the others, person i (who
knows the function h) is able to rank any two of his own strategies with
regard to his own welfare. For each e, we therefore have a game, denoted
Ge, and we may choose a solution concept—say Nash equilibrium—to
identify the game’s equilibria. The scheme /S, hS implements a goal co-
rrepondence G from E ¼ E1 ? En to A if h (s)AG (e) whenever
s ¼ (s1, y , sn) is a Nash equilibrium of the game Ge.
49
Groves and Ledyard (1977) is the classic paper on public-good provision in an organization whose
members differ with regard to their private valuation of the good , using a mechanism that appears to be
informationally cheap, although no formal claim about its informational minimality is made. By con-
trast, informational cost is formally studied in a ‘‘nonincentive’’ paper by Sato (1981). That paper finds a
lower bound to the message space size for mechanisms used by an economy which seeks an efficient
allocation of resources to the production of public goods, if we assume that the agents voluntarily follow
the mechanism’s rules.
50
See, for example, three papers by Melumad et al. (1992, 1995, 1997).
51
See, e.g., Ziv (2000).
266 T. Marschak
Now consider the equilibrium strategy profiles of the game Ge. As before,
let m denote (m1, y , mn) and let M denote M1 ? Mn. Define the
correspondence
mi ðei Þ ¼ fm 2 M : for some e 2 E,
m is the step-T announcement for ðs; e Þ,
s is an equilibrium of Ge ; and ei ¼ ei g.
Let h~ be an outcome function from M to A, with the following property:
~
h(m) ¼ a if there is an S such that (1) a ¼ h (s) and (2) for some eAE, s is
an equilibrium strategy profile of the game Ge and m is the step-T message for
~ is a (privacy-preserving) mechanism on
(e, s). The triple /M, (m1, y , mn), hS
E in our previous sense. Moreover it realizes the goal correpondence G.
One can ask: among the G-implementing mechanisms (obtained from a
G-implementing game form in the manner just described), which ones have
a minimal message-space size? Or, less ambitiously, is there a useful lower
bound to such a mechanism’s message-space size? The message-space re-
quirements for implementation of a goal correspondence are certainly going
to be harsher, in general, than the requirements for realization alone. ‘‘
How much harsher?’’ is a difficult question, and very few papers have
addressed it.52
3.2 Models in which the organization’s mechanism is partly designed by its

self-interested members, who bear some of the informational costs
3.2.1 A model in which the organization’s decentralized self-interested

members choose their search efforts, a ‘‘Decentralization Penalty’’ results,
and the effect of improved search technology on that penalty can be traced
Consider a three-person organization53 consisting of two Managers,
called 1 and 2, and a Headquarters (HQ). There is a changing external
environment e ¼ (e1, e2). Manager i, i ¼ 1, 2, observes ei, whose set of
possible values is Ei. HQ does no observing. Whenever there is a new value
of e ¼ (e1, e2), HQ learns something about it from the managers, though
their knowledge of the new e is imperfect. Having received reports about the
new e from the managers, HQ chooses an action a. The organization then
collects a payoff W (a, e). [Example: The organization makes a product
which it markets in two locations. The variable ei is the product’s price in
location i next week. The action a has two components a1 and a2, where ai
is the amount to be shipped to location i for sale next week. The payoff W
(a, e) is the profit earned next week.]
Manager i always learns something about a new ei in the same way. He
has a finite partitioning of Ei, which he has chosen once and for all out of
52
Three of them are Reichelstein (1984), Reichelstein and Reiter (1988), and Tian (1990).
53
Studied in Marschak (2004).
some set of available partitionings. Whenever ei changes, he conducts a

search to find that set of his partitioning which contains the new ei. Let s1,
s2 denote the two chosen partitionings. Manager i reports the set he finds,
say SiAsi, to HQ. Then HQ chooses an action â (S1, S2) which maximizes
the conditional expectation E(W (a, e)|e1AS1, e2AS2), where the symbol e
again denotes expectation. Let O ^ (S1, S2) denote the value of that condi-
tional expectation when the maximizing action â is used. The highest at-
tainable expected gross performance of the two chosen partitionings, which
we simply call gross performance (for brevity), is the expected value of O;^
where the expectation is taken over all the possible pairs (S1, S2). Our
symbol for gross performance is just O. Thus
^ 1 ; S 2 Þ.
Oðs1 ; s2 Þ ¼ Eðs1 ;s2 Þ2s1 s2 OðS
Even though the technology of search may be advanced, the managers’
search has a cost, since it requires human time and human expertise. We
suppose that the cost of searching the partitioning si is measured by y C(si),
where y and the values taken by the function C are positive real numbers.
When search technology improves, y drops. Consider two definitions of C(si).
The first is simply the number of sets in si. The second is the ‘‘Shannon
content’’ of si, i.e.,
X
ðprobability of Si Þ ðlog2 ðprobability of S i ÞÞ:
si 2si
The number-of-sets measure ignores the fact that some sets occur more
frequently than others. But it is an appropriate measure if the searchers who
assist the manager have to maintain their ability to distinguish between the
sets. That may require substantial training, and the number-of-sets cost
measure may be viewed as the opportunity cost of the investment made in
such training. As the technology of search improves, the training required
to distinguish among a given number of sets becomes less costly. On the
other hand, the Shannon content is sensitive to the set probabilities. Using
the most elementary of the theorems in the Information Theory which
Shannon founded (the noiseless coding theorem),54 one can show that if y is
small then the partitioning’s Shannon content approximately equals the
average number of steps required when a searcher follows a well-chosen
binary tree to find the set in which his current ei lies. Then y times the
Shannon content is (approximately) the average amount of dollars paid for
searcher time when search is efficient. That drops when search technology
improves (i.e., when y drops).55
54
See, for example, Abramson (1963, pp. 72–74).
55
Once the two partitionings s1,s2 have been specified, they define—using our previous terminology—
a speak-once-only mechanism with HQ as the action taker. For every e, the action-taker computes the
^ 1 ; S2 Þ when e1AS1,e2AS2.
function F (e) which takes the value aðS
268 T. Marschak
We now compare a centralized organization with a decentralized one.

Here we use the term ‘‘decentralized’’ in a new way, appropriate to the
discussion of incentives. In the decentralized organization the managers are
self-interested and are free to make choices that were not available to them
in the centralized case. In the centralized case, monitoring (or perhaps
socialization) insures that each manager strikes the correct balance between
his search cost and the gross performance that his search permits. So the
chosen partitionings, say s1 ,s2 are first best. They maximize
Oðs1 ; s2 Þ y Cðs1 Þ y Cðs2 Þ.
In the decentralized case, there is no monitoring and each self-interested
Manager i is free to choose his preferred si. He bears the associated cost,
namely yC(si). But he is rewarded with a share of the expected gross per-
formance O. So the two decentralized managers play a sharing game in
which Manager 1 chooses s1, Manager 2 chooses s2, and each Manager i
collects an (average) payoff equal to
r Oðs1 ; s2 Þ y Cðsi Þ,
where r, each manager’s share, is greater than zero and not more than 12: In
the decentralized case, the chosen partitionings comprise a Nash equilib-
rium of the sharing game.
The sharing-game interpretation of decentralization seems very natural if
one seeks a workable and plausible model. ‘‘Profit-sharing’’ is, after all, the
oldest (and simplest) of the schemes that one sees when one looks at de-
centralized real-world attempts to reconcile individual incentives with or-
ganizational goals. Such schemes will not perform as well as first-best
choices would, but they may be the best practical alternative if the only way
to ensure first-best choices is to engage in intrusive policing, or to adopt (in
a real-world setting) some version of the sophisticated but informationally
costly monitoring schemes that theorists have proposed.
Now let (sy1, sy2) denote a pair of decentralized partitionings, chosen at an
equilibrium of the sharing game. Our central interest is the Decentralization
Penalty when y is the level of search technology. The Penalty associated with
s1 , s2 ,sy1, sy2 is
PðyÞ ¼ ½Oðs1 ; s2 Þ yCðs1 Þ yCðs2 Þ
½Oðsy1 ; sy2 Þ yCðsy1 Þ yCðsy2 Þ.
In a shirking equilibrium of the decentralized sharing game, the managers’
total search expenditures are less than the first-best expendures. In a squan-
dering equilibrium the reverse is true. It turns out that under plausible as-
sumptions a search technology improvement (a drop in y) decreases the
Decentralization Penalty associated with a squandering equilibrium, but its
effect on the penalty associated with a shirking equilibrium is not so clear-cut.
First consider a ‘‘short-run’’ setting, where a finite collection of possible

partitionings is available to each manager, and the technology parameter y
changes but its values lie in some interval [R, S], where 0oRoS. The
interval is sufficiently small that when y drops within that interval there is
no change in the decentralized managers’ partitioning choices or in the first-
best partitionings. For all y in the interval [R, S], the former partitionings
are sy1,sy2 and the latter are s1 ,s2 . We can rearrange the expression for the
associated Decentralization Penalty to obtain:
PðyÞ ¼ Oðs1 ; s2 Þ Oðsy1 ; sy2 Þ þ y½ðCðsy1 Þ þ Cðsy2 ÞÞ ðCðs1 Þ þ Cðs2 ÞÞ.
Examining the term in square brackets, we immediately see that if squan-

dering occurs at the decentralized partitionings sy1,sy2, then a drop in
y decreases the Decentralization Penalty, but if shirking occurs at the de-
centralized partitionings, then a drop in y increases the Decentralization
Penalty.
Now let us turn to a ‘‘long-run,’’ where an infinity of possible partition-
ings is available and, in addition, the first-best and decentralized (sharing-
game equilibrium) partitionings change whenever y changes. Let Di(y)>0
identify Manager i’s first-best partitioning and let Diy (y )>0 identify the
decentralized Manager i’s partitioning at an equilibrium of the sharing
game. Let y C (D) be the cost of the partitioning D, where C is increasing.
Then for a given y, the Decentralization Penalty is

PðyÞ ¼ OðD1 ðyÞ; D2 ðyÞÞ yCðD1 ðyÞÞ yCðD2 ðyÞÞ

OðDy1 ðyÞ; Dy2 ðyÞÞ yCðDy1 ðyÞÞ yCðDy2 ðyÞÞ .
Suppose the functions Di, Diy are differentiable.56 For every y, the first
best partitionings Di have to satisfy the first-order condition for the max-
imization of OyC1yC2, and the decentralized partitionings Diy have to
satisfy the first-order condition for D1 to be a best reply (in the sharing
game) to D2, and vice versa. When we compute the derivative P0 , while
inserting the first-order conditions, we obtain:
h i
P0 ðyÞ ¼ CðDy1 ðyÞÞ þ CðDy2 ðyÞÞ CðD1 ðyÞÞ þ CðD2 ðyÞÞ

1 y0 0 y y0 0 y
y 1 D1 ðyÞC ðD1 ðyÞÞ þ D2 ðyÞC ðD2 ðyÞÞ .
r
56
The differentiability assumption is made for analytic convenience. Typically one would want each of
the available partitionings to be identified by a positive integer T. (Thus Manager i’s external variable ei
might have the interval [A, B] as its support, with 0oAoB, and T might identify the partitioning in
which that interval is divided into T subintervals of equal length.) Under very plausible assumptions
about O, the finding that the derivative P0 is negative (when we treat Di,Diy as continuous and differ-
entiable rather than integer-valued) implies that for sufficiently small y, P remains a decreasing function
when we restrict the functions Di,Diy to integer values.
270 T. Marschak
Here primes denote derivatives and r, with 0orr1/2, is each manager’s

share of the gross performance O in the sharing game which the managers
play when they are decentralized. Now suppose we know the following:
8
>
> when search technology improves ðy dropsÞ;
>
>
>
> the decentralized managers
>
<
ð#Þ spend more on search; i:e:; costlier partitions
>
>
>
> are chosen at the equilibria
>
>
>
: of the sharing game
0
Then Dyi ðyÞo0; i ¼ 1; 2 for all y>0. Since C is increasing and 1/r1>0, the
second of the two terms in square brackets is negative or zero. If there is
squandering at the decentralized partitionings, then the first of the two terms
in square brackets is positive and hence the entire expression for P0 is positive,
i.e., the Decentralization Penalty shrinks when IT improves. If there is shirk-
ing, then the sign of P0 is indeterminate unless we make further assumptions.
Some of the scarce empirical literature on the effect of IT improvement
on organizational structure suggests that as IT improves, the case for ‘‘de-
centralizing’’ an organization becomes stronger.57 The model of managers
who engage in search is a first attempt to see why this might be so. If the IT
improvement in question is the explosive growth of efficiently searched
databases, and if ‘‘decentralized’’ managers can reasonaby be modeled as
players of a sharing game, then it is of considerable interest to find the
conditions under which improved search technology indeed implies a drop
in the Decentralization Penalty. An explicit formula for the Penalty has
been found for certain cases.58 But there are also examples of a W and a
probability distribution on the ei in which the decentralized sharing
game has a squandering equilibrium.59 The effect of a drop in y on
the Decentralized Penalty associated with squandering is strikingly unam-
biguous, both in the ‘‘short-run’’ setting, and (provided the statement
(#) holds) in the long-run setting as well. So it would be valuable to un-
derstand the squandering phenomenon much better. For a very wide class
57
See the papers mentioned in the Section 1: Bresnahan et al. (2000, 2002).
58
Suppose, in particular, that the payoff function W has the linear/quadratic structure discussed in
Section 2.6, where the techniques of the Theory of Teams were considered. For that W, person-by-
person-satisfactoriness is both necessary and sufficient for a team action rule to be best for a given
information structure. That allows us to find the best action rules used by HQ once each manager tells
HQ the set in which his current local environment lies. Suppose further that each ei is uniformly
distributed on a closed interval, that each Manager i divides his interval into Ti subintervals of equal
length, and that each choice of Ti defines one of the manager’s possible partitionings. It then turns out
that (1) in the unique equilibrium of the sharing game, each manager shirks (chooses a Ti that is lower
than the first-best Ti), (2) in the ‘‘long-run’’ setting (where a small change in the technology parameter y
leads to new equilibrium values of the Ti as well as new first-best values) the Decentralization Penalty
indeed drops when y drops. That is true for the number-of-sets cost measure as well as the Shannon cost
measure.
59
In the examples found so far, there is a finite set of possible values for each ei.
of sharing games a sufficient condition for ruling out squandering is com-

plementarity: an increment in Manager 1’s search expenditures improves
(or does not damage) Manager 2’s ‘‘marginal productivity,’’ i.e., the incre-
ment in O due to an increment in Manager 2’s search expenditures.60 On
the other hand, a class of functions O has been found for which the
following is true: if O exhibits sufficiently strong ‘‘anticomplementarity’’,
then there are squandering equilibria in the sharing game defined by O
and the shares ri. (Anticomplementarity means that an increment
in Manager 1’s search expenditures damages Manager 2’s marginal pro-
ductivity.)
In an informal way, one can imagine that the much discussed but seldom
modeled phenomenon of ‘‘information overload’’ can indeed lead to an-
ticomplementarity and to squandering. As Manager 1 increases his search
effort and the level of detail in his report to HQ, he greatly diminishes the
value of Manager 2’s report, because Manager 2’s report then becomes
somewhat redundant, i.e., its marginal contribution is small. The managers
might then be trapped in an equilibrium where Manager 1’s partitioning is
very costly, Manager 2’s partitioning is cheap, and the sum of the costs is
higher than the cost sum for a first-best partitioning pair. A rough con-
jecture would be as follows:
Conjecture (*). For sets GDE1, HDE2, let aðG; ^ HÞ denote a value of a
which maximizes the conditional expected value of the payoff W (a, e1, e2),
given that e1AG, e2AH. Given the probability distribution of e ¼ (e1, e2),
the payoff function W, and partitionings s1, s2 let us measure the marginal
contribution of s2 by

^ 1 ; Q2 Þ; e1 ; e2 Þje1 2 Q1 ; e2 2 Q2 Þ
EQ1 2s1 ;Q2 2s2 EðW ðaðQ

EQ1 2s1 EðW ðaðQ^ 1 ; E 2 Þ; e1 ; e2 Þje1 2 Q1 Þ :
Then if the marginal contribution of s2 is ‘‘sufficiently small but not too

small,’’ (s1,s2) is a squandering equilibrium of the sharing game.
Here ‘‘but not too small’’ seems appropriate, since if s2 makes an
extremely small marginal contribution, then Manager 2 finds that when
60
Consider n-person sharing games in which each person i chooses a strategy xi, bears the nonde-
creasing cost ci (xi), and receives a reward Ri (z) when the organization earns the revenue z ¼ A (x1, y ,
xn), where A is nondecreasing in each of its arguments. Suppose the functions Ri obey a ‘‘nondecreasing
residual’’ property: when z increases by any positive amount D, the sum of the rewardsPdoes not rise by
more than D. (That is satisfied, for example in the ‘‘balanced-budget’’ case where z ¼ Ri(z) as well as
the ‘‘constant-share’’ case Ri(z) ¼ ri z which we have been considering.) Suppose A obeys comple-
mentarity: if we look at the increment in A when any player spends more, we find that the increment
does not drop when some other player spends more. Then in any equilibrium no player squanders
relative to an efficient (first-best) (x1, y , xn), where A minus the cost sum is maximal. Every player
spends the efficient amount or less. The argument holds whether the strategy sets are finite or are
continua. The argument is provided in Courtney and Marschak (forthcoming).
272 T. Marschak
compared to the one-set partitioning (the cheapest partitioning), his share

of s2‘s extra revenue fails to justify its extra cost.
Research challenge #9. To better understand the squandering phenom-
enon, find probability distributions on the local environments, and payoff
functions W, such that
Condition (#) holds: when search technology improves, the decentral-
ized managers spend more on search at the sharing-game equilibrium.
(Then, in the ‘‘long-run’’ setting, a search-technology improve-
ment diminishes the Decentralization Penalty associated with squan-
dering)
Conjecture (*) holds.
3.2.2 The effect of search-technology improvement on the Coordination

Benefit
In addition to studying the Decentralization Penalty, one can conduct a
parallel investigation of the effect of improved search technology on the
Coordination Benefit. That might be motivated by informal suggestions, in
some empirical studies, that improved IT leads to greater ‘‘lateral commu-
nication.’’61 To define coordination in our two-manager model, let the or-
ganization’s action have two components: a ¼ (a1, a2), where ai is associated
with Manager i. Then coordination for the centralized organization (where
managers are loyal team members and are not self-interested) means that
HQ chooses ai as a function of both Managers’ search results, while no
coordination means that each ai is chosen as a function of i’s results only.
The Coordination Benefit is the improvement in expected net payoff—i.e., O
minus the sum of the search costs—when we move from no coordination to
coordination. To define the concept in the decentralized (sharing-game)
case, suppose that HQ no longer exists. Instead, the self-interested Manager
i chooses ai himself. Coordination now means that i bases his choice on both
manager’s search results (after ‘‘lateral communication’’ occurs), and no
coordination means that he bases it on his own results only. Two different
sharing games are thus defined. The Coordination Benefit is the improve-
ment in net payoff when we move from an equilibrium of the first game to
an equilibrium of the second. Classes of probability distributions and payoff
fucntions have been found for which it is indeed the case that when search
technology improves (y drops), the Coordination Benefit rises, in both the
centralized and the decentralized cases.62
61
Again, see Bresnahan et al. (2002).
62
Suppose, once again, that the payoff function W has the linear/quadratic structure and each ei is
uniformly distributed on a closed interval. Then the methods of the Theory of Teams again allow us to
find explicit formulae for the action rules used by HQ in the centralized/coordinated and centralized/
uncoordinated cases, and the rules used by each manager in the decentralized/coordinated and decent-
ralzied/uncoordinated cases. In all cases one finds (for the number-of-sets cost measure and for the
Shannon cost measure) that the Coordination Benefit rises when y drops.
3.2.3 Extending the model so that managers may be decentralized with

regard to ‘‘production’’ as well as search; studying the effect of a search-
technology improvement on the Decentralization Penalty when the ‘‘degree of
decentralization’’ is varied
Suppose that in our two-manager model the organizational payoff (or
gross profit) function W takes the following form:
W ¼ V ða; eÞ K 1 ða1 Þ K 2 ða2 Þ.
For the action a ¼ (a1, a2), the function Ki is Manager i’s local production
cost, while V (a, e), which may reflect externalities between the two man-
agers, is the gross revenue earned by the organization when the environ-
ment is e. We modify our model of the decentralized organization. Each
manager i now chooses (1) a partioning of Ei (as before) and (2) a rule ai
which assigns a value of ai to every possible search result. He receives a
share r V of V, but has to pay a share g of his production cost, with 0
ogo1. Thus his net payoff in the game is the expected value of r V minus
the expected value of g Ki (when he uses the rule ai) minus the search cost
for his chosen partitioning. In the first-best (centralized) situation, by con-
trast, the partitionings and the rules ai are chosen so as to maximize the
expected value of V–K1–K2 minus the total search costs. We can call g the
degree of decentralization. As g rises toward one, each manager bears more
and more responsibility for his own production costs. We can again study
the Decentralization Penalty, i.e., the amount by which the expected value
of [V K1K2] ( the total search costs )] in the decentralized case falls
short of its first-best value. The Penalty depends on g and on y. We now
have a new question.
Research challenge #10. When is it the case that a search technology

improvement (a drop in y) ‘‘justifies’’ a higher degree of decentralization?
When is it the case, in other words, that the value of g which minimizes
the Decentralization Penalty for ȳ exceeds the value of g which minimizes
the Penalty for y4ȳ?
This is a subtle question. In answering it, a key role is played by another
question: For a given level of search technology ( a given y), is the value (to
a manager) of another dollar spent on search higher for low g or for high g?
What is intriguing about the latter question is that one’s off-the-cuff in-
tuition can go either way. One could very roughly argue that ‘‘when g goes
up, each manager ends up with less for every action, so learning a little
more about ei becomes less valuable to the manager.’’ But one could just as
well argue, in an equally primitive way, that ‘‘when g goes up, it becomes
more important for each manager to get his action just right, so learning
more about ei becomes more valuable.’’
One could start learning about the matter by studying a one-man-
ager version. To give it a concrete setting, let the organization consist of a
274 T. Marschak
manufacturer and a franchised retailer. The manufacturer can produce and

ship the product quantity q to the retailer at a cost C (q). He charges the
franchised retailer a fee of g for each unit shipped. We can interpret g as the
‘‘degree of decentralization’’; the higher g is set, the higher the retailer’s
contribution to the manufacturing cost. Once the manufacturer chooses g,
the retailer is entirely on her own. She sells the received product in a local
market. The demand curve she faces is
P ¼ e Q,
where P is the price, Q the quantity, e>0 and Qre. But e is a random
variable. It changes each ‘‘week’’ and its possible values comprise an interval
[A, B] with B>A. The retailer has to commit to the quantity she orders for
next week’s sales before she knows next week’s demand curve. By spending
money on search (market research) she can learn about next week’s e. She
chooses a T-interval partitioning of [A, B] from a set of available partition-
ings and searches to find the interval in which next week’s e lies. Denote the T
intervals I1, y , It, y , IT. The search costs the retailer the amount y T. Once
the current interval, say It, has been found, the retailer chooses a quantity
Tg
Q^ ðtÞ so as to maximize the following conditional expectation:
pðt; gÞ ¼ E½ðe QÞ Q g Q j e 2 I t .
The retailer chooses T so as to maximize

X
ðprob: of I t Þ pðt; gÞ yT:
t2f1;...;Tg
^ yÞ denote the retailer’s chosen T. Then the manufacturer’s ex-

Let Tðg;
pected profit depends on both g and y. It is
X ^ ^
ðprob: of I t Þ ðg QTðg;yÞ;g ðtÞ CðQTðg;yÞ;g ðtÞÞÞ:
^
t2f1;...;Tðg;yÞg
He chooses the franchise fee g so as to maximize expected profit. We ask:

When is it the case that a drop in y leads the manufacturer to raise g? When is
it the case that the drop leads the manufacturer to lower g?63
63
Preliminary exercises suggest that both can happen. Going back to the general one-manager prob-
lem, consider two examples: Vða; eÞ ¼ ea_ and K (a) ¼ a2; (2) V ¼ J1/ae, K ¼ b In a, where J>0, b>0,
a>0. In both cases, the Manager collects (for given a, e) the amount VgK, where 0rgr1. In both
cases, let e be uniformly distributed on [A, B] and let the partitioning defined by T consist of T equal-
length intervals. The partitioning costs the Manager yT. It turns out that if we keep y fixed, then in the
first example, raising g leads the manager to choose a lower value of T, while in the second example it
leads him to choose a higher value.
3.3 Networks of self-interested decision-makers, who bear the network’s

informational costs
There is an abundant literature on the formation of networks, where

every node is occupied by a self-interested person, who obtains information
from those persons to whom he is linked.64 That literature usually takes as
given a ‘‘value function’’ defined on the possible networks. The function
expresses the network’s collective performance. Much of the literature then
considers alternative ‘‘allocation rules.’’ Such a rule assigns a share of the
network’s value to each of its participants. Suppose we specify a rule, and
we also specify the cost of forming a link, borne by the two persons who
form the link. Then we have a game, in which each person chooses the links
to be formed with others, and each person collects his share of value (for the
resulting network) minus his link costs. The equilibria of these games are
studied, with particular attention paid to the structure of the equilibrium
networks. (When is the equilbrium network a tree? When is it a ring?) The
literature has developed a great many interesting results about equilibrium
networks.
In this framework, we could, in principle, study the impact of improved
IT. Given a value function and an allocation rule, how does a drop in link
costs (due to IT improvement) change the structure of the equilibrium
networks?
Suppose, however, that we venture in a new direction and we no longer
take the value function and the allocation rule as exogenously given. In-
stead we let each of n persons, say person j, observe some external random
variable ej. In addition, person j learns the current value of ek, for all
persons k who are neighbors of j, i.e., there is a link between j and k. Finally,
we let j be a decision-maker. He responds to the current value of his own
and his neigbors’ external environments by taking an action. He then col-
lects a gross revenue Uj, which depends on his action, the others’ actions,
and the current value of e ¼ (e1, y , en). Let Vj denote the expected value of
Uj. Person j’s net payoff is Vj minus his share of the cost of his links. (In one
version we assume that a link’s cost is shared equally between its two users.)
In our network formation game, each person chooses his intended neigh-
bors and he also chooses a rule that tells him, what action to choose for
each value of his intended neighbors’ and his own current external vari-
ables. When all n persons have made these choices, a network emerges. That
network has a link between j and some k if and only if both j and k have
chosen the link.
64
Two surveys are Jackson (2003a, 2003b).
276 T. Marschak
Research challenge #11. A collection of related questions awaits explo-

ration:
Which networks and action-rule n-tuples are stable, i.e., they have the
property that if a player changes his intended neighbors or his action
rule (while no one else makes any changes) then his expected payoff
drops or stays the same? Pn
Which networks and action-rule n-tuples are efficient, i.e., i ¼ 1 Vi
minus the total link costs is as high as possible?
When is a stable network efficient and vice versa?
When is a stable network connected (there is a path between any two
persons)? When is an efficient network connected?
When is a stable network a tree and when is it a ring? When is an
efficient network a tree and when is it a ring?
For the case of a tree, does its number of tiers shrink as link costs go
down—i.e., can we substantiate the classic claim that an improvement
in IT indeed leads to the ‘‘flattening of hierarchies’’?
When does an inefficient stable network exhibit shirking (total link
costs are less than in an efficient network) and when does it exhibit
squandering (total link costs exceed those of an efficient network)?
Clearly this is an ambitious agenda. The existing literature has made

progress on some of these questions, because it suppresses the decision-
making role of each person and takes network value and its allocation as
given. The agenda is even more challenging in the proposed new research
path, where each person is explicitly a decision-maker. Nevertheless,
progress has been made in the following case.65 Suppose the highest at-
tainable value of the expected-revenue function Vj has the property that it
depends only on the number of j’s neighbors.
Here is a simple example where this is indeed so. Let each person j have
one unit of a product to sell. Each of the n persons is located at a market
where the product can be sold. The random variable ei is the product’s
current price at location i. Given what he knows about the prices at his
neighbors’ locations, person j has to choose the location where he will sell
his product. That can be his own location, it can be the location of a
neighbor, or it can be the location of a nonneighbor. Now suppose each ei
takes just two values: zero and one, each with probability 12; whatever all the
other ek (k6¼i) may be. Suppose person i is given a neighbor set, say the set
Ni. He wants to choose a selling location so as to maximize the expected
value of his revenue. It is clear that he may as well choose his own location
if he observes ei ¼ 1 or if no location k in the set Ni has ek ¼ 1. But if he sees
ei ¼ 0 and some neighbor k has ek ¼ 1, then he chooses the location k. His
highest attainable expected revenue depends only on jN i j; the number of his
65
The results given here are due to Xuanming Su.
neighbors. It is
jN i j
1
hðjN i jÞ ¼ 1 .
2
The function h has a fortunate ‘‘diminishing marginal product property’’: it
increases more and more slowly as jN i j increases.
When the number of neighbors is all that matters, and each person’s
highest attainable expected revenue depends on the number of his neighbors
through an increasing function h which (as in the above example) has
the diminishing marginal product property, then a fairly complete analysis
is possible. Let J be the link cost, and let g (J) denote the largest
value of |Ni| for which h (|Ni|+1)h (|Ni|)ZJ. Let half of the cost of each
link be paid by each of its users. The results one can establish include the
following:
P
Let us measure a network’s net performance by n1 ¼ 1 h(|Ni|) ( total
link costs). If a network is stable and its net performance is maximal
among all stable networks, then it is efficient.
If n1rg(J), then both stable and efficient networks are connected.
If g(J)on1/2og(J/2), then an efficient network is connected but a
stable network need not be.
The net performance of an efficient network rises as J drops (i.e., IT
improves), but more and more slowly.
In an inefficient stable network there is shirking, not squandering.
One might view the requirement that ‘‘only the number of neighbors
matters’’ as very strong. One might guess, in particular that it rules out
externalities, wherein one person’s action affects the payoffs collected
by others. (There are no externalities in our selling-location example.) But
this need not be so, as the following example shows: each person j again
observes a random variable ej and again learns the current ek for every
neighbor k. In response, he chooses an action aj (a real number). When the
action vector a ¼ (a1, y , an) has been chosen, then for a given
e ¼ (e1, y ,en), the entire organization obtains a payoff
X
n X
n X
n
W ða; eÞ ¼ ei ai qij ai aj ,
i¼1 i¼1 i¼1
where the matrix ((qij)) is positive definite (which insures that there is a
unique W-maximing a). person i’s own revenue is a share of W, namely
r W, where 0or 1n: Thus there are externalities between the actions. Let
the ei be independently distributed and let each be normally distributed with
mean zero. For any given network, consider the possible action rules for
each person j. The rule assigns a value of aj to each value of (ej, eN j ), where
eN j denotes the external variables observed by j’s neighbors and hence
278 T. Marschak
learned by j as well. Because of the linear/quadratic payoff structure we

may use, once again, the methods of the Theory of Teams. It turns out, in
spite of the externalities, that in the rule n-tuple which achieves the highest
expected value of W attainable by the network, each person’s action de-
pends only on the number of his neighbors. If we now turn to the network-
formation game, we find that for a given network, each person’s highest
attainable expected payoff (the highest attainable expected value of r W),
again depends only on the number of his neighbors, whatever action rules
the n1 persons may have chosen. Moreover the function h, which gives
that highest attainable expected payoff, has the required diminishing mar-
ginal product property.
4 Organizational models in which the primitive is a ‘‘task’’, ‘‘problem’’,

‘‘project’’, or ‘‘item’’
There are formal models of organization which do not explicitly consider

the actions the organization takes in response to a changing environment.
Instead the model supposes that ‘‘tasks’’ ‘‘problems,’’ ‘‘projects,’’ or ‘‘in-
formation items’’ flow in from the outside world, and the organization has
to process them. They become the primitives of the model and are not
further explained. One seeks to characterize organizations which process
the flow efficiently, in some appropriate sense. Just as in the approaches we
have discussed, processing costs (or capacities) are part of the model. To
illlustrate, we briefly discuss three studies that share this approach but are
otherwise rather different.66
In Sah and Stiglitz (1986), the organization has to judge ‘‘projects.’’
A project has a net benefit x, where x is a random variable with known
distribution. The organization consists of evaluators. Once he has received
a project, an evaluator either chooses to reject it or to pass it on to
another evaluator for further evaluation. The total expected benefit
of the project portfolio selected by a ‘‘hierarchy’’ is compared to that
achieved by a ‘‘polyarchy.’’ In a hierarchy, the evaluators are arranged in
two ‘‘bureaus,’’ called 1 and 2. All projects first flow to a member of Bureau
1, who either rejects it or passes it onto Bureau 2, where a final judgment is
made. The only projects reviewed by Bureau 2 are those it receives from
Bureau 1. In a polyarchy, the organization is divided into two ‘‘firms.’’ Each
receives half of the projects. If it accepts the project no further evaluation
occurs; if it rejects the project then the other firm performs a final eval-
uation. The quality of the evaluators can be varied, by changing the con-
ditional distribution of x given the ‘‘reject’’ judgment. Costs could be
66
Other studies, which use similar primitives and pay some attention to informational costs, include
the following Keren and Levhari (1983), Beckman (1983), Arenas et al. (2003), and Visser (2000).
attached to quality improvement but the 1986 paper only sketches how that
might be done.67
In Bolton and Dewatripont (1994), the main primitive is a ‘‘cohort
of M information items’’ received by the organization from an external
source. Each item is a ‘‘type’’ of information about the outside world. All
cohorts yield the same ‘‘value’’ R to the organization, once they are ‘‘proc-
essed.’’ (‘‘Processing’’ is another primitive, not further explained.) In order
for the organization to realize the cohort’s value, at least one agent
must receive the entire cohort. (Thus far the problem can be restated as
the choice of a speak-once-only mechanism, with an action-taker who needs
to know the entire cohort.) While one can study delay (time until at least
one agent knows the entire cohort), the paper emphasizes another question.
It supposes that economies of scale are achieved when a given agent
processes a given type more and more frequently. A network of agents
is given and each can be made a specialist in one or more items; he processes
only those. For each given network, one seeks an assignment of types
to agents so as to minimize ‘‘the total labor time spent per processed
cohort.’’ In efficient networks, the labor time for a best assignment is
minimal. Some suggestive properties of efficient networks are worked
out.
In Garicano and Rossi-Hansberg (2005), the agents who form organ-
izations belong to a ‘‘knowledge economy.’’ In each time period each
agent receives a problem whose level of difficulty can vary, and solves it if
its difficulty is below her level of knowledge. (‘‘Problem,’’ ‘‘level of diffi-
culty,’’ ‘‘solve,’’ and ‘‘level of knowledge’’ are primitives, not further ex-
plained.) Each problem is identified by a value of Z, a nonnegative
number; a higher Z means the problem is harder. The population of pos-
sible problems has a known probability distribution over the possible
values of Z. Each agent is endowed with a ‘‘cognitive ability’’ a, a random
variable with known distribution. By incurring a cost, an agent can learn
to handle all problems of difficulty up to a given level z. The cost is
increasing in z and decreasing in a. An agent receives an income, which is
increasing in the proportion of problems the agent is able to solve. An
agent seeks to maximize income minus learning costs. But high-ability
agents can help low-ability agents, and that leads to the formation of
organizations. The structure of those organizations in an equilibrium of
the economy is studied.68
67
In a further paper (Sah and Stiglitz, 1988), the organization becomes an n-person ‘‘committee.’’ All n
persons judge each project and it is accepted if krn members judge it favorably. This time the trade-off
between portfolio quality and cost is central. The cost measure is simply n itself.
68
Another model in which ‘‘task’’ is a primitive is developed in Malone (1987) and Malone and Smith
(1988). These papers study a number of cost measures for a given organizational structure.
280 T. Marschak
5 Concluding remarks
We have followed a modeling path with four elements: environment,

action, goal, and informational cost. The models describe, in detail,
the process by which the organization reaches new actions when the en-
vironment changes, i.e., they describe the mechanism that the organization
chooses. This is not an easy research path. A great deal more work needs to
be done before one can judge whether the effort is worthwhile. In the
present state of knowledge there is a formidable gulf between the propo-
sitions that can be proved and the complexities of organization and infor-
mation as they appear in the real world. Yet without theory, it is hard to
make sense of that reality and to see why casual claims, like those about the
impact of IT on organizations, might or might not be true. Perhaps models
that omit one or more of our four elements, or forego the detailed
description that the concept of mechanism permits, or use entirely different
primitives, may prove more useful guides—for the time being—to persons
who are able to observe real organizations and assemble new datasets.
But surely all four our elements arise in real organizations, and real or-
ganizations follow some procedure in choosing new actions. It seems in-
evitable that future empirical work will eventually try to examine that
procedure in detail and will deal, in one way or another, with all four of our
elements.
Comparing mechanisms with regard to informational cost is particularly
tricky. Minimal message-space size provides one fundamental way of judg-
ing the complexity of an organizational goal, and it tells us, in a preliminary
and abstract way, how expensive one goal is relative to another if the goal is
to be met by an organization which is decentralized in the sense that each
mmber privately observes some aspect of the environment. But delay, the
number of persons, and the individual communication and computation
burden that each person faces are all important as well. Modeling of those
costs is still in an early stage.
We have looked primarily at work conducted by economic theorists. But
there are parallel efforts in computer science and in artificial intelligence.
Theoretical research that bridges disciplines is finally emerging.69 There are
also many parallel efforts by social scientists who are not economists (e.g.,
persons in the Organizational Behavior field). A review of that literature
69
As noted in Section 2.2.8, the computer science specialty called Communication Complexity deals
with minimal privacy-preserving mechanisms (protocols) for an n-person organization whose goal is the
computation of an action that depends on n variables, each privately observed by one of the n persons.
Both communication-complexity results and economic-theory results are central to the bridge-building
paper by Nisan and Segal (2005) on the allocation of many objects among many persons, mentioned in
Section 2.2.8. At present it seems fair to say that there is very much more work by theoretical computer
scientists influenced by economic ideas, than work by economic theorists influenced by computer science
ideas. In particular, computer scientists have explored the merits of market-based approaches to com-
putational problems. Many papers which do so could be cited. A few of them are Deng et al. (2002),
Walsh et al. (2003), Walsh and Wellman (2003).
would doubtless paint a very different picture as to what has been learned
or could be learned about the effect of IT advances on organizational
structure. Economic theorists are endowed (or perhaps burdened!) with a
certain point of view when they approach such a challenging question. That
point of view has thoroughly permeated this chapter.
References
Abelson, H. (1980). Lower bounds on information transfer in distributed computations. Journal of the
ACM 27, 384.
Abramson, N. (1963). Information Theory and Coding. McGraw-Hill, New York.
Aizpura, J., A. Manresa (1995). A decentralized and informationally efficient mechanism realizing fair
outcomes in economies with public goods. Economic Design 1, 141.
Alon, N., D. Kleitman (1986). Covering a square by small perimeter rectangles. Discrete and Com-
putational Geometry 1, 1.
Arenas, A., L. Cabrales, et al. (2003). Optimal information transmission in organizations: search and
congestion’’. Working Paper, Universitat Pompeu Fabra.
Apostol, T. (1957). Mathematical Analysis. Addison-Wesley, Reading, MA.
Beckman, M. (1983). Tinbergen Lectures on Organization Theory. Springer, Heidelberg.
Bolton, P., M. Dewatripont (1994). The firm as a communication network. Quarterly Journal of
Economics CIX, 809.
Bresnahan, T., E. Brynjolfsson, L. Hitt (2000). Technology, organization, and the demand for skilled
labor, in: M. Blair, T. Kochan (eds.), The New Relationship: Human Capital in the American
Corporation, Brookings Institution, Washington.
Bresnahan, T., E. Brynjolfsson, L. Hitt (2002). Information technology, workplace organization, and
the demand for skilled labor. Quarterly Journal of Economics CXVII, 339.
Calsamiglia, X. (1977). Decentralized resource allocation and increasing returns. Journal of Economic
Theory 14, 263.
Calsamiglia, X., A. Kirman (1998). A unique informationally efficient decentralized mechanism with fair
outcomes. Econometrica 61, 1147.
Chander, P. (1983). On the informational size of message spaces for efficient resource allocation proc-
esses. Econometrica 51, 919.
Courtney, D., T. Marschak (Forthcoming). Shirking and squandering in sharing games, Topics in
Theoretical Economics, Berkeley Electronic Press.
Creti, A. (2001). Firms’ organization and efficient communication networks. The Manchester School
69, 77.
Deng, X., C. Papadimitriou, M. Safra (2002). On the complexity of equilibria. Symposium on the
Theory of Computing.
Garicano, L., E. Rossi-Hansberg. (2005). Organization and inequality in a knowledge economy. Work-
ing paper.
Groves, T., J. Ledyard (1977). Optimal allocation of public goods. Econometrica 45, 783.
Harris, M., A. Raviv (2002). Organization design. Management Science 48, 852.
Hurwicz, L. (1960). Optimality and efficiency in resource allocation, in: K. Arrow, S. Karlin,
P. Suppes (eds.), Mathematical Methods in the Social Sciences, Stanford University Press, Stanford,
CA.
Hurwicz, L. (1986). On informational decentralization and efficiency in resource allocation mechanisms,
in: S. Reiter (ed.), Studies in Mathematical Economics, Mathematical Association of America,
Washington.
Hurwicz, L. (1997). On the dimensional requirements of informationally decentralized Pareto-satisfac-
tory processes, in: K. Arrow, L. Hurwicz (eds.), Studies in Resource Allocation Processes, Cam-
bridge University Press, New York.
282 T. Marschak
Hurwicz, L., T. Marschak (1985). Discrete allocation mechanisms: dimensional requirements for re-
source allocation mechanisms when desired outcomes are unbounded. Journal of Complexity 1, 264.
Hurwicz, L., T. Marschak (2003a). Comparing finite mechanisms. Economic Theory 21, 783.
Hurwicz, L., T. Marschak (2003b). Finite allocation mechanisms: approximate Walrasian versus ap-
proximate direct revelation. Economic Theory 21, 545.
Hurwicz, L., T. Marschak (2004). The informational efficiency of finite price mechanisms, in: C. Alipr-
antis K.J. Arrow, P. Hommand, F. Kubler, H.-M. Wu, N.C. Yannelis (eds.), Assets, Beliefs, and
Equilibria in Economic Dynamics, Springer, Heidelberg.
Hurwicz, L., S. Reiter (2006). Designing Economic Mechanisms. Cambridge University Press, New York.
Hurwicz, L., H. Weinberger (1990). A necessary condition for decentralization and an application to
intertemporal allocation. Journal of Economic Theory 51, 313.
Ishikida, T., T. Marschak (1996). Mechanisms that efficiently verify the optimality of a proposed action.
Economic Design 2, 33–68.
Jackson, M.O. (2001). A crash course in implementation theory. Social Choice and Welfare 18, 655.
Jackson, M.O. (2003). On the formation of networks and groups, in: B. Dutta, M.O. Jackson (eds.),
Networks and groups: models of strategic formation, Springer, Heidelberg.
Jackson, M.O. (2003a). The stability and efficiency of economic and social networks, in: B. Dutta, M.
Jackson (eds.), Networks and Groups: Models of Strategic Formation, Springer, Heidelberg.
Jackson, M.O. (2003b). A survey of models of network formation: stability and efficiency, in: G.
Demange, M. Wooders (eds.), Group Formation in Economics: Networks, Clubs, and Coalitions,
Cambridge University Press, New York.
Jackson, M.O. (2004). A survey of models of network formation: stability and efficiency, in: G. De-
mange, M. Wooders (eds.), Group formation in networks, clubs, and coalitions, Cambridge Uni-
versity Press, New York.
Jordan, J. (1982). The competitive allocation process is informationally efficient uniquely. Journal of
Economic Theory 28, 1.
Jordan, J. (1987). The informational requirements of local stability in decentralized allocation
mechanisms, in: T. Groves, R. Radner, S. Reiter (eds.), Information, Incentives, and Economic
Mechanisms: Essays in Honor of Leonid Hurwicz, University of Minnesota Press, Minneapolis.
Jordan, J. (1995). Information flows intrinsic to the stability of economic equilibrium, in: J. Ledyard
(ed.), The Economics of Informational Decentralization: Complexity, Efficiency, and Stability,
Kluwer, Boston.
Jordan, J., D. Xu (1999). On the communication complexity of expected profit maximization. Journal of
Economic Theory 86, 185.
Kaganovitch, M. (2000). Decentralization of intertemporal economies with discounted maximin cri-
terion. International Economic Review 41, 1021.
Karchmer, M. (1989). Communication Complexity: A New Approach to Circuit Depth. MIT Press,
Cambridge, MA.
Keren, M., D. Levhari (1983). The internal organization of the firm and the shape of average costs. Bell
Journal of Economics XL, 474.
Kushilevitz, E., N. Nisan (1997). Communication Complexity. Cambridge Universisty Press, New York.
Leavitt, H., T. Whistler (1958). Management in the 1980s. Harvard Business Review 36, 6.
Leontieff, W. (1947). A note on the interrelation of subsets of independent variables of a continuous
function with continuous first derivatives. Bulletin of the American Mathematical Society 53, 343.
Lovász, L. (1990). Communication complexity: a survey, in: B.H. Korte (ed.), Paths, Flows, and VSLI
Layouts, Springer, Berlin.
Malone, T. (1987). Modeling coordination in organizations and markets. Management Science 53, 1317.
Malone, T., S. Smith (1988). Modeling the performance of organizational structures. Operations Re-
search 36, 421.
Manresa, A. (1993). An infinite dimensional extension of the theory of decentralized mechanisms.
Mathematical Social Science 26, 157.
Marschak, T. (1959). Centralization and decentralization in economic organizations. Econometrica
27, 399.
Marschak, T. (1972). Computation in organizations: the comparison of price mechanisms and other
adjustment processes, in: C.B. McGuire, R. Radner (eds.), Decision and Organization, North-
Holland, Amsterdam.
Marschak, T. (1987). Price versus direct revelation: informational judgments for finite mechanisms, in:
T. Groves, R. Radner, S. Reiter (eds.), Information, Incentives, and Economic Mechanisms: Essays
in Honor of Leonid Hurwicz, University of Minnesota Press, Minneapolis.
Marschak, T. (2004). Information technology and the organization of firms. Journal of Economics and
Management Strategy 13, 473.
Marschak, J., R. Radner (1972). The Economic Theory of Teams. Yale University Press, New Haven.
Marschak, T., S. Reichelstein (1995). Communication requirements for individual agents in networks
and hierarchies, in: J. Ledyard (ed.), The Economics of Informational Decentralization: Complexity,
Efficiency, and Stability, Kluwer, Boston.
Marschak, T., S. Reichelstein (1998). Network mechanisms, informational efficiency, and hierarchies.
Journal of Economic Theory 44, 106.
Melumad, N., D. Mookherjee, S. Reichelstein (1992). A theory of responsibility centers. Journal of
Accounting and Economics 15, 445.
Melumad, N., D. Mookherjee, S. Reichelstein (1995). Hierarchical decentralization of incentive con-
tracts. RAND Journal of Economics 26, 654.
Melumad, N., D. Mookherjee, S. Reichelstein (1997). Contract complexity, incentives, and the value of
delegation. Journal of Economics and Management Strategy 6, 257.
Mount, K., S. Reiter (1974). The informational size of message spaces. Journal of Economic Theory
8, 161.
Mount, K., S. Reiter (1977). Economic environments for which there are Pareto-satisfactory mecha-
nisms. Econometrica 45, 821.
Mount, K., S. Reiter (1987). On the existence of a locally stable dynamic process with a statically
minimal message space, in: T. Groves, R. Radner, S. Reiter (eds.), Information, incentives, and
economic mechanisms, University of Minnesota Press, Minneapolis.
Mount, K., S. Reiter (1996). A lower bound on computational complexity given by revelation mech-
anisms. Economic Theory 7, 237.
Mount, K., S. Reiter (1998). A modular network of bounded rationality, in: M. Majumdar (ed.),
Organizations with Incomplete Information, Cambridge University Press, New York.
Mount, K., S. Reiter (2002). Comunication and Complexity in Economic Behavior and Organizations.
Cambridge University Press, New York.
Nisan, N., I. Segal (2005). The communication requirements of efficient allocations and supporting
Lindahl prices. Journal of Economic Theory 129, 192.
Osana, H. (1978). On the informational size of message spaces for resource allocation processes. Journal
of Economic Theory 17, 66.
Osana, H. (2005). Externalities do not necessarily require larger message spaces for realizing Pareto-
efficient allocations. Review of Economic Design 19, 227.
Radner, R. (1993). The organization of decentralized information processing. Econometrica
62, 1109.
Radner, R., T. Van Zandt (1995). Information processing in firms and returns to scale, in: J. Ledyard
(ed.), The Economics of Informational Decentralization: Complexity, Efficiency, and Stability,
Kluwer Publishing Company, Boston.
Reichelstein, S. (1984). Incentive compatibility and informational requirements. Journal of Economic
Theory 32, 384.
Reichelstein, S., S. Reiter (1988). Game forms with minimal message spaces. Econometrica 53, 661.
Sah, K., J. Stiglitz (1986). The architecture of economic systems: hierarchies and polyarchies. American
Economic Review LXXVI, 716.
Sah, K., J. Stiglitz (1988). Committees, hierarchies, and polyarchies. Economic Journal 98, 451.
Sato, F. (1981). On the informational size of message spaces for resource allocation processes in econ-
omies with publicg goods. Journal of Economic Theory 24, 48.
Saari, D., C. Simon (1978). Effective price mechanisms. Econometrica 46, 1097.
284 T. Marschak
Segal, I (2004). The communication requirements of social choice rules and supporting budget sets.
Working Paper.
Serrano, R. (2004). The theory of implementation of social choice rules. SIAM Review 46, 377.
Stoenescu, T. (2004). Decentralized resource allocation mechanisms in networks. Ph.D. dissertation, De-
partment of EECS-Systems, University of Michigan.
Tian, G. (1990). Completely feasible and continuous Nash implementation of the Lindahl correspond-
ence with a message space of minimal dimension. Journal of Economic Theory 51, 443.
Tian, G. (2004). A unique informationally efficient allocation mechanism in economies with consump-
tion externalities. International Economic Review 45, 79.
Van Zandt, T. (1998). Organizations with an endogenous number of information processing agents, in:
M. Majumdar (ed.), Organizations with Incomplete Information, Cambridge University Press,
New York.
Van Zandt, T. (1999). Real-time decentralized information processing as a model of organizations with
boundedly rational agents. Review of Economic Studies 66, 633.
Van Zandt, T. (2003a). Real-time hierarchical resource allocation. Discussion Paper.
Van Zandt, T. (2003b). Real-time hierarchical resource allocation with quadratic payoffs. Discussion
Paper.
Visser, B. (2000). Organizational communication structure and performance. Journal of Economic Be-
havior and Organization 42, 231.
Walker, M. (1977). On the informational size of message spaces. Journal of Economic Theory 15, 366.
Walsh, W., M. Wellman (2003). Decentralized supply chain formation: a market protocol and com-
petitive equilibrium analysis. Journal of Artifical Intelligence Research 19, 513.
Walsh, W., M. Yokoo, K. Hirayama, M. Wellman (2003). On market-inspired approaches to propos-
itional satisfiability. Artificial Intelligence 144, 125.
Williams, S. (2007). Communication in Mechanism Design: A Differential Approach. Cambridge
University Press, New York.
Ziv, A. (2000). Information technology and optimal firm structure. Journal of Accounting Research
38, 297.
Chapter 5
Open Source Software: The New Intellectual Property

Paradigm
Stephen M. Maurer and Suzanne Scotchmer

University of California, Berkeley, CA 94720-7320, USA
Abstract
Open source methods for creating software rely on developers who volun-
tarily reveal code in the expectation that other developers will reciprocate.
Open source incentives are distinct from earlier uses of intellectual property,
leading to different types of inefficiencies and different biases in R&D in-
vestment. The open source style of software development remedies a defect of
intellectual property protection, namely, that it does not generally require or
encourage disclosure of source code. We review a considerable body of sur-
vey evidence and theory that seeks to explain why developers participate in
open source collaborations instead of keeping their code proprietary, and
evaluates the extent to which open source may improve welfare compared to
proprietary development.
1 Introduction
Open source software, which burst on the innovation scene in the mid-
1990s, is produced in a completely different way than other commercial
products. Workers are usually unpaid; management and direction are lim-
ited; and legal restrictions on using the product are modest (Lerner and
Tirole, 2004). The open source style of development has various features,
but it generally involves software developers making their source code
available free-of-charge to end users and improvers, usually subject to
285
286 S. M. Maurer and S. Scotchmer
license restrictions such as GPL and BSD.1 Developers often work on code
provided by others.
The open source movement is a substantial phenomenon. LINUX runs
on 29 million machines (The Linux Counter, 2005) and Apache runs on 70
million servers (Netcraft, 2005). Despite this demonstrated success, survey
evidence indicates that the nature of open source activities is changing
rapidly (Comino et al., 2005). For example, the survey data of Ghosh et al.
(2002) show that there is a secular shift from ‘‘hobbyist’’ contributions to
‘‘commercial’’ contributions. It is still unclear to what extent open source
will supplant proprietary methods for software development, let alone
branch out into other information goods such as pharmaceuticals or ge-
ographic data. In this essay, we provide a snapshot of the emerging open
source phenomenon, and a discussion of how scholars have tried to make
sense of it.
There are several natural questions to ask about the phenomenon.
Among them,
How do open source methods provide sufficient incentive to invest in
software, given that users do not pay innovators?
What is it about computer software, if anything, that calls for a new
invention paradigm? Which other inventive activities share these
features?
Does the efficacy of open source depend on licenses (e.g., BSD, GPL)
and, indirectly, the underlying availability of intellectual property
protection?
Does the market need a coordination mechanism to collectively choose
open source over more traditional ways of exploiting intellectual
property, and does it always do so when open source is the better
choice?
In what circumstances does open source work better than traditional
intellectual property incentives or other funding schemes, such as
public sponsorship?
In Section 2, we lay out the various arguments for why open source
works as an incentive scheme, and compare it to more traditional uses of
intellectual property. In Section 3, we focus on how open source collabo-
rations are organized. In Section 4, we turn to some of the observable
welfare consequences of organizing R&D in this fashion, and in Section 5
we discuss some of the gaps in what is known about open source, and its
potential to organize invention in other arenas.
1
The General Purpose License (GPL) is a ‘‘viral’’ license that obligates a further developer of the
software to make it available under the same license. In general, there are no restrictions on use or an
obligation to pay, but in some versions there is an obligation for attribution. The Berkeley Software
Distribution (BSD) license originated through UNIX development at the University of California,
Berkeley and is not viral. It requires users to give attribution credit to the University but does not
prohibit commercial use or development. For a more complete history, see, for example, Weber (2004).
Ch. 5. Open Source Software: The New Intellectual Property Paradigm 287
2 Incentives for R&D
Not all R&D environments call for the same incentive structure. For
example, R&D environments differ in the extent to which ideas for invest-
ments are scarce or common knowledge, the extent to which disclosure of
progress is inevitable or important, the extent to which innovation proceeds
cumulatively through the efforts of many contributors, and the extent to
which natural restrictions on entry protect innovators.
The open source movement emerged to support an industrial product
(software) for which disclosure of code is particularly useful, but not re-
quired by intellectual property law. Copyrights for software can be regis-
tered without fully revealing the source code, and source code is typically
not included in software patents.2 Source code is typically not disclosed in
either regime.
Of course, code can be released under license, but here the nature of the
innovative environment matters. If ‘‘ideas are scarce’’ in the sense that each
idea for an improvement occurs to a single, random person (Scotchmer,
2004, Chapter 2), and ideas depend on prior disclosure, then traditional
protection through patents and copyrights may cripple inventive activity.
This is because rights holders do not know to whom they should license
and disclose. In the open source regime, full disclosure is automatic, and
thus encourages new ideas and re-use of the code by developers who cannot
be identified in advance. The surprise is that this can be done while still
preserving incentives. In this section we explore those incentives, but
first compare open source with the more traditional way of exploiting
intellectual property.
Open source communities have been well studied with survey instru-
ments. When asked their motives, survey respondents cite various incentives
including: own use benefits, complementarity with proprietary products
sold in the market, signaling, education, and social psychological motives
such as altruism or simple enjoyment. In terms of the technical problems
that contributors seek to address, Ghosh et al. (2002) report that 39.8%
are trying to improve the products of other developers. Fewer are trying
to realize a good product idea (27%), or trying to solve a problem that
could not be solved by proprietary software (29.6%). Among contrib-
utors at SourceForge, the top three reasons for participating in open
source communities include ‘‘work functionality’’ (33.8%) and ‘‘non-work
functionality’’ (29.7%)3 (Lakhani and Wolf, 2005).
To control for the possibility that these responses are sensitive to the
number and phrasing of the questions, Lakhani and Wolf (2005) use factor
2
See Lemley et al. (2002) at 204-205 (for patents), Samuelson (1984), and U.S. Copyright Office (2002)
(for copyrights).
3
The most common response is ‘‘intellectually stimulating’’ (44.9%). Ideological responses (‘‘Beat
Proprietary Software’’) are much less frequent (11.9%) (Lakhani et al., 2002; Lakhani and Wolf, 2005).
analysis to group the responses into four classes: workers who are primarily
motivated by education/intellectual stimulation (‘‘Learning and Fun’’—
29%), by non-work user needs (‘‘Hobbyists’’—27%), by work-related user
needs (‘‘Professionals’’—25%), and by feelings of obligation/community
(‘‘Community Believers’’—19%). Significantly, the two user needs categories
comprise about one-half of all respondents.
Not surprisingly, different communities report different incentives. Surveys
of the embedded-LINUX community find that most hardware firms release
code in order to continue receiving similar donations from others (61.4%),
benefit from other participants’ efforts to find and fix bugs (59.9%), to be
known as a good player in the open source community (58.9%) and because
they hope that others will develop their code further (57.7%). Employees
working for software companies report broadly similar motives except that
they tend to place slightly more emphasis on marketing (e.g., signaling and
reputation). This effect is larger for small, young companies than for older
and more established firms (Henkel, 2005b).
2.1 Intellectual property and open source
Because open source puts knowledge (software) in the public domain, it

would not perform well as an incentive mechanism in the usual innovation
environment where the objective is to prevent imitation. To the contrary,
the point of putting the knowledge in the public domain is to encourage
imitation. As we will see in this section, open source works in environments
where the knowledge created (a) is complementary to some other good
whose profitability is immune to imitation, such as human capital or an-
other proprietary product, or (b) where the motives to invent are intrinsic
and have nothing to do with appropriating value.4
We begin our discussion by comparing open source and the ordinary use
of intellectual property in the two environments where open source is
mainly used: where innovations are complementary and where innovation
is cumulative.
Following Scotchmer (2004), we distinguish between having ideas, which
are random and costless, and creating innovations, which require invest-
ment. Let the idea of firm i be summarized by an indicator of its commercial
value and its private cost, (vi,ci). We describe complementarity by the fol-
lowing profit function, for each firm i, where n is the total number of
contributions to the open source project, and f is positive, increasing and
bounded:
vi f ðnÞ ci .
4
For discussion of how economists have treated the design of intellectual property in regard to these
issues, see Scotchmer (2004). For a broader set of references that also includes legal scholarship, see
Menell and Scotchmer (forthcoming).
Complementarity produces a type of network effect: contributors prefer

to contribute to open source projects with many participants. Even though
each developer’s code is freely usable by others, the commercial value vi is
assumed to survive. The commercial value must therefore be due to some
other proprietary product that embeds the software, and not directly to the
software itself.
To compare open source with patents, assume that if firms keep their
contributions proprietary, they cross-license symmetrically at some price ‘:
Then with n contributors, each firm earns licensing revenue (n–1)‘: How-
ever, each firm also has licensing obligations in amount (n–1)‘; so there is
no net burden due to licensing. GPL leads to the same end without
imposing the license fees ‘ and without the burden of negotiating.
Notice, however, that participating in such a scheme might not be the
best thing from either a private or social point of view if most of the users
are not developers so that the open source community confers benefits on
non-reciprocating third parties. If such use constitutes the bulk of the social
benefits, and if development costs are relatively high, a better incentive
scheme would involve royalties. Royalties from third parties may be nec-
essary to cover costs.
Now suppose that the software contributions are cumulative. For exam-
ple, suppose that a set of proprietary products embed a common software
product, such as an operating system, which can be improved sequentially
by other developers who find bugs or see opportunities for improve-
ment. Suppose further that each improvement increases the quality of the
embedded software and that ideas for improvement occur randomly to
members of the community to whom the product has been disclosed.
In the cumulative context, traditional intellectual property (IP) has the
same problem as with complements. If ideas for improvement are scarce, so
that a potential improver cannot be identified in advance, then licensing on
a one-by-one basis to disclose the prior code will not work very well. But, as
with complements, if each member makes the same number of contribu-
tions as every other, at least in expectation, then their expected receipts and
payouts from licensing will be equal for each contributor, and equivalent to
making no payments at all under a GPL scheme.
However, the cumulative context is where we see that a viral GPL license
may be useful. A GPL system locks the proprietors into a royalty-free
exchange of knowledge. Unlike the case of complements, knowledge ex-
change is not symmetric because later developers can free ride on earlier
ones, but not vice versa. Absent GPL, a developer might be tempted to free
ride on earlier developers and make his code proprietary so that he can
impose royalties on later developers. With GPL, he can only revert to
royalties by building his code from the ground up. This may be more costly
than using prior code and accepting the GPL.
Because the whole point of making code open is to encourage imitation
and use, the code itself will not be a profit center for its developer. In the
remainder of this section, we discuss why open source works as an incentive

system, namely, that the benefits come from own use, other products sold in
the market, signaling, and education. We also discuss social psychology
motives that claim to transcend traditional economic incentives. Despite
broad similarities, these incentives lead to different behaviors and welfare
implications.
2.2 Own use
User innovation has been documented as far back as Victorian times.5

By the late 20th century it was ubiquitous: examples include everything
from machine tools to windsurfing (Lerner and Tirole, 2002a; Harhoff
et al., 2003; Von Hippel, 2004). The modern open source movement
extends this user innovation model to software (Raymond, 1999). The
Apache web server collaboration provides a leading example.
Writing code for own use is not entirely about volunteer labor. Eighty six
percent of contributors who participate in open source projects for work
reasons are paid for at least some of their work. Not surprisingly, paid
contributors spend almost twice as much time as volunteers—10.3 hours
compared to 5.7 hours (Lakhani and Wolf, 2005). Corporations that
employ own use strategies to make or improve products that will be sold to
mass market products can afford to bear more cost than individuals. At the
same time, own use includes substantial non-commercial activity. Lakhani
and Wolf (2005) report that 27% of SourceForge contributors write code
for ‘‘non-work needs.’’ Since 10.9% of all SourceForge listings involve
games (Comino et al., 2005), ‘‘hobbyist’’ motives are apparently important.
Own-use incentives may lead to underprovision of code, since the in-
vesting party does not appropriate the benefits conferred on third parties.
While reciprocity within the open source community may avoid duplica-
tion, it does not solve this basic deficiency in incentives. This is also true in
the simple model of complementarity that we started with, where every
contributor is concerned with own use. The user will not invest if his ci is
larger than his own benefit P vif(n), even if the cost is smaller than the social
benefit ½ f ðnÞ f ðn 1Þ j nj .
2.3 Complementary goods and services
We have already stressed that open source incentives will not work if
the open source software must itself be a profit center. This is because
5
The practice of creating new information for one’s own use is as old as mankind. Scholars have long
argued that the apparent perfection of European folklore reflects accumulated interactions (and incre-
mental improvements) of storytellers and audiences. See, e.g., Bettelheim (1976). The high ranking of
Homer’s Iliad and Odyssey among Western ‘‘great books’’ is an obvious example.
imitation is its very lifeblood. Instead, the open source activity must be
complementary with something that remains proprietary.
West and Gallagher (2004) refer to open source as ‘‘pooled R&D.’’ In
particular, companies share code to test software, fix bugs, and to get im-
provements, feedback, and extensions (Rossi and Bonaccorsi, 2003, 2005),
all of which they would otherwise have to do independently with substantial
duplicated costs. Contributors can afford to cooperate in this way because
the open source software is bundled into different goods and services that
are mostly non-rival in the market and allow individual contributors to
appropriate benefits. Typical complements include proprietary operating
systems; proprietary applications programs; hardware; documentation;
distribution through trusted and convenient brand name channels; bun-
dling open source software into convenient, ready-to-use packages; tech
support and warranties; custom software services; consulting, education
and training; remote services; complete data solutions; making applica-
tions more reliable for particular applications or libraries; and organizing
fairs and conferences (Raymond, 1999; Ghosh et al., 2002; Hawkins, 2002;
Harhoff et al., 2003; O’Mahoney, 2003; Varian and Shapiro, 2003;
West, 2003; Dahlander, 2004; West and Gallagher, 2004; Henkel, 2005b).
Complements are particularly important for server software, desktop/
client software, enterprise solutions, IT consulting, IT services, and the
embedded software used in appliances like DVDs and cell phones (Ghosh
et al., 2002).
Commercial firms are less likely to participate in open source devel-
opment where competition among them is significant (von Hippel, 2002;
Harhoff et al., 2003). Harhoff et al. (2003) present a model with two users
who practice in-house innovation, and are imperfect competitors selling
products that are enhanced by the open source product.6 Even though each
firm’s rival would benefit from a disclosure of its code, the firm may nev-
ertheless disclose it in the hope that a third-party manufacturer will develop
the disclosed product still further. Harhoff et al. find that, if competition,
technology spillovers, and the cost of adopting disclosed improvements are
high, there is an equilibrium in which neither rival discloses. However, if the
cost of adopting the manufacturer’s improvement is less than its benefits,
there is also equilibrium where both disclose, provided competition and
spillovers are low enough.
Henkel (2005a) explores a model in which two firms each need two dis-
tinct technologies to manufacture their products. If firms cannot share
information, each must invest in both technologies. This raises entry costs
and makes monopoly more likely. If they choose open source, there are
Nash equilibria where each firm specializes in one technology and obtains
6
Harhoff et al. claim that this symmetric duopoly analysis remains qualitatively correct for mod-
erately large or asymmetric oligopolies.
the other through free riding. Henkel finds that firms with similar technol-
ogy needs disclose even where competition is strong. However, firms may or
may not share information where their needs are different. Henkel finds
equilibria in which both firms disclose. Each user performs whatever R&D
generates the most value for itself and free rides otherwise. In this game,
heterogeneous needs suppress the temptation to free ride but still produce
useful technology spillovers for the entire industry.7
Models with competition require two strong assumptions. First, each
game assumes payoffs in which the parties can earn (and split) non-zero
economic profits. Rivals are protected against competition and entry by
some unspecified means. The key issue of open source—appropriability is
introduced as an assumed parameter. Second, the models assume that par-
ties cannot negotiate licenses with one another. In our view, this makes any
comparison with intellectual property suspect, since licensing would also
allow the firms to avoid duplication. The authors argue that their licensing
assumption is justified by high transactions costs, the legal difficulty of
patenting minor (but cumulatively important) innovations, and the alleged
weakness of patents and trade secrets (Von Hippel, 2002; Harhoff et al.,
2003; Henkel, 2005a).8
Empirical studies of the ‘‘embedded LINUX’’ used in proprietary elec-
tronic devices like DVDs and cell phones (Henkel, 2005b) provide a market
example in which firms can decide between keeping code proprietary and
mutual disclosure. In this case, loopholes in the GPL license give manu-
facturers the power to keep certain classes of code confidential if they want
to.9 Despite this, roughly, half of all industry members (49%) participate in
at least limited sharing, and this fraction is growing. In general, firms with
strong complements tend to release code more readily. For example, 34.5%
of hardware companies release code but only 28.6% of software houses do.
More surprisingly, small firms reveal substantially more code than large
ones. Henkel (2005b) argues that these firms would prefer to develop code
in-house, but lack the resources to do so. Small companies frequently rely
on open source communities to fix bugs and improve software (Rossi and
Bonaccorsi, 2005).
7
Henkel analogizes this result to a jukebox, in which multiple patrons with heterogeneous tastes
produce a stream of music that benefits everyone.
8
We note that evidence showing that most licenses earn modest royalties is equally consistent with the
proposition that licensing is efficient.
9
The reason for the loophole is that GPL says that customers are entitled to source code, but confers
no broader right on the general public. In the case of embedded LINUX, the customers tend to be a
small number of device manufacturers who have good reason to keep their software secret. Other tactics
for evading GPL disclosure requirements include releasing drivers as loadable binary modules rather
than source code (done at least sometimes by 53.1% of industry respondents); tying revealed software to
secret or copyrighted code; releasing code after a delay of up to 18 months (35.7% of firms); and re-
designing software architecture so that functions are moved to the (proprietary) application layer
(Henkel, 2005b).
2.4 Signaling
Open source software grew out of an academic environment in which

researchers write articles to improve their career prospects. There is extensive
anecdotal evidence that high-profile open source programmers can trade on
their reputations to gain job offers, shares in commercial companies, and
possibly access to venture capital (Raymond, 1999; Kogut and Metiu, 2000;
Lerner and Tirole, 2002a). Hann et al. (2004) argue that star programmers
are an order of magnitude more productive than their peers, so there is much
to signal. The study of the Apache collaboration by Roberts et al. (2006)
shows that signaling benefits grow stronger as workers rise in rank.
A key feature of traditional patent or copyright incentives is that the
private value of the right increases with the social value of the contribution.
Hence, the prospect of winning an intellectual property right creates a
screening mechanism. When a potential inventor has an idea for an inno-
vation, he will compare its cost to a correlate of its social value before
deciding to invest.
With signaling incentives, the benefits of the R&D investment are still
correlated with the social value of the investment, but less so. Instead of
investing in the products with most consumer value, contributors will
choose projects that showcase their technical virtuosity (Lerner and Tirole,
2002a). Signaling incentives, therefore, explain why open source projects
tend to involve server operating systems, programming languages, and
other applications aimed at sophisticated users (Schmidt and Schnitzer,
2002). Osterloh (2002) argues on this basis that open source collaborations
listen to their smartest users while, for example, Microsoft listens to the
dumbest. Similarly, Kollock (1999) argues that, as a consequence of signa-
ling, mass market software tends to be underserved, and that open source
software ignores ‘‘duller, more complex, but no less useful public goods.’’
Signaling is also weak for such useful tasks as reporting bugs, submitting
comments, suggesting new functionality, preparing documentation, build-
ing easy-to-use interfaces, providing technical support, ensuring backwards
compatibility, and writing programs for utilitarian tasks like power man-
agement or wizards (Schmidt and Schnitzer, 2002; Osterloh and Rota, 2004;
Rossi, 2004). Empirically, the sole Apache rank open to bug hunters has no
measurable impact on salary (Hann et al., 2004).
Signaling is not confined to individuals. Case studies and survey evidence
show that computer companies also participate in open source in order to
build reputation, although the number of such companies seems to be small
(Ghosh et al., 2002; Dahlander, 2004). The Henkel and Tins (2004) study of
embedded LINUX finds that 45.4% of manufacturing companies disclose
code in order to add to technical reputation, and 32.6% disclose to get on
mailing lists as a form of marketing.
As in other incentive mechanisms, signaling can create agency problems,
such as hiding errors or claiming credit for the work of others. Johnson
(2004) argues that open source peer reviewers may collude to hide flaws in
each other’s code. The only empirical evidence of agency problems we have
found is due to Gandal and Ferschtman (2005), who point out that signa-
ling incentives are likely to be more significant for licenses such as GPL
that ban commercialization than for those such as BSD that allow it. They
find that SourceForge contributors submit 2.9 times more lines of code to
BSD-type licenses than to GPL-type licenses, and interpret this data as
evidence that signaling incentives become less important once contributors
deliver enough code to obtain formal credit.
Signaling incentives can also have an impact on code architecture. Schmidt
and Schnitzer (2002) speculate that increased modularity makes individual
contributions more visible.10 Assuming that modularity does promote signa-
ling, Weber (2000) argues that open source members may engage in ‘‘stra-
tegic forking’’ to become a leader, that is, unnecessarily splitting a
collaboration. Dalle and David (2003) similarly hypothesize that program-
mers gain more reputation by launching new code than by contributing to an
existing project; by working on early releases rather than later ones; and by
working on frequently called modules (e.g., kernels) instead of applications.
Dalle and David point out, somewhat optimistically, that these effects can be
beneficial if, for example, working on new modules is socially more valuable
than extending existing ones.
There is extensive evidence that signaling works. Programmers often
receive job offers, stock, and other benefits (Lerner and Tirole, 2002a).
Many programmers reportedly believe that being a member of the
LINUX community ‘‘commands a $10,000 premium on annual wages.’’
(Kogut and Meitu, 2000). Statistical studies by Hann et al. (2004) confirm
that each promotion above the lowest rank boosts Apache programmer
salaries by 13.3–29.3%. Similarly, surveys by Bonnacorsi and Rossi
(2003, 2005) and Henkel (2005b) confirm that many commercial firms use
open source collaborations to find new workers.11 Finally, Lakhani and
Wolf (2005) and Lakhani et al. (2002) use factor analysis on their survey
to sort respondents into four groups, including ‘‘professionals’’ (25%)
who are motivated by signaling (‘‘gaining status’’) as well as solving work
needs. This group is only slightly smaller than their largest group (29%),
‘‘learning and fun.’’
It is harder to know how powerful the signaling incentive is. Lakhani and
Wolf (2005) compare reported incentives against the number of hours
worked each week, and find that signaling (‘‘reputation’’) has only about
one-third as much impact as the most powerful predictor, creativity. How-
ever, this average figure probably obscures the importance of signaling
10
Baldwin and Clark (2003) disagree, arguing that large numbers of modules dilute the superiority of
any one contribution.
11
This pattern is not universal. Henkel and Tins (2004) report that only 11.5% of hardware man-
ufacturers participate in embedded LINUX in order to find potential employees.
within specific projects. For example, Roberts et al. (2006) find that high-
status Apache volunteers contribute more code than other members.
2.5 Education
Due to its emphasis on peer review, open source provides particularly a

good vehicle for education (Lakhani and Wolf, 2005). Education incen-
tives are closely aligned with signaling, but focus on objective skills rather
than the perceptions of others. For this reason, they avoid the free rider
and agency problems referred to above. Education incentives explain why
surveys routinely find roughly one-fifth of all open source volunteers are
students.12
Education also looms large in how the collaborators report their own
motives. In the survey of Lakhani and Wolf (2005), improving skill
(41.3%) is the second most common incentive reported by open source
volunteers, behind intellectual stimulation (44.9%). Other responses are
markedly less important. In the survey of Ghosh et al. (2002), 70.5% of
respondents report that they are motivated by learning new skills and
67.2% are motivated by sharing knowledge and skills. These are by far the
most commonly cited incentives.13 In the factor analysis of Lakhani and
Wolf (2005) and Lakhani et al. (2002), the largest group is ‘‘learning and
fun’’ at 29%.
2.6 Achieving network externalities and denying them to others
Achieving a favorable market position through network externalities is

one of the most important strategic goals for companies in the new econ-
omy. Schmidt and Schnitzer (2002) argue that the importance of network
effects and switching costs ‘‘is largely independent of whether the software
is proprietary or open source.’’ Strategies for solidifying a market position
can be loosely grouped into four categories.
Achieving common standards. Adopting a single, industry-wide open
source standard for software fosters a common pool of skilled workers,
reduces costs associated with unnecessary versioning, increases the total
number of programmers submitting bug reports and extensions, and avoids
transactions cost associated with intellectual property such as ‘‘patent
12
Lakhani and Wolf (2005) report 19.5% of all open source collaborators are students. Hertel et al.
(2003) report a 23% figure, while Ghosh et al. (2002) report 21%.
13
Henkel and Tins (2004) report similar results for the embedded LINUX industry. Among devel-
opers who work for software companies, the most commonly reported motivations are getting better
personal skills (66.7%), recognition (60%), feedback to boost performance in their current job (56.9%),
and demonstrating their skills to future employers (41.7%). Results for developers working for hardware
companies are broadly similar. Not surprisingly, education and signaling incentives are even more
important for contributors who work for universities, non-profits, and hobbyists. The most common
survey responses for this group include obtaining feedback to improve personal skills (75.0%), im-
proving technical reputation (53.6%), and demonstrating skills to future employers (50.0%).
thickets’’ and ‘‘anticommons’’ effects (Ghosh et al., 2002; Lerner and

Tirole, 2004; Rossi, 2004; Rossi and Bonaccorsi, 2005). Such benefits are
particularly attractive for competitively supplied code that is likely to earn
low profit margins in any case (West and O’Mahoney, 2005).
Market penetration. Releasing code in an open source format facilitates
customer acceptance by (a) making it impossible for manufacturers to raise
prices at a later date, (b) creating a community of developers who will
continue to support the code even if the original author abandons it, and
(c) releasing information about APIs that dramatically reduces consumer
switching costs if companies fail to deliver on their promises (Raymond,
1999; Varian and Shapiro, 2003).
Influencing standards. Manufacturers often join open source collabora-
tions in order to steer code in directions that favor their own technology. If
first mover advantages are strong, companies may even race to reveal code
in order to preempt alternatives and gain a permanent advantage (von
Hippel, 2002; Harhoff et al., 2003; Varian and Shapiro, 2003; West and
Gallagher, 2004). The effect of these dynamics is ambiguous. For every
company that hopes to influence code, others may fear that the open source
collaboration will be hijacked so that it no longer supports their needs
(Ghosh et al., 2002; Lerner and Tirole, 2002b). Such fears may persuade
would-be participants that it is better not to support a project in the first
place. Alternatively, companies may decide that joining open source
collaborations is the best way to detect and prevent hijacking.
Blocking market dominance by others. Finally, companies may support
open source as a counterweight to dominant proprietary standards such as
Windows (Kogut and Metiu, 2000). Individuals may similarly decide that
open source is the best way to prevent large corporations from control-
ling the tools on which their livelihoods depend. O’Mahoney (2003) says,
‘‘Informants spoke of their contributions as investment in their future tools:
they are creating code that they will never have to pay someone to use
again.’’ Kogut and Meitu (2000) report that LINUX contributors are
frequently motivated by ‘‘a common fear of Microsoft.’’
Is any of this important? Henkel and Tins (2004) report that only 30% of
manufacturers in the embedded LINUX industry reveal their code in hopes
of making it an industry standard.
2.7 Social psychology
So far, we have concentrated on monetizable rewards. But everyday

experience confirms that charitable voluntarism can also be a significant
force. Experimental economics routinely finds that people contribute
more to the provision of public goods than self-interest alone can explain
(Kogut and Metiu, 2001). In some cases, voluntarism may also offer
quality advantages. For example, blood from volunteer donors contains
fewer contaminants (e.g., hepatitis) than blood purchased on the open

market (Titmuss, 1972).
Social psychology is the natural lens to look at non-monetizable incen-
tives. Following Lakhani and Wolf (2005), we distinguish between extrinsic
motives (i.e., doing an activity for some separable consequence) and in-
trinsic motives based on enjoyment or a sense of obligation or community
(Osterloh et al., 2003a; Lakhani and Wolf, 2005). Extrinsic motivations
include the desire for reputation within the open source community, ‘‘ego
boost,’’ ‘‘feelings of personal efficacy,’’ or other similar incentives (Ray-
mond, 1999; Weber, 2000, 2004). By contrast, intrinsic motivations do not
require an audience. They include creative pleasure, sometimes character-
ized by a ‘‘flow state’’ in which the individual loses track of time (Lakhani
and Wolf, 2005), the desire to be part of a team (Hann et al., 2004), the
ability to express creativity, and experiencing satisfaction and accomplish-
ment (Benkler, 2002; Roberts et al., 2006). Intrinsic motivations also
include altruistic incentives, identifying with a particular group, or ideo-
logical opposition to proprietary software and software makers14 (Kollock,
1999; Hertel et al., 2003; Osterloh et al., 2003b; Rossi and Bonaccorsi,
2005).
In choosing projects, extrinsic motivations may lead volunteers to take
account of the social benefits they confer, but intrinsic motivations do not.
Volunteers will join projects that place a premium on creativity rather than
algorithmic solutions; are neither too easy nor too difficult; are challenging,
fun, and simple to learn; and are interesting, fast-moving and even glam-
orous (Kollock, 1999; Dahlander and Magnusson, 2005). Social psychol-
ogy incentives may also be unusually strong for groups such as teenagers
(Benkler, 2002). This may explain why open source contributors are over-
whelmingly young, male, and single (Ghosh et al., 2002, Lakhani and Wolf,
2005). Henkel and Tins (2004) and Hertel et al. (2003) report, respectively,
that open source contributors are 98% and 96% male.
Commentators claim that social psychology incentives work best when
contributors are only asked to contribute small amounts to the public good
(Baldwin and Clark, 2003; Osterloh et al., 2003b; Osterloh and Rota, 2004).
Nevertheless, even weak incentives can sometimes be important. For ex-
ample, Baldwin and Clark (2003) argue that the decision to reveal previ-
ously written code is a Prisoners Dilemma game. Here, intrinsic motivation
14
The most famous suggestion for obligation/community incentive is due to Raymond (1999), who
argued that open source was driven by a post-modern ‘‘gift culture’’ in which social status is determined
‘‘not by what you control but by what you give away.’’ This view represents an elaboration of an older
anthropology literature in which gift giving creates the ‘‘compulsion to return a gift,’’ confers status and
power on the giver, fosters kinship-like relations in which each person ‘‘takes what they need and gives
what they can’’; and encourages gift recipients to praise the giver and feel solidarity with the community
(Bergquist and Ljungberg, 2001; Zeitlyn, 2003). One problem with Raymond’s (1999) original formu-
lation argument is that it claimed that gift culture was characteristic of a postscarcity society. As Weber
points out, this argument ignores the fact that ‘‘time and brainspace of smart, creative people are not
abundant’’ (Weber, 2000, 2004).
may provide sufficient reward to cover the small (but non-trivial) cost of
communication.15 Similarly, Weber (2000) argues that shared culture and
norms help to suppress open source forking. Even weak norms may be
amplified if network externalities favor a single, dominant standard. This
can happen if the average programmer favors strong economies of scale,
dislikes porting to multiple versions, and wants to minimize conflict (Lerner
and Tirole, 2002a).
The significance of social psychological incentives can also change over
the life of a project. Osterloh and Rota (2004) and Franck and Jungwirth
(2002) argue that social psychological incentives are most significant for
young projects where monetary or reputation incentives are comparatively
weak. On the other side, Lerner and Tirole (2002a) argue that visibility of
early contributions encourages reputation seekers to join at an early stage.
Conversely, social psychology incentives are expected to weaken as the
original volunteers lose energy or mature projects become less fast paced
and interesting (Lerner and Tirole, 2002b; Dahlander and Magnusson,
2005). Social psychology motives may also be less important for company-
sponsored projects, which usually start with a large mass of mature code
that offers little sense of ownership or creative excitement (West and
O’Mahoney, 2005). Finally, ideologically motivated volunteers may decide
that the best way to accomplish their goals is to start as many new col-
laborations as possible. If so, their best strategy may be to leave projects as
soon as signaling or monetary incentives kick in (Franck and Jungwirth,
2002).
A virtue of enjoyment as an incentive is that free riding does not destroy
it, at least in principle (von Hippel, 2002). Open source collaborations based
on social psychology incentives, therefore, escape the ‘‘game of chicken’’
dynamics that cause delay under other incentives. Bitzer et al. (2004)
explore games in which developers gain utility from a combination of con-
ventional own-use incentives and two social psychology incentives (fun and
pleasure from giving gifts). They find that if the value of social psychology
incentives exceeds expected costs, open source members start to produce
code immediately. The situation is different, however, if volunteers need
both types of incentives to cover costs. In this new equilibrium, one member
develops code immediately while the others wait. Bitzer et al. also explore
an extended model in which agents have a finite life because they expect to
change jobs and/or see their human capital become obsolete. In this richer
model, whichever agent expects to receive open source payoffs for the
longest time realizes that she cannot win a waiting game and starts to write
code immediately. Based on this analysis, Bitzer et al. argue that real-world
open source development should be swift. They also predict that open
15
Other incentives, including signaling, may also play a role. Alternatively, Baldwin and Clark (2003)
argue that open source projects may resemble multistage games without a definite time horizon. This
would obviate the problems associated with one-shot Prisoners Dilemma games.
source members will often be unusually talented and well educated, place
unusual emphasis on own use, gift, and play, be patient (have a low
discount rate for future benefits) and also young (hence, have long-time
horizons).
Many observers claim that external incentives can crowd out intrinsic
ones, particularly when the new mechanisms are accompanied by moni-
toring, control, or time pressure (Osterloh et al., 2003b; Osterloh and Rota,
2004). For now, there is little empirical evidence of this. Lakhani and Wolf
(2005) report that mixing salary with creativity and political goals does not
reduce work effort.
Social psychology theorists also predict that intrinsic incentives will
decline if there is a widespread perception that third parties are profiting
from the community’s efforts. Such perceptions are said to make otherwise
willing volunteers feel exploited (Kollock, 1999; Osterloh et al., 2003b).
Once again, empirical evidence is limited. Putative cures for crowding out
include demonstrating that rules exist and are enforced through social
pressure; demonstrated transparency and procedural fairness; encouraging
self governance and allowing volunteers to choose their own projects; pro-
moting group identity; providing clearly articulated goals; and making sure
that contributions are visible so that members know their contributions are
being reciprocated (Kollock, 1999; Kogut and Metiu, 2001; O’Mahoney,
2003; Osterloh et al., 2003b). One problem with these suggestions is that
they are more or less identical to the recommendations that one would
expect from an analysis of games based on ‘‘own use’’ or ‘‘signaling’’ re-
wards. It would be interesting to know whether social psychology literature
implies any distinctive predictions.16
Survey responses suggest that different social psychology incentives have
wildly different strengths. For example, Lakhani and Wolf (2005) and
Lakhani et al. (2002) find that intellectual stimulation is the most commonly
cited incentive (44.9%) among Sourceforge developers. However, other
social psychology incentives appear to be less substantial. These include
believing that code should be open (33.1%), feeling an obligation to repay
the community (28.6%), and deriving enjoyment from a team enterprise
(20.3%). Finally, ideological motives are extremely weak. Only a handful of
respondents (11.3%) report that they participate in open source code in
order to beat proprietary software and only half of these (5.4%) feel so
strongly that they would ‘‘never’’ participate in a closed source project.
Ghosh et al. (2002) similarly find that social psychology motivations tend to
be reported less often than own use, education, or signaling. They find that
the most common social psychology motives include participating in a
‘‘new form of cooperation’’ (37.2%), participating in the open source scene
16
Licenses that block commercialization are said to promote social psychology incentives. See
Section 4.
(35.5%), believing that software should not be a proprietary good (37.9%),

and limiting the power of large software companies (28.9%).
The problem with such self-reported responses is that they may convey
politically correct opinions more than actual incentives.17 When survey data
are combined with information on work effort, intrinsic motives decline
sharply in importance. For example, Roberts et al. (2006) find that intrinsic
motivations are a statistically insignificant predictor of how much software
Apache volunteers contribute. Similarly, Hertel et al. (2003) find that social
psychology incentives and identification with the LINUX community are
poor predictors of how many hours members will work. On the other hand,
Lakhani and Wolf (2005) report that a personal sense of creativity has the
largest impact on hours worked per week. They find that this impact is
twice as large as enjoyment or receiving a salary and three times larger than
reputation incentives.18
Referring again to the cluster analysis of Lakhani and Wolf (2005), the
smallest of the four clusters (19%) consists of respondents who are prima-
rily motivated by obligation or community-based intrinsic motivations.
Within this group, large majorities report that open source is their most
creative experience (61%) and that they lose track of time (i.e., experience
‘‘flow states’’) while programming (73%). Similar majorities either strongly
(42%) or somewhat agreed (41%) that the hacker community provided
their primary sense of identity.
3 Stability and organizational issues
3.1 Who contributes, and how much?
The amount of effort expended by open source volunteers varies widely

from individual to individual. Open source projects typically begin with a
single programmer making a large investment (Raymond, 1999). Even after
additional volunteers join the collaboration, much of the work continues to
be done by small minorities. Ghosh et al. (2002) report that 10% of
Sourceforge developers create 74% of code. Similarly, Mockus et al. (2002)
tell us that 15 core developers provide 83% of all Apache contributions,
while Von Krogh et al. (2003) report that about 1% of Freenet’s members
17
The appeal of ‘‘politically correct’’ responses is particularly evident for embedded LINUX. Despite
obvious business motivations, Henkel and Tins (2004) report that more than 90% of the developers
claim to be motivated by a desire to give code back to the community. This was true whether the
developers worked for hardware companies (93%), software companies (91.7%), or non-profits and
universities (92.6%).
18
Social psychology motivations seem to have little effect on corporate behavior. A survey of 146
Italian open source companies found that ideological statements had no measurable effect in predicting
whether a company would contribute to open source programs. The authors conclude that corporations
express such sentiments in order to please developers (Rossi and Bonaccorsi, 2005).
account for 50% of all developer e-mails.19 This disproportionate reliance

on small minorities is particularly evident for new code. Bergquist and
Ljungberg (2001) report that almost all new functionality is provided by
small groups, while Von Krogh et al. (2003) report that 13% of Freenet
developers provide 53% of all new code. In open source tech support,
2% of tech support providers supplied 50% of all answers (Lakhani and
Von Hippel, 2000).
Small contributors are nevertheless important. This is particularly true
for bug reporting.20 In many open source collaborations, these functions
are almost entirely performed by volunteers (Bergquist and Ljungberg,
2001). Mockus et al. (2002) estimate that 87% of the Apache members who
report bugs submit just one report. Hann et al. (2004) report that many
Apache volunteers who report bugs have just one encounter with the
project. While less glamorous than creating new code, these activities may
be more valuable than creating code in the first place. According to Bessen
(2004), testing, debugging, and maintenance account for 82% of software
costs.21
Finally, firms face special obstacles in trying to convince open source
volunteers to form communities around their products. In particular, firms
must persuade volunteers that the code has value as an open source project
and is not simply being abandoned because it is technically flawed or losing
market share (Lerner and Tirole, 2002b; West and Gallagher, 2004). Per-
haps the best way for companies to demonstrate that the code has promise
is to make high-profile investments in building an open source collabo-
ration. This can be done by supplying personnel, offering infrastructure
like user tool kits, providing awards and other recognition mechanisms
for contributors, bundling open source code with company products, and
providing coordination functions like recruiting potential contributors,
integrating their efforts, building online forums and mailing lists, and
developing governance structures (Kogut and Meitu, 2000; West, 2003;
West and Gallagher, 2004; West and O’Mahoney, 2005).
3.2 Who pays?
The next big question is who pays for open source. Firms participate
in open source communities almost as much as individuals do. Roughly
half of all open source workers are directly or indirectly supported by
19
The small size of core groups may be enforced by technology. Mockus et al. (2002) argue that cores
larger than 12–15 members find it almost impossible to coordinate their actions. Once this point is
reached, the number of incompatibilities generated by new software quickly becomes unmanageable.
20
Bug patching represents an intermediate case: the number of people who fix bugs is an order of
magnitude larger than those who write new code but an order of magnitude smaller than those who
merely report bugs (Kogut and Metiu, 2001; Mockus et al., 2002).
21
Kogut and Metiu (2001) similarly report that maintenance activities account 50–80% of all software
costs. See also Raymond (1999).
corporations.22 Ghosh et al. (2002) report that 54% of respondents were

paid for open source work; Lakhani and Wolf (2005) report that 55% of
respondents contributed code during work hours; and Hertel et al. (2003)
report that 43% of LINUX kernel developers sometimes or always receive
salary for their work.
Such data probably understate the importance of corporate support,
since other survey data suggest that paid open source contributors work
more hours, are older and more educated, and devote more time to com-
munication and coordination activities. In the data of Lakhani and Wolf
(2005), paid workers spend roughly twice as much time on open source
projects as unpaid ones. Similarly, Roberts et al. (2006) show that salary is
an important predictor of effort. Similar arguments can be found in Ghosh
et al. (2002) and Kogut and Meitu (2000).
3.3 Why licenses?
Most open source communities assume that restrictive licenses like GPL
are beneficial or at least unavoidable. However, the need for licenses is not
entirely obvious nor, assuming that licenses are needed, is it clear which
restrictions are necessary or desirable.23 From a welfare standpoint, the best
way to ensure use and re-use of software would be to place it in the public
domain without any license at all. This strategy would also be simpler to
implement than the elaborate licenses that open source actually uses. This
section describes five possible reasons why open source licenses might, after
all, be necessary. Intriguingly, most can be satisfied with licenses that are
significantly less restrictive than GPL.
Symbolism. Dahlander and Magnusson (2005) argue that licenses are
sometimes chosen for symbolic reasons. Presumably, the need for symbol-
ism is linked to social psychology incentives that would erode in the pres-
ence of private gain, Lakhani and Wolf (2005), or the absence of rules
enforcing reciprocity Kollock (1999). By now, there is reason to be skeptical
of these explanations. For example, Kogut and Metiu (2001) argued five
years ago that the comparatively restrictive Apache license posed a threat to
the ‘‘generalized reciprocity that characterizes the community culture.’’
Despite this, Apache has thrived.
22
Approximately, one-third (32%) of the world’s 25 largest software companies engage in significant
open source activities. IBM reputedly spent $1 billion on open source projects in 2001 (Ghosh et al.,
2002).
23
Like most standards, choice of license exhibits strong network effects. open source collaborations
that adopt widespread preexisting licenses face fewer legal impediments to sharing and merging code.
They also save the substantial costs associate with writing and learning a new license (Lerner and Tirole,
2002b). The fact that three-fourths (72%) of all SourceForge collaborations use GPL suggests that
network effects are substantial. In this view, the dominance of GPL is little more than an historical
accident—neither inevitable nor optimal.
Lerner and Tirole (2002b) present a more nuanced argument based on the
hypothesis that social psychology incentives and GPL licenses only become
necessary when alternative incentives (e.g., ego gratification, signaling, and
own use) are weak. This hypothesis may explain why GPL-style licenses are
much less prevalent among projects aimed at developers and system ad-
ministrators than for games and other applications aimed at end users.
Similarly, the symbolic use of licenses may be particularly tempting when
open source collaborations with corporate partners place social psychology
incentives under threat. Many commentators argue that ideologically
motivated licenses and norms pose formidable barriers for most firms
(Bonnacorsi and Rossi, 2003; West, 2003; Dahlander, 2004; Dahlander and
Magnusson, 2005). That said, the experience of Red Hat and several other
companies shows that GPL licenses do not pose insurmountable barriers to
business (Osterloh et al., 2002; Weber, 2004).
Protecting complementary investments. We stressed above that incentives
for participation arise partly through proprietary complements, whether
human capital or commercial products. The license may be integral to this,
for example, by preventing users from removing an author’s name (Franck
and Jungwirth, 2002). Licenses can also shield programmers from potential
liability (Gomulkiewicz, 1999; Bonaccorsi and Rossi, 2003).
Preventing forking and hijacking. GPL-style blanket prohibitions on
commercialization can keep open source collaborations from forking or
being diverted in unintended directions (‘‘hijacking’’). However, this prob-
lem can also be addressed by alternative and less intrusive measures such as
giving a trusted leader the exclusive right to decide which changes and
extensions become part of the official code (Bonaccorsi and Rossi, 2003;
O’Mahoney, 2003), social pressure (‘‘flaming’’ and ‘‘shunning’’) (Raymond,
1999), trademark (O’Mahoney, 2003), and the use of charisma or astute
political skills (Raymond, 1999; Weber, 2000). Furthermore, it is unclear
how much danger forking poses in any case. Despite several historical ex-
amples (see Lerner and Tirole, 2002a; Varian and Shapiro, 2003), only
11.6% of embedded-LINUX respondents see forking as a threat (Henkel
and Tins, 2004). Even if no special steps are taken, Weber (2000) argues
that network effects tend to suppress forking in any case.24
LINUX provides an instructive example of how measures short of GPL-
style prohibitions can be used to prevent forking. Although leaders decide
which code carries the trademarked LINUX name, users remain free to
develop unbranded versions. From the point of view of consumer sover-
eignty, this may be better than to prohibit forking altogether.
24
Weber argues that forkers necessarily forfeit the benefits of a large community. Thus, they cannot
offer followers a built-in audience needed to support signaling incentives or generate bug reports.
Community members may also resist forking on the theory that it opens the door to further splintering
and balkanization.
Ideology. Franck and Jungwirth (2002) argue that licenses reassure

activists that their donated labor will not be diverted to commercial ven-
tures that enrich open source leaders.25 More broadly, activists may adopt
GPL-style licenses in order to force programmers to choose between writing
non-commercial extensions of code and no code at all. Such tactics pre-
sumably increase open source software production while reducing the total
stock of open plus-proprietary extensions in the market.
Stabilizing open source against IP incentives. We argue above that, absent
GPL, an opportunistic developer will be tempted to free ride on earlier
developers and make his code proprietary so he can impose royalties on
later ones. This argument for why GPL may be necessary in the cumulative
context is similar to an argument of Hall (2004) and Gambardella and Hall
(2005). They point out that a community member who ‘‘cheats’’ by making
his code proprietary gets a discrete jump in income, but only reduces the
community’s output of publicly available by an infinitesimal amount. This
leads to a Prisoner’s Dilemma in which each community member might
choose proprietary rights even though they would prefer a collective strat-
egy that preserved open source. The game can be stabilized if actors are
required to make decisions within large, organized groups. Norms, defer-
ence to lead researchers, and GPL-style licenses all serve this function. The
net result is to stabilize open source so that proprietary code is no longer an
‘‘absorbing state.’’
3.4 Why leadership?
Open source communities are often romanticized as collections of atom-

istic volunteers who self-organize in minimally hierarchical environments.
At the same time, open source leaders like Linus Torvalds have cult status.
Clearly, self-organization has its limits. In general, open source collabora-
tions rely on leadership to solve a variety of information, agency, and
coordination functions.
Information problems are particularly important in the early stages of
open source collaborations, when would-be contributors must decide
whether the proposed project is feasible, interesting, and potentially useful.
The easiest way for leaders to demonstrate these qualities is to provide
working software, even if it is incomplete and flawed (Lerner and Tirole,
2002a).26 Survey respondents report that the most important leadership
functions are providing an initial code base (48.6%), writing code (34.3%),
and creating a promise/vision (32.3%) (Lakhani et al., 2002). After an open
25
Franck and Jungwirth (2002) further analogize GPL’s restrictions to the ‘‘non-distribution’’ clauses
that prevent traditional charities from diverting surpluses to managers or outsiders.
26
Leadership is especially important where the open source collaboration is built around donated
corporate software. Open source volunteers frequently interpret the corporation’s decision to forego
intellectual property rights as evidence that the code is worthless (Lerner and Tirole, 2002b).
source collaboration has been established, information problems change.

However, there must still be a procedure to decide which contributions
should be included in new releases. Trusted leaders play a key role in this
process (Lerner and Tirole, 2002a).
Leadership also mitigates agency problems. For example, corporations
can commit to keep code in the public domain or to highlight individual
contributions by surrendering control to outside leaders (Lerner and Tirole,
2004). Following Max Weber, successful leaders must persuade volunteers
that their objectives are congruent and not polluted by ego, commercial
incentives, or biases (Weber, 2000).
Finally, leadership accomplishes a variety of coordination functions. For
example, leaders let volunteers know which projects are worth supporting
and, conversely, which constitute forking or abandonment. More generally,
someone must make basic architecture choices (e.g., modularity) and co-
ordinate work by volunteers. These functions mitigate the delays and
inefficiencies that arise when volunteers interact through decentralized
games (Kogut and Meitu, 2000; Lerner and Tirole, 2002a).
3.5 Network effects
Proprietary software creates significant network externalities, and it is

hard to see why open source software would be any different (Schmidt and
Schnitzer, 2002). For example, there is anecdotal evidence that small de-
clines in the popularity of open source projects can snowball into collapse
(Lerner and Tirole, 2002a).
There are several reasons to believe that open source collaborations
should exhibit significant network effects. First, Weber (2000) argues that
large open source communities have an intrinsic advantage over small ones,
since they are more likely to include ‘‘outliers who have a high level of
interest and surplus resources of time and mindspace.’’ These outliers may
reflect variance within a single population of contributors or else the pres-
ence of distinct subpopulations within a given open source collaboration.27
Second, open source projects need a critical mass of participants. If a
competing proprietary system siphons off volunteers and ideas, open source
becomes less exciting (Lerner and Tirole, 2002a). Third, some important
open source incentives (e.g., ego gratification, signaling) scale with the size
of the audience (Lerner and Tirole, 2002a). Finally, bug identification and
bug fixing also scale with the number of users.
With network externalities, it is natural to expect multiple equilibria and/
or a winner-take-all dynamic (Weber, 2000; Lerner and Tirole, 2002a).
Surveys of open source projects are consistent with this intuition. Healy and
Schussman (2003) observe, ‘‘It seems clear that for every successful open
27
One frequently overlooked cost factor is competing uses of workers’ time. Hertel et al. (2003) report
that ‘‘tolerance for time losses’’ is an important predictor of effort.
source project there are thousands of unsuccessful ones.’’28 Indeed, most

open source projects are very small scale. Ghosh et al. (2002) find that 29%
of projects have just one author, only 45% had more than two, and only
1% had more than 50%.29 Similarly, most open source projects never pro-
duce much code.30 Healy and Schussman comment ‘‘Raymond’s image of
the bazaar does not capture the fact that the typical project has one de-
veloper, no discussion or bug reports, and is not downloaded by anyone.’’
In this environment, agonizing about the choice of license or modular
architecture is largely symbolic. Project success may depend on accident as
much as quality.
An interesting angle explored by Mustonen (2005) is that proprietary
software firms can use the open source community to enhance their own
networks.
4 Efficiency implications
As pointed out earlier, the open source mechanism does not include a
means to appropriate benefits conferred on third parties. We would there-
fore expect underprovision. Other incentives, not based on appropriability
(e.g., signaling and education), not only mitigate this problem but also pose
a risk of overprovision. In this section, we discuss some of those issues, and
take the opportunity to contrast the inefficiencies of open source against the
inefficiencies incurred by the more ordinary use of intellectual property.
4.1 Disclosure of code
As we have discussed before, a failing of proprietary software is that

proprietors almost never make their source code available to users. Open
code enhances welfare in several respects. It is easier to adapt and reuse, and
therefore may retain its value rather than becoming obsolete (Raymond,
1999). It facilitates finding, diagnosing, and fixing bugs (Schmidt and
Schnitzer, 2002). And it reduces entry costs for firms that supply custom-
ization and support services, increasing the likelihood that such services will
be competitively supplied (Hawkins, 2002).
28
Provocatively, Madey et al. (2005) find a power law distribution for the number of developers
involved in open source projects. The reasons for this distribution are obscure.
29
See also, Comino et al. (2005), reporting that 80% of SourceForge projects have at most two
developers. Ninety nine percent of projects have 16 or fewer developers. Healy and Schussman (2003)
report that the median number of developers in SourceForge databases is one and that projects in the
95th percentile have only five active developers. Ghosh and Prakash (2000) report that 76% of all
projects in their survey have just one author and only 2% have more than five.
30
According to Comino et al. (2005), 80% of SourceForge projects show no activity since registration.
Healy and Schussman (2003) find little or no programming activity in more than half of all SourceForge
projects.
But if these advantages are so substantial, why does not the proprietary
software industry exploit them for profit? Although formal intellectual
property protection does not require disclosure, neither does it prevent it.
Given better enforcement (stronger protection), proprietary firms might
decide that the protection against copying that comes from keeping the
code closed is no longer necessary. From this standpoint, the failings of
the proprietary software industry arise from the weakness of intellectual
property, not from its strength.31 In fact, proprietary firms already share
source code to some extent. Examples include, giving developer toolkits
to licensees (Bessen, 2004), sharing source code with selected developers
(Microsoft’s ‘‘Shared Source Initiative’’), and promoting code reuse inside
Microsoft itself (Lerner and Tirole, 2002a).
In yet another twist, Lewis and Talley (2006) point out that developers in
the open source community may also withhold their code for strategic
reasons, namely, to encourage others to invest time and effort.
4.2 Meeting users’ needs
We distinguish here between the incentive to meet the needs of users, and
the ability to do so. Obviously, own-use incentives are directed to the user’s
needs, even if not to the needs of third parties. However, a programmer’s
incentive to signal his proficiency, or to become a better programmer, or to
participate in a community of altruistic providers, may not be. Since those
incentives are not based on appropriating value from users, there is no need
for the innovative activity to track user needs particularly closely.
Lakhani and Wolf (2005) tell us that about 58% of all volunteers are IT
professionals. Despite their proficiency in writing code, there is no obvious
reason they would do mundane tasks useful to third parties like testing
for usability (Lakhani and Wolf, 2005), learning unfamiliar programming
languages and architectures (Kogut and Meitu, 2000, Von Krogh et al.,
2003), deciphering complex commercial code (West and O’Mahoney, 2005),
or meeting the needs of highly specialized audiences such as lawyers or
accountants (Schmidt and Schnitzer, 2002). Skilled programmers get most
benefit from creating tools for other IT professionals. See West (2003),
discussing Internet service providers; Johnson (2002) arguing that open
source produces more utilities than end-user applications; and Comino
et al. (2005) reporting that two-thirds of all SourceForge projects involve
software languages, systems, internet code, communications and multi-
media tools, or scientific software.
However, there is a countervailing argument, namely, that open source
communities have closer contact with their users than owners of proprietary
software, and therefore have a better ability to meet their needs (von Hippel,
31
This can be overstated. If intellectual property over-rewards proprietary software, then it creates
deadweight loss that would better be avoided.
2002; Henkel and von Hippel, 2005). Mockus et al. (2002) point out that
programmers in large proprietary software projects frequently do not know
the domain for which they are writing. User feedback is particularly valuable
where consumer needs cannot be reduced to a few simple criteria of merit
(Kogut and Metiu, 2000; von Hippel, 2005). Such information is even more
valuable if workers are ‘‘lead users’’ who understand needs, risks, and even-
tual market size before manufacturers do. In this view, open source allows
user-developers to play a larger role in developing and extending the products
they use (Kogut and Metiu, 2000; Varian and Shapiro, 2003).32
4.3 Deadweight loss and pricing
Proprietary prices are generally above competitive levels. The proprietary

price reduces consumption of the patented good and causes users to shift
their use to less-preferred substitutes. Proprietary pricing can also lead to
costly R&D to invent around patents, and may reduce the incentive for
second-generation innovators either directly through royalties or indirectly
by forcing them to invent around (Henkel and von Hippel, 2005). Open
source avoids these problems by dispensing with intellectual property. Of
course, that leaves the puzzle of where the developers get their rewards,
which has been the subject of most of this essay.
4.4 Training and using programmers
Proprietary software firms find it hard to observe the competence of a

programmer, and therefore cannot tailor wages to programmers’ marginal
productivities (Gaudeul, 2004; Johnson, 2004).33 We, therefore, expect in-
ferior programmers to self-select into private sector jobs that offer a high
average wage. Once hired, low-quality programmers also have various in-
centives to hide errors. If they report bugs, they will likely be asked to fix
them34 (Johnson, 2004). If programmers collude in not reporting each
other’s failings, firms will find it hard to know whether they have good code
(few bugs). This undercuts the value of peer review (Johnson, 2004).
Proprietary firms may be forced into suboptimal architectures to control
these agency problems. For example, they may substitute top-down super-
vision for a more stimulating environment in which programmers can
32
Proprietary firms try to replicate these advantages by using in-house developer networks to appraise
and debug software (Kogut and Metiu, 2001). Microsoft’s Shared Source Initiative similarly lets select
customers view, customize, and patch an open version of Windows (Lerner and Tirole, 2004).
33
Gaudeul (2004) also suggests alternative reasons why wages could be higher than marginal product.
These include legal costs associated with enforcing intellectual property, organizational costs associated
with setting up and managing a firm, and compensation for restricting the developer’s choice of projects.
34
This result requires the admittedly ‘‘strong assumption’’ that contracts are written in a way that pays
programmers a flat wage no matter how many tasks are discovered and assigned ex post (Johnson,
2004).
devise their own projects and work in parallel on many projects, recognizing
that many should fail (Kogut and Metiu, 2000). A management-intensive
way of organizing the development effort reduces worker satisfaction as well
as worker effectiveness, and firms must offer a wage premium to overcome it.
By contrast, open source programmers can join whichever projects fit their
personal interests (Mockus et al., 2002; Rossi et al., 2004). Most open source
incentives tie rewards to the actor’s own efforts, and therefore avoid the
management problems we have identified for proprietary firms.35 Other
things equal, we expect talented programmers to self-select away from the
private sector and into open source.36
Mockus et al. (2002) and Rossi et al. (2004) take this argument one step
further by arguing that talented programmers also self-select into whichever
open source project promises to make the best use of their skills. Perhaps
the clearest articulation is due to Benkler (2002), who argues that open
source selects ‘‘the best person to produce a specific component of a project,
all abilities and availabilities to work on the specific module within a specific
time frame considered’’ (emphasis original). Such claims are almost certainly
overstated. While there are surely selection effects, there is no obvious way
to aggregate information in order to match programmers efficiently with
projects. The same difficulty arises for patents and prizes (Scotchmer 2004,
Chapter 2).
We can ask the same question for labor practices as we asked for disclosure
of source code: If open source practices are best, why do not proprietary
firms emulate them? The answer, at least to some extent, is that they do.
First, some proprietary firms deliberately foster reputation incentives by
attaching programmers’ names to code. Compared to open source, however,
the results are ambiguous. On the one hand, giving credit attracts and elicits
more effort from talented workers. On the other, the strategy increases the
risk that star employees will be hired away by competitors (von Hippel, 2002;
Lerner and Tirole, 2002a). Second, many firms try to create work environ-
ments that respect motives like reciprocity, altruism, and being ‘‘part of a
team’’ (Lerner and Tirole, 2002a). Finally, some firms have experimented
with decentralization. Microsoft’s efforts to have employees challenge each
other’s ideas are one step in this direction (Kogut and Meitu, 2000).
So far, we have treated programmers’ self-selection decisions as immu-
table. This is reasonable for incentives based on complementary goods and
services, since programmers who join open source communities in order to
build their own businesses will presumably stay there. On the other hand,
reputation and signaling incentives are meaningless unless programmers
35
This is clearly not true to the extent that corporations pay employees to work on a particular project.
Even here, however, open source may have significant advantages to the extent that paid workers also
have non-cash incentives and/or must interact with members who have such incentives.
36
We note that this effect is not unique to open source. Proprietary firms frequently reduce agency
problems by adopting incentives—notably prizes and patents—that tie worker rewards to measurable
results.
eventually return to the private sector. In practice, reviewing code is

expensive, and firms prefer to use open source hierarchies and ranking
systems as a proxy for quality (Lerner and Tirole, 2002). Weber (2000)
argues that open source rankings are inherently trustworthy because
programmers know that the best way to build reputation is to work with
high-quality collaborators and exclude everyone else.
4.5 Free riding
If ideas are not scarce, that is, if any good idea is likely to be had and
implemented by someone else—it is tempting to let someone else bear the
development cost. In this environment, open source programmers who are
otherwise willing to write code may wait in hopes that someone else will do
the job first (Raymond, 1999). Users of code who do not contribute are free
riders.
The ability to free ride can reduce welfare. Scholars have used game
theory to explore these results. In general, they find equilibria with pure
strategies in which some developers invest and others free ride, and other
equilibria with mixed strategies in which each developer works with some
probability and development sometimes fails.37 Communities that play
mixed strategies will deliver code more slowly or less reliably than propri-
etary software companies do. (Johnson, 2002; Baldwin and Clark, 2003;
Bitzer et al., 2004).
Patent incentives have the opposite problem. Much of the literature on
patent incentives is concerned with racing, given that only the winner will
have rights. Depending on the commercial value of the patent, racing may
lead to too much investment. (See Scotchmer, 2004, Chapter 4, for an
overview.)
4.6 Modularity and the organization of the research effort
The degree to which software is modularized affects its suitability for the
open source style of development. This suggests that leaders can make open
source collaborations stronger by designing more modular architectures.
For example, Benkler (2002) claims that modularity (‘‘granularity’’) is the
key to making open source projects viable. His conjecture is supported by
survey data suggesting that own-use programmers spend relatively few
hours on open source, and tend to lose interest as soon as their own narrow
needs are addressed (Roberts et al., 2006).
Johnson (2002) makes a more systematic investigation of the role of
modularity in open source. He considers a game in which each developer
37
Open source members clearly understand this logic. Hertel et al. (2003) find that LINUX kernel
members work harder when they believe that their contributions are ‘‘highly important’’ to further
progress.
can choose to create one module based on his own individual ratio of
personal benefit to cost. In this model, each developer invests if the added
probability of success provided by his own investment, weighted by his
personal valuation of the module, outweighs his own personal cost. The
probability that at least one member succeeds depends on the number of
members who decide to invest. Johnson finds that adding a programmer to
the pool, with randomly drawn cost and benefit for each module, typically
increases the likelihood that the code will be supplied. Assuming that the
ratio of personal benefit to cost is bounded, he finds that the total number
of volunteers who write code in any particular period approaches a limit as
their numbers increase. Conversely, free riding increases. Johnson shows
that the total investment is smaller (produces less software) than the effort
needed to maximize social value in expectation. Intuitively, this result follows
from the fact that each programmer’s decision to contribute depends only on
her own personal benefits, and does not reflect spillovers to other users.
Johnson also investigates how variance in (personal) benefit/cost ratios
affect the optimality of open source compared to single-source develop-
ment. Suppose that each of k modules is necessary for completion of the
whole. If a single programmer must create the entire project, the code will
not be written unless the ratio of total benefit to total cost over all k
modules is greater than one for at least one developer. Apply the same test
on a module-by-module basis. In this case, the maximum benefit/cost ratio
could be greater than one for each module separately even though the
maximum ratio of the totals was less than one and vice versa. For this
reason, a modularized project can succeed where a non-modularized project
fails. Nevertheless, Johnson argues that the success rate of modular projects
grows with the number of developers because programmers can self-select
into working on the modules for which they are most proficient. The main
drawback is that the project may founder if costs are badly distributed
among modules.
Baldwin and Clark (2003) explore how modularization fares when different
programmers can communicate to avoid redundant work. Since workers
share costs, a collective effort with adequate communication is always pref-
erable to coding in isolation. Free riding may increase, however, if commu-
nications are so slow that programmers cannot monitor which modules have
been already written or if the number of developers exceeds the number of
modules.38 Even in this environment, increasing the number of modules re-
duces free riding. Similarly, systems with large potential payoffs will normally
attract more workers per module provided that there is enough randomness
in outcomes and if costs are small enough.39 Developers may intentionally
duplicate each other’s efforts in order to obtain a higher best outcome.
38
This may be a realistic concern for very large projects like LINUX or Apache. The problem is
mitigated if most members are paid to participate by small group of employers.
39
See also, Johnson (2002).
The degree to which programmers can coordinate their efforts, either to

avoid duplication or to reinforce efforts to achieve particularly valuable
modules, depends on how isolated they are. Ghosh et al. (2002) report that
17.4% of open source developers have no regular contact with the rest of
the community and 67.9% had regular contact with less than five other
members. Only 17.3% had regular contact with more than 10 other mem-
bers. In principle, better and more frequent communication could improve
the performance. Corporate support, where it exists, may help fill this gap
by supporting the ‘‘professional elite’’ who maintain regular contact with 50
or more developers(Ghosh et al., 2002).
5 Open source and proprietary software
Open source limits the market power of proprietary code by providing

competition and the threat of entry (Bessen, 2004; Henkel and von Hippel,
2005). There are two possible scenarios in which this can happen: first, open
source and proprietary code can compete head-to-head within a single
market. Second, they can occupy separate niche markets. Commentators
have explored both scenarios.
5.1 Competition between open source and proprietary software
One explanation for why proprietary software can survive in competition

with open source products is that users prefer it, either because it is better
code or because it is user friendly. Another reason might be that proprietary
software becomes entrenched by network effects (Casadesus-Masanell and
Ghemawat, 2006).
We have already uncovered several reasons why proprietary firms might
have a quality advantage. First, open source incentives probably capture
less social value than the patent monopoly does, so that certain projects will
not be developed in that community (Schmidt and Schnitzer, 2002). Second,
if open source relies on signaling incentives, and a programmer receives no
benefit from the unglamorous task of making the code user friendly, no one
will bother. Third, we have seen that open source contributors may delay
writing or improving code in hopes that others will step in instead. Since
corporations suffer the opposite coordination problem—they want to be
first to market with a good product—we expect them to act faster whenever
it is profitable to write good code.
There is some evidence that commercial code has a quality edge over
open source, at least for some users. Consulting studies report that the
up-front costs of purchasing Windows are approximately offset by lower
system and administrator costs compared to LINUX. For now, it is unclear
whether the cancellation is exact. Varian and Shapiro (2003) suggest that the
life-cycle costs of owning LINUX may be 10–15% cheaper than Windows,

but remark that this difference is not ‘‘striking.’’ We remark, though, that
such studies face serious methodological challenges including (a) difficulties
in comparing non-comparable software, (b) poor metrics for reliability and
quality, (c) sensitivity to IT wages, and (d) limited information about
how much additional IT effort is needed to support LINUX compared to
Windows (Robert Frances Group, 2002; Cybersource, 2004; Metagroup,
2005).
Mustonen (2003) uses a quality differentiation model to show that low-
quality open source can sometimes coexist with high-quality proprietary
products. The model assumes that both the programmer and user markets
are segmented. Mustonen finds that high-quality programmers self-select
into the open source community where they are better rewarded for their
skills even though the proprietary sector continues to produce superior
products by hiring large numbers of low-quality programmers. He also
assumes that users with high demand are willing to pay a high price for
superior software. Provided that consumers can install software at low cost,
Mustonen finds an equilibrium in which consumers who place a low val-
uation on software always choose open source over proprietary code and
the proprietors set prices based on demand from high-valuation consumers.
If installation costs are high, Mustonen finds that proprietary software
dominates for large markets that value high-quality software. Conversely,
proprietary firms tend to abandon markets that are small or which place a
low value on quality.40
Kuan (2001) presents a model where proprietary companies offer differ-
ent versions of their software to high- and low-quality consumers. Absent
perfect price discrimination, she finds that low-value consumers always
switch to open source. However, the willingness of high-value consumers
and consumer/programmers to switch to open source is less clear, depend-
ing on how many such individuals exist. In general, Kuan finds that the
analysis is ‘‘very undetermined.’’
Casadesus-Masanell and Ghemawat (2006) construct a dynamic model
of entrenchment through network effects in which a proprietary product
and open source product coexist. Starting from a situation where the
proprietary product has a network advantage over the (intrinsically sim-
ilar) open source product, they ask whether the open source product can
supplant it. Although the price of the proprietary product will be higher in
each period, the proprietor will choose the path of prices so as to maintain
a market advantage through the network advantage. Unless network
effects are weak, proprietors never lower prices enough to drive out the
40
Bessen (2004) similarly predicts that open source may crowd out proprietary software in cases where
markets are small or only modest effort is needed to develop an initial product. Where markets grow
rapidly, open source may use this advantage in small markets to gain a first-mover advantage over
proprietary competitors.
open source product entirely. Open source and proprietary products

therefore coexist forever.
5.2 Market segmentation
Even if open source and proprietary code do not compete directly,

they may serve different niche markets. Bessen (2004) notes that packaged
software like Microsoft Windows accounts for only about 30% of all soft-
ware spending. The remainder is customized software created by and for
individual users, often in open source communities. However, creating
customized software products is expensive. Bessen argues that, because it
is difficult to write an enforceable contract for customized software in
advance, proprietary firms must negotiate a price after the software pack-
age is created. But with a single customer, the customer will have a lot of
bargaining power (Aghion and Tirole, 1994). Anticipating this, the propri-
etary firm will be loathe to invest and would-be customers will be forced to
assemble the software from open source modules. Bessen argues that users
who place a low value on quality or have highly unique needs will choose
open source. Users whose needs can be satisfied with simpler applications
will continue to use proprietary products.
Gaudeul (2004) explores how market segmentation can arise from
developers’ profit-maximizing choices about how best to exploit their code.
The developer’s first choice is to copyright the software and hire developers
to implement it, collecting proprietary profit. This option will not be
attractive, though, if wages are too high. The innovator’s second choice is
to obtain development efforts at zero cost by choosing a GPL license.
However, GPL strategies are only feasible if the project has sufficient value
for the open source community to participate. The innovator’s only other
choice is to purchase developed software on the open market. This can be
done by offering developers a BSD license. Gaudeul argues that GPL is
usually worse from a welfare point of view; since, software is developed
with lower probability than it would be in a proprietary regime where
corporations do all work immediately or a BSD regime in which program-
mers race for proprietary rights. However, these differences will likely be
minor where development costs are small; furthermore, there may be cases
in which rising costs reduce the value of proprietary software faster than
they undermine GPL incentives. If so, GPL may sometimes generate more
social welfare than copyright incentives ‘‘for medium level[s] of c[ost].’’41
41
Gaudeul recognizes that his two-stage model is relatively simple. He adds that more complex models
would include wasteful duplication due to racing; GPL developers’ ability to earn income by selling
complements; and the possible advantages of GPL in multi-stage games where developers continue to
add new ideas over time.
6 Limitations and extensions
6.1 Limits to open source software
Open source incentives do not perform all tasks equally well. For this
reason, otherwise useful projects can be limited by open source’s weakness
in supplying specific components such as documentation or tech support.
Henkel and Tins (2004) report that lack of documentation is the most
commonly reported obstacle (57.5%) to embedded LINUX.42 Gambardella
and Hall (2005) similarly argue that open source incentives are poorly
suited to downstream tasks like assembling knowledge into unified, user-
friendly formats. GPL licenses may be counterproductive in this situation if
they displace intellectual property rights needed to extract full value from
society’s investment in knowledge.
Many commentators assume that open source networks are larger than
corresponding corporations and, for that reason, draw on more widely
scattered ideas and expertise (Kogut and Metiu, 2001; Benkler, 2002;
Gambardella and Hall, 2005).43 Wide dissemination is said to be partic-
ularly important in fast changing or complex products (Dahlender, 2004).
As we have seen, however, most open source collaborations are far smaller
than corporations. Arguments based on the superior ability of open source
to elicit knowledge should therefore be viewed with caution except, per-
haps, in the case of Open Science.
Finally, we have seen that open source works best for modularized
software. However, it is not clear whether the number of modules is a free
parameter. Lerner and Tirole (2002a) speculate that the ability to break
projects into modules may be a technical accident peculiar to UNIX that
will likely fade as open source moves to other languages. Modularity may
also be limited by the ability of a small central group to screen submissions
for quality and consistency.
6.2 Beyond software: drug discovery, geographic information systems, and

Wikipedia
Unlike previous eras, 20th century R&D is specialized in delivering

complex systems that required more man-years of effort than a human life
span could supply. Examples included airplanes, automobiles, rockets and
spacecraft, automobiles, pharmaceuticals, silicon chips, and most recently
large computer programs. Prior to open source, all such projects relied on
42
Open source sometimes works surprisingly well despite seemingly implausible incentives. For ex-
ample, Lakhani and Von Hippel (2000) report that open source volunteers sometimes provide tech
support in order to gain information about bugs. Weber (2000) argues that open source collaborations
can accomplish even mundane tasks if a handful of members have unusual preferences.
43
This is related to our argument above that disclosure shakes loose the ‘‘scarce ideas’’ that might
otherwise remain untapped.
organizing large numbers of contract researchers within rigid, top-down

hierarchies. From Moscow to Los Angeles, the scene was strikingly similar:
hundreds and even thousands of engineers sitting at drafting tables or, more
recently, computer terminals. One of the most striking aspects of LINUX
is that it is possible to organize at least one form of complex invention—
large operating systems—a different way. It is, therefore, tempting to ask
whether the open source model can be extended to other information
goods. Because of their high information content, the most natural tech-
nologies are online reference works (e.g., Wikipedia) drug discovery and
geographic information systems (GIS).
The Internet hosts innumerable blogs and wikis where volunteers collect
and process information for a wider audience. There are also more struc-
tured sites that invite members to perform defined tasks like counting
Martian craters, ranking the most interesting news stories, and even writing
encyclopedia entries (Benkler 2002; Von Hippel, 2005). In many ways, these
projects resemble and extend an earlier tradition of large nuclear and par-
ticle physics databases that have relied on worldwide networks of volunteer
editors since the 1940s (Maurer, 2003). Like those earlier ‘‘big science’’
projects, the quality of at least some online resources seems to be high.
Recent peer review tests of the Wikipedia online encyclopedia suggest that
its accuracy is comparable to its most prestigious IP-supported counterpart,
Encyclopedia Britannica (Giles, 2005). In some ways, blogs and wikis may
actually be a more favorable environment than software for open source
since each article can be written and enjoyed as a freestanding ‘‘module.’’
The ability of authors to derive value is largely independent of how well
their article interacts with other modules or, indeed, whether other articles
are written at all. This is very different from open source, in which vol-
unteers cannot derive value unless they are eventually able to combine their
parts to create a single, unified product.
Drug discovery and GIS are much closer to the software model. Weber
(2000, 2004) and Burk (2002) suggest that deciphering and annotating
the genome might be organized as an open source project. More recent
observers have usually been skeptical. Lerner and Tirole (2004) suggest that
many biotechnology tasks cannot be broken up into modules and, in any
case, are expensive. Furthermore, there may not be enough ‘‘sophisticated
users who can customize the molecules to their own needs’’ (Lerner and
Tirole, 2004). Similarly, Kogut and Metiu (2001) assert ‘‘A molecule y is
not modular’’ because ‘‘changing atoms drastically alters its pharmaceutical
properties.’’44
Maurer (2006) argues that these objections assume that drug discovery is
an indivisible black box. In fact, there are approximately one dozen distinct
44
The existence of ‘‘me-too’’ drugs, in which drug companies change existing patented compounds just
enough to avoid patent infringement suggests that molecules may be more ‘‘modular’’ than Kogut and
Metiu suspect.
subtasks between basic science and the delivery of a completed drug. Pro-
prietary companies routinely break the black box apart by offering different
incentives for each specific substep. Examples include buying and selling
drug ideas (‘‘external innovation’’), offering prizes for recalcitrant chemical
engineering problems, and outsourcing preclinical and clinical trials to
outside entities. There is no obvious reason why open source collaborations
could not similarly perform one or more substeps. Scholars have suggested
open source collaborations for a wide variety of tasks including basic
science, using database (bioinformatics) tools to find the gene sequence
‘‘targets’’ that code for disease, computational design of drugs for specific
targets, in vitro chemistry and biology experiments designed to validate
proposed targets and drugs, and clinical trials (Benkler, 2002; Maurer et al.,
2004; Maurer, 2006; Von Hippel, personal communication). While con-
vincing examples of open source biology do not yet exist,45 it is reasonable
to think that incentives based on education, signaling, and ideology should
appeal to biologists just as much as they do for computer scientists.
The deeper question is whether open source biology can coexist with
the powerful incentives offered by conventional patents (Cohen, 2005). One
solution is to create open source biology collaborations in fields where
patent incentives are weak, for example, tropical disease research (Maurer
et al., 2004). More speculatively, much of the risk and expense associated
with clinical-phase trials involves documentation costs that pharmaceutical
companies incur to convince a skeptical FDA that their data is unbiased.
This problem is aggravated by outsourcing, which gives contract research-
ers obvious incentives to suppress and even falsify data to keep test pro-
grams alive. As noted by Titmuss (1972), non-market solutions avoid these
problems by relying on volunteers who have nothing to gain by lying. Since
most open source incentives similarly suppress agency problems, conven-
tional pharmaceutical companies might decide that funding open source
clinical trials was a more cost-effective way to convince FDA that their
drugs were effective than conventional contract research. In this case, pat-
ent incentives would reinforce open source biology instead of competing
with it.
Finally, GIS present a second industry where open source methods may
be acquiring a foothold. Like biology, the technology is highly computer-
ized and depends on users to notice and correct errors. Some government
consortia already style themselves ‘‘open source GIS’’ and invite users to
submit reports when mapping data turns out to be inaccurate (National
45
Commentators sometimes argue that bioinformatics software projects and/or science collaborations
that adopt open source-like licensing terms deserve the label ‘‘Open Source Biology’’ (Boettinger and
Burk, 2004 describing HapMap license; Rai, 2005). The former are indistinguishable from other types of
software, while the latter invariably turn out to be grant-supported collaborations that have adopted
open source-like licenses. Whether such collaborations should be called ‘‘open source’’ is, of course, a
matter of semantics. Suffice to say, they do not seem fundamentally different from traditional big science
projects dating back to the 1930s.
Research Council, 2004). In theory, proprietary firms could similarly

mencourage users to report errors. Open source GIS may, however, have an
inherent advantage in fostering social psychology incentives (e.g., altruism)
among users.
How much further can the open source model be extended? Our dis-
cussion suggests several factors. First, where contributors make comple-
mentary contributions for their own use, open source projects are more
likely to go forward when each contributor’s individual cost is small. This
suggests that projects with large, costly, and indivisible modules are
disfavored. Examples include projects where volunteers would have to build
physical prototypes or conduct physical experiments. Second, we hypoth-
esize that most workers in most industries find incentives like ideology,
education, and signaling much weaker than IP rewards. If contributors turn
to traditional IP instead of open source, open source collaborations may
never acquire the critical mass of workers they need to get started. Instead,
we expect open source projects to arise in markets where traditional IP
incentives are weak.46 Weaknesses of traditional IP incentives may occur
because users are poor (e.g., users of drugs for neglected diseases), because
extreme ease of copying undercuts the value of formal IP rights (e.g., soft-
ware, data), or because licensing is impractical due to market imperfections
like high transactions costs or the presence of ‘‘experience goods.’’ Finally,
we have seen that public domain sharing of innovation makes sense for
cumulative innovation environments in which ‘‘ideas are scarce,’’ so that
the owners and would-be improvers of IP cannot readily find each other.
In such systems, GPL-style open source stabilizes the public domain by
preventing improvers from taking ideas private. The phenomenon will
normally be more important for high risk, cutting-edge research (e.g., early-
stage drug discovery) that puts a high value on clever new ideas compared
to well-understood fields where successful development is more or less
assured once funds are invested (e.g., automobiles).
7 Conclusion
Open source comprises not one but an entire suite of incentives. In gen-
eral, each has separate and distinct welfare implications. Furthermore, the
importance of, say, ‘‘signaling’’ or ‘‘own-use’’ incentives vary significantly
across and even within projects. While generalizations are difficult, most
open source incentives reduce agency problems and deadweight loss com-
pared to patents, and accelerate discovery through automatic disclosure.
Against these virtues, open source incentives often lead to an undersupply
46
In principle, policymakers can encourage open source by reducing IP benefits. They should not,
however, abolish IP entirely. Open source incentives can certainly supplement intellectual property;
because they tend to undersupply goods, however, they should not replace it.
of goods relative to the patent system. Open source may also be less re-
sponsive to certain users, especially when those users are non-programmers.
Because of undersupply, open source can only be a partial solution: it is
not viable, and cannot operate in every environment where patent incentives
do. Where it works, however, it is often superior.
Acknowledgments
We thank the Toulouse Network on Information Technology for finan-

cial support, and Terry Hendershott for thoughtful comments.
References
Aghion, P., J. Tirole (1994). The management of innovation. Quarterly Journal of Economics 109, 1185.
Baldwin, C., K. Clark (2003). Does code architecture mitigate free riding in the open source devel-
opment model? Mimeo, Harvard University Business School.
Benkler, Y. (2002). Coase’s Penguin, or Linux and the nature of the firm. Yale Law Journal 112, 369.
Bergquist, M., J. Ljungberg (2001). The power of gifts: organizing social relationships in open source
communities. Information Systems Journal 11, 305.
Bessen, J. (2004). Open source software: free provision of complex public goods. Mimeo, Boston Uni-
versity Law School.
Bettelheim, B. (1976). The uses of enchantment: the meaning and importance of fairy tales. Knopf, New
York.
Bitzer, J., W. Schrettl, P. Schroder (2004). Intrinsic motivation in software development. Free University
of Berlin Discussion Paper 2004/19.
Boettinger, S., D. Burk (2004). Open source patenting. Journal of International Biotechnology Law 1,
221.
Bonnacorsi, A., C. Rossi (2003). Licensing schemes in the production and distribution of open source
software: an empirical investigation. Mimeo, Sant’ Ana School for Advanced Studies Institute for
Informatics and Telematics.
Bonnacorsi, A., C. Rossi. (2003). Contributing to the common pool of resources in open source soft-
ware: A comparison between individuals and firms. Mimeo. Sant’ Ana School of Advanced Studies
and Institute for Informatics and Telematics, Pisa.
Burk, D. (2002). Open source genomics. Boston University Journal of Science and Technology Law 8,
254.
Casadesus-Masanell, R., P. Ghemawat (2006). Dynamic mixed duopoly: a model motivated by Linux
vs. Windows. Management Science.
Cohen, W. (2005). Does open source have legs?, in: R. Hahn (ed.), Intellectual Property Rights in
Frontier Industries, AEI-Brookings Press, Washington, DC.
Comino, S., F. Manenti, M. Parisi (2005). From planning to mature: on the determinants of open source
take off. Mimeo, University of Trento, Department of Economics.
Cybersource (2004). Linux vs. Windows: total cost of ownership comparison (self-published).
Dahlander, L. (2004). Appropriating returns from open innovation processes: a multiple case study of
small firms in open source software. Mimeo, Chalmers University of Technology Deptartment of
Industrial Dynamics.
Dahlander, L., M.G. Magnusson (2005). Relationships between open source software companies and
communities: observations from Nordic firms. Research Policy 34, 481.
Dalle, J.-M., P.M. David (2003). The allocation of software development resources in ‘‘open source’’
production mode. SIEPR Discussion Paper No. 02-27.
Franck, E., C. Jungwirth (2002). Reconciling investors and donors—the investment structure of open
source. University of Zurich Lehrschule fur Unternehmensfuhrung und, politik.
Gambardella, A., B. Hall (2005). Proprietary vs. public domain licensing of software and research
products. NBER Working Paper 11120.
Gandal, N., C. Ferschtman (2005). Open source projects: output per contributor and restrictive licens-
ing. Mimeo, University of Tel Aviv.
Gaudeul, A. (2004). Open source software development patterns and license terms. Mimeo, University
of Toulouse.
Ghosh, R., R. Glott, B. Kriger, G. Robles (2002). Free/Libre and open source software: survey
and study. Mimeo, University of Maastricht Institute of Infonomics and Berlecon Research
GmbH.
Ghosh, R., Prakash, V. (2000). The Orbiten free software survey. First Monday Issue 5 (Number 7).
Giles, J. (2005). Internet encyclopedias go head to head. Nature 438, 900.
Gomulkiewicz, R.W. (1999). How copyleft uses license rights to succeed in the open source software
revolution and the implications for Article 2B. Houston Law Review 36, 179.
Hall, B. (2004). Incentives for knowledge production with many producers. Cambridge University
ESRC Centre for Business Research Working Paper 292.
Hann, I.-H., J. Roberts, S. Slaughter, R. Fielding (2004). An empirical analysis of economic returns to
open source participation. Mimeo, Carnegie Mellon University.
Harhoff, D., J. Henkel, E. von Hippel (2003). Profiting from voluntary information spillovers: how users
benefit by freely revealing their innovations. Research Policy 32, 1753.
Hawkins, R.E. (2002). The economics of open source for a competitive firm: Why give it away?
Netnomrics 6, 103.
Henkel, J. (2005a). The jukebox mode of innovation—a model of commercial open source development.
Mimeo, Technische Universitat, Munich.
Henkel, J. (2005b). Selective revealing in open innovation processes: the case of embedded Linux.
Mimeo, Technische Universitat, Munich.
Henkel, J., E. von Hippel (2005). Welfare implications of user innovation. Journal of Technology
Transfer 30, 73.
Henkel, J., M. Tins (2004). Munich/MIT survey: development of embedded Linux. Mimeo, University
of Munich Institute for Innovation Research, Technology Management, and Entrepreneurship.
Hertel, G., S. Niedner, S. Herrmann (2003). Motivation of software developers in open source projects:
an internet-based survey of contributions to the Linux Kernel. Research Policy 32, 1159.
Johnson, J.P. (2002). Open source software: private provision of a public good. Journal of Economics
and Management Strategy 24, 637.
Johnson, J.P. (2004). Collaboration, peer review and open source software. Mimeo, Cornell University
Johnson School of Management.
Kogut, B., Metiu, A. (2000). The emergence of e-innovation: insights from open source software de-
velopment. University of Pennsylvania Wharton School Reginald H. Jones Center Working Paper
WP 00-11.
Kogut, B., A. Metiu (2001). Open-source software development and distributed innovation. Oxford
Review of Economic Policy 17, 248.
Kollock, P. (1999). The economics of online cooperation: gifts and public goods in cyberspace, in: M.
Smith, P. Kollock (eds.), Communities in Cyberspace, Routledge, London.
Kuan, J. (2001). Open source software as consumer integration into production. University of California
Haas School of Business, Mimeo.
Lakhani, K., E. von Hippel (2000). How open source software works: ‘‘free’’ user-to-user assistance.
MIT Sloan School of Management Working Paper No. 4117.
Lakhani, K., R. Wolf (2005). Why hackers do what they do: understanding motivation and effort in
free/open source software projects, in: J. Feller, B. Fitzgerald, S. Hissam, K. Lakhani (eds.), Per-
spectives in Free and Open Source Software, MIT, Cambridge.
Lakhani, K., R. Wolf, J. Bates, and C. DiBona (2002). The Boston Consulting Group Hacker Survey
Release 0.73 (self-published).
Lemley, M., P. Menell, R. Merges, P. Samuelson (2002). Software and internet law 2nd edition. Aspen
Law and Business, New York.
Lerner, J., J. Tirole (2002a). Some simple economics of open source. Journal of Industrial Economics 52,
197.
Lerner, J., J. Tirole (2002b). The scope of open source licensing. Journal of Law, Economics, and
Organization 21, 20.
Lerner, J., Tirole, J. (2004). The economics of technology sharing: open source and beyond. NBER
Working Paper 10956.
Lewis, T.R., E. Talley (2006). Open source innovation. Duke University, Mimeo.
Madey, G., V. Freeh, R. Tynan (2005). Understanding open source as a self-organizing process. Mimeo,
University of Notre Dame Department of Computer Science and Engineering.
Maurer, S. (2003). New institutions for doing science: from databases to open source biology. Mimeo,
European Policy for Intellectual Property Conference.
Maurer, S. (2006). Choosing the right incentive strategy for R&D in neglected diseases. World Health
Organization Bulletin 84, 376.
Maurer, S., A. Rai, A. Sali (2004). Finding cures for tropical disease: is open source an answer?. Public
Library of Science: Medicine 1, 56.
Menell, P., S. Scotchmer (forthcoming). Intellectual property, in: M. Polinsky and S. Shavell (eds.),
Handbook of Law and Economics. Elsevier, Amsterdam.
Metagroup (2005). File, web, and database server administration (self-published).
Mockus, A., R.T. Fielding, J.D. Herbsleb (2002). Two case studies of open source software develop-
ment: Apache and Mozilla. ACM Transactions on Software Engineering and Methodology 11, 309.
Mustonen, M. (2003). Copyleft—the economics of Linux and other open source software. Information
Economics and Policy 15, 99.
Mustonen, M. (2005). When does a firm support substitute open source programming? Journal of
Economics and Management Strategy 14, 121.
National Research Council (2004). Licensing Geographic Data and Services. National Academy Press,
Washington, DC.
Netcraft (2005). Web server survey turns ten (self-published).
O’Mahoney, S. (2003). Guarding the commons: how community managed software projects protect
their work. Research Policy 32, 1179.
Osterloh, M. (2002). Open source software: new rules for the market economy. Mimeo, University of
Zurich Institute for Research in Business Administration.
Osterloh, M., S. Rota (2004). Trust and community in open source software production. Analyse und
Kritik: Zeitschift fur Sozialtheorie 26, 279.
Osterloh, M., S. Rota, B. Kuster (2003a). Trust and commerce in open source—a contradiction?, in: O.
Petrovic, M. Fallenbock, Ch. Kittle, M. Ksela (eds.), Trust in the Network Economy, Springer,
Vienna, p. 129.
Osterloh, M., S. Rota, B. Kuster (2003b). Open source software production: climbing on the shoulders
of giants. Mimeo, University of Zurich Institute for Research in Business Administration.
Rai, A. (2005). Open and collaborative research: a new model for biomedicine, in: R. Hahn (ed.),
Intellectual Property Rights in Frontier Industries, AEI-Brookings Press, Washington, DC.
Raymond, E.S. (1999). The Cathedral and the Bazaar: Musings on Linux and Open Source by an Ac-
cidental Revolutionary. O’Reilly, Sebastopol, CA.
Robert Frances Group (2002). Total cost of ownership for Linux in the enterprise (self-published).
Roberts, J., I.-H. Hann, S. Slaughter (2006). Understanding the motivations, participation and per-
formance of open source software developers: a longitudinal study of the Apache projects. Man-
agement Science.
Rossi, M. (2004). Decoding the free/open source (’F/OSS’) puzzle, a survey of theoretical and empirical
contributions. Mimeo. Department of Political Economy, Siena.
Rossi, M.A. (2004). Decoding the ‘free/open source (F/open source) software puzzle’: a survey of
theoretical and empirical contributions. University of Sienna Department of Political Economy
Working Paper No. 424.
Rossi, C., A. Bonaccorsi (2005). Intrinsic vs. extrinsic incentives in profit-oriented firms supplying open
source products and services. First Monday Issue 10(No. 5).
Samuelson, P. (1984). CONTU revisited: the case against copyright protection in machine-readable
form. Duke Law Review 1984, 663–769.
Schmidt, K., M. Schnitzer (2002). Public subsidies for open source? Some economic policy issues of the
software market. University of Munich, Mimeo.
Scotchmer, S. (2004). Ideas and Incentives. MIT, Cambridge.
The Linux Counter (2005). Available at http://www.counter.li.org/estimates.php (self-published).
Titmuss, R.M. (1972). The Gift Relationship: From Human Blood to Social Policy. Vintage, New York.
Varian, H.R., C. Shapiro (2003). Linux adoption in the public sector: an economic analysis. Mimeo,
University of California Haas School of Business.
von Hippel, E. (2002). Open source projects as horizontal innovation networks—by and for users. MIT
Sloan School of Management Working Paper No. 4366-02.
von Hippel, E. (2004). Democratizing innovation. MIT, London.
von Krogh, G., S. Spaeth, K. Lakhani (2003). Community, joining, and specialization in open source
software innovation: a case study. Research Policy 32, 1217.
Weber, S. (2000). The political economy of open source software. Berkeley Roundtable on the Inter-
national Economy Working Paper 140.
Weber, S. (2004). The Success of Open Source. Harvard University Press, Cambridge.
West, J. (2003). How open is open enough? Melding proprietary and open source platform strategies.
Research Policy 32, 1259.
West, J., S. Gallagher (2004). Key challenges of open innovation: lessons from open source software.
Mimeo, San Jose State College of Business.
West, J., S. O’Mahoney (2005). Contrasting community building in sponsored and community founded
open source projects, in: R. H. Sprague (ed.), IEEE Proceedings of the 38th Annual Hawaii Inter-
national Conference on System Sciences. IEEE Computer Society Press, Los Alamitos, CA.
Zeitlyn, D. (2003). Gift economies in the development of open source software: anthropological re-
flections. Research Policy 32, 1287.
Chapter 6
Information, Search, and Price Dispersion
Michael R. Baye
Kelley School of Business, Indiana University, Bloomington, IN, USA
John Morgan
Haas School of Business and Department of Economics, University of California, Berkeley, CA,
USA
Patrick Scholten
Department of Economics, Bentley College, Waltham, MA, USA
Abstract
We provide a unified treatment of alternative models of information acqui-

sition/transmission that have been advanced to rationalize price dispersion in
online and offline markets for homogeneous products. These different frame-
works—which include sequential search, fixed sample search, and clearing-
house models—reveal that reductions in (or the elimination of) consumer
search costs need not reduce (or eliminate) price dispersion. Our treatment
highlights a ‘‘duality’’ between search-theoretic and clearinghouse models of
dispersion and shows how auction-theoretic tools may be used to simplify
(and even generalize) existing theoretical results. We conclude with an over-
view of the burgeoning empirical literature. The empirical evidence suggests
that price dispersion in both online and offline markets is sizeable, pervasive,
and persistent—and does not purely stem from subtle differences in firms’
products or services.
1 Introduction
Simple textbook models of competitive markets for homogeneous prod-

ucts suggest that all-out competition among firms will lead to the so-called
323
324 M.R. Baye et al.
‘‘law of one price.’’ Yet, empirical studies spanning more than four decades
(see Table 1a and b) reveal that price dispersion is the rule rather than the
exception in many homogeneous product markets. The observation that the
prices different firms charge for the same product often differ by 30 percent
or more led Hal Varian to suggest that ‘‘the ‘law of one price’ is no law at
all’’ (Varian, 1980, p. 651). This chapter provides a unified treatment of
several theoretical models that have been developed to explain the price
dispersion observed in homogeneous product markets and surveys the bur-
geoning empirical literature (including the studies summarized in Ta-
ble 1a and b), which documents ubiquitous price dispersion. A key
motivation for this chapter is to dispel the erroneous view that the Inter-
net—through its facilitation of dramatic declines in consumer search
costs—will ultimately lead to the ‘‘law of one price.’’
When confronted with evidence of price dispersion, many are quick to
point out that even in markets for seemingly homogeneous products, subtle
differences among the ‘‘services’’ offered by competing firms might lead
them to charge different prices for the same product. Nobel Laureate
George Stigler’s initial response to wags making this point was philosoph-
ical: ‘‘y [While] a portion of the observed dispersion is presumably at-
tributable to such difference[s] y it would be metaphysical, and fruitless, to
assert that all dispersion is due to heterogeneity’’ (Stigler, 1961, p. 215).
Thirty-five years later, the literature has amassed considerable support for
Stigler’s position. As we shall see in Sections 2 and 3, there is strong the-
oretical and empirical evidence that much (and in some markets, most) of
the observed dispersion stems from information costs—consumers’ costs of
acquiring information about firms and/or firms’ costs of transmitting in-
formation to consumers.
As Fig. 1 reveals, research on information, search, and price dispersion
has become increasingly important since the publication of Stigler’s seminal
article on the Economics of Information. Until about 1998, most studies
focused on environments where consumers incur a positive cost of obtain-
ing each additional price quote. Search costs in these studies consist of
consumers’ opportunity cost of time in searching for lower prices (the so-
called ‘‘shoe-leather’’ costs), plus other costs associated with obtaining price
quotes from competing firms (such as the incremental cost of the postage
stamps or phone calls used in acquiring price information from firms).
Consumers in these environments weigh the cost of obtaining an additional
price quote against the expected benefits of searching an additional firm. As
we discuss in Section 2.1, equilibrium price dispersion can arise in these
environments under a variety of market conditions and search strategies
(including sequential and fixed sample search).
While marginal search costs are useful in explaining price dispersion in
some markets, in many online markets incremental search costs are very
low—and in some cases, zero. For example, price comparison sites and
shopbot technologies create environments where consumers may obtain a
Ch. 6. Information, Search, and Price Dispersion 325
Table 1
Measures of price dispersion reported in the literaturea
Study Data Product market Intervals of Dispersion

period estimated measure
price
dispersion
measures
(a) Offline markets

Bailey (1998) 1997 Books 13.2% Standard deviation
Books 10.4% Standard deviation
Compact discs 17.6% Standard deviation
Compact discs 11.0% Standard deviation
Software 7.1% Standard deviation
Software 8.1% Standard deviation
Borenstein and Rose 1986 U.S. Airline 0.018–0.416 Gini coefficient
(1994)
Carlson and 1976 Consumer sundries 3.3–41.4% Coefficient of
Pescatrice (1980) variation
Eckard (2004) 1901–2001 Baking powder, 3.1–10.1% Coefficient of
sugar, salt—1901 variation
Baking powder, 0.0–13.4% Coefficient of
sugar, salt—2001 variation
Friberg et al. (2001) 1999 Books $54.00– Range
$122.00
Books $21.94– Standard deviation
$76.20
Compact discs $20.00– Range
$40.00
Compact discs $12.91– Standard deviation
$23.86
Books (Sweden) $19.00– Range
$58.00
Compact discs $21.00– Range
(Sweden) $46.00
Lach (2002) 1993–1996 Refrigerator 4.9% Coefficient of
(Israel) variation
Chicken, flour, 11.4–19.7% Coefficient of
coffee (Israel) variation
Marvel (1976) 1964–1971 Regular gasoline $0.048 Range
Regular gasoline $0.015 Standard deviation
Premium gasoline $0.048 Range
Premium gasoline $0.017 Standard deviation
Pratt et al. (1979) 1975 Various products 4.4–71.4% Coefficient of
and services variation
Various products 11.0– Range
and services 567.0%
Various products 7.2–200.0% Value of
and services information
price
dispersion
measures
Roberts and Supina 1963–1987 Wood products 13.8–90.2% Coefficient of

(2000) variation
Fabrics 18.8–78.1% Coefficient of
variation
Coffee 14.3–25.1% Coefficient of
variation
Ready-mixed 13.2–37.2% Coefficient of
concrete variation
Newsprint 4.5–8.2% Coefficient of
variation
Gasoline 6.2–11.8% Coefficient of
variation
Tinplate steel cans 25.0–31.0% Coefficient of
variation
Pan bread 26.0–49.6% Coefficient of
variation
Corrugated 21.8–39.6% Coefficient of
shipping variation
containers
Scholten and Smith 1976–2000 Consumer 3.3–41.4% Coefficient of
(2002) sundries—1976 variation
Consumer 1.6–42.0% Coefficient of
sundries—2000 variation
sundries—2000 variation
Sorensen (2000) 1998 Prescription drugs $13.17 Range
Prescription drugs 22.0% Coefficient of
variation
Stigler (1961) 1953 Anthracite coal $3.46 Range
Anthracite coal $1.15 Standard deviation
1959 Identical $165.00 Range
automobiles
Identical $42.00 Standard deviation
automobiles
Villas-Boas (1995) 1985–1987 Coffee 21.5% Coefficient of
variation
price
dispersion
measures
(b) Online markets

only
Ancarani and 2002 Books (Italy) h4.26–h4.84 Standard deviation
Shankar (2004)
Books (Italy) h20.00– Range
h22.88
Compact discs h2.29–h2.79 Standard deviation
(Italy)
Compact discs h11.82– Range
(Italy) h14.75
Arbatskaya and 1998 Mortgage interest >0.25 Range
Baye (2004) rates
Arnold and Saliba 2001 Textbooks 10.7–52.6% Range
(2002)
Textbooks 3.5–10.0% Coefficient of
variation
Textbooks 0.2–12.5% Price gap
Baye et al. (2003) 2000–2001 Consumer $123.88– Range
electronics $143.15
Baye et al. (2004a) 2000–2001 Consumer 9.1–9.7% Coefficient of
electronics variation
Consumer 3.79–5.38% Gap
electronics
Baye et al. (2004b) 1999–2001 Consumer 57.4% Range
electronics
Consumer 12.5% Coefficient of
Baylis and Perloff 1999 Cameras $342.00 Range
(2002)
Scanners $106.00 Range
Brynjolfsson and 1998–1999 Books 33.0% Range
Smith (2000)
Compact discs 25.0% Range
Chevalier and 2001 Books 8.1–2.3% Range
Goolsbee (2003)
Clay et al. (2001) 1999–2000 Books 27.7% Coefficient of
variation
Books $7.62 Range
Clay et al. (2003) 1999 Books 10.0–18.0% Coefficient of
variation
Clemons et al. (2002) Travel $8.03–13.40 Range
price
dispersion
measures
Ellison and Ellison 2000–2001 Memory modules 5.9–29.0% Range

(2004)
Gatti and Kattuman 2002 Consumer 3.0–15.3% Coefficient of
(2003) electronics variation
(France)
(Italy)
(Netherlands)
(Spain)
(Sweden)
electronics (UK) variation
(Denmark)
Consumer 7.8–47.4% Range
electronics
(France)
electronics
(Italy)
electronics
(Netherlands)
electronics
(Spain)
electronics
(Sweden)
electronics (UK)
12.8–42.9% Range
price
dispersion
measures
Consumer
electronics
(Denmark)
electronics
(France)
electronics
(Italy)
electronics
(Netherlands)
electronics
(Spain)
electronics
(Sweden)
electronics (UK)
electronics
(Denmark)
Hong and Shum 2002 Books $8.19–27.05 Range
(2006)
Books 6.2–8.5% Coefficient of
variation
Janssen et al. (2005) 2004 Keyboards $6.50– Range
$91.67
Keyboards 8.0–52.0% Coefficient of
variation
Pan et al. (2002) 2000 Books 15.0% Coefficient of
variation
Compact discs 15.4% Coefficient of
variation
DVDs 12.7% Coefficient of
variation
PDAs 11.8% Coefficient of
variation
Software 11.7% Coefficient of
variation
Consumer 9.6% Coefficient of
price
dispersion
measures
Pan et al. (2003) 2000–2003 Consumer 9.8–11.7% Coefficient of

electronics and variation
Books
Books 33.3–48.9% Range
Compact discs 22.2–51.0% Range
DVDs 30.7–43.7% Range
Computers 15.0–34.4% Range
Software 19.0–35.6% Range
electronics
Smith and 1999 Books 28.0–33.0% Value of
Brynjolfson (2001) information
Books $6.29– Standard deviation
$10.51
a
Includes studies comparing offline and online price dispersion.
list of the prices that different sellers charge for the same product. Despite
the fact that this information is available to consumers in seconds, ulti-
mately at the cost of a single ‘‘mouse click,’’ the overwhelming empirical
finding is that even in these environments, price dispersion is pervasive and
significant—the law of one price is egregiously violated online. In Section
2.2, we examine an alternative line of theoretical research where marginal
search costs are not the key driver for price dispersion. Our theoretical
analysis concludes in Section 2.3 with a discussion of alternative behavioral
rationales for price dispersion (including bounded rationality on the part of
firms and/or consumers).
Section 3 provides a more detailed overview of the growing empirical
literature. As one might suspect based on the trend in Fig. 1 and the
research summarized in Table 1a and b, most empirical studies of price dis-
persion postdate the Internet and rely on online data. Our view is that this is
more an artifact of the relative ease with which data may be collected
in online markets—not an indication that price dispersion is more important
(or more prevalent) in online than offline markets. For this reason, we have
attempted to provide a balanced treatment of the literatures on online and
offline price dispersion. As we shall argue, the overwhelming conclusion of
both literatures is that price dispersion is not purely an artifact of product
heterogeneities.
Fig. 1. Percentage of articles published in the American Economic Review, Journal of

Political Economy, and Econometrica on information, search or price dispersion.
2 Theoretical models of price dispersion
This section presents alternative models that have been used to rationalize
the price dispersion observed in both offline and online markets. One
approach is to assume that it is costly for consumers to gather information
about prices. In these ‘‘search-theoretic’’ models, consumers searching
for the best price incur a positive cost of obtaining each additional price
quote. Representative examples include Stigler (1961), Rothschild (1973),
Reinganum (1979), MacMinn (1980), Braverman (1980), Burdett and Judd
(1983), Carlson and McAfee (1983), Rob (1985), Stahl (1989, 1996), Dana
(1994), McAfee (1995), Janssen and Moraga-González (2004), as well as
Janssen et al. (2005).
A second approach de-emphasizes the marginal search cost as a source
for price dispersion. Instead, consumers access price information by con-
sulting an ‘‘information clearinghouse’’ (e.g., a newspaper or an Internet
price comparison site); e.g., Salop and Stiglitz (1977), Shilony (1977),
Rosenthal (1980), Varian (1980), Narasimhan (1988), Spulber (1995), Baye
and Morgan (2001), and Baye et al. (2004a).1 The distinguishing feature of
‘‘clearinghouse models’’ is that a subset of consumers gain access to a list of
prices charged by all firms and purchase at the lowest listed price. In the
1
A third approach de-emphasizes consumer search and mainly focuses on whether price dispersion
can arise when consumers ‘‘passively’’ obtain price information directly from firms (as in direct mail
advertisements); cf. Butters (1977), Grossman and Shapiro (1984), Stegeman (1991), Robert and Stahl
(1993), McAfee (1994), and Stahl (1994). A related marketing literature examines similar issues, ranging
from loyalty and price promotion strategies to channel conflicts and the Internet; see Lal and Villas-
Boas (1998), Lal and Sarvary (1999), Raju et al. (1990), and Rao et al. (1995).
earliest of these models, equilibrium price dispersion stems from ex ante

heterogeneities in consumers or firms. For example, in the Varian and
Salop-Stiglitz models, some consumers choose to access the clearinghouse
to obtain price information, while others do not. In Shilony et al., some
consumers are loyal to a particular firm (and thus will buy from it even if it
does not charge the lowest price), while other consumers are ‘‘shoppers’’
and only purchase from the firm charging the lowest price. Spulber (1995)
shows that equilibrium price dispersion arises even when all consumers can
costlessly access the clearinghouse—provided each firm is privately in-
formed about its marginal cost. Baye and Morgan (2001) offer a clearing-
house model that endogenizes not only the decisions of firms and
consumers to utilize the information clearinghouse (in the previous clear-
inghouse models, firms’ listing decisions are exogenous), but also the fees
charged by the owner of the clearinghouse (the ‘‘information gatekeeper’’ to
consumers and firms who wish to access or transmit price information.
They show that a dispersed price equilibrium exists even in the absence of
any ex ante heterogeneities in consumers or firms.
In this section, we provide an overview of the key features and ideas
underlying these literatures.
2.1 Search-theoretic models of price dispersion
We begin with an overview of search-theoretic approaches to equilibrium

price dispersion. The early literature stresses the idea that, when consumers
search for price information and search is costly, firms will charge different
prices in the market. There are two basic sorts of models used: models with
fixed sample size search and models where search is sequential. We will
discuss each of these in turn.
The search models considered in this subsection are all based on the
following general environment. A continuum of price-setting firms (with
unit measure) compete in a market selling an identical (homogeneous)
product. Firms have unlimited capacity to supply this product at a constant
marginal cost, m. A continuum of consumers is interested in purchasing the
product. Let the mass of consumers in the market be m, so that the number
of customers per firm is m. Each consumer has a quasi-linear utility func-
tion, u (q)+y, where q is the quantity of the homogeneous product and y is
the quantity of some numeraire good whose price is normalized to be unity.
This implies that the indirect utility of a consumer who pays a price p per
unit of the product and who has an income of M is
Vðp; MÞ ¼ uðpÞ þ M
where u ( ) is nonincreasing in p. By Roy’s identity, note that the demand
for the product of relevance is q (p) –u0 (p).
To acquire the product, a consumer must first obtain a price quote from a
store offering the product for sale. Suppose that there is a search cost, c, per
price quote.2 If, after obtaining n price quotes, a consumer purchases q (p)
units of the product from one of the firms at price p per unit, the consumer’s
(indirect) utility is
V ¼ uðpÞ þ M cn
The analysis that follows focuses on posted price markets where con-
sumers know the distribution of prices but do not know the prices charged
by particular stores.3
2.1.1 The Stigler model
Stigler (1961) considers the special case of this environment where:
1. Each consumer wishes to purchase KZ1 units of the product; that is,
q(p) ¼ –u0 (p) ¼ K;
2. The consumer’s search process is fixed sample search—prior to
searching, consumers determine a fixed sample size, n, of firms from
whom to obtain price quotes and then buy from the firm offering the
lowest price; and
3. The distribution of firms’ prices is given by an exogenous nondegen-
erate cdf F(p) on ½p; p:
Stigler assumes that a consumer chooses a fixed sample size, n, to min-

imize the expected total cost (expected purchase cost plus search cost) of
purchasing K units of the product:
E½C ¼ KE½pðnÞ
min þ cn
where E½pðnÞ
min ¼ E½minfp1 ; p2 ; . . . ; pn g; that is, the expected lowest price
quote obtained from n draws from F. Since the distribution of the lowest of
n draws is F ðnÞ
min ðpÞ ¼ 1 ½1 F ðpÞ ;
n
Z p
E½C ¼ K pdF ðnÞ
min ðpÞ þ cn
p
" Z #
p
n
¼ K pþ ½1 F ðpÞ dp þ cn
p
2
In what follows, we assume that consumers have identical search costs. Axell (1977) offers a model of
price dispersion with heterogeneous search costs.
3
This assumption is relaxed in Rothschild (1974), Benabou and Gertner (1993), and Dana (1994),
where buyers learn about the distribution of prices as they search, and in Rauh (1997), where buyers’
search strategies depend on only finitely many moments of the distribution of prices. Daughety (1992)
offers an alternative search-theoretic model of equilibrium price dispersion that results from informa-
tional asymmetries and a lack of price precommitment on the part of firms.
where the second equality obtains from integration by parts. Notice that the
term in square brackets reflects the expected purchase price, which is a
decreasing function of the sample size, n. However, since each additional
price observation costs c>0 to obtain, an optimizing consumer will choose
to search a finite number of times, n ; and thus will generally stop short of
obtaining the best price ðpÞ in the market.
The distribution of transaction prices is the distribution of the lowest of
n draws from F, that is,

ðn Þ
F min ðpÞ ¼ 1 ð1 FðpÞÞn .
From this, Stigler concludes that dispersion in both posted prices and
transactions prices arises as a consequence of costly search.
How do transactions prices and search intensity relate to the quantity
of the item being purchased (or equivalently, to the frequency of purc-
hases)?4 Stigler’s model offers sharp predictions in this dimension. Note
that the expected benefit to a consumer who increases her sample size from
n –1 to n is

E½BðnÞ ¼ E½pðn1Þ
min E½p ðnÞ
min K, (1)
which is decreasing in n. Furthermore, the expected benefit from search are

greater for products bought in greater quantities or more frequently, that is,
equation (1) is increasing in K. Since the cost of the nth search is independent
of K while the expected benefit is increasing in K, it immediately follows that
the equilibrium search intensity, n ; is increasing in K. That is, consumers
obtain more price quotes for products they buy in greater quantities (or more
frequently).
Despite the fact that the Stigler model assumes each individual inelas-
tically purchases K units of the product, a version of the ‘‘law of demand’’
holds: each firm’s expected demand is a nonincreasing function of its price.
To see this, note that a firm charging price p is visited by mn consumers and
nn 1
offers the lowest price with probability ð1 F ðpÞÞ Thus, a representative
firm’s expected demand when it charges a price of p is
n
QðpÞ ¼ mnn Kð1 FðpÞÞn 1
(2)
which is decreasing in p.
4
K may be related to purchase frequency as follows. Suppose prices are ‘‘valid’’ for T periods, and the
consumer wishes to buy one unit every trT periods, that is, t represents a consumer’s purchase fre-
quency. Then the total number of units purchased during the T periods is KT/t. Thus, an increase in
purchase frequency (t) is formally equivalent to an increase in K in the model above.
The Stigler model implies that both the expected transactions price
(Proposition 1) as well as the expected total costs inclusive of search costs
(Proposition 2) are lower when prices are more dispersed (in the sense of a
mean preserving spread).5
Proposition 1. Suppose that a price distribution G is a mean preserving

spread of a price distribution F. Then the expected transactions price of a
consumer who obtains n>1 price quotes is strictly lower under price dis-
tribution G than under F.
Proof. Let D ¼ E F ½pðnÞ ðnÞ

min E G ½pmin be the difference in the expected
transactions price under F compared to G. We will show that for all n>1,
D>0. Using the definition of E½pðnÞ min ;
Z 1 Z 1
n1
D¼ pnð1 F ðpÞÞ dF ðpÞ tnð1 GðtÞÞn1 dGðtÞ
1 1
Let u ¼ F(p) and u ¼ G(p), so that du ¼ dF(p), du ¼ dG(p), p ¼ F –1(u),

and t ¼ G –1(u).
Then
Z 1 Z 1
n1
D¼n F 1
ðuÞð1 uÞ du n G 1 ðuÞð1 uÞn1 du
0 0
Z 1
¼n ðF 1 ðuÞ G 1 ðuÞÞð1 uÞn1 du
0
Since G is a mean preserving spread of F, there exists a unique interior

point u ¼ F(EF[P]) such that F 1 ðun Þ ¼ G 1 ðun Þ: Further, for all
uoun ; F 1 ðuÞ G 1 ðuÞ40 and for all u4un ; F 1 ðuÞ G1 ðuÞo0: Thus
Z u
D¼n ðF 1 ðuÞ G1 ðuÞÞð1 uÞn1 du
0
Z 1
1 1 n1
þ ðF ðuÞ G ðuÞÞð1 uÞ du .
u
5
R1 Rz
G is a mean preserving spread of F if (a) 1 ½GðpÞ F ðpÞdp ¼ 0 and ðbÞ 1 ½GðpÞ F ðpÞdp 0;
with strict inequality for some z. Note that (a) is equivalent to the fact that the means of F and G are
equal. Together, the two conditions imply that F and G cross exactly once (at the mean) on the interior
of the support.
Next, notice that (1 – u)n–1 is strictly decreasing in u; hence,

Z un
D4n ðF 1 ðuÞ G 1 ðuÞÞð1 u Þn1 du
0
Z 1
1 1 n1
þ ðF ðuÞ G ðuÞÞð1 u Þ du
u
Z 1
n1
¼ nð1 u Þ ðF 1 ðuÞ G 1 ðuÞÞdu
0
¼0
where the last equality follows from the fact that F and G have the same
mean. ’
Proposition 2. Suppose that an optimizing consumer obtains more than one
price quote when prices are distributed according to F, and that price dis-
tribution G is a mean preserving spread of F. Then the consumer’s expected
total costs under G are strictly less than those under F.
Proof. Suppose that, under F, the optimal number of searches is n : Then
the consumer’s expected total cost under F is

E½C F ¼ E F ½pðnÞ
min K cn

4E G ½pðnÞ
min K cn

E½C G ,
where the strict inequality follows from Proposition 1, and the weak
inequality follows from the fact that n searches may not be optimal
under the distribution G. ’
At first blush, it might seem surprising that consumers engaged in fixed
sample search pay lower average prices and have lower expected total costs
in environments where prices are more dispersed. The intuition, however, is
clear: In environments where prices are more dispersed, the prospects for
price improvement from search are higher because the left tail of the price
distribution—the part of the distribution where ‘‘bargains’’ are to be
found—becomes thicker as prices become more dispersed.
2.1.2 The Rothschild critique and Diamond’s paradox

While Stigler offered the first search-theoretic rationale for price dis-
persion, the model has been criticized for two reasons. First, as pointed
out in Rothschild (1973), the search procedure assumed in Stigler’s model
may not be optimal. In fixed sample search, consumers commit to a fixed
number, n, of stores to search and then buy at the lowest price at the
conclusion of that search. A clear drawback to such a strategy is that it

fails to incorporate new information obtained during search, such as an
exceptionally low price from an early search. Indeed, once the best price
quote obtained is sufficiently low, the benefit in the form of price im-
provement drops below the marginal cost of the additional search. As we
will see below, sequential search results in an optimal stopping rule such
that a consumer searches until she locates a price below some threshold,
called the reservation price. Second, the distribution of prices, F, is ex-
ogenously specified and is not based on optimizing firm behavior. In fact,
in light of equation (2), a representative firm with constant marginal cost
of m enjoys expected profits of
pðpÞ ¼ ðp mÞQðpÞ.
That is, absent any cost heterogeneities, each firm faces exactly the same
expected profit function. Why then, would firms not choose the same
profit-maximizing price or, more generally, how could the distribution of
prices generated by profit-maximizing firms be consistent with the price
distribution over which consumers were searching? In short, Rothschild
pointed out that it is far from clear that information costs give rise to an
equilibrium of price dispersion with optimizing consumers and firms; in
Stigler’s model, only one side of one market, the consumers, are acting in
an optimizing fashion consistent with equilibrium. For this reason, Rot-
hschild criticized the early literature for its ‘‘partial–partial equilibrium’’
approach.
Diamond (1971) advanced this argument even further—he essentially
identified conditions under costly search where the unique equilibrium in
undominated strategies involves all firms charging the same price—the
monopoly price. Diamond’s result may be readily seen in the following
special case of our environment where:
1. Consumers have identical downward sloping demand, i.e., u00 ðpÞ ¼

q0 ðpÞo0;
2. Consumers engage in optimal sequential search;
3. A firm acting as a monopoly would optimally charge all consumers the
unique monopoly price, p ; and
4. A consumer who is charged the monopoly price earns surplus sufficient
to cover the cost of obtaining a single price quote, that is, uðp Þ4c:
In this environment, all firms post the monopoly price and consumers
visit only one store, purchase at the posted price, p ; and obtain surplus
uðp Þ c40: Given the stopping rule of consumers, each firm’s best re-
sponse is to charge the monopoly price; given that all firms charge p ; it is
optimal for each consumer to search only once. To see that this is the
unique equilibrium in undominated strategies, suppose to the contrary

that there is an equilibrium in which some firm posted a price below the
monopoly price (clearly, pricing above the monopoly price is a dominated
strategy). Let p0 be the lowest such posted price. A firm posting the lowest
price could profitably deviate by raising its price to the lower of p or
p0 +c. Any consumer visiting that firm would still rationally buy from it
since the marginal benefit of an additional search is smaller than c—the
marginal cost of an additional search. Thus, such a firm will not lose any
customers by this strategy and will raise its earnings on each of these
customers.
The Diamond paradox is striking: even though there is a continuum of
identical firms competing in the model—a textbook condition for perfect
competition—in the presence of any search frictions whatsoever the mo-
nopoly price is the equilibrium. Rothschild’s criticism of the Stigler model,
along with the Diamond paradox, spawned several decades of research into
whether costly search could possibly generate equilibrium price disper-
sion—a situation where consumers are optimally gathering information
given a distribution of prices, and where the distribution of prices over
which consumers are searching is generated by optimal (profit-maximizing)
decisions of firms.
2.1.3 The Reinganum model and optimal sequential search

Reinganum (1979) was among the first to show that equilibrium price
dispersion can arise in a sequential search setting with optimizing consum-
ers and firms. Reinganum’s result may be seen in the following special case
of our environment where:
1. Consumers have identical demands given by u0 (p) ¼ q(p) ¼ Kpe,

where e o1 and K>0;
2. Consumers engage in optimal sequential search;
3. Firms have heterogeneous marginal costs described by the atomless
distribution G(m) on ½m; m;
4. A consumer who is charged the monopoly price by a firm with the
highest marginal cost, m; earns surplus to cover the cost of
sufficient

obtaining a single price quote, that is, u 1þ m 4c:
Reinganum shows that, under these assumptions, there exists a dispersed

price equilibrium in which firms optimally set prices and each consumer
engages in optimal sequential search. To establish this, we first show how
one derives the optimal reservation price in a sequential search setting.
Suppose consumers are confronted with a nondegenerate distribution of
prices F(p) on ½p; p; that is atomless, except possibly at p: Consumers engage
in optimal sequential search with free recall. If, following the nth search, a
consumer has already found a best price z min(p1, p2,y,pn), then, by
making an additional search, such a consumer expects to gain benefits of

Z z
BðzÞ ¼ ðuðpÞ uðzÞÞdFðpÞ
p
Z z
¼ u0 ðpÞFðpÞdp;
p
where the second equality obtains through integration by parts. Using

Leibnitz’ rule, we have
B0 ðzÞ ¼ u0 ðzÞF ðzÞ
¼ Kz F ðzÞ40. ð3Þ
Thus, the expected benefits from an additional search are lower when the
consumer has already identified a relatively low price. Since search is costly
(c>0), consumers must weigh the expected benefits against the cost of an
additional search. The expected net benefits of an additional search are
hðzÞ BðzÞ c.
If the expected benefits from an additional search exceed the additional cost,
h (z)>0, it is optimal for the consumer to obtain an additional price quote.
If h (z)o0, the consumer is better off purchasing at the price z than ob-
taining an additional price quote.
A consumer’s optimal sequential search strategy may be summarized as
follows: Rp
Case 1. hðpÞo0 and p uðpÞdF ðpÞoc: Then the consumer’s optimal strat-
egy is to not search.
Rp
Case 2. hðpÞo0 and p uðpÞdF ðpÞdp c: Then the consumer’s optimal
strategy is to search until she obtains a price quote at or below the res-
ervation price, r ¼ p:
Case 3. hðpÞ 0: Then the consumer’s optimal strategy is to search until
she obtains a price quote at or below the reservation price, r, where r solves
Z r
hðrÞ ¼ ðuðpÞ uðrÞÞdF ðpÞ c ¼ 0. (4)
p
Equation (4) represents a price at which a consumer is exactly indiffe-

rent between buying and making an additional search. To see that such a
price is uniquely defined by this equation, notice that hðpÞ ¼ co0; hðpÞ
0; and h0 (z) ¼ B0 (z)>0. A consumer who observes a price that exceeds r will
optimally ‘‘reject’’ that price in favor of continued search, while a consumer
who observes a price below r will optimally ‘‘accept’’ that price and stop
searching.
Case 1 is clearly not economically interesting as it leads to the absence
of any market for the product in the first place. Case 2 arises when the
expected utility of purchasing the product exceeds the cost of an initial
search, but the distribution of prices is sufficiently ‘‘tight’’ relative to search

costs to make additional searches suboptimal. Most of the existing search
literature, including Reinganum, restricts attention to Case 3, as we shall do
hereafter.
The reservation price defined in equation (4) has several interesting com-
parative static properties. Totally differentiating equation (4) with respect
to r and c, and using equation (3) reveals that
dr 1 1
¼ ¼ 40.
dc qðrÞF ðrÞ Kr F ðrÞ
Thus, an increase in search costs leads to a higher reservation price: other
things equal, the range of ‘‘acceptable’’ prices is greater for products with
higher search costs. Note that, for the special case when q(r) ¼ 1, dr/dc
¼ 1/F(r)>1. In this case, a one unit increase in search costs increases the
range of acceptable prices by more than one unit—that is, there is a ‘‘mag-
nification effect’’ of increases in search costs.6
Reinganum avoids Rothschild’s criticism and the ‘‘Diamond paradox’’ by
introducing firm cost heterogeneities. Since each firm j differs in its marginal
cost, mj, prices will differ across firms even when they price as monopolists.
Suppose that a fraction 0rlo1 of firms price above r, and recall that
there are m consumers per firm. A representative firm’s expected profit when
it prices at pj is
( m
ðpj mj Þqðpj Þ 1l if pj r;
Epj ¼
0 if pj 4r:
Ignoring for a moment the fact that a firm’s demand is zero if it prices
above r, note that profit-maximization implies the first-order condition
h i m
ðpj mj Þq0 ðpj Þ þ qðpj Þ ¼ 0.
1l
Standard manipulation of the first-order condition for profit-maximization
implies that firm j ’s (unconstrained) profit-maximizing price is a constant
markup over its cost:

pj ¼ mj .
1þ
Suppose that firms simply ignore the consumer’s reservation price, r, and
price at this markup. This would imply that consumers face a distribution
^
of posted prices FðpÞ ¼ Gðpð1 þ Þ=Þ on the interval ½m =ð1 þ Þ; m=ð1 þ
Þ: Given this distribution of prices, optimizing consumers would set a
6
In general, there may be either a magnification or an attenuation effect of a one unit increase in the
cost of search.
reservation price, r, such that

Z r
hðrÞ ¼ ^ c ¼ 0:
ðuðpÞ uðrÞÞd FðpÞ
p
Furthermore, if rom=ð1 þ Þ; firms charging prices in the interval

ðr; m=ð1 þ Þ would enjoy no sales. Since the elasticity of demand is con-
stant, firms that would maximize profits by pricing above r in the absence of
consumer search find it optimal to set their prices at r when consumers
^
search.7 Thus, the distribution of prices, FðpÞ; is inconsistent with optimizing
behavior on the part of firms. In fact, given the reservation price r, optimizing
behavior on the part of firms would imply a distribution of prices
(
F^ ðpÞ if por
FðpÞ ¼
1 if p ¼ r:
To establish that this is, in fact, an equilibrium distribution of prices one

must verify that consumers facing this ‘‘truncated’’ distribution of prices have
no incentive to change their reservation price. Given this truncated distribu-
tion of prices, the net expected benefits of search are
Z r
hðrÞ ¼ ðuðpÞ uðrÞÞdF ðpÞ c
p
Z r
¼ ðuðpÞ uðrÞÞd F^ ðpÞ þ ½1 FðrÞ½uðrÞ
^ uðrÞ c
p
Z r
¼ ðuðpÞ uðrÞÞd F^ ðpÞ c ¼ 0,
p
where the last equality follows from the fact that r is the optimal reservation
price when consumers face the price distribution F^ : In short, Reinganum’s
assumptions of downward sloping demand and cost heterogeneity give rise to
an equilibrium of price dispersion with optimizing consumers and firms.
Note that downward sloping demand and cost heterogeneities together
play a critical role in generating equilibrium price dispersion in this envi-
ronment. To see that both assumptions are required, suppose first that costs
are heterogeneous but that each consumer wished to purchase one unit of
the product, valued at u. In this case, given a reservation price of r u; all
firms would find it optimal to price at r, and the distribution of prices would
be degenerate. Of course, a reservation price of rou is inconsistent with
7
Reinganum assumes that m̄ m =ð1 þ Þ; which guarantees that firms who would otherwise price
above r find it profitable to price at r.
optimizing behavior on the part of consumers. To see this, suppose that a

consumer was unexpectedly presented with a price p0 ¼ r þ d where doc.
According to the search strategy, such a consumer is supposed to reject this
price and continue searching; however, the benefit from this additional
search is less than the cost. Thus, a consumer should optimally accept a
price p0 rather than continuing to search. The upshot of this is that the only
equilibrium reservation price is r ¼ u: However, these are precisely the
conditions given in Case 1; hence, the only equilibrium is where no con-
sumers shop at all.8
If demand were downward sloping but firms had identical marginal costs of
m, each firm would have an incentive to set the same price, pn ¼
minfr; m=ð1 þ Þg; given the reservation price. This leads back to Case 2 and
one obtains the Diamond paradox: all firms charge the monopoly price, pn ¼
m=ð1 þ Þ: Indeed, in the environment above, a limiting case where the dis-
tribution of marginal costs converges to a point is exactly the Diamond model.
Finally, we examine how the variance in the distribution of posted (and
transactions) prices varies with search costs. Note that, in equilibrium, the
variance in prices is given by
s2 ¼ E½p2 ðE½pÞ2
Z r Z r !2
¼ p2 dF ðpÞ pdF ðpÞ
p p
Z Z !2
r r
¼ p2 f^ðpÞdp þ ð1 F^ ðrÞÞr2 pf^ðpÞdp þ ð1 F^ ðrÞÞr ;
p p
where f^ðpÞ is the density of F^ ðpÞ: Hence,

Z !
r
ds2
¼ 2rð1 F^ ðrÞÞ 2 pf^ðpÞdp þ ð1 F^ ðrÞÞr ð1 F^ ðrÞÞ
dr p
¼ 2½1 F^ ðrÞðr E½pÞ 0

with strict inequality if rom=ð1 þ Þ: Thus, we have:
Conclusion 1. In the Reinganum model, a reduction in search costs de-

creases the variance of equilibrium prices.
As we will see below, however, this is not a general property of search-

theoretic models of price dispersion.
8
Carlson and McAfee (1983) show that if one introduces heterogeneities in consumer search costs, a
dispersed price equilibrium may exist provided that individual consumers have perfectly inelastic (in
contrast to downward sloping) demand.
2.1.4 Remarks on fixed versus sequential search

It is useful to highlight some key differences between sequential and
fixed sample size search. With sequential search, the number of searches
is a random variable from a geometric distribution, and the expected
number of searches, given a distribution of prices F(p) and reservation
price r, is
1
E½n ¼ .
FðrÞ
In contrast, with fixed sample size search, consumers commit up front to n
searches. Both types of search have advantages and disadvantages, and
indeed Morgan and Manning (1985) have shown that both types of search
can be optimal in different circumstances. The key advantage of sequential
search is that it allows a searcher to economize on information costs—
the decision-maker weighs the expected benefits and costs of gathering
additional price information after each new price quote is obtained. If an
acceptable price is obtained early on, the expected gains from additional
searches are small and there is no need to pay the cost of additional
searches. The primary advantage of fixed sample size search is that it allows
one to gather information quickly. Consider, for instance, a firm that re-
quires raw materials by the end of the week. If it takes a week for a raw
materials vendor to provide a price quote, sequential search would permit
the firm to obtain a price quote from only a single vendor. In this case, fixed
sample size search is optimal—the firm commits to obtain quotes from n
vendors, where n is chosen by the firm to minimize expected costs as out-
lined above in our discussion of the Stigler model.
2.1.5 The MacMinn model

In light of the fact that there are instances in which fixed sample size
search is optimal, one may wonder whether equilibrium price dispersion can
arise in such a setting. MacMinn (1980) provides an affirmative answer to
this question. MacMinn’s result may be seen in the following special case of
our environment where:
1. Consumers have unit demand with valuation u;
2. Consumers engage in optimal fixed sample size search; and9
3. Firms have privately observed marginal costs described by the atom-
less distribution G(m) on ½m; m; where mou:
At the time, MacMinn derived equilibrium pricing by solving a set of
differential equations under the special case where G is uniformly distrib-
uted. However, subsequent to his paper, a key finding of auction theory, the
9
MacMinn also provides a version of the model that is valid for optimal sequential search.
Revenue Equivalence Theorem (Myerson, 1981) was developed.10 Using the

revenue equivalence theorem, we can generalize MacMinn’s results to ar-
bitrary cost distributions.
To see this, notice that when consumers optimally engage in a fixed
sample size search consisting of n firms, each firm effectively competes with
n 1 other firms to sell one unit of the product. Of these n firms, the firm
posting the lowest price wins the ‘‘auction.’’
Using the revenue equivalence theorem, one can show that the expected
revenues to a firm with marginal cost m in any ‘‘auction’’ where the firm
charging the lowest price always wins and the firm with the highest mar-
ginal cost earns zero surplus is
Z m
nn 1 n
RðmÞ ¼ mð1 GðmÞÞ þ ð1 GðtÞÞn 1 dt. (5)
m
In the MacMinn model, expected revenues are simply a firm’s posted

price, p(m), multiplied by the nprobability it charges the lowest price, which, in
equilibrium, is ð1 GðmÞÞn 1 : Using the fact that RðmÞ ¼ pðmÞð1
nn 1
GðmÞÞ ; substituting into equation (5), and solving for p(m) yields the equi-
librium pricing strategy of a firm with marginal cost m when consumers sample
n firms
Z m n
1 GðtÞ n 1
pðmÞ ¼ m þ dt. (6)
m 1 GðmÞ
Notice that, after integration by parts, we can rewrite equation (6) to obtain the
familiar formula for equilibrium bidding in reverse first-price auctions
h n n
i
ðn 1Þ
pðmÞ ¼ E mmin jmðn
min
1Þ
m , (7)
n
where mðn 1Þ
min is the lowest of nn 1 draws from the distribution G.
For the special case where G is uniformly distributed, the equilibrium
pricing strategy simplifies to
nn 1 1
pðmÞ ¼ m þ n m. (8)
nn n
Notice that the equilibrium pricing strategy gives rise to a distribution of
posted prices, F(p), induced by the distribution of costs, that is,
FðpÞ ¼ GðpðmÞÞ.
10
See Klemperer (1999) for a nontechnical survey of auction theory including the revenue equivalence
theorem. McAfee and McMillan (1988) establishes an equivalence between search and auctions in a
mechanism design context.
For this to be an equilibrium distribution of prices, it must be optimal for

consumers to sample n firms. That is,
n n
E½Bðn þ1Þ
oc E½Bðn Þ
where the expression E½BðnÞ ; as previously defined in equation (1) when

K ¼ 1, is the expected benefit from increasing the number of price quotes
obtained from n –1 to n. As in the Stigler model, a reduction in search costs
increases the optimal sample size n (so that consumers optimally sample
more firms).
Thus, MacMinn shows that, provided search costs are low enough, a
dispersed price equilibrium exists. This not only leads to ex post differences
in consumers’ information sets (different consumers sample different firms
and so observe different prices), but induces a degree of competition among
firms (since they are competing against at least one other firm, whose cost
they do not know). As in the Reinganum model, the level of price dispersion
depends on the dispersion in firms’ costs. For the special case where costs
are uniformly distributed, the variance in equilibrium prices ðs2p Þ is given by
n
2 n 1 2 2
sp ¼ sm , (9)
nn
where n is the optimal number of searches by consumers and s2m is the
variance in firm’s costs.
Two interesting results emerge from the model. First, the variance in
prices increases as the variance in firms’ marginal costs increases. This result
is intuitive. Somewhat counterintuitively, note that as the sample size in-
creases, the variance in equilibrium prices increases. This implies that, tak-
ing into account the interaction between consumers and firms in this fixed
sample size search model, dispersion varies inversely with search costs.
Conclusion 2. In the MacMinn model, a reduction in search costs increases
the variance of equilibrium prices.
This conclusion is in contrast to Conclusion 1, where precisely the op-
posite implication is obtained in the Reinganum sequential search model.
This highlights an important feature of search-theoretic models of price
dispersion: Depending on the model, a reduction in search costs may be
associated with higher or lower levels of price dispersion. In the Reinganum
model, a reduction in search costs reduces the reservation price of con-
sumers and thus induces marginal ‘‘high-cost’’ firms to reduce their prices
from their monopoly price to the reservation price. Since the monopoly
prices of low-cost firms are below the reservation price, their prices remain
unchanged; lower search costs thus reduce the range of prices. In the
MacMinn model, lower search costs induce consumers to sample more
firms before purchasing—in effect, each firm competes with more rivals. As
a consequence, the optimal amount of ‘‘bid shading’’ (pricing above mar-

ginal cost) is reduced, thus increasing the level of price dispersion.
2.1.6 The Burdett and Judd model

Burdett and Judd (1983) were the first to show that equilibrium price
dispersion can arise in a search-theoretic model with ex ante identical con-
sumers and firms.11 Burdett and Judd’s main result may be seen in the
following special case of our environment where:
1. Consumers have unit demand up to a price u;

2. Consumers engage in optimal fixed sample size search;12
3. Each firm has constant marginal cost, m, and would optimally charge
all consumers the unique monopoly price, p ¼ u; and
4. A consumer who is charged the monopoly price earns surplus suffi-
cient to cover the cost of obtaining a single price quote.13
In the Burdett and Judd model, an equilibrium consists of a price dis-
tribution F( p) (based on optimal pricing decisions by firms) and an
optimal search distribution oyn 41 n¼1 ; where oyn 4n¼1 is the distribution
1
of the number of times a consumer searches in the population. Thus, yi is

the probability that a consumer searches (or alternatively, the fraction
of consumers that search) exactly i firms. If y1 ¼ 1, then all consumers
sample only one firm. If y1 ¼ 0, then all consumers sample at least two
firms, and so on. Consumers purchase from the firm sampled that offers the
lowest price.
We begin by studying optimal search on the part of consumers given a
price distribution F( p). Recall that the expected benefit to a consumer who
increases her sample size from n –1 to n is
E½BðnÞ ¼ E½pðn1Þ ðnÞ
min E½pmin
as in the Stigler model. Moreover, the expected benefit schedule is strictly

decreasing in n. Thus, an optimal number of price quotes, n, satisfies
E½Bðnþ1Þ oc E½BðnÞ .
First, consider the case where all consumers obtain two or more price
quotes, that is, where y1 ¼ 0. In this case, the optimal pricing strategy on
11
Janssen and Moraga-González (2004) provide an oligopolistic version of the Burdett and Judd
model.
12
Burdett and Judd also provide a version of the model that is valid under optimal sequential search.
13
These assumptions are satisfied, for example, when
8
<1
> if pou;
pu
qðpÞ ¼ 1 k if u p u þ k;
>
:0 if p4u þ k
and k>c/2.
the part of firms is to price at marginal cost (the Bertrand paradox) since
each firm is facing pure price competition with at least one other firm and
all firms are identical. Of course, if all firms are pricing at marginal cost,
then it would be optimal for a consumer to sample only one firm, which
contradicts the hypothesis that y1 ¼ 0. Thus, we may conclude that, in any
equilibrium y1>0.
Next, consider the case where consumers all obtain exactly one price
quote. In that case, each firm would optimally charge the monopoly price,
p ¼ u: Hence, y1 a1 in any dispersed price equilibrium.
From these two arguments it follows that, in any dispersed price equi-
librium, y1 2 ð0; 1Þ: In light of the fact that consumers’ expected benefits
from search are decreasing in the sample size, it follows that a consumer
must be indifferent between obtaining one price quote and obtaining two
price quotes. That is, in any dispersed price equilibrium
E½Bð1Þ 4E½Bð2Þ ¼ c4E½Bð3Þ 4 4E½BðnÞ .
Thus, in any dispersed price equilibrium, y1, y2>0, while yi ¼ 0 for all i>2.
Let y1 ¼ y and y2 ¼ 1– y.
We are now in a position to characterize an atomless dispersed price
equilibrium. First, note that since yA(0, 1), there is a positive probability
that a firm faces no competition when it sets its price. Thus, if firm i charges
the monopoly price, it earns expected profits of
E½pi jpi ¼ u ¼ ðu mÞ my.
In contrast, a firm choosing some lower price ‘‘wins’’ when its price is below
that of the other firm a consumer has sampled. Thus, if firm i charges a
price piru, it earns expected profits of
E½pi jpi u ¼ ðpi mÞ mðy þ ð1 yÞð1 F ðpi ÞÞÞ.
Thus, for a given distribution of searches, equilibrium price dispersion re-

quires that the distribution of firm prices, F( ), satisfies
ðu mÞ
y þ ð1 yÞð1 F ðpÞÞ ¼ y
ðp mÞ
or
ðu pÞ ðyÞ
FðpÞ ¼ 1 , (10)
ðp mÞ ð1 yÞ
which is a well-behaved atomless cumulative distribution having support
[m+y(u–m), u].
Finally, it remains to determine an equilibrium value of y. Since each

consumer must be indifferent between searching one or two firms,
E½Bð2Þ ¼ c.
Notice that, when y ¼ 0, or y ¼ 1, E [B(2)] ¼ 0, while E [B(2)]>0 for all yA
(0, 1). Burdett and Judd show that E [B(2)] is quasi-concave; thus, when c is
sufficiently low, there are generically two dispersed price equilibria—one
involving a relatively high fraction of consumers making two searches, the
other with a relatively low fraction of consumers.14
To summarize, Burdett and Judd show that equilibrium price dispersion
can arise even when all firms and consumers are ex ante identical. In the
equilibrium price distribution, all firms charge positive markups. A fraction y
of consumers do not comparison shop—they simply search at one store and
purchase. The remaining fraction of consumers are ‘‘shoppers’’—these con-
sumers search at two stores and buy from whichever offers the lower price.
2.2 Models with an ‘‘Information Clearinghouse’’
In search-theoretic models, consumers pay an incremental cost for each

additional price quote they obtain. These models are relevant, for example,
when consumers must visit or phone traditional sellers in order to gather
information about prices. They are also relevant in online environments
where consumers must search the web sites of individual retailers to gather
information about the prices they charge.
An alternative class of models is relevant when a third party—an infor-
mation clearinghouse—provides a subset of consumers with a list of prices
charged by different firms in the market. Examples of this environment
include newspapers which display prices different stores charge for the same
product or service and online price comparison sites.
In this section we provide a general treatment of clearinghouse models,
and show that these models are surprisingly similar to those that arise under
fixed sample size search. One of the key modeling differences is that clear-
inghouse models tend to be oligopoly models; thus, there is not a contin-
uum of firms in such settings.
Where possible, we shall use the same notation as in the previous
section; however, for reasons that will become clear when we compare
clearinghouse models with the search models presented above, we now
let n denote the number of firms in the market. The general treatment
that follows relies heavily on Baye and Morgan (2001) and Baye et al.
(2004a).
Consider the following general environment (which we will specialize
to cover a variety of different models). There is a finite number, n>1, of
14
There is a nondispersed price equilibrium where all consumers search once and all firms charge the
monopoly price.
price-setting firms competing in a market selling an identical (homogeneous)

product. Firms have unlimited capacity to supply this product at a constant
marginal cost, m. A continuum of consumers is interested in purchasing the
product. This market is served by a price information clearinghouse. Firms
must decide what price to charge for the product and whether to list this price
at the clearinghouse. Let pi denote the price charged by firm i. It costs a firm
an amount f Z0 if it chooses to list its price. All consumers have unit
demand with a maximal willingness to pay of u>m.15 Of these, a mass, S>0,
of the consumers are price-sensitive ‘‘shoppers.’’ These consumers first con-
sult the clearinghouse and buy at the lowest price listed there provided this
price does not exceed u. If no prices are advertised at the clearinghouse or all
listed prices exceed u, then a ‘‘shopper’’ visits one of the firms at random and
purchases if its price does not exceed u. A mass, LZ0, of consumers per firm
purchase from that firm if its price does not exceed u. Otherwise, they do not
buy the product at all.
It can be shown that if L>0 or f>0, equilibrium price dispersion arises
in the general model—provided of course that f is not so large that firms
refuse to list prices at the clearinghouse. More precisely,
Proposition 3. Let 0 foðn 1Þðu mÞS=n: Then, in a symmetric equi-
librium of the general clearinghouse model:
1. Each firm lists its price at the clearinghouse with probability
n 1=ðn1Þ
n1 f
a¼1 .
ðu mÞS
2. If a firm lists its price at the clearinghouse, it charges a price drawn from
the distribution
n 1= ðn1Þ !
1 f þ ðu pÞL
FðpÞ ¼ 1 n1 on ½p0 ; u,
a ðp mÞS
where
L n=ðn 1Þ
p0 ¼ m þ ðu mÞ þ f.
LþS LþS
3. If a firm does not list its price at the clearinghouse, it charges a price
equal to u.
4. Each firm earns equilibrium expected profits equal to
1
Ep ¼ ðu mÞL þ f.
n1
15
Baye and Morgan (2001) consider an environment with downward sloping demand.
Proof. First, observe that if a firm does not list its price at the clearing-
house, it is a dominant strategy to charge a price of u.
Next, notice that aA(0,1] whenever
nf
o1.
ðn 1Þðu mÞS
This condition holds, since foðn 1Þðu mÞS=n:
Notice that p0>m, provided that L>0 or f>0. In this case, it can be
shown that F is a well-defined, atomless cdf on [p0, u]. When L ¼ 0 and
f ¼ 0, notice that p0 ¼ m. In this case, the symmetric equilibrium dis-
tribution of prices is degenerate, with all firms pricing at marginal cost
(the Bertrand paradox outcome).
Next, we show that, conditional on listing a price, a firm can do no
better than pricing according to F. It is obvious that choosing a price
above or below the support of F is dominated by choosing a price in the
support of F. A firm choosing a price p in the support of F earns expected
profits of
! !
Xn1
n1 i n1i i
EpðpÞ ¼ ðp mÞ L þ ði Þa ð1 aÞ ð1 F ðpÞÞ S f.
i¼0
Using the binomial theorem, we can rewrite this as:

EpðpÞ ¼ ðp mÞ L þ ð1 aFðpÞÞn1 S f
f
¼ ðu mÞL þ
n 1;
where we have substituted for F to obtain the second equality. Since a
firm’s expected profits are constant on [p0, u], it follows that the mixed
pricing strategy, F, is a best response to the other n–1 firms pricing based
on F.
When f ¼ 0, it is a weakly dominant strategy to list. It remains to show
that when f>0 and a 2 ð0; 1Þ; a firm earns the same expected profits
regardless of whether it lists its price. But a firm that does not list earns
expected profits of

S n1
Ep ¼ ðu mÞ L þ ð1 aÞ
n
f
¼ ðu mÞL þ ,
n1
which
equals
the expected profits earned by listing any price
p 2 p0 ; u : ’
We are now in a position to examine the many well-known clearinghouse
models that emerge as special cases of this general environment.
2.2.1 The Rosenthal model

Rosenthal (1980) was among the first to show that equilibrium price
dispersion can arise in a clearinghouse environment when some consumers
have a preference for a particular firm. Under his interpretation, each firm
enjoys a mass L of ‘‘loyal’’ consumers. Rosenthal’s main results may be
seen in the following special case of the general clearinghouse model:
1. It is costless for firms to list prices on the clearinghouse: f ¼ 0; and
2. Each firm has a positive mass of loyal consumers: L>0.
Since f ¼ 0, it follows from Proposition 3 that a ¼ 1; that is, all of the n

firms advertise their prices with probability one. Using this fact and Prop-
osition 3, the equilibrium distribution of prices is

ðu pÞ L 1=ðn1Þ
FðpÞ ¼ 1 on ½p0 ; u, (11)
ðp mÞ S
where
L
p0 ¼ m þ ðu mÞ .
LþS
The price dispersion arising in the Rosenthal model stems from ex-
ogenous differences in the preferences of consumers. While shoppers view
all products as identical and purchase at the lowest listed price, each firm is
endowed with a stock of L loyals. The equilibrium price dispersion arises
out of the tension created by these two types of consumers. Firms wish to
charge u to extract maximal profits from the loyal segment, but if all firms
did so a firm could slightly undercut this price and gain all of the shoppers.
One might imagine that this ‘‘undercutting’’ argument would lead to the
Bertrand outcome. However, once prices get sufficiently low, a firm is better
off simply charging u and giving up on attracting shoppers. Thus, the only
equilibrium is in mixed strategies—firms randomize their prices, sometimes
pricing relatively low to attract shoppers and other times pricing fairly high
to maintain margins on loyals.
It is interesting to examine the equilibrium transactions prices in the
market. Loyal customers expect to pay the average price charged by firms:
Z u
E½p ¼ pdFðpÞ,
p0
while shoppers expect to pay the lowest of n draws from F(p); that is, the
expected transaction price paid by shoppers is
Z u
ðnÞ
E½pmin ¼ pdF ðnÞ
min ðpÞ,
p0
where F ðnÞ
min ðpÞ is the cdf associated with the lowest of n draws from F.
How do transactions prices vary with the number of competing firms?

Rosenthal’s striking result is that, as the number of competing firms in-
creases, the expected transactions prices paid by all consumers go up. As we
shall see below, the result hinges on Rosenthal’s assumption that entry
brings more loyals into the market. Indeed, the fraction of shoppers in the
market is S/(S+nL) and it may readily be seen that as n becomes large,
shoppers account for an increasingly small fraction of the customer base of
firms. As a consequence, the incentives to compete for these customers is
attenuated and prices rise as a result. The key is to recognize that increases
in n change the distribution of prices, and this effect as well as any order
statistic effect associated with an increase in n must be taken into account.
Formally, notice that the equilibrium distribution of prices, F, is stoch-
astically ordered in n. That is, the distribution of prices when there are
n+1 firms competing first-order stochastically dominates the distribution
of prices where there are n firms competing. This implies that the trans-
actions prices paid by loyals increase in n. To show that the transactions
prices paid by shoppers also increase in n requires a bit more work; how-
ever, one can show that the same stochastic ordering obtains for the cdf
F ðnÞ
min ðpÞ:
Finally, it is useful to note the similarity between the Rosenthal version
of the clearinghouse model and the search-theoretic model of Burdett
and Judd. In Burdett and Judd, even though there is a continuum of firms,
each consumer only samples a finite number of firms (one or two). Further,
in Burdett and Judd, a fixed fraction of consumers per firm, my, sample
only a single firm. In effect, these consumers are ‘‘loyal’’ to the single firm
sampled, while the fraction (1–y)m of customers sampling two firms are
‘‘shoppers’’—they choose the lower of the two prices. For this reason, when
n ¼ 2 in the Rosenthal model, the equilibrium price distribution given in
equation (11) is identical to equation (10) in Burdett and Judd model
(modulo relabeling the variables for loyals and shoppers).
2.2.2 The Varian model

Varian (1980) was among the first to show that equilibrium price dis-
persion can arise in a clearinghouse environment when consumers have
different ex ante information sets.16 Varian interprets the S consumers as
‘‘informed consumers’’ and the L consumers as ‘‘uninformed’’ consumers.
Thus a mass, S, of consumers choose to access the clearinghouse while
others, the mass L per firm, do not. Varian’s main result may be seen in the
following special case of the general clearinghouse model:
1. It is costless for firms to list prices on the clearinghouse: f ¼ 0; and
2. The total measure of ‘‘uninformed’’ consumers lacking access to the
16
Png and Hirshleifer (1987), as well as Baye and Kovenock (1994), extend the Varian model by
allowing firms to also engage in price matching or ‘‘beat or pay’’ advertisements.
clearinghouse is U>0; hence, each firm is visited by L ¼ U/n of these

consumers.
Again, since f ¼ 0, it follows that a ¼ 1 and hence all n firms advertise

their prices at the clearinghouse. Using this fact and setting L ¼ U/n in
Proposition 3, the equilibrium distribution of prices is
1=ðn1Þ
ðu pÞ Un
FðpÞ ¼ 1 on ½p0 ; u;
ðp mÞ S
where
U=n
p0 ¼ m þ ðu mÞ .
U=n þ S
The fact that this atomless distribution of prices exists whenever there is
an exogenous fraction of consumers who do not utilize the clearinghouse
raises the obvious question: Can this equilibrium persist when consumers
are making optimal decisions? Varian shows that the answer to this ques-
tion is yes—provided different consumers have different costs of accessing
the clearinghouse. The easiest way to see this is to note that the value of
information provided by the clearinghouse is the difference in the expected
price paid by those accessing the clearinghouse, E½pðnÞmin ; and those not, E
[p], that is;
VOI ðnÞ ¼ E½p E½pðnÞ
min , (12)
where VOI denotes the value of (price) information contained at the clear-
inghouse. Suppose consumers face a cost of accessing the information
provided by the clearinghouse. Note that this cost is essentially a fixed cost
of gaining access to the entire list of prices, not a per price cost as in the
search-theoretic models considered above. Varian assumes that the cost to
type S and L consumers of accessing the clearinghouse is kS and kL, with
kSokL. Then provided ksrVOI (n)okL, type S consumers will optimally
utilize the clearinghouse, while the type L consumers will not. In short, if
different consumers have different costs of accessing the clearinghouse,
there exists an equilibrium of price dispersion with optimizing consumers
and firms. In such an equilibrium, informed consumers pay lower average
prices than uninformed consumers.
It is important to emphasize that when one endogenizes consumers’ de-
cisions to become informed in the Varian model, the level of price disper-
sion is not a monotonic function of consumers’ information costs. When
information costs are sufficiently high, no consumers choose to become
informed, and all firms charge the ‘‘monopoly price,’’ u. When consumers’
information costs are zero, all consumers choose to become informed, and
all firms price at marginal cost in a symmetric equilibrium—the Bertrand
paradox. Thus, for sufficiently high or low information costs, there is no
price dispersion; for moderate information costs, prices are dispersed on the
nondegenerate interval ½p0 ; u: A similar result obtains in Stahl (1989), which
is related to Varian as follows. Stahl assumes a fraction of consumers have
zero search costs and, as a consequence, view all firms’ prices and purchase
at the lowest price in the market. These consumers play the role of S in
Varian’s model (informed consumers). The remaining fraction of consum-
ers correspond to the L’s in the Varian model, but rather than remaining
entirely uninformed, these consumers engage in optimal sequential search in
the presence of positive incremental search costs. Stahl shows that when all
consumers are shoppers, the identical firms price at marginal cost and there
is no price dispersion. When no consumers are shoppers, Diamond’s par-
adox obtains and all firms charge the monopoly price. As the fraction of
shoppers varies from zero to one, the level of dispersion varies continu-
ously—from zero to positive levels, and back down to zero.
Conclusion 3. In general, price dispersion is not a monotonic function of
consumers’ information costs or the fraction of ‘‘shoppers’’ in the market.
How does the number of competing firms affect transactions prices? In
the Rosenthal model, we saw that increased ‘‘competition’’ led to higher
expected transactions prices for all consumers. In the Varian model, in
contrast, the effect of competition on consumer welfare depends on whether
or not the consumer chooses to access the clearinghouse. Morgan et al.
(forthcoming) show that as n increases, the competitive effect predictably
leads to lower average transactions prices being paid by informed consum-
ers. However, the opposite is true for uninformed consumers—as the
number of competing firms increases, firms face reduced incentives to cut
prices in hopes of attracting the ‘‘shoppers’’ and, as a consequence, the
average price charged by a firm, which is also the average price paid
by an uninformed consumer, increases. If one views the clearinghouse as
representing access to price information on the Internet, then one can in-
terpret the price effect as one consequence of the so-called ‘‘digital divide’’;
see Baye et al. (2003). Consumers with Internet access are made better off
by sharper online competition while those without such access are made
worse off.
2.2.3 The Baye and Morgan model

All of the above models assume that it is costless for firms to advertise
their prices at the clearinghouse. Baye and Morgan (2001) point out that, in
practice, it is generally costly for firms to advertise their prices and for
consumers to gain access to the list of prices posted at the clearinghouse. For
example, newspapers charge firms fees to advertise their prices and may
choose to charge consumers subscription fees to access any posted infor-
mation. The same is true of many online environments. Moreover, the
clearinghouse is itself an economic agent, and presumably has an incentive to
endogenously choose advertising and subscription fees to maximize its own
expected profits. Thus, Baye and Morgan examine the existence of dispersed
price equilibria in an environment with optimizing consumers, firms, and a
monopoly ‘‘gatekeeper’’ who controls access to the clearinghouse.
Specifically, Baye and Morgan consider a homogeneous product envi-
ronment where n identical, but geographically distinct, markets are each
served by a (single) local firm. Distance or other transaction costs create
barriers sufficient to preclude consumers in one market from buying this
product in another market; thus each firm in a local market is a monopolist.
Now imagine that an entrepreneur creates a clearinghouse to serve all
markets. In the Internet age, one can view the clearinghouse as a virtual
marketplace—through its creation, the gatekeeper expands both consumers’
and firms’ opportunities for commerce. Each local firm now has the option
to pay the gatekeeper an amount f to post a price on the clearinghouse in
order to gain access to geographically disparate consumers. Each consumer
now has the option to pay the gatekeeper an amount k to shop at the
clearinghouse and thereby purchase from firms outside the local market.
The monopoly gatekeeper first sets k and f to maximize its own expected
profits. Given these fees, profit-maximizing firms make pricing decisions
and determine whether or not to advertise them at the clearinghouse. Sim-
ilarly, consumers optimally decide whether to pay k to access the clearing-
house. Following this, a consumer can simply click her mouse to research
prices at the clearinghouse (if she is a subscriber), visit the local firm, or
both. With this information in hand, a consumer decides whether and from
whom to purchase the good.
Baye and Morgan show that the gatekeeper maximizes its expected
profits by setting k sufficiently low that all consumers subscribe, and
charging firms strictly positive fees to advertise their prices. Thus, Baye and
Morgan’s main results may be seen in the following special case of the
general clearinghouse model:
1. The gatekeeper optimally sets positive advertising fees: f>0; and
2. The gatekeeper optimally sets subscription fees sufficiently low such
that all consumers access the clearinghouse, that is, L ¼ 0.
Under these conditions, using Proposition 3, we obtain the following
characterization of equilibrium firm pricing and listing decisions: Each firm
lists its price at the clearinghouse with probability
n 1=ðn1Þ
n1 f
a¼1 2 ð0; 1Þ.
ðu mÞS
When a firm lists at the clearinghouse, it charges a price drawn from the
distribution
n 1=ðn1Þ !
1 n1 f
FðpÞ ¼ 1 on ½p0 ; v;
a ðp mÞS
where
n
ðn1Þ f
p0 ¼ m þ .
S
When a firm does not list its price, it charges a price equal to u, and each
firm earns equilibrium expected profits equal to
1
Ep ¼ f.
n1
Notice that na represents the aggregate demand by firms for advertising
and is a decreasing function of the fee charged by the gatekeeper. Prices
advertised at the clearinghouse are dispersed and strictly lower than un-
advertised prices (u).
Several features of this equilibrium are worth noting. First, equilibrium
price dispersion arises with fully optimizing consumers, firms, and endog-
enous fee-setting decisions on the part of the clearinghouse—despite the
fact that there are no consumer or firm heterogeneities and all consumers
are ‘‘fully informed’’ in the sense that, in equilibrium, they always purchase
from a firm charging the lowest price in the global market. Second, while
equilibrium price dispersion in the Varian model is driven by the fact that
different consumers have different costs of accessing the clearinghouse,
Baye and Morgan show that an optimizing clearinghouse will set its fees
sufficiently low that all consumers will rationally access the clearinghouse.
Equilibrium price dispersion arises because of the gatekeeper’s incentives to
set strictly positive advertising fees. Strikingly, despite the fact that all
consumers use the gatekeeper’s site and thus purchase at the lowest global
price, firms still earn positive profits in equilibrium. In expectation, these
profits are proportional to the cost, f, of accessing the clearinghouse.
Conclusion 4. In the Baye and Morgan model, equilibrium price dispersion
persists even when it is costless for all consumers to access the information
posted at the gatekeeper’s site. Indeed, price dispersion exists because it is
costly for firms to transmit price information (advertise prices) at the
gatekeeper’s site.
Why does the gatekeeper find it optimal to set low (possibly zero) fees for
consumers wishing to access information, but strictly positive fees to firms
who wish to transmit price information? Baye and Morgan point out that
this result stems from a ‘‘free-rider problem’’ on the consumer side of the
market that is not present on the firm side. Recall that the gatekeeper can
only extract rents equal to the value of the outside option of firms and
consumers. For each side of the market, the outside option consists of the
surplus obtainable by not utilizing the clearinghouse. As more consumers
access the site, the number of consumers still shopping locally dwindles and
the outside option for firms is eroded. In contrast, as more firms utilize the
clearinghouse, vigorous price competition among these firms reduces listed
prices and leads to a more valuable outside option to consumers not using
the clearinghouse. Thus, to maximize profits, the gatekeeper optimally
subsidizes consumers to overcome this ‘‘free-rider problem’’ while capturing
rents from the firm side of the market. No analogous ‘‘free-rider problem’’
arises on the firm side; indeed greater consumer participation at the clear-
inghouse increases the frequency with which firms participate (a increases)
and hence permits greater rent extraction from firms.
2.2.4 Models with asymmetric consumers

In general, little is known about the general clearinghouse model with
asymmetric consumers.17 However, for the special case of two firms, results are
available. Narsimhan (1988) analyzes the case where there are two firms, one
of whom has more loyal customers than the other, and where the cost of listing
at the clearinghouse is zero. He shows that the unique equilibrium involves
mixed strategies in which the firm with more loyal customers prices less ag-
gressively than its rival. As a result, the rival (with fewer loyals) earns equi-
librium profits that exceed the profits it would earn by exclusively serving its
loyal customers at the monopoly price. The firm with more loyals, in contrast,
earns equilibrium profits that equal the profits it could earn by exclusively
serving its loyal customers at the monopoly price. Interestingly, shoppers are
harmed by these asymmetries – they pay, on average, higher prices than would
be the case were the loyal customers divided equally between the firms. Further
theoretical work on clearinghouse models with consumer asymmetries and
positive listing fees would be a useful addition to the literature.
2.5.5 Cost heterogeneities and the Spulber model

Spulber (1995) considers a situation where consumers have access to the
complete list of prices and buy from the firm offering the lowest price. Of
course, in such a setting, if firms were identical one would immediately
obtain the Bertrand outcome. To generate price dispersion, Spulber exam-
ines the situation where firms have heterogeneous costs and consumers have
downward sloping demand. However, the main economic intuition under-
lying the model may be seen through the following adaptation of our gen-
eral clearinghouse framework for the unit demand case:
1. All consumers are shoppers: S>0 and L ¼ 0;
2. There is no cost to advertise prices on the clearinghouse: f ¼ 0; and
3. Firms have privately observed marginal costs described by the atom-
less distribution G (m) on ½m; m:
Since there are no costs to advertise prices, all firms list prices on
the clearinghouse. Each firm faces competition from n –1 other firms with
17
For specific clearinghouse models, some results are available. For instance, Baye et al. (1992)
characterize all equilibria in a version of the Varian model in which firms have asymmetric numbers of
consumers.
random marginal costs. Since the firm charging the lowest price wins the
entire market, firms are effectively competing in an auction in which their
own costs are private information. For the special case of unit demand, the
equilibrium price for a firm is again the familiar expression from a first-
price auction
pðmÞ ¼ E½mðn1Þ ðn1Þ
min jmmin m, (13)
where mðn1Þ
min is the lowest of n–1 draws from the distribution G.
There are several noteworthy features of this equilibrium. First, equilib-
rium firm pricing entails positive markups despite the fact that all con-
sumers are ‘‘shoppers’’ and have a complete list of prices. Intuitively, there
is a trade-off between lowering one’s price to attract shoppers and the
profitability of this price. In equilibrium, this results in a markup which
depends on the number of competing firms. As the number of firms grows
large, the equilibrium markup becomes small. Second, notice that cost het-
erogeneity leads to equilibrium price dispersion despite the fact consumers
are identical and all consumers are purchasing at the lowest price.
It is interesting to compare the Spulber model, which occurs in the clear-
inghouse framework, with the search-theoretic framework of MacMinn.
Notice that when the number of competing firms in Spulber, n, is equal to the
optimal fixed sample size for consumers in the MacMinn model, n ; the
equilibrium distribution of prices, equations (13) and (7), are identical in the
two models. That is, cost heterogeneities are sufficient to generate price dis-
persion in oligopoly models where all consumers obtain complete price in-
formation, as well as in models where a continuum of firms compete but each
consumer only obtains price quotes from a finite number n of these firms.
2.3 Bounded rationality models of price dispersion
Several recent papers have emphasized that bounded rationality can also
lead to price dispersion. The idea is to relax the Nash equilibrium assump-
tion—which requires that each decision-maker in the market is choosing an
action (be it a price or a search strategy) that is a best response to given
actions of other market participants. Two equilibrium concepts—quantal
response equilibrium (McKelvey and Palfrey, 1995) and epsilon equilibrium
(Radner, 1980)—are particularly useful because they nest the standard
Nash equilibrium concept as a special case.
In a quantal response equilibrium (QRE), the likelihood that a particular
firm sets a specific price depends on the expected profits arising from that price
(see Lopez-Acevedo, 1997). A firm’s price is determined by a stochastic de-
cision rule, but prices leading to higher expected profits are more likely to be
charged. Of course, each firm’s expected profits from different pricing deci-
sions depend on the probability distributions of other players’ prices. A QRE
requires that all firms hold correct beliefs about the probability distributions
of other players’ actions. The nondegenerate distributions of prices resulting in

a QRE may be viewed as shocks to firms’ profit functions. Alternatively,
nondegenerate price distributions might stem from decision errors by firms.
Such errors may arise from limitations in managers’ cognitive processing
abilities or ‘‘bugs’’ in dynamic pricing algorithms used by Internet retailers.
In an e-equilibrium, the prices charged by each firm are such that no firm
can gain more than e in additional profits by changing its price. Such an
equilibrium may arise because of cognitive or motivational constraints on
the part of firms. For example, if it is costly to reprogram dynamic pricing
algorithms, managers may not be willing to incur these economic or psychic
costs when the resulting gain is small (less than e).
Recently, Baye and Morgan (2004) applied QRE and e-equilibrium con-
cepts to pricing games and showed that only a little bounded rationality is
needed to generate the patterns of price dispersion documented in labo-
ratory experiments as well as observed on Internet price comparison sites.
In a similar vein, Rauh (2001) shows that price dispersion can arise when
market participants make small but heterogeneous mistakes in their beliefs
about the distribution of prices. Ellison (2005) provides a more detailed
treatment of recent advances along these lines.
2.4 Concluding remarks: theory
Despite a slow start, there are now a variety of models that can be used
to rationalize equilibrium price dispersion in online and offline markets.
We conclude our theoretical discussion with the following general obser-
vations.
1. There is not a ‘‘one-size-fits-all’’ model of equilibrium price dispersion;

different models are appropriate for analyzing different market
environments. For instance, search-theoretic models are most appro-
priate for analyzing environments where consumers must visit different
stores or firms’ web sites to gather price information. Clearinghouse
models are appropriate when consumers are able to access a list of
prices (e.g., in a newspaper or at a price comparison site).
2. The distribution of prices is determined by the interaction of all market
participants—firms, consumers and, in the case of clearinghouse mod-
els, information gatekeepers. As a consequence, the level of price dis-
persion depends on the structure of the market—the number of sellers,
the distribution of costs, consumers’ elasticities of demand, and so on.
3. Reductions in search costs may lead to either more or less price
dispersion, depending on the market environment. Furthermore,
the elimination of consumer search costs need not eliminate price
dispersion.
4. Depending on the market environment, heightened competition (in-
creases in the number of firms) can increase or decrease the level of
dispersion. Moreover, in some models, heightened competition of this

form leads to higher transactions prices paid by all consumers. In
other models, the effect of increased competition on the welfare of
consumers depends on which side of the ‘‘digital divide’’ a consumer
resides.
5. Price dispersion is not purely an artifact of ex ante heterogeneities in
firms or consumers. While differences in firms’ costs or base of loyal
consumers (stemming from firms’ branding efforts, differential service
qualities, or reputations) can contribute to equilibrium price disper-
sion, such differences are not necessary for equilibrium price disper-
sion.
6. Thanks to the Internet, information gatekeepers are playing an in-
creasingly important role in the economy. In their attempt to max-
imize profits and enhance the value of information provided by their
sites, information gatekeepers have an incentive to charge fees for their
services that induce equilibrium price dispersion.
7. A little bounded rationality goes a long way in explaining price dis-
persion.
3 Empirical analysis of price dispersion
We now turn to the empirical literature on price dispersion. In Section 3.1,

we discuss some of the strengths and weaknesses of commonly used metrics
for measuring price dispersion in online and offline markets. Section 3.2
provides an overview of the empirical literature and highlights empirical
evidence suggesting that information costs (either on the consumer or firm
side of the market) contribute to price dispersion; that is, dispersion is not
purely an artifact of subtle product heterogeneities.
3.1 Measuring price dispersion
The equilibrium models of price dispersion presented above each imply

non-degenerate distributions of prices, F(p), on some interval ½p; p: Given
such a distribution, a standard measure of dispersion is the variance in
prices. For each model of equilibrium price dispersion, this measure can be
directly computed. For instance, in the MacMinn model, if firms have
uniformly distributed marginal costs, the variance in prices is
n
2 n 1 2 ðm m Þ2
sp ¼ .
nn 12
Notice that one is then in a position to test comparative static predictions of
the model using this measure. In a similar manner, expressions for the
variance in prices may be derived from the other models previously pre-
sented.
A number of authors use the sample variance to measure price dispersion

(e.g., Pratt et al., 1979; Ancarani and Shankar, 2004). The obvious advan-
tage is that it uses all available data. A drawback of this measure is ap-
parent when comparing dispersion across products or over time. For
instance, suppose that, during an inflationary period, the marginal costs of
all firms in the MacMinn model increased by a factor g>1. In that case, the
new variance would simply scale-up the original variance by a factor g2.
Thus, this measure of price dispersion would change even though the un-
derlying real economics of the situation are the same after the inflationary
period.
For this reason, if one wishes to compare levels of price dispersion either
across different products or across time, one must standardize the data
in some fashion. An alternative is to use the coefficient of variation,
CV ¼ sp/E[P] (or its sample analogue), which is homogenous of degree
zero in the level of prices. The CV is particularly useful when comparing
levels of price dispersion over long periods of time (e.g., Scholten and
Smith, 2002; Eckard, 2004) or across different products (e.g., Carlson and
Pescatrice, 1980; Sorensen, 2000; Aalto-Setälä, 2003; Baye, 2004a,b). An
added advantage is that, unlike some methods of standardization, the co-
efficient of variation may preserve the comparative static predictions of the
model of interest. For instance, in the MacMinn model, equation (8) implies
that the expected price is

nn 1 m þ m m
E½p ¼ þ n,
nn 2 n
and thus the coefficient of variation is
1 ðnn 1Þðm mÞ
CV ¼ pffiffiffi n .
3 ðn 1Þðm þ mÞ þ 2m
One may verify that this statistic is, like the variance, decreasing in search
costs, but, unlike the variance, this statistic does not change with a mul-
tiplicative shift in firms’ costs.
Another widely used measure of price dispersion is the (sample) range;
see, for instance, Pratt et al. (1979) and Brynjolfsson and Smith (2000).
Letting pðnÞ ðnÞ
min and pmax denote, respectively, the lowest and highest of n
observed prices drawn from F, then the range is
ðnÞ
RðnÞ ¼ pðnÞ
max pmin .
Given the equilibrium distribution of prices implied by a particular the-

oretical model, comparative static predictions about changes in the range
are possible based on the behavior of the highest and lowest order statistics.
That is, one can perform comparative static analysis on the expected
range18:
ðnÞ
E½RðnÞ ¼ E½pðnÞ
max E½pmin .
Unfortunately, all of the above measures of price dispersion suffer from a

potential theoretical defect. Suppose that n>2 firms compete in a classical
homogeneous product Bertrand setting. Under standard conditions there
will exist a unique symmetric equilibrium where all firms price at marginal
cost. But in addition, there are asymmetric equilibria where two firms price
at marginal cost and the remaining n –2 firms price strictly above marginal
cost. Thus, price dispersion can arise in a classical Bertrand environment.
Yet, the apparent price dispersion is arguably not economically relevant
because the unique transactions price is marginal cost.
To remedy this theoretical defect, Baye et al. (2004a) propose a measure
called ‘‘the gap,’’ which they define to be the difference between the two
lowest prices in the market. Letting pðnÞ2 denote the second-lowest price
realization from n draws from F, the (sample) gap is defined as19
GðnÞ ¼ pðnÞ ðnÞ

2 pmin .
The classical Bertrand model (as well as textbook models of perfect com-
petition) implies that the gap between the two lowest prices is zero in any
equilibrium (symmetric or otherwise). All of the oligopoly models of price
dispersion discussed above, in contrast, imply a positive gap. An additional
property of the gap is that it gives greater weight to low prices, which, in the
absence of quantity data, one might reasonably assume lead to more sales
than higher prices. The key disadvantage, shared by the range, is that it
relies purely on extreme values of the data. Hence, the range and gap are
more sensitive to outliers and other forms of ‘‘noise’’ than measures that use
all the available data, such as the sample variance and coefficient of var-
iation.
In addition to these measures, the value of information (VOI) defined
earlier in equation (12) can also be used as a gauge of dispersion. This
measure, which is simply the difference between the average observed price
and the lowest observed price, is zero in the absence of any price dispersion
but otherwise positive. The principal advantage of this measure of disper-
sion is that it has a very intuitive interpretation: its value indicates the
18
To facilitate comparisons across different products or over time, it is sometimes useful to normalize
the range by dividing it by the minimum or average price; see Baye et al. (2004b) and Brynjolfsson and
Smith (2000).
19
As with the range, one can perform comparative static analyses for any of the theoretical models
using the expected gap, and it is sometimes useful to normalize the gap by dividing by the lowest price.
In this formulation, the gap represents the difference between the two lowest prices expressed as a
percentage of the lowest price realization.
amount of money a consumer saves by purchasing at the best price rather

than from a randomly selected firm in the market.
3.2 Price dispersion in the field
If price dispersion stems from frictions related to the acquisition and

transmission of information (as implied by the models in Section 2) rather
than subtle differences in firms’ service levels, observed levels of dispersion
should systematically depend on ‘‘environmental factors’’ present in the
models. For example, in his seminal article on the economics of informa-
tion, George Stigler advanced the following hypotheses:
y dispersion itself is a function of the average amount of search, and this in turn is a function of
the nature of the commodity:
1. The larger the fraction of the buyer’s expenditures on the commodity, the greater the savings
from search and hence the greater the amount of search.
2. The larger the fraction of repetitive (experienced) buyers in the market, the greater the
effective amount of search (with positive correlation of successive prices).
3. The larger the fraction of repetitive sellers, the higher the correlation between successive
prices, and hence, the larger the amount of accumulated search.
4. The cost of search will be larger, the larger the geographic size of the market.
Stigler (1961, p. 219)
Stigler’s hypotheses offer a useful guide for understanding the empirical

literature on price dispersion. Much of this literature tests Stigler’s hy-
potheses by examining whether search intensity (proxied by variables that
affect the benefits and costs of search) is correlated with levels of price
dispersion. As we have seen, however, when one takes Rothschild’s crit-
icism into account, an increase in search intensity can lead to increases or
decreases in the level of equilibrium price dispersion, depending on the
model. Thus, one challenge for empirical researchers is choosing a model
that closely approximates the ‘‘data generating’’ environment. A second
challenge is to control for factors outside of the model that might influence
levels of dispersion. A third challenge arises because firm optimization is
absent in Stigler’s model, but is clearly present in the data. For this reason,
a number of empirical studies look beyond Stigler’s hypotheses to test hy-
potheses derived from specific search-theoretic or clearinghouse models of
equilibrium price dispersion. We provide a broad overview of these and
related strands of the literature below.
3.2.1 Dispersion and the ‘‘benefits’’ of search

The search-theoretic models presented in Section 2 imply that search
intensity depends, in part, on the consumer’s demand for a product. In the
Stigler model, demand is represented by the parameter, K. The greater is K,
the greater the expected benefits of search and hence the greater the search
intensity. Stigler’s first two hypotheses are based on the notion that the
share of an item in a consumer’s overall budget and the frequency with
which an item is purchased are good proxies for K.
Dispersion for ‘‘cheap’’ versus ‘‘expensive’’ items. Stigler (1961) provides
casual evidence in support of his first hypothesis—that dispersion is lower
for items that account for a large expenditure share of a searcher’s con-
sumption bundle (‘‘expensive items’’) than those that account for a smaller
expenditure share (‘‘cheap items’’). Government coal purchases are a small
percentage of the overall government budget, while a household’s expen-
ditures on an automobile comprise (in 1961 as well as today) a much larger
percentage of its overall budget. Stigler obtained different sellers’ prices for
two homogeneous products—anthracite-grade coal to be sold to the gov-
ernment and an automobile to be sold to a household. Prices for anthracite
coal ranged from $15.46 to $18.92, with an average price of $16.90 and a
standard deviation of $1.15. Prices for the automobile (based on what
Stigler called ‘‘an average amount of higgling’’) ranged from $2,350 to
$2,515, with an average price of $2,436 and standard deviation of $42.
Stigler’s data thus tend to support his first conjecture: If one calculates the
implied coefficient of variation based on Stigler’s figures, the coefficient of
variation for coal (which makes up a small percentage of the government’s
budget) is 14.7 percent, while that for an automobile (which makes up a
large percentage of a household’s budget) is 1.7 percent.
Pratt et al. (1979) observe a similar pattern in a cross-section of consumer
products sold in Boston in the 1970s. They obtain the following regression
result by regressing the sample (log) standard deviation of prices for a given
item on the sample (log) mean price for the same item:
ln s ¼ 1:517 þ 0:892 ln E½p. (14)
Straightforward manipulation of equation (14) reveals that a 1 percent
increase in the mean price of an item decreases the coefficient of variation
by 10.8 percent. Thus, the Pratt, Wise, and Zeckhauser data also suggest
that, empirically, the coefficient of variation is lower for more expensive
items than cheaper items. However, equation (14) also highlights that
the relationship depends crucially on the measure of price dispersion
used: if one were to use the standard deviation to measure price dispersion,
equation (14) implies that a 1 percent increase in the mean price of a
product leads to a 0.892 percent increase in dispersion, as measured by the
standard deviation.
A number of other authors have reported similar patterns in online and
offline markets, both in the US and in Europe for products ranging from
consumer sundries and electronic products to gasoline; cf. Marvel (1976),
Carlson and Pescatrice (1980), Clay and Tay (2001), Scholten and Smith
(2002), Johnson (2002), Gatti and Kattuman (2003), and Aalto-Setälä
(2003). More recently, Eckard (2004) compares price dispersion for staple
products in 1901 and 2001, and reports coefficients of variation in 2001 that
are almost twice those based on data from 1901. Eckard argues that one
reason for the increased dispersion is that his sample consists of staple items
(such as sugar and baking powder) that accounted for a much larger share
of household budgets in 1901 than in 2001.
Dispersion and Purchase Frequency. In his second hypothesis, Stigler ar-
gues that in markets where there are more repetitive or experienced buyers,
the greater is the amount of effective search. Unfortunately, it is difficult to
directly test this hypotheses, since in most markets there is no direct (ob-
jective) measure of ‘‘buyer experience’’ or ‘‘purchase frequency’’ to use in
examining its impact on levels of price dispersion. A number of the studies
mentioned above, however, provide casual evidence that purchase fre-
quency impacts the level of price dispersion (cf. Pratt et al. 1979; Carlson
and Pescatrice, 1980).
Sorensen (2000), however, has provided a very ‘‘clean’’ and elegant test of
Stigler’s second hypothesis. His analysis is based on data from the market
for prescription drugs. The unique aspect of this market is that purchase
frequency—the typical dosage and duration of therapy for a given prescrip-
tion drug—may be objectively measured. A consumers’ benefit per search is
clearly highest for frequently purchased drugs, and, Sorensen argues, this
should lead to greater search and lower price dispersion. His empirical
analysis identifies a strong inverse relationship between purchase frequency
and price dispersion. For example, after controlling for other factors (which
together explain about one-third of the variation in prices), Sorensen finds
that the price range for a drug that must be purchased monthly is about 30
percent lower than if it were a one-time therapy. Importantly, Sorensen
shows that the results are qualitatively similar when alternative measures of
price dispersion (such as the standard deviation) are used.
3.2.2 Dispersion and the ‘‘cost’’ of search

Researchers studying the empirical relationship between search costs and
price dispersion have faced obstacles similar to those of researchers focus-
ing on the benefit side of the search equation. First, the predicted impact of
search costs on levels of dispersion depends not only on the model, but also
on the metric used for measuring dispersion. Second, search costs are
generally unobservable. Some of the more influential papers in the area are
ones that have devised innovative methods of dealing with these problems.
One important example is Brown and Goolsbee (2002). Their starting
point is the Stahl (1989) model of equilibrium price dispersion, which as we
noted in Section 2, predicts that price dispersion is initially an increasing
function of the fraction of ‘‘shoppers’’ who enjoy zero search costs, but
after a threshold, is a decreasing function of the fraction of shoppers.
Brown and Goolsbee point out that the Stahl model closely matches the
market for term-life insurance during the 1992–1997 period. Consumers
who did not have an Internet connection arguably had to search sequenti-
ally to obtain price quotes from different insurance agents, while those with
Internet access could use web sites such as Quickquote.com to ‘‘costlessly’’

identify the company offering the lowest annual premium. In their data,
variation in the fraction of ‘‘shoppers’’ (those who research insurance on-
line) stems not only from the general rise in Internet penetration during the
1990s, but more importantly, from variation in the growth rates in Internet
usage across different groups of policyholders. Brown and Goolsbee regress
the standard deviation in residuals (obtained from a price regression that
controls for observable characteristics of people and policy types) on a
cubic function of their proxy for the fraction of ‘‘shoppers.’’ Consistent
with the prediction of the Stahl model, price dispersion initially rises as the
fraction of shoppers increases, but starts to decline once the fraction of
consumers researching insurance online exceeds about 5 percent.
A similar approach is implicit in a number of papers that have compared
levels of dispersion in online versus offline markets (cf. Brynjolfsson and
Smith, 2000; Carlton and Chevalier, 2001; Scholten and Smith, 2002; An-
carani and Shankar, 2004). The basic premise is that search costs are lower
in online (search entails clicks) versus offline (search entails travel costs)
markets.20 In general, since different search models make different predic-
tions about the impact of reductions in search costs on levels of price
dispersion, it is not too surprising that the findings of this literature are
decidedly mixed; for some products, dispersion is lower in online markets;
for other products, dispersion is actually higher online.21
Along these same lines, a number of studies compare average prices in
online versus offline markets. The idea is that search costs are lower online,
thus affecting not only the range or variance in prices, but also the mean
price (and hence the coefficient of variation through both the mean and
variance). Scott-Morton et al. (2001) find that prices are lower in online
markets for automobiles. Consumers who purchase a car through the In-
ternet referral service Autobytel.com reduce their purchase price by ap-
proximately 2.2 percent. A potentially confounding explanation for this
price difference is that the consumers who choose to shop online may also
be skilled ‘‘higglers,’’ to use Stigler’s phrase, and thus the price difference
might purely reflect a difference in the negotiating skills of consumers
across the two channels. Interestingly, Zettelmeyer et al. (2004) provide
evidence that this is not the case: consumers who purchase automobiles
online are not typically those who negotiate well in the traditional channel.
20
The view that online search is either more prevalent or cheaper than offline search is a matter of
some debate; see, for instance, Adamic and Huberman (2001) and Johnson et al. (2004). Bakos (1997)
was among the first to advance a theoretical argument that when the cost of price information is close to
zero, equilibrium price is close to marginal cost. More recently, however, Harrington (2001) has argued
that Bakos’ results are flawed. Finally, the Internet itself also offers opportunities for obfuscation (see
Ellison and Ellison, 2004) or unobserved lack of inventories (see Arnold and Saliba, 2002) that can raise
search and/or transactions costs relative to offline markets.
21
One may speculate that once shipping costs are accounted for, price dispersion online vanishes. This
is not the case; cf. Brynjolfsson and Smith (2000), Smith and Brynjolfson (2001), Pan et al. (2002),
Ancarani and Shankar (2004), Brynjolfsson et al. (2004), and Dinlersoz and Li (2005).
There are a number of other studies, however, that find equal or higher
prices online (cf. Bailey, 1998; Erevelles et al., 2001; Goolsbee, 2001; Clem-
ons et al., 2002; Clay, et al., 2003). Further studies distinguish price levels
depending on whether the retailer is a solely online or ‘‘multichannel’’ (cf.
Tang and Xing, 2001; Chevalier and Goolsbee, 2003).
An alternative approach is to ‘‘recover’’ search costs using structural
parameters from a particular model of price dispersion. For example, Hong
and Shum (forthcoming) obtain search cost estimates using restrictions
imposed by theoretical search models and assuming that observed price
dispersion is an equilibrium phenomenon arising from heterogeneous con-
sumer search costs. Their estimation technique is applied to online price
data on four economics and statistics textbooks. They obtain search cost
estimates ranging from $1.31 to $29.40 for these items. A similar approach
can be used in clearinghouse models. Villas-Boas (1995) uses the theoretical
density function implied by the Varian (1980) clearinghouse model to ob-
tain estimates of the number of shoppers in the offline coffee and saltine
cracker markets. More recently, Baye et al. (2005) used a theoretical clear-
inghouse model as the basis for estimating the fraction of ‘‘shoppers’’ in an
online market for PDAs in the UK. Their results suggest that about 13
percent of the consumers in this market are shoppers.
3.2.3 Dispersion and the number of sellers

The oligopoly models presented in Section 2 reveal that equilibrium dis-
tributions of prices, and hence levels of dispersion, vary with the number of
sellers competing in the market. The direction in which prices move as a
consequence of a change in the number of sellers is, however, model spe-
cific, as we saw in the Varian and Rosenthal models. Thus, examining the
relationship between price dispersion and the number of competing sellers
not only provides a test of whether informational factors play a role in
generating observed price dispersion, but also in making distinctions among
the various theory models.
For instance, Baye et al. (2004a) examine the theoretical and empiri-
cal relationship between the number of competitors and the level of price
dispersion in clearinghouse models. They show that the theoretical
relationship between number of competitors and the level of price disper-
sion in clearinghouse models is, in general, ambiguous, due to competing
‘‘order statistic’’ and ‘‘strategic’’ effects. Through a calibration displayed in
Fig. 2, they show that the impact of the number of sellers on price dis-
persion depends on the variant of the model. As the figure shows, in the
Varian model (where firms’ information transmission costs do not drive
price dispersion), the expected gap between the two lowest prices is initially
increasing in the number of sellers, and then declines. In contrast, in the
Baye and Morgan model (where firms’ information transmission costs are
the main driver of price dispersion), the expected gap is monotonically
decreasing in the number of firms. Based on online data from a popular
Fig. 2. Theoretical and empirical relationship between price dispersion measured by per-
centage gap and the number of competing sellers listing prices on a clearinghouse site.
price comparison site for consumer electronics products, and controlling for
other factors contributing to price dispersion, they find an inverse relation
between the gap and the number of online sellers. This relationship is de-
picted as the dotted ‘‘observed’’ line in Fig. 2. As the figure reveals, the
nonmonotonicity predicted by the Varian model, as well as the relatively
flat relationship between the gap and number of firms predicted in the
calibrated version of the Rosenthal model, is absent in the data. Specifi-
cally, in markets served by between two and four firms, the average gap (as
a percentage of the lowest price) is about 14 percent. The average percent-
age gap falls to about 3 percent in markets with five to ten firms, and is less
than 1 percent in markets with more than 10 firms.
More broadly, several empirical papers have suggested that the amount
of price dispersion observed in the market depends on various measures of
the numbers of competitors. Marvel (1976) reports that an increase in the
number of competitors (measured by ln(HHI)) reduces the range in the
price of gasoline. Barron et al. (2004) study the structural determinants of
price dispersion in the retail gasoline industry in four geographic locations,
and provide empirical evidence that, controlling for station-level charac-
teristics, an increase in station density decreases both price levels and
price dispersion.22 Borenstein and Rose (1994) investigate the relationship
22
See also Png and Reitman (1994).
between dispersion among airfares and the number of competitors or flight

density. They find that dispersion in fares increases on routes with lower
flight density or more competition. Thus, there is evidence that the number
of sellers matters for price dispersion.
3.2.4 Dispersion and price persistence

Varian (1980) was the first to distinguish between what he referred to as
‘‘spatial’’ and ‘‘temporal’’ price dispersion. Under spatial price dispersion,
different firms charge different prices at any point in time, but a firm’s
position in the distribution of prices does not change over time. Absent
random cost shocks, spatial price dispersion arises in the Reinganum,
MacMinn, and Spulber models. In contrast, with temporal price dispersion,
firms charge different prices at each point in time, but their position in
the distribution of prices changes over time. Temporal price dispersion
arises in the general clearinghouse model (and various special cases) as
well as in the Burdett and Judd model. Varian critiques models of spatial
price dispersion, arguing that if consumers can learn from experience
that some firms persistently offer lower prices than other firms, then models
of spatial price dispersion suggest a ‘‘convergence hypothesis’’: price dis-
persion should diminish over time due to the positive correlation in suc-
cessive prices (to use Stigler’s terminology) and cumulative search
information. This has led to a number of studies that examine whether
there is any evidence for the convergence hypothesis and whether the tem-
poral price dispersion predicted by the clearinghouse models is, in fact,
present in the data.
Using monthly store-level price data from Israel, and after controlling for
observed and unobserved product heterogeneities, Lach (2002) finds some
evidence of temporal price dispersion. Lach estimates month-to-month
transitions among quartiles by firms, that is, the probability that a firm
offering a price in a given quartile at the start of the month is still offering a
price in the same quartile at the end of the month. His estimates suggest
that the probability of remaining in the same quartile is 78 percent for firms
selling refrigerators and 71 percent for firms selling flour. These probabil-
ities are somewhat lower for firms selling chicken (51 percent) and coffee (43
percent). When the transition period is extended to six months instead of
one month, the probability of remaining in the same quartile is considerably
lower—falling to around 30–35 percent.
Roberts and Supina (2000) suggest that structural differences in firms’
costs account for a considerable portion of price dispersion in the
offline sector—as predicted by a variety of search-theoretic models. Using
plant-level US Census data, they find some evidence for price persistence.
The evidence is strongest in the tails of the distribution: high-price firms
tend to persistently charge high prices, and low-price firms tends to per-
sistently charge low prices. A variety of other studies also suggest that
heterogeneities either across firms or across markets impact price dispersion
in online markets (cf. Smith et al., 1999; Clay et al., 2001; Smith and
Brynjolfson, 2001; Chen and Hitt, 2002; Resnick and Zeckhauser, 2002;
Brynjolfsson et al., 2004). In all cases, however, even after controlling for
various heterogeneities, economically significant levels of price dispersion
remain.
There is also evidence that online prices exhibit temporal price dispersion.
For instance, Baye et al. (2004b) examine turnover of the identity of the
low- and high-price firms using a dataset consisting of 36 popular consumer
electronics products sold over a 19-month period. They find considerable
evidence for month-to-month changes in the identity of the low-price firms,
but some evidence of persistence in the identity of high-priced firms. Sim-
ilarly, Iyer and Pazgal (2003) collect biweekly price data on music CDs,
movie videos, and books from five price comparison sites: MySimon,
BottomDollar, EvenBetter, Bsilly, and Pricescan during the period April–
October 2000 and find empirical results suggesting that no single firm con-
sistently charges the low price.
Finally, Baye et al. (2004a) examine the convergence hypothesis of price
dispersion using a dataset consisting of over four million daily price ob-
servations for over 1,000 consumer electronics products sold on a popular
Internet price comparison site over an eight-month period. Even allowing
for a nonlinear relationship between observed price dispersion and time,
they find no evidence for the convergence hypothesis in this market—the
level of price dispersion remained stable over the period.
3.3 Concluding remarks: empirics
We conclude with four simple observations.
1. As is evident from the studies highlighted in Table 1, price dispersion

is ubiquitous and persistent. Regardless of the particular product
(tinplate cans or PDAs), the venue in which they are sold (online or
offline, in the US or abroad), or the time period (1901 or 2005), the
inescapable conclusion from the empirical literature is a validation of
Stigler’s and Varian’s initial observations: Information remains a val-
uable resource, and the law of one price is still no law at all.
2. Theory is useful for understanding dispersion data, and dispersion
data is useful for discriminating among alternative theoretical models.
3. The relationship between price dispersion and economic primitives is
often sensitive to the measure of price dispersion used.
4. Despite the widespread adoption of inventions such as the automobile,
the telephone, television, and the Internet, price dispersion is still the
rule rather than the exception in homogeneous product markets. Re-
ductions in information costs over the past century have neither re-
duced nor eliminated the levels of price dispersion observed in
homogeneous product markets.
Acknowledgments
We owe a special debt to Michael Rauh and Felix Várdy for providing us
with detailed hand-written comments on earlier drafts. We also thank Ville
Aalto-Setälä, Fabio Ancarani, Maria Arbatskaya, Judy Chevalier, Karen
Clay, Woody Eckard, Sara Fisher Ellison, Xianjun Geng, Rupert Gatti,
Jose Moraga Gonzalez, Joe Harrington, Terry Hendershott, Ganesh Iyer,
Maarten Janssen, Ronald Johnson, Ken Judd, Ramayya Krishnan, Sol
Lach, Rajiv Lal, Preston McAfee, Xing Pan, Jeff Perloff, Ivan Png, Ram
Rao, Jennifer Reinganum, Nancy Rose, Venky Shankar, Jorge Silva-Risso,
Michael Smith, Alan Sorensen, Dan Spulber, Mark Stegeman, Beck Taylor,
Miguel Villas-Boas, Xiaolin Xing, and Richard Zeckhauser for encourage-
ment, helpful comments, and suggestions. Of course, we are responsible for
any shortcomings that remain in this offering.
References
Aalto-Setälä, V. (2003). Explaining price dispersion for homogeneous grocery products. Journal of
Agricultural and Food Industrial Organization 1, Article 9, 1–14.
Adamic, L.A., B.A. Huberman (2001). The web’s hidden order. Communications of the ACM 44(9),
55–59.
Ancarani, F., V. Shankar (2004). Price levels and price dispersion within and across multiple
retailer types: further evidence and extension. Journal of the Academy of Marketing Science 32(2),
176–187.
Arbatskaya, M., M.R. Baye (2004). Are prices ‘sticky’ online? Market structure effects and asymmetric
responses to cost shocks in online mortgage markets. International Journal of Industrial Organization
22(10), 1443–1462.
Arnold, M.A., C. Saliba (2002). Price dispersion in online markets: the case of college textbooks.
Working Paper.
Axell, B. (1977). Search market equilibrium. Scandinavian Journal of Economics 79(1), 20–40.
Bailey, J. (1998). Electronic commerce: prices and consumer issues for three products: books, compact
discs, and software. Organization Economics Co-Operation Development 98(4), 3–22.
Bakos, Y. (1997). Reducing buyer search costs: implications for electronic marketplaces. Management
Science 43(2), 1676–1692.
Barron, J.M., B.A. Taylor, J.R. Umbeck (2004). Number of sellers, average prices, and price dispersion.
International Journal of Industrial Organization 22, 1041–1066.
Baye, M.R., R. Gatti, P. Kattuman, J. Morgan, (2005). Estimating firm-level demand at a price com-
parison site: accounting for shoppers and the number of competitors. Working Paper.
Baye, M.R., D. Kovenock (1994). How to sell a pickup truck: beat or pay advertisements as facilitating
devices. International Journal of Industrial Organization 12(1), 21–33.
Baye, M.R., D. Kovenock, C. de Vries (1992). It takes two to tango: equilibria in a model of sales.
Games and Economic Behavior 4, 493–510.
Baye, M.R., J. Morgan (2001). Information gatekeepers on the internet and the competitiveness of
homogeneous product markets. American Economic Review 91(3), 454–474.
Baye, M.R., J. Morgan (2004). Price dispersion in the lab and on the internet: theory and evidence. Rand
Journal of Economics 35(3), 449–466.
Baye, M.R., J. Morgan, P. Scholten (2003). The value of information in an online consumer electronics
market. Journal of Public Policy and Marketing 22(1), 17–25.
Baye, M.R., J. Morgan, P. Scholten (2004a). Price dispersion in the small and in the large: evidence from
an internet price comparison site. Journal of Industrial Economics 52(4), 463–496.
Baye, M.R., J. Morgan, P. Scholten (2004b). Temporal price dispersion: evidence from an online
consumer electronics market. Journal of Interactive Marketing 18(4), 101–115.
Baylis, K., J.M. Perloff (2002). Price dispersion on the internet: good firms and bad firms. Review of
Industrial Organization 21, 305–324.
Benabou, R., R. Gertner (1993). Search with learning from prices: does increased uncertainty lead to
higher markups?. Review of Economic Studies 60(202), 69–94.
Borenstein, S., N.L. Rose (1994). Competition and price dispersion in the U.S. airline industry. Journal
of Political Economy 102(4), 653–683.
Braverman, A. (1980). Consumer search and alternative market equilibria. Review of Economic Studies
47, 487–502.
Brown, J.R., A. Goolsbee (2002). Does the internet make markets more competitive? Evidence from the
life insurance industry. Journal of Political Economy 110(3), 481–507.
Brynjolfsson, E., A.A. Dick, M.D. Smith, (2004). Search and product differentiation at an internet
shopbot. Working Paper.
Brynjolfsson, E., M.D. Smith (2000). Frictionless commerce? A comparison of internet and conven-
tional retailers. Management Science 46(4), 563–585.
Burdett, K., K.L. Judd (1983). Equilibrium price dispersion. Econometrica 51, 955–969.
Butters, G.R. (1977). Equilibrium distributions of sales and advertising prices. Review of Economic
Studies 44, 465–491.
Carlson, J.A., R.P. McAfee (1983). Discrete equilibrium price dispersion. Journal of Political Economy
91(3), 480–493.
Carlson, J.A., D.R. Pescatrice (1980). Persistent price distributions. Journal of Economics and Business
33, 21–27.
Carlton, D., J. Chevalier (2001). Free riding and sales strategies for the internet. Journal of Industrial
Economics 49(4), 441–461.
Chen, S., L.M. Hitt (2002). Measuring switching costs and the determinants of customer retention in
internet-enabled business: a study of the online brokerage industry. Information Systems Research
13(3), 255–274.
Chevalier, J., A. Goolsbee (2003). Measuring prices and price competition online: Amazon.com and
BarnesandNoble.com. Quantitative Marketing and Economics 1, 203–222.
Clay, K., R. Krishnan, E. Wolff (2001). Prices and price dispersion on the web: evidence from the online
book industry. Journal of Industrial Economics 49, 521–539.
Clay, K., R. Krishnan, E. Wolff, D. Fernandes (2003). Retail strategies on the web: price and non-price
competition in the online book industry. Journal of Industrial Economics 50(3), 351–367.
Clay, K., C.H. Tay (2001). The cross-country price differentials in the online textbook market. Working
Paper.
Clemons, E.I., L. Hahn, M. Hitt (2002). Price dispersion and differentiation in online travel: an em-
pirical investigation. Management Science 48, 534–549.
Dana, J.D. (1994). Learning in an equilibrium search model. International Economic Review 35, 745–771.
Daughety, A. (1992). A model of search and shopping by homogeneous consumers without price
precommitment by firms. Journal of Economics and Management Strategy 1(3), 455–473.
Diamond, P. (1971). A model of price adjustment. Journal of Economic Theory 3, 156–168.
Dinlersoz, E.M., H. Li, (2005). The shipping strategies of internet retailers: evidence from internet book
retailing. Working Paper.
Eckard, E.W. (2004). The ‘Law of One Price’ in 1901. Economic Inquiry 42(1), 101–110.
Ellison, G., (2005). Bounded rationality in industrial organization. MIT Working Paper.
Ellison, G., S. Ellison, (2004). Search, obfuscation, and price elasticities on the internet. MIT Working
Paper.
Erevelles, S., E. Rolland, S. Srinivasan. (2001). Are prices really lower on the internet?: an analysis of the
vitamin industry. Working Paper.
Friberg, R., M. Ganslandt, M. Sandström (2001). Pricing strategies in e-commerce: bricks vs. clicks.
Working Paper.
Gatti, R., P. Kattuman (2003). Online price dispersion within and between seven European countries.
Advances in Applied Microeconomics 12, 107–143.
Goolsbee, A. (2001). Competition in the computer industry: online versus retail. Journal of Industrial
Economics 49(4), 498–499.
Grossman, G.M., C. Shapiro (1984). Informative advertising with differentiated products. Review of
Harrington, J.E. (2001). Comment on ‘reducing buyer search costs: implications for electronic mar-
ketplaces’. Management Science 47(12), 1727–1732.
Hong, H., M. Shum (2006). Using price distributions to estimate search costs. Rand Journal of Eco-
nomics 37(2), 258–276.
Iyer, G., A. Pazgal (2003). Internet shopping agents: virtual colocation and competition. Marketing
Science 22(1), 85–106.
Janssen, M., J.L. Moraga-González (2004). Strategic pricing, consumer search, and the number of firms.
Review of Economic Studies 71(4), 1089–1118.
Janssen, M., J.L. Moraga-González, M. Wildenbeest (2005). Truly costly sequential search and
oligopolistic pricing. International Journal of Industrial Organization 23, 451–466.
Johnson, E.J., W. Moe, P. Fader, S. Bellman, J. Lohse (2004). On the depth and dynamics of world wide
web shopping behavior. Management Science 50(3), 299–308.
Johnson, R.N. (2002). Search costs, lags and prices at the pump. Review of Industrial Organization 20,
33–50.
Klemperer, P. (1999). Auction theory: a guide to the literature. Journal of Economic Surveys 13,
227–286.
Lach, S. (2002). Existence and persistence of price dispersion: an empirical analysis. Review of Economics
and Statistics 84(3), 433–444.
Lal, R., M. Sarvary (1999). When and how is the internet likely to decrease price competition. Marketing
Science 18(4), 485–503.
Lal, R., M. Villas-Boas (1998). Price promotions and trade deals with multiproduct retailers. Man-
agement Science 44(7), 935–949.
Lopez-Acevedo, G. (1997). Quantal response equilibria for posted offer markets. Estudios Económicos
12, 95–131.
MacMinn, R.D. (1980). Search and market equilibrium. Journal of Political Economy 88(2), 308–327.
Marvel, H.P. (1976). The economics of information and retail gasoline price behavior: an empirical
analysis. Journal of Political Economy 84, 1033–1080.
McAfee, R.P. (1994). Endogenous availability, cartels, and merger in an equilibrium price dispersion.
Journal of Economic Theory 62, 24–47.
McAfee, R.P. (1995). Multiproduct equilibrium price dispersion. Journal of Economic Theory 67, 83–105.
McAfee, R.P., J. McMillan (1988). ’’Search mechanisms’’. Journal of Economic Theory 44, 99–123.
McKelvey, R., T. Palfrey (1995). Quantal response equilibria for normal form games. Games and
Economic Behavior 10, 6–38.
Morgan, J., H. Orzen, M. Sefton (2006). An experimental study of price dispersion. Games and Eco-
nomic Behavior 54(1), 134–158.
Morgan, P., R. Manning (1985). Optimal search. Econometrica 53(4), 923–944.
Myerson, R. (1981). Optimal Auctions. Mathematics of Operations Research 6, 58–63.
Narasimhan, C. (1988). Competitive Promotional Strategies. Journal of Business 61(4), 427–449.
Pan, X., B.T. Ratchford, V. Shankar (2002). Can price dispersion in online markets by explained
by differences in e-tailer service quality?. Journal of the Academy of Marketing Science 30(4),
433–445.
Pan, X., V. Shankar, B. Ratchford (2003). The evolution of price dispersion in internet retail markets.
Advances in Applied Microeconomics 12, 85–105.
Png, I.P.L., D. Hirshleifer (1987). Price discrimination through offers to match price. Journal of Business
60(3), 365–384.
Png, I.P.L., D. Reitman (1994). Service time competition. Rand Journal of Economics 25(4), 619–634.
Pratt, J.W., D.A. Wise, R. Zeckhauser (1979). Price differences in almost competitive markets. Quarterly
Journal of Economics 93(2), 189–211.
Radner, R. (1980). Collusive behavior in oligopolies with long but finite lives. Journal of Economic
Theory 22, 136–156.
Raju, J.S., V. Srinivasan, R. Lal (1990). The effects of brand loyalty on competitive price promotional
strategies. Management Science 36(3), 376–404.
Rao, R.C., R.V. Arjunji, B.P.S. Murthi (1995). Game theory and empirical generalizations concerning
competitive promotions. Marketing Science 14(3), G89–G100.
Rauh, M.T. (1997). A model of temporary search market equilibrium. Journal of Economic Theory
77(1), 128–153.
Rauh, M.T. (2001). Heterogeneous beliefs, price dispersion, and welfare improving price controls.
Economic Theory 18(1), 577–603.
Reinganum, J.F. (1979). A simple model of equilibrium price dispersion. Journal of Political Economy
87, 851–858.
Resnick, P., R. Zeckhauser (2002). Trust among strangers in internet transactions: empirical analysis of
eBay’s reputation system. Advances in Applied Microeconomics 11, 127–158.
Rob, R. (1985). Equilibrium price distributions. Review of Economic Studies 52(3), 487–504.
Robert, J., D.O. Stahl (1993). Informative price advertising in a sequential search model. Econometrica
61, 657–686.
Roberts, M.J., D. Supina (2000). Output price and markup dispersion in micro data: the roles of
producer heterogeneity and noise. Advances in Applied Microeconomics 9, 1–36.
Rosenthal, R.W. (1980). A model in which an increase in the number of sellers leads to a higher price.
Econometrica 48(6), 1575–1580.
Rothschild, M. (1973). Models of market organization with imperfect information: a survey. Journal of
Rothschild, M. (1974). Searching for the lowest price when the distribution of prices is unknown.
Journal of Political Economy 82(4), 689–711.
Salop, S.C., J.E. Stiglitz (1977). Bargains and ripoffs: a model of monopolistically competitive price
dispersion. Review of Economic Studies 44, 493–510.
Scholten, P., A. Smith (2002). Price dispersion then and now: evidence from retail and e-tail markets. in:
M.R. Baye (ed.), The Economics of the Internet and E-commerce. Advances in Applied Microeco-
nomics, 11, 63–88.
Scott-Morton, F., F. Zettelmeyer, J. Silva-Risso (2001). Internet car retailing. Journal of Industrial
Economics 49(4), 501–519.
Shilony, Y. (1977). Mixed pricing in oligopoly. Journal of Economic Theory 14, 373–388.
Smith, M.D., J. Bailey, E. Brynjolfsson (1999). Understanding digital markets: reviews and assessments,
in: E. Brynjolfsson, B. Kahin (eds.), Understanding the Digital Economy, MIT Press, Cambridge,
MA.
Smith, M.D., E. Brynjolfsson (2001). Consumer decision-making at an internet shopbot: brand still
matters. Journal of Industrial Economics 49(4), 541–558.
Sorensen, A. (2000). Equilibrium price dispersion in retail markets for prescription drugs. Journal of
Spulber, D. (1995). Bertrand competition when rivals’ costs are unknown. Journal of Industrial Eco-
nomics 43(1), 1–11.
Stahl, D.O. (1989). Oligopolistic pricing with sequential consumer search. American Economic Review
79, 700–712.
Stahl, D.O. (1994). Oligopolistic pricing and advertising. Journal of Economic Theory 64, 162–177.
Stahl, D.O. (1996). Oligopolistic pricing with heterogeneous consumer search. International Journal of
Industrial Organization 14, 243–268.
Stegeman, M. (1991). Advertising in competitive markets. American Economic Review 81, 210–223.
Stigler, G. (1961). The economics of information. Journal of Political Economy 69(3), 213–225.
Tang, F., X. Xing (2001). Will the growth of multi-channel retailing diminish the pricing efficiency of the
web?. Journal of Retailing 77, 319–333.
Varian, H.R. (1980). A model of sales. American Economic Review 70, 651–659.
Villas-Boas, M. (1995). Models of competitive price promotions: some empirical evidence from the
coffee and saltine crackers markets. Journal of Economics and Management Strategy 4(1), 85–107.
Zettelmeyer, F., F. Scott-Morton, J. Silva-Risso (2004). Cowboys or cowards: why are internet car
prices lower? Working Paper.
Chapter 7
Behavior-Based Price Discrimination and Customer

Recognition$
Drew Fudenberg
Harvard University
J. Miguel Villas-Boas
University of California, Berkeley
Abstract
When firms are able to recognize their previous customers, they may be able
to use their information about the consumers’ past purchases to offer differ-
ent prices and/or products to consumers with different purchase histories.
This article surveys the literature on this ‘‘behavior-based price discrimina-
tion.’’
1 Introduction
When firms have information about consumers’ previous purchases, they

may be able to use this information to offer different prices and/or products
to consumers with different purchase histories. This sort of ‘‘behavior-
based price discrimination’’ (BBPD) and use of ‘‘customer recognition’’
occurs in several markets, such as long-distance telecommunications, mo-
bile telephone service, magazine or newspaper subscriptions, banking serv-
ices, credit cards, and labor markets; it may become increasingly prevalent
with improvements in information technologies and the spread of e-com-
merce and digital rights management.
This article focuses on models of ‘‘pure’’ BBPD, in which past pur-
chases matter only for their information value, and do not directly alter
$
NSF grant SES-04-26199 provided financial support for some of this work.
377
378 D. Fudenberg and J. M. Villas-Boas
consumers’ preferences. We do make some comparisons with switching-cost

models, where past purchases do have a direct effect, but we say almost
nothing about traditional models of third-degree price discrimination,
where firms can base their prices on observable and exogenous character-
istics of the consumers.
One recurrent theme throughout the article is the possibility that firms
may face a commitment problem: although having more information helps
the firm extract more surplus with its current prices, consumers may an-
ticipate this possibility, and so alter their initial purchases. Thus, as in the
related literatures on bargaining, durable-good monopoly, and dynamic
mechanism design,1 the seller may be better off if it can commit to ignore
information about the buyer’s past decisions. A second theme is that, as in
traditional models of third-degree price discrimination,2 more information
may lead to more intense competition between firms. Thus, even if each firm
would gain by being the only one to practice BBPD, industry profits can fall
when all of the firms practice it. Third, and related, firms would often gain
from using long-term contracts when they are able to do so as, for example,
in the market for cell-phone services. The last implication is somewhat
unfortunate from the analyst’s perspective: The welfare implications of
BBPD seem to be ambiguous, and to depend on many aspects of the market
structure.
Section 2 examines a monopoly supplier of a single, non-durable good.
We start with a simple two-period model, and then consider the infinite-
horizon models of Hart and Tirole (1988) and Schmidt (1993), where all
consumers are infinitely lived, and Villas-Boas (2004), where there are
overlapping generations of consumers who live only two periods. We com-
pare this situation with that of a durable-goods monopolist. Then we con-
sider the use of long-term contracts, and relate the resulting outcome again
to that in models of durable-goods monopoly. We also discuss the case
where the consumer’s preferences vary over time, as in Kennan (2001) and
Battaglini (2005). Finally, we consider the situation where the monopolist
sells more than one good, which we use as a benchmark when studying
BBPD with multiple firms; we also compare this with a monopolist seller of
two goods in a model of switching costs.
Section 3 studies BBPD with two firms, each still selling a single good. In
these models, firms can try to ‘‘poach’’ their rivals’ customers by giving new
customers special ‘‘introductory’’ prices. We begin with Fudenberg and
Tirole’s (2000) analysis of a two-period model of competition in short-term
contracts, and its extension by Chen and Zhang (2004) and Esteves (2004)
to other distributions of consumer types, where other insights emerge. Next,
we discuss Villas-Boas’ (1999) model of poaching in an infinite-horizon
1
See, for example, Stokey (1981), Bulow (1982), Fudenberg and Tirole (1983), Baron and Besanko
(1984), Freixas et al. (1985), and Sobel (1991).
2
See, e.g., Thisse and Vives (1988).
Ch. 7. Behavior-Based Price Discrimination and Customer Recognition 379
model with overlapping generations of consumers, each of whom lives only

for two periods, where firms cannot distinguish between new consumers
and old ones who bought from their rival. We then return to the two-period
setting to present Fudenberg and Tirole’s analysis of competition in simple
long-term contracts, meaning that firms offer both a ‘‘spot’’ or one-period
price and also a long-term commitment to supply the good in both periods.
Finally, we compare the predictions of these models to models of switching
costs, where past decisions are directly payoff relevant, and may also pro-
vide information, as in Chen (1997) and Taylor (2003).
Section 4 discusses models where each firm can produce multiple versions
of the same product. We begin with Fudenberg and Tirole (1998), and
Ellison and Fudenberg (2000), who study the provision of ‘‘upgrades’’ by a
monopolist in a setting of vertical differentiation, where all customers agree
that one good is better than the other. We then consider the work of Zhang
(2005) on endogenous product lines in a Hotelling-style duopoly model of
horizontal differentiation. Finally, we discuss the papers of Levinthal and
Purohit (1989), Waldman (1996), and Nahm (2004) on the introduction of a
new product in models with anonymous consumers and a frictionless sec-
ond-hand market. Although these papers do not consider behavior-based
pricing, the analysis of the anonymous case is an important benchmark for
the effects of behavior-based pricing.
Section 5 briefly discusses three related topics: privacy, credit markets,
and customized pricing. We discuss the work of Taylor (2004a) and Ca-
lzolari and Pavan (2005) on consumer privacy. If consumers are not my-
opic, they will realize that information revelation can reduce their future
surplus; in some cases, this can give firms an incentive to try to protect
consumer privacy. In credit markets, lenders may learn about the ability of
their borrowers, their customers, to repay loans; this information can then
be used by the firms in future loans to those customers. In this case what a
firm learns about its previous customers relates to the cost of providing the
customer with a given contract, as opposed to the customer’s willingness to
pay, which is the focus of most of the work we discuss. Our presentation
here is based in large part on Dell’Ariccia et al. (1999), and Dell’Ariccia and
Marquez (2004); we also discuss Pagano and Jappelli (1993), and Padilla
and Pagano (1997, 2000). Finally, for completeness, we briefly present the
case of competition when firms already have information about consumer
tastes, starting from the initial work of Thisse and Vives (1988). Section 6
presents concluding remarks.
2 Monopoly
We begin with the case of a monopolist who can base prices to its con-
sumers on their past purchase history. For example, in newspaper or mag-
azine subscriptions, firms with market power may offer different rates
depending on the consumers’ past purchase behavior.3 We start by con-

sidering a two-period model to illustrate some of the effects that can be
present, discussing the role of commitment, and of forward-looking con-
sumers. Then, we consider the case of overlapping generations of consum-
ers (Villas-Boas, 2004), and discuss the case when consumers are long lived
(Hart and Tirole, 1988; Schmidt, 1993). We consider the effect of long-term
contracts and the relationship to the durable-goods and bargaining liter-
ature in that setting (Hart and Tirole). We also discuss the case where the
consumer’s preferences vary over time, as in Kennan (2001) and Battaglini
(2005), who study short-term and long-term contracts, respectively. Finally,
we consider the situation where the monopolist sells more than one good, as
in Section 5 of Fudenberg and Tirole (2000), which will be an important
benchmark case for the next section on competition, and discuss the differ-
ences between purely informational behavior-based price discrimination
and price discrimination when previous purchases have a direct impact on
consumer preferences as in models of switching costs.
2.1 Two-period model
Base model
Consider a monopolist that produces a non-durable good at zero-mar-
ginal cost in each of two periods. A continuum of consumers with mass
normalized to one is in the market in each of the two periods. In each
period each consumer can use one unit or zero units of the good; no con-
sumer has any additional gain from using more than one unit in each
period. The consumer preferences are fixed across the two periods. The
consumers’ valuation for the good is represented by a parameter y distrib-
uted in the line segment [0,1] with cumulative distribution function F(y) and
density f(y). We assume throughout that p[1F( p)] is strictly quasi-concave
in p (which is the condition necessary for the existence of a unique local
maximum in the static monopoly case). The assumption on the support of
the distribution is without loss of generality relative to any compact inter-
val. Hart and Tirole (1988) and Villas-Boas (2004) consider the case of the
two-point distribution. Schmidt (1993) considers the case of any discrete
number of types.4 Here, we present the case of a continuum of consumer
types, and note differences with the two-type case when they arise. In order
to obtain some of the sharper results we will sometimes restrict attention to
the uniform distribution, with f(y) ¼ 1,8 y.
3
See, for example, ‘‘Publications are Trying New Techniques to Win over Loyal Readers,’’ The New
York Times, January 4, 1999, p. C20.
4
We restrict attention to the case in which the consumers are the only parties with private information.
It would also be interesting to investigate what happens when the monopolist has also some private
information, and the consumers may learn what price offers they will get in the future from the offers
made by the firm in the past. From the literature on ‘‘reputation effects’’ we expect that this could allow
the firm to obtain higher profits.
Each consumer is endowed with the same y in both periods. This val-
uation y represents the gross utility the consumer enjoys from using the
good in one period. Therefore, the net utility per period of a consumer of
type y purchasing the good at price p in one period is yp. The lifetime
utility of a consumer is the discounted sum of the net utilities of the two
periods the consumer is in the market with discount factor dC with
0rdCo1. In the first period, the monopolist chooses one price a to be
charged to all consumers (the monopolist cannot distinguish among them,
and all consumers prefer a lower price). In the second period, the mon-
opolist chooses two prices: a price ap to be charged to the previous cus-
tomers of the firm, and a price an to be charged to the consumers that did
not buy in the first period, the new customers.
The monopolist wants to maximize the expected discounted value of its
profits, using a discount factor dF with 0rdFo1. Except where expressly
noted we restrict attention to the case in which dF ¼ dC, and then, the
discount factor is denoted by d.
Given that there is a continuum of consumers, each of them realizes that
his decision does not affect the prices charged by the monopolist in the next
period. Then a consumer of type y just entering the market decides to buy in
the first period if y a þ dC max½y ap ; 0 dC max½y an ; 0: From this
inequality one can then obtain directly that given dCo1, if a type y^ chooses
to buy in the first period then all the types y4y^ also choose to buy in the
first period. That is, the consumers that buy for the first time in the second
period value the product by less than any of the consumers that buy in the
first period.
In order to compute the type of the marginal consumer it is helpful to
consider the pricing decision of the monopolist with respect to its previous
customers. Define p arg maxp p½1 FðpÞ; the price that maximizes the
profit in one period when the consumers do not have any reason to refrain
from buying, that is, they buy if their valuation y is greater than the price
charged. This is the monopoly price in the static case, or if the monopolist is
not able to recognize its previous customers or price differently to them.
Denoting y^ as the type of the marginal consumer in the first period, if
^y4p the monopolist sets ap ¼ y: ^ If, on the other hand yop^
; the mon-
^
opolist sets ap ¼ p : That is, ap ¼ max½y; p ; the marginal consumer in the
first period does not get any surplus in second period. This is the ‘‘ratchet
effect’’ of the consumers being hurt (i.e., being charged a higher price) by
revealing, even if partially, their types (Freixas et al., 1985).
The marginal consumer in the first period is then determined by
y^ a ¼ dC max½y^ an ; 0, (1)

which results in
y^ ¼ a if a an
a dC an
y^ ¼ a if a4an .
1 dC
This expression for y^ shows an important aspect of the market dynamics.
If prices are expected to increase, each consumer does not have any reason
to behave strategically and buys if his valuation is above the current price.
If, on the other hand, prices are expected to decrease, some consumers will
behave strategically, not being identified in the first period, and being able
to get a better deal in the second period.
No customer recognition
Consider first as a benchmark the case of no customer recognition, in
which the monopolist cannot price discriminate in the second period be-
tween the consumers that bought, and did not buy, in the first period. The
optimal price charged in each period is then p arg max p½1 FðpÞ;
generating a profit in each period equal to p ½1 F ðp Þ: Note that, obvi-
ously, there is no price variation through time. For the uniform distribution
example we have p ¼ 1=2; a profit per period of 1/4, and a total profit of
(1+d)/4.
Customer recognition and behavior-based price discrimination

Consider now the case in which the monopolist is able to recognize the
previous customers, as in Hart and Tirole (1988), Schmidt (1993), and
Villas-Boas (2004).5 For example, an internet store may be able to recognize
returning customers through cookies installed in their computer, and charge
them different prices. In this setting, the monopolist can identify in the
second period two different groups of consumers: those who have bought in
the first period, and those who have not bought in the first period. In the
second period the monopolist can charge two different prices. The price
paid by the monopolist’s previous consumers, ap, and the price paid by
consumers who have not bought previously, an.
Given that the marginal consumer buying the product in the first period is
y; ^ ¼ maxðp ; yÞ
^ the optimal prices in the second period are a ðyÞ ^ and
p
^ ¼ arg maxa an ½F ðyÞ

an ðyÞ ^ F ðan Þ.
n
The marginal consumer in the first period, y^ ¼ yðaÞ; ^ is determined by

^
^yðaÞ ¼ ða dC a ðyðaÞÞÞ=ð1 dC Þ: In order to obtain the optimal first period
n
5
See also Acquisti and Varian (2005) for results focusing on the role of commitment (see below) and
the effect of being able to offer enhanced services. The possibility of enhanced services is also covered in
Section 4.
price, a ; the monopolist then maximizes

8 9
^
< maxap ap ½1 F ðmax½ap ; yðaÞÞ =
^
maxa a½1 F ðyðaÞÞ þ dF (2)
^
: þa ðyðaÞÞ½F ^
ðyðaÞÞ ^
F ða ðyðaÞÞÞ ;
n n
where the first term represent the first-period profit, and the second term
represents the second-period profit, both from the consumers who bought
in the first period and from the consumers who did not buy in the first
^
period. Under the assumption that y4p
; which is satisfied in equilibrium,
the first-order condition that defines the optimal a is then
1 F ðyÞ ^ y^ 0 þ dF y^ 0 ½1 F ðyÞ
^ a f ðyÞ ^ f ðyÞ
^ y^ þ f ðyÞa
^ ðyÞ
^ ¼ 0. (3)
n
Note that for dC ¼ dF ¼ d the marginal consumer buying the product in
the first period has a higher valuation than if there were no customer rec-
ognition. To see this note that the first derivative of the objective 0
function
above evaluated at yðaÞ ^ ¼ p is equal to f ðp Þp ½1 ð1 dÞy^ after substi-
tuting
0
for 1 F ðp 0Þ p f ðp0 Þ ¼ 0 and p ð1 dÞ ¼ a dan ðp Þ: Given that
y^ ¼ 1=ð1 d þ dan Þ and an 40; that derivative is positive, which means
that the monopolist should increase a, which implies a higher valuation of
the marginal consumer than p : One can also obtain for dC ¼ dF that the
present value of profits is y½1 ^ F ðyÞ ^ þ da ½1 Fða Þ; which is strictly be-
n n
low the present value of profits under no customer recognition, as p
uniquely maximizes p[1F(p)]. The intuition of this result is that the mar-
ginal consumers refrain from buying in their first period in the market
because they know that they can get a lower price in the next period. This
result of lower profits with customer recognition does not hold if the con-
sumers are myopic, while the monopolist is forward looking (or dF large as
compared to dC).
For the uniform distribution example one can obtain an ðyÞ ^ ¼ y=2;
^ ^
yðaÞ ¼
2
2a=ð2 dÞ; and a ¼ ð4 d Þ=ð8 þ 2dÞ: One can also easily check that, as
argued above, the present value of profits is lower than in the no customer
recognition case for all d. One can also get that 2/(4+d) consumers buy in
both periods, while (2+d)/(8+2d) consumers only buy in the second pe-
riod. As consumers become more strategic (greater d) the number of con-
sumers buying in both periods decreases, as the consumers wait for future
deals, and consequently, the number of consumers that only buy in the
second period increases.
The main ideas from these results can also be obtained with a two-type
distribution as presented in the references listed above.
The role of commitment

A crucial feature in the previous section is that the monopolist could not
commit in the first period to its second-period price. This led the consumers
to refrain from buying in the first period, because the marginal consumers
knew that if they bought in the first period they would get zero surplus in
the second period. One could then wonder what would happen if the mono-
polist were able to commit in the first period to its second-period prices. For
example, in the market for cellular phone services firms are sometimes able
to commit to prices for some future periods. In this case one can then apply
the revelation principle, giving incentives for consumer types to reveal
themselves in the first period. That is, we suppose that consumers announce
their valuations in the first period, and are then assigned a price and a
consumption plan for the two periods, such that consumers announce their
valuation truthfully. Without commitment, the firm could change the utility
(or consumption) a consumer gets in the second period given what the firm
learns in the first period.
In a durable-good context, Stokey (1979) shows that when firms can
commit to the time path of prices, and dC ¼ dF, the monopolist commits to
having the same price in all periods, which ends up being the static mono-
poly price. Hart and Tirole (1988) show that the same conclusion applies
when the firm can engage in behavior-based price discrimination: the op-
timal policy is to forgo the ability to price discriminate and simply charge
the static monopoly price in every period.6 Villas-Boas (2004) shows that
the result also applies when there are overlapping generations of consumers.
To see this in the model presented here, note that if the monopolist is able
to commit to the second-period prices for the consumers who bought in the
first period, ap, and who did not buy in the first period, an, the most that it
can get is ap ½1 F ðap Þ þ dan ½1 F ðan Þ which is maximized when ap ¼
an ¼ p ; with a first-period price a ¼ p ; no price discrimination. Note also
that commitment allows the monopolist to be better off.
Note that when the monopolist is more forward-looking than the con-
sumers, dF>dC, the firm may then choose to price discriminate, cutting
prices through time.
2.2 Overlapping generations of consumers
The two-period model above is able to highlight some of the effects under
customer recognition and behavior-based price discrimination, but since it
focuses on the effects of the beginning of the market (in the first period) and
the end of the market (in the second period), it potentially may not get at
some of the effects in an on-going market.
Consider then a market where there is an infinitely lived monopolist
facing overlapping generations of consumers as in the previous section
(Villas-Boas, 2004). Each generation lives for two periods, and in each
6
Acquisti and Varian (2005) derive the same result. The result can also be seen as the same as in Baron
and Besanko (1984) who show that in a dynamic single-principal single-agent relationship with constant
types over time the optimal long-term contract under full-commitment consists in a sequence of static-
optimal contracts.
period there are two generations of consumers in the market (each of mass
one for a total mass in the market of two in each period), one in its first
period in the market, the other in its second and final period in the market.
Assume further that 1 F ðpÞ 2pf ðpÞ ¼ 0 has only one solution in the set
of real numbers. This last assumption is not necessary, but simplifies the
presentation of the results.7
Note first that if the monopolist is not only able to recognize whether a
consumer bought in the past, but also his ‘‘age,’’ all the results that we
obtained in the previous section (including the equilibrium prices) apply
directly, the monopolist charging three prices in each period: one price for
the customers that are just arriving into the market; one price for the
consumers who are in their second period in the market and bought the
product in the previous period; and finally one price for the consumers who
are in their second period in the market and did not buy the product in the
previous period.
However, in many situations, a firm may not be able to recognize a
consumer’s ‘‘age,’’ and therefore have to charge the same price to both
consumers that are just entering the market and consumers that have been
in the market in the previous period, but did not buy the product. Note also
that this has the realistic feature of the monopolist knowing more about the
consumers that bought the product in the previous period than about the
new customers. In terms of the notation of the previous section, not rec-
ognizing the customers’ age means that a ¼ an.
In order to concentrate on the dynamic effects of customer recognition
we focus the analysis on the Markov perfect equilibria (MPE; Fudenberg
and Tirole, 1991, p. 513) of this game, i.e., equilibria in which the actions
in each period depend only on the payoff-relevant state variables in
that period. In this particular game the payoff-relevant state variable in
each period is the stock of previous customers of the monopolist in each
period.
From the analysis in the previous section, we know that in each
period the consumers just arriving in the market who buy the product
in that period are the ones with the highest valuation. That is, in a period t,
the payoff-relevant state variables can be summarized by the type of the
marginal consumer entering the market in period t1 who chooses to buy
in period t1, denoted by y^ t : The computation of y^ t is exactly as in the
previous section. In what follows, let at be the price charged to new cus-
tomers in period t, and at be the price charged to previous customers in
period t.
Denoting as y^ tþ1 ¼ yða
^ t Þ the marginal consumer purchasing in period t
^
given price at, and V ðyt Þ the net present value of the monopolist’s profits
from period t onwards if the marginal consumer purchasing in period t1
7
This assumption is implied by the condition 3f ðpÞ þ 2pf 0 ðpÞ40 which is satisfied for distributions
close to the uniform or the truncated normal with sufficiently large variance.
had valuation y^ t we can write the monopolist’s problem as

Vðy^ t Þ ¼ maxat at ½1 Fðmax½at ; y^ t Þ
þ max ^ aðy^ t Þ½1 F ðyðað
aðyt Þ
^ y^ t ÞÞÞ þ max½Fðy^ t Þ F ðaðy^ t ÞÞ; 0
^ y^ t ÞÞÞ.
þ dV ðyðað ð4Þ
The function aðy^ t Þ is the price to charge the new customers in period t if
the marginal consumer purchasing in period t1 has valuation y^ t : The
right-hand side of (4) is composed of three terms. The first term is the profit
from repeat buyers. The second term is the profit from first-time buyers
which are either new in the market, 1 F ðyðað ^ y^ t ÞÞÞ; or in their second period
^ ^
in the market, max½0; Fðyt Þ F ðaðyt ÞÞ: The set of new buyers who are in
their second period in the market has only positive measure if aðy^ t Þoy^ t : The
third term represents the net present value of profits from the next period
onwards.
The MPE is then characterized by the functions V ðy^ t Þ; aðy^ t Þ satisfying (4)
and yða ^ t Þ satisfying yða^ t Þ ¼ max½at ; ðat daðyða
^ t ÞÞÞ=ð1 dÞ: Note also that
^
if aðyt Þ yt then aðyt Þ is a constant (the case of y^ t small) because the max-
^ ^
imization in (4) is independent of y^ t : This also means that if for a certain y^ t
the optimal aðy^ t Þ y^ t then aðxÞ y^ t ; 8x y^ t : If, on the other hand,
aðy^ t Þoy^ t then aðy^ t Þ is increasing in y^ t because the objective function is
supermodular in y^ t and aðy^ t Þ:
No constant prices in equilibrium

We now show that the general prices are not constant through time.
Suppose that we are in the steady state, with the monopolist charging the
price ā to the new customers in every period. Then, because prices are not
going to decrease and the marginal consumer gets always zero surplus in the
second period, all consumers with valuation above ā buy in the current
^
period, that is, yðāÞ ¼ ā: Then, we also know that aðxÞ ¼ ā; 8x ā: Let a^
^ 2af^ ðaÞ^ ¼ 0; and note that aop^
be defined by 1 F ðaÞ :
If ā4a;^ a small price cut da from ā attracts all consumers with valuation
y ā da; and the effect on the present value of profits is f1 F ðāÞ
2āf ðāÞ þ d min½1 FðāÞ āf ðāÞ; 0gda; which is always positive. Then, ā4a^
cannot be an equilibrium. The intuition is that if the candidate constant
price is not low enough the monopolist gains from cutting prices in the next
period to attract the consumers of the older generation that have a lower
valuation for the good.
Consider now ā a; ^ and a deviation where the monopolist chooses in the
current period t, at ¼ dā þ ð1 dÞp ; followed by atþ1 ¼ ā: That is, in pe-
riod t the monopolist charges a price above ā and in period t+1 returns to
the equilibrium price ā: Once ā is charged, the consumers believe that
no lower price is charged in the future, and all the consumers with the
valuation above ā buy the product. In period t, under the deviation, the
marginal consumer buying the product can be computed to be y^ tþ1 ¼ p :

The present value of profits from this deviation is then
p ½1 F ðp Þð1 þ dÞ þ ½dā þ ð1 dÞp ½1 Fðp Þ

þ dā½1 F ðāÞ þ F ðp Þ FðāÞ
þ d2 V ðāÞ,
while the equilibrium present value of profits can be represented by p ½1

F ðp Þð1 þ dÞ þ ā½1 FðāÞð1 þ dÞ þ d2 V ðāÞ: The difference between the
former and the latter can then be obtained to be ð1 dÞfp ½1 Fðp Þ
ā½1 F ðāÞg, which is greater than zero because p maximizes p[1F(p)].
Then, this deviation is profitable and the monopolist charging always āoa^
cannot also be an equilibrium. That is, if the monopolist charges a sufficiently
low price that it does not have the incentive to cut prices in the next period
(to attract the consumers of the older generation that have a lower valuation
for the good) then it would gain from deviating and charging a high price for
one period in order to identify the consumers that value more the good in the
incoming generation. This shows that there are going to be price fluctuations
in any MPE.8
Let us briefly note that if the analysis is not restricted to MPE one can
obtain subgame perfect equilibria in which prices are constant through time
(as in Ausubel and Deneckere, 1992) at the level obtained when future price
commitments are possible, which is also the case with no customer recog-
nition. In such a case, a deviation by the monopolist is ‘‘punished’’ with the
equilibrium path in the MPE.
Price cycles in equilibrium

Let us now present an equilibrium with price cycles, for the particular
case where f(y) ¼ 1, the uniform case. We restrict attention to smooth
equilibria—equilibria where in steady state the prices being chosen by the
monopolist result from the maximization of a smooth concave function. As
noted below, there are equilibria of this type if d is sufficiently small.9 In the
steady state, the monopolist alternates between high and low prices for the
new customers, denoted by ah and a‘ ; respectively.
If in period t the marginal consumer from the previous generation, y^ t ; is
high, the monopolist charges a low price in order to attract not only the new
generation consumers but also the old generation consumers who did not
buy in the previous period. If, on the other hand, in period t the marginal
consumer from the previous generation, y^ t ; is low, the monopolist charges a
8
Villas-Boas (2004) shows that this same argument also goes through in a two-type distribution for
some parameter values. However, because in a two-type distribution continuous deviations may not be
possible, under some parameter values there are equilibria with constant prices through time.
9
When d-0 all the equilibria converge to the equilibrium presented here.
high price targeted only at the new generation of consumers. In this case, we
can see that V ða‘ Þ ¼ V ‘ is independent of a‘ :
One can then obtain that for d small there is an MPE where the beha-
vior of the game settles in steady state into a price cycle alternating
between ðat ¼ ð8 d2 Þ=ð16 þ 2dÞ; at ¼ 1=2Þ and ðat ¼ ð6 þ dÞ=ð16 þ
2dÞ; at ¼ ð4 þ dÞ=ð8 þ dÞÞ: The prices for the new customers are always
lower than the prices to the previous customers. However, both prices
fluctuate in opposite directions: the price for the new customers is high
when the price for the previous customers is low, and vice versa. The
monopolist charges a high price to the new customers when it had in the
previous period a high demand of new customers. Then, it has relatively
small demand from new customers of 4/(8+d) (all from the new genera-
tion), and a large demand from the previous customers, 1/2. In the next
period, the monopolist charges a low price to the new customers attracting
all the customers from the new generation that have a valuation higher than
the price being charged (with mass (10+d)/(16+2d)), plus the consumers
from the previous generation that waited for the low price in this period,
with mass (2+d)/(16+2d), for a total demand of new customers of (6+d)/
(8+d). The demand from the previous customers is equal to all the new
customers of the previous generation, 4/(8+d). Profits in each of the al-
ternating periods can also be immediately obtained.10
It is also interesting to check the effect of the discount factor on prices,
demands, and profits. In the periods in which the monopolist charges a high
price to the new customers, an increase in the discount factor decreases that
price, the demand from new customers, and therefore profits from new
customers. In the periods in which the monopolist charges a low price to the
new customers, an increase in the discount factor increases that price, the
price to the previous customers, the demand from new customers, and
profits from new customers, and decreases the demand from the previous
customers. The average per period profit decreases with an increase in the
discount factor.
An increase in the discount factor makes the customers more willing to
wait for price cuts. This means that in periods in which the monopolist
charges a high price to new customers, the monopolist has less overall
demand, which makes it lower its price, and results in lower profits. Given
that the marginal customer buying the product has now a greater valuation,
10
The condition on d being small is important because if d were high, more consumers would wait for
the lower prices in the future, which means that there is less advantage for the monopolist to charge a
high price. That is, if d were high, after charging supposedly the lowest price (in steady state), (6+d)/
(16+2d), the monopolist would gain from cutting the price even further (and ‘‘surprising’’ some of its
previous customers). One can check that if do1/2 there is no such profitable deviation. One can also
check that when d is high there is an equilibrium with prices alternating between high and low prices for
the new customers, with similar properties to the ones of the equilibrium presented here, and where the
low price is such that the monopolist does not want to cut the price even further (for d-1 the low price
converges to 1/3).
in the next period the profits are greater, and the monopolist chooses to
charge a greater price to the new customers. However, if one computes the
effect of a higher discount factor on the normalized discounted profit (the
constant profit that would yield the same net present value of profits), one
finds that profits decrease in the discount factor. This is because with a
higher discount factor, consumers are ‘‘more strategic’’, i.e., in the periods
in which the monopolist charges a high price more consumers refrain from
buying.
It is also interesting to compare the equilibrium profits with the case in
which the monopolist is not able to recognize its customers from the current
period on. One can then obtain, as in the previous section, that the average
per period profit without customer recognition is higher than if the mono-
polist were able to recognize its customers.
Comparing the equilibrium profits with the case in which the monopolist
is able to recognize both the previous customers and the consumers’ age one
obtains that the monopolist is hurt by being able to recognize the con-
sumers’ age in addition to recognizing its previous customers. The result is
interesting because it reinforces the idea that the monopolist having more
information (in this case the consumers’ age) ends up hurting the mono-
polist. The intuition is that when the monopolist recognizes the consumers’
age in the market, it offers an even lower price to the consumers that
do not buy the product in their first period in the market, which makes
consumers refrain even more from buying in the first period.
2.3 Long-lived consumers
The longer consumers are in the market, the more information they po-
tentially can give about their preferences through their decisions to buy or
not to buy at different prices. This means that the firm’s policy with respect
to its previous customers is exponentially more complicated with the
number of periods that a consumer has been in the market. Hart and Tirole
(1988) consider the perfect Bayesian equilibrium of this case of long-lived
consumers with a two-type distribution, {y1,y2} with y1oy2, and only one
generation of consumers. They find that in equilibrium, if d>1/2, there is
no price discrimination when the horizon tends to infinity, with the mono-
polist always charging the low price (the valuation of the low type). The
intuition for this result is that if a high-valuation consumer y2 were to buy
the product at a higher price, he would reveal that he has high valuation
and will have zero surplus from that period onwards. If there were a price
strictly above the lowest valuation y1 for which the high-valuation con-
sumer would buy the product with positive probability (such that after that
price, if there were no purchase, the monopolist would charge a price y1
forever), a high-valuation consumer buying the product would be better off
deviating, not buying the product, and getting a low-valuation price from
then on. By buying the product the high-valuation consumer would get a
surplus of at most y2y1, while if the high-valuation consumer waited for

one period (and not be identified as a high valuation consumer) he would
get a surplus approaching ½d=ð1 dÞðy2 y1 Þ; which is greater than y2y1
for d>1/2.
Schmidt (1993) considers the case with any discrete number of consumer
types,11 fy1 ; y2 ; . . . ; yn g with y1 oy2 o oyn ; while restricting attention to
the MPE.12 He finds, as in Hart and Tirole, that, for d>1/2, there is no
price discrimination when the horizon tends to infinity, with the monopolist
always charging the low price y1 (the valuation of the low type).
The method of proof used in Schmidt allows us to better understand the
relation of this result with the general results on reputation (e.g., Kreps et
al., 1982; Fudenberg and Levine, 1989). The proof is similar to the one in
Fudenberg and Levine (1989) on the reputation of a long-term player facing
a sequence of short-term players. Schmidt first shows that if there is a price
strictly above y1 on the equilibrium path, then there is a strictly positive
minimum probability of that price being accepted and revealing a consumer
type with valuation strictly above y1. He then shows that because types
y>y1 can build a reputation for being of type y1, they will do so. That is,
the no-discrimination equilibrium can be seen as a response of the mono-
polist to the consumers’ threat to build a reputation that they have the
lowest valuation for the product if the price is above y1. In Fudenberg and
Levine’s model, the type that a consumer would like to be seen as is type y1.
Given the greater structure of the game considered here (in comparison to
the general class of games considered in Fudenberg and Levine), Schmidt is
able to extend the results of Fudenberg and Levine to the case of two long-
run players, and characterize the equilibrium actions (while Fudenberg and
Levine only characterize payoffs). Schmidt looks at a long finite horizon
game using backward induction, which is what allows him to show that y1
acts like a Fudenberg-Levine ‘‘commitment type’’ and rejects all prices
above y1.
Kennan (2001) considers the case in which the consumer types can change
randomly through time, but with positive serial correlation. He then finds
that we can then have stochastic price cycles because (no) purchases in-
dicate a high- (low-) consumer valuation and are followed by a high (low)
price.
It is interesting to discuss in this context of long-lived consumers what
happens if the firm is allowed to offer long-term contracts, and the rela-
tionship of behavior-based price discrimination with the results from the
durable-goods and bargaining literatures.
11
He considers that it is the monopolist that is the party that has private information (on her costs).
We present here the result in terms of private information of the consumers.
12
The Markov assumption is necessary for the case of any n to guarantee that the continuation
payoffs are the same for a price equal or below y1 (with n ¼ 2 this can be shown without the Markov
assumption).
Long-term contracts
Suppose that the firm would be able to offer a contract to a customer
committing itself to a sequence of prices for the future to be charged to that
consumer. Note that the effect of this possibility is that a consumer would
know now that he would not be taken advantage of in the future for
revealing his high valuation. Hart and Tirole (1988) consider this situation
with the possibility of renegotiation, such that the firm might be able to
offer different contracts in the future.13 For example, in the market for
cellular phone service firms can offer long-term contracts, and can change
which long-term contracts to offer in the future. Hart and Tirole show that
in such a setting with two consumer types, the firm might now be able to sell
to the high-valuation consumers at a price above the lowest price. The
intuition is that with a long-term contract the monopolist has greater ability
to price discriminate. It can get the high-valuation consumer to buy
the product at an average price per period above the lowest price (low
type valuation), because it commits to this average price with a long-term
contract.
For example, if the monopolist offers a long-term contract at an average
per-period price p>y1, the surplus for the high-valuation consumer if he
accepts the contract is ðy2 pÞ=ð1 dÞ: If this consumer decides not to buy
in this period, the most the consumer is able to get is dðy2 y1 Þ=ð1 dÞ; if
the monopolist offers in the next period a contract with an average per-
period price of y1 (the monopolist will never offer a lower average per-
period price). Then, if p ¼ dy1 þ ð1 dÞy2 the high-valuation consumers
accept the contract, and the monopolist is able to sell to such consumer at a
price strictly above y1. As shown in Hart and Tirole, the equilibrium long-
term contract is for the monopolist to offer a contract in a number of initial
periods with average per-period price strictly above dy1 þ ð1 dÞy2 ; such
that type y2 randomizes between accepting and not accepting the contract,
and then, after a certain number of periods, the monopolist offers a con-
tract with average per-period price y1, and both types accept the contract.
However, this possibility of selling to the high-valuation consumers with
an average per-period price strictly above y1 is not possible without a long-
term contract. Without a long-term contract a high-valuation consumer
gets zero surplus after revealing his type, and therefore, must be offered a
price below the low-type valuation to accept buying.14 But then the low-
valuation consumer would also buy the product, and, therefore, no infor-
mation would actually be revealed about the type of the customer buying
the product. Hart and Tirole then show that, because of this greater ability
13
Laffont and Tirole (1990) consider a two-period version of such contracts with continuous con-
sumption per period in the context of procurement.
14
One can see this as a high-valuation consumer maintaining the reputation that he may have a low
valuation. See the discussion above.
to price discriminate a firm is better off when it has the ability to offer a
long-term contract.
It turns out that this effect of long-term contracts does not occur if the
consumer lives only for two periods, with the second period being the last
period. In the two-period model presented above it turns out that the in-
troduction of long-term contracts does not have any effect, and the equi-
librium with long-term contract is exactly the same as the equilibrium
without long-term contracts. This is because the zero surplus obtained by
the marginal consumer after revealing his type only lasts for one period.
Battaglini (2005) considers the case of infinitely lived consumers where
the preferences change through time following a Markov process, as in
Kennan (2001), but allowing for continuous consumption. A consumer’s
per-period utility in period t is ytqp, for q units bought at price p. The
monopolist’s cost of selling q units is ð1=2Þq2 : For future reference, note that
the efficient quantity to be sold in period t is qe ðyt Þ ¼ yt : The marginal
benefit yt in period t is private information of the consumer, can only take
one of two values, fy; ȳg; with ȳ4 y; and evolves through according to a
Markov process. The transition probabilities between states are in (0,1),
and are denoted by Prðytþ1 jyt Þ: Types are assumed to be positively corre-
lated over time, PrðȳjȳÞ Prðȳj yÞ and Prðy j yÞ Prðy jȳÞ: At date 0 the
monopolist has a prior m that the consumer’s type is ȳ and a prior 1m that
the consumer’s type is y :
Battaglini computes the optimal long-term contract. First, he shows that
under commitment the optimal contract always involves the efficient quan-
tity being supplied if in the history of the relationship (including the current
period) there has been a period in which the marginal benefit has been equal
to ȳ: That is, with varying types we have the result that a long-term contract
supply is at the efficient level in finite time (which is not the case for fixed
types). The intuition for this result has to do with the role of the quantity
distortions in the contract. Distortions are introduced only to extract more
surplus from higher types, and therefore, there is no reason not to offer the
highest type the efficient quantity. After any history the rent that must be
paid to a high type to reveal himself is independent of the future quantities.
That is, the monopolist is the residual claimant on the surplus generated on
histories after a high-type report, and therefore the quantities that follow
such report are the efficient ones. In addition, Battaglini finds that if the
history has never had a period where the buyer had type ȳ; the quantity
distortion vanishes through time as the initial state has less and less infor-
mation about the current buyer’s type.
Battaglini then considers the case in which the contract can be renego-
tiated, and shows that under general conditions the contract with commit-
ment is renegotiation-proof, and when these conditions fail, the contract is
renegotiation-proof after a finite amount of time.
Battaglini’s analysis relies heavily on the assumption that there are
only two types. As noted in the paper, with n types, then the conditional
distribution for each type is represented by a n1 vector, each type has n1
characteristics, and we would need to solve a multidimensional-screening
problem. It would be interesting to investigate further this n-type case, even
if putting some structure on the conditional distribution for each type.
Relationship to durable goods and bargaining

The strategic behavior of consumers when firms practice behavior-based
price discrimination is related to settings where a monopolist sells a du-
rable-goods, or to settings where two parties bargain, and in which one of
the parties has private information. Here, we first briefly discuss some of
the forces present in a market where a monopolist sells a durable good, or
in a bargaining situation between two parties, in which one party has pri-
vate information. Then, we relate the durable-goods setting with the be-
havior-based price discrimination model. (For some discussion of the
durable-good monopoly literature see, for example, Chapter 1.5 in Tirole,
1988. For a survey of the bargaining literature see, for example, Chapter 10
in Fudenberg and Tirole, 1991.)15
Durable goods and bargaining. Consider the two-period model above, but
suppose now that the monopolist sells a product in the first period that lasts
for the two periods. Let A be the price of the durable in the first period.
Denoting y^ as the type of the marginal consumer in the first period, the
surplus of this consumer is ð1 þ dC Þy^ A when buying in the first period,
and is dC ðy^ an Þ if waiting for the second period.
The marginal consumer in the first period is then determined by
y^ ¼ A dC an (5)
Given that the marginal consumer buying the product in the first period
^ the optimal price in the second period is a ðyÞ
is y; ^ ¼ arg maxa an ½F ðyÞ
^
n n
^ ^
F ðan Þ: Using this we then have that (5) defines y as a function of A, yðAÞ:
In order to obtain the optimal first period price, A, the monopolist then
maximizes
^
maxA Að1 FðyðAÞÞÞ ^
þ dF an ðyðAÞÞ½F ^
ðyðAÞÞ ^
F ðan ðyðAÞÞÞ (6)
where the first term represent the first-period profit, and the second term
represents the second-period profit. The first order condition that defines
the optimal A is then
1 F ðyÞ ^ y^ 0 þ dF y^ 0 f ðyÞa
^ A f ðyÞ ^ ðyÞ
^ ¼ 0. (7)
n
Note that for dC ¼ dF the marginal consumer buying the product in
the first period has a higher valuation than if the firms were selling a
15
For early work on the durable-goods monopolist problem see, for example, Stokey (1981) and
Bulow (1982).
non-durable good. To see this note that the first derivative of the objective 0
function above evaluated at yðAÞ ^ ¼ p is equal to f ðp Þp ½1 y^ after
substituting for 1 F ðp Þ p f ðp Þ ¼ 0 and p ¼ A dan ðp Þ: Given that
^y0 ¼ 1=ð1 þ da0 Þ and a0 40; that derivative is positive, which means that
n n
the monopolist should increase A, which implies a higher valuation of
the marginal consumer than p : One can also obtain for dC ¼ dF that the
present value of profits is y½1 ^ F ðyÞ ^ þ da ½1 Fða Þ; which is strictly
n n
below the present value of profits under no customer recognition, as p
uniquely maximizes p½1 F ðpÞ: The intuition of this result is that the
marginal consumers refrain from buying in their first period in the market
because they know that they can get a lower price in the next period.
For the uniform distribution example one can obtain an ðyÞ ^ ¼ y=2;
^ ^
yðAÞ ¼
2
2A=ð2 þ dÞ; and A ¼ ð2 þ dÞ =ð8 þ 2dÞ: One can also get that 2/(4+d)
consumers buy in the first period, while (2+d)/(8+2d) consumers buy in
the second period.
The model above can also represent a bargaining situation where there is
a single buyer, and if the buyer does not take the first offer A then he is
offered an in the second period. In such a setting one can then obtain that
private information of the buyers leads to an inefficient outcome for some
consumer types (if rejection occurs in the first period).
In a durable-goods setting, if new generations of consumers come into the
market in every period, there are incentives for the monopolist to raise its
price in order to try to extract more surplus from the consumers who have a
high valuation and who have entered the market most recently. This can
then generate price cycles in which prices come down to clear the demand
from low-valuation consumers, and then go up to better extract the surplus
from the consumers with high valuation who just entered the market. This
setting is analyzed in Conlisk et al. (1984), and Sobel (1984, 1991). Al-
though having the flavor of the results in Subsection 2.2 for overlapping
generations of consumers and behavior-based price discrimination, and as
also discussed below, the results are different, as we can have price cycles in
the behavior-based price discrimination model, but constant prices in the
corresponding durable goods model.
In some situations the monopolist may also have some private informa-
tion regarding its costs, so that the price offers can potentially reveal some
information about the monopolist’s costs. Fudenberg and Tirole (1983),
with a bargaining set-up, characterize the set of equilibria in two-period
games when the monopolist and the buyer each have two potential
types (two-sided incomplete information). They show that this additional
private information may lead to a continuum of perfect Bayesian equilibria.
Ausubel and Deneckere (1992) consider the infinite horizon version of
this two-sided incomplete information model, and show that we may have
(stationary) equilibria in which prices stay high, and the seller tries to
maintain a reputation of having a high cost.
Relationship of durable goods to behavior-based price discrimination. When

a monopolist is selling a durable good through time, consumers refrain
from buying in the initial periods because they foresee that the monopolist
may cut its price in future periods. In such a setting consumers may prefer
to forsake the benefits of the product if buying earlier, with the lower price
if buying later. On the other hand, with customer recognition and behavior-
based price discrimination for a non-durable, consumers refrain from buy-
ing in the initial periods because they foresee that the monopolist may cut
its price in future periods, and therefore they will be identified as lower
valuation consumers and get lower prices for the future. This difference
between durable-goods and behavior-based price discrimination for non-
durables leads to different consumer surplus effects from purchasing the
product, and therefore different market implications.
When buying a durable-good the consumer pays a price and gets a benefit
of using the product for the duration of the product’s life. Consumers for
whom the present value of future benefits is greater than that price may be
willing to purchase the product. However, under customer recognition and
behavior-based price discrimination for a non-durable good, the marginal
consumer buying the product pays the price and gets a benefit in the current
period, but then gets zero surplus in all future periods. Therefore, in order
for a consumer to be willing to buy, the initial price must be so low, that
even consumers with very low valuation may be willing to buy the product.
For an infinite horizon with two types, Hart and Tirole (1988) show then
that the durable-good case is better for the monopolist than the case of a
non-durable with the ability to recognize customers and price discriminate
according to past behavior (Hart and Tirole consider this possibility in
terms of rental of the durable good).
In the long-lived consumers with two consumer types model that they
consider, Hart and Tirole also find that the durable-good case is exactly the
same as when the monopolist can offer long-term contracts (and different
from short-term contracts), as the separation between types can be done ex-
ante. In the two-period model considered above the durable-good case
results in exactly the same outcome as the long-term contract case, and
generates exactly the same outcome as the customer recognition case for a
short-term sales of a non-durable, or equivalently with short-term rentals of
a durable. This is because, in a two-period model, the consumer surplus
effects of purchasing a durable-good are the same as purchasing a non-
durable with customer recognition, as the zero surplus of the marginal
consumers under customer recognition lasts only for one period.
In the case of overlapping generations of consumers, with consumers
living for two periods (and without the ability to recognize the customer’s
age) selling a durable good may not generate price cycles, as selling the
durable good for a consumer that only uses the product for one period
requires a much lower price than selling the durable good for a consumer
who uses the product for two periods (Villas-Boas, 2004). That is, with
overlapping generations of consumers, selling a durable good does not yield
the same outcome as selling a non-durable with customer recognition (with
or without long-term contracts).
Thus, in general, the sale of a durable good is not the same as a sequence
of short-term rentals. Although the price falls over time, the price a con-
sumer faces is not based directly on its own past behavior. Loosely speak-
ing, the commitment involved in selling a durable good lets the monopolist
commit to not use behavior-based pricing.
2.4 Two-good monopoly
In order to serve as an introduction to the next section on competi-

tion, and to serve as a benchmark, consider now the case of a monopoly
selling two goods, A and B. The presentation here follows closely part of
Section 5 of Fudenberg and Tirole (FT, 2000). To focus on the interaction
between the two goods we set up preferences such that consumers buy a
unit of one of the goods in every period. Indexing the relative preferences
for B over A as y, let the valuation per period of a consumer of type y be
vy/2 if the consumer buys good A, and v+y/2 if the consumer buys good
B, with v ‘‘large’’ and y distributed in ½y; ȳ; where y ¼ ȳo0; with cumu-
lative distribution function F(y), strictly positive density f(y), and F(y) is
symmetric about zero and satisfies the monotone hazard rate (MHR) con-
dition that f(y)/[1F(y)] is strictly increasing in y. The parameter v is as-
sumed large, such that the monopolist chooses prices such that all
consumers buy one unit of one of the goods in every period. For this
reason, we consider that the monopolist’s production will be constant
across the pricing regimes , so that the costs of production are a constant
that can be ignored.
Let a and b be the prices charged in the first period for products A and B,
respectively, a and b be the prices charged in the second period for products
A and B, respectively, for consumers who bought the same product in the
previous period, and a^ and b^ be the prices charged in the second period for
products A and B, respectively, for consumers that bought a different
product in the previous period.
Consider first the case in which long-term contracts are not available.
Then the firm will charge a ¼ a ¼ b ¼ b ¼ v, for a present value of profits
of (1+d)v. Note that consumers do not switch products.
Consider now the case in which long-term contracts are available (with
commitment not to renegotiate). Then, in the first period the monopolist
can offer four product consumption experiences: product A in both the
periods, product A followed by product B, product B followed by product
A, and product B in both periods. By symmetry it is enough to analyze the
interval ½y; 0: As argued in Fudenberg and Tirole, incentive compatibility
requires that consumers in an interval ½y; y ^ choose product A in both

periods (which we call AA), and consumers in the interval ½y; ^ 0 choose
product A followed by product B (which we call AB). In order for the type 0
to be indifferent between buying and not buying, and between the ‘‘switch-
ing’’ bundles AB and BA, it must be that the price for each of these bundles
is (1+d)v. Indifference for type y^ between bundles AA and AB requires that
the price of AA be above the price of AB by dy: ^ Thus, it is as if the
monopolist first sold all consumers a ‘‘switching’’ bundle at price ð1 þ dÞv;
and then offered an ‘‘upgrade’’ to AA or BB for a premium of dy: ^ The
present value of profits is then ð1 þ dÞv 2dyF ðyÞ; where the optimal y^
^ ^
^ þ yf
satisfies F ðyÞ ^ ðyÞ
^ ¼ 0: Note that the optimum has some consumers
switching products across periods. Since consumer preferences are the same
in both periods, this switching is inefficient; it is used to extract rents for the
privilege of not switching. For the uniform distribution one can obtain
y^ ¼ y =2; so that one half of the consumers switch products from the first to
the second period.
Fudenberg and Tirole also show that the monopolist can do better than
the above deterministic menu, with a randomized menu where consumers in
^ y
½y; ^ get a (1/2, 1/2) randomization between products A and B. This
allows the monopolist to extract a greater surplus from the consumers that
get no ‘‘switching’’. Again, as in the deterministic menu, we have some
inefficient (stochastic) switching by some consumers.
When we reach the second period, as stated above, ‘‘switching’’ consum-
ers are consuming a product that is not the best for them. This means that
there are gains to be made from the monopolist renegotiating the product
that is offered to those consumers. This renegotiation may then affect the
choices of consumers in the first period (and the monopolist’s offers). It
would be interesting to investigate whether we would still have inefficient
switching in equilibrium if the monopolist can offer long-term contracts
subject to renegotiation.
In order to compare also with the next section consider now the case of
switching costs, where a monopolist sells two products in each of two suc-
cessive periods, consumers have the same valuation v for each product per
period and incur in a cost s if they change products from the first to the
second period. It is clear that in this situation the best the monopolist can
do is extract v per period per consumer (with a price equal to v), and there is
no switching products from the first to the second period. This can be
accomplished either with short or long-term contracts.
Consider now in the model above (with heterogeneous consumers), the
role of the introduction of switching costs s (suppose s small). The price of
the switching bundles can then be at most ð1 þ dÞv ds and indifference for
type y^ between bundles AA and AB requires that the price of AA be above
the price of AB by dðs yÞ:^ Thus, comparing with the no switching costs
case, the price of the switching bundle is now lower, but the premium to
upgrade to the non-switching bundle became now greater. The present
value of profits is now ð1 þ dÞv 2dyF ^ ðyÞ

^ ds½1 2F ðyÞ;
^ where the opti-
^ ^ ^ ^
mal y satisfies FðyÞ þ ðy sÞf ðyÞ ¼ 0: Note that the optimum has some
consumers switching products across periods, but the number of switching
consumers is decreasing in the switching cost s. Note that switching also is
inefficient, and that the monopolist profit and welfare are decreasing in the
switching cost s. For the uniform distribution one can obtain y^ ¼ ðy þsÞ=2;
so that less than one half of the consumers switch products from the first to
the second period.
3 Competition
Several new issues arise in models of behavior-based price discrimination

with multiple firms. Starting with the most obvious, firms can try to
‘‘poach’’ their rivals’ customers by giving them special ‘‘introductory’’
prices.16 This raises the questions of how much switching we should expect
to occur, and of its efficiency consequences. At a more theoretical level, we
have already seen that in equilibrium a monopolist without commitment
power can be made worse off by the ability to condition the price it charges
a customer on that customer’s past decisions, because consumers will fore-
see this condition and adjust their earlier behavior. The same sort of fore-
sight can operate in models with multiple firms, but now its impact on profit
is a priori ambiguous, because of the interactions between the customers’
behavior (basically the elasticity of demand) and the equilibrium in the
pricing decisions of the firms. Furthermore, while a monopolist with com-
mitment power can always do at least as well when behavior-based dis-
crimination is possible (by committing itself to ignore past behavior in
setting prices), a group of oligopolists with commitment power can all be
worse off if all of them become able to discriminate based on past customer
behavior, as the better information may lead to more intense price com-
petition (see Subsection 5.3 below).17 For this reason, while each firm has a
dynamic incentive to adjust its prices so that it learns more about the
consumers and can better segment the market, the firm also has an incentive
to reduce the information that is obtained by its rivals.
The way that these various effects combine to determine equilibrium
prices and allocations depends on the nature of preferences and on the form
of market competition. The first part of this section considers Fudenberg
and Tirole (FT, 2000)’s analysis of a two-period model of competition in
short-term contracts, and some variations on the distributions of consumer
types studied by Chen and Zhang (2004) and Esteves (2004). The second
16
In 1994, about 20% of all U.S. households changed their provider of long-distance telephone
services (Schwartz, 1997).
17
Of course, a single oligopolist with commitment power who is given the ability to condition prices
on customer history cannot be made worse off, provided that none of the other firms are allowed to have
this ability.
part discusses Villas-Boas’ (1999) extension of the two-period model to an

infinite horizon with overlapping generations of consumers, each of whom
lives only for two periods. We then return to the two-period setting to
present Fudenberg and Tirole’s analysis of competition in simple long-term
contracts, meaning that firms offer both a ‘‘spot’’ or one-period price and
also a long-term commitment to supply the good in both periods. Finally,
we compare the predictions of these models of ‘‘pure price discrimination,’’
where the past matters only for the information it provides about prefer-
ences, to models of switching costs, where past decisions are directly payoff
relevant, and may also provide information, as in Chen (1997) and Taylor
(2003).
3.1 Two periods, short-term contracts
Following FT, suppose that there are two firms, A and B, who produce
non-durable goods A and B, respectively, at constant marginal cost c. There
are two periods, 1 and 2; each period a consumer can either consume a unit
of good A or a unit of good B or neither, but not both. There is a con-
tinuum of consumers, whose preferences are quasi-linear in money and are
indexed by y 2 ½y; ȳ; where y ¼ ȳo0: The consumption utility from goods
A and B is vy/2 and v+y/2, respectively, so that y measures the con-
sumer’s preference for good B over good A. There is a known distribution F
on y, which is assumed to be symmetric about 0. Fudenberg and Tirole
assume that F is smooth, with density f, and that F is symmetric and that it
satisfies the MHR property that f ðyÞ=½1 F ðyÞ is strictly increasing in y;
their sharpest results are for the special case of the uniform distribution.
Esteves (2004) considers the case where F has a two-point support;18 Chen
and Zhang (2004) assume that F is concentrated on the three points y; 0; ȳ:
Fudenberg and Tirole assume that all agents use a common discount factor
d; the other papers suppose that firms use discount factor dF while con-
sumers use the possibly different discount factor dC.
With simple short-term contracts, and no commitment power, each firm
will offer a single first-period price, which we denote a and b, respectively.
In the second period, each firm can offer two prices, one to its own past
customers and another price to all others. (We will assume that the res-
ervation value is high enough that all consumers purchase in the first pe-
riod, so that consumers who didn’t purchase from firm A must have
purchased from firm B.19) Note that if firms do not observe the identities of
their customers, there is no link between the periods, and the equilibrium
18
Esteves supposes that the two mass points are in the interval [0,1], symmetric about the point 1/2; to
map her notation to ours suppose that the mass points are at yA ¼ t(2xA1) and yB ¼ yA, and that
v ¼ v0 tð1 xA Þ; where v0 is the reservation value in her notation.
19
Chen and Zhang consider an extension of their model to the case where consumers with y ¼ 0 have
lower reservation values; in this case not all consumers purchase in the first period.
Fig. 1. Second-period competition given the first period cut-off.
reduces to two repetitions of the static equilibrium. Our question is how the
prices and efficiency of the equilibrium with short-term contracts and cus-
tomer poaching compare to that of the static benchmark.
Under FT’s assumptions, the static one-period problem is well behaved:
each firm’s objective function, pi ¼ F ðpj pi Þðpi cÞ; is strictly quasi-con-
cave, so that firms are never willing to randomize, and the game has a
unique equilibrium, namely
F ð0Þ
pA ¼ pB ¼ þ c.
f ð0Þ
In the case of a uniform distribution, this simplifies to p ¼ c þ ðȳ yÞ
=2 ¼ c þ ȳ; so that each firm’s profit is ȳ=2: Moreover, in the uniform case
the dynamic equilibrium is also in pure strategies, and can be characterized
with first-order conditions. With the discrete supports specified in the other
two papers, the static equilibrium is in mixed strategies, which makes the
calculations more complex and the intuition more subtle. For this reason we
use the FT case for exposition, and try to explain informally the effects of
the other distributional assumptions.
Analysis of the two-period model under the MHR assumption

A standard argument shows that at any pair of first-period prices such
that all consumers purchase and both firms have positive sales, there is a
cut-off y such that all consumers with types yoy purchase from firm A in
the first period.20 Given this cut-off, the second period game is as depicted
in Fig. 1: consumers to the left of y lie in ‘‘firm A’s turf’’ and the consumers
on the right lie in firm B’s. On firm A’s turf, firm A offers price a, while firm
^ on B’s turf B charges b and A charges a^ : Thus, a consumer
B offers price b;
20
To deal with out-of-equilibrium beliefs, we suppose that if first period prices are such that no
consumer is expected to buy from firm A, a consumer who unexpectedly does purchase from A is
assumed to have type y; and similarly a consumer who unexpectedly purchases from B is assumed to
have type ȳ:
on firm A’s turf will stick with good A if yoa b; ^ and otherwise will

switch to good B. If y is very near the endpoint y; then A’s turf is very
small, and consists only of consumers with a strong preference for A, and
firm A can charge the monopoly price in this market and not lose any sales
to firm B. The paper shows that this occurs when y oy ; where y is the
‘‘isoelastic point’’ where F(y )+f(y )y ¼ 0, so that marginal revenue
equals 0. In this case firm A sets a ¼ v y =2 and sells to everyone on its
turf, while firm B sets b^ ¼ c: Otherwise, both firms will have positive sales in
each market, which implies that the ‘‘poacher’s’’ price in a market must be
lower than the incumbent’s.
The intuition for this result comes from the fact that on the interior
of firm A’s turf, its second-period reaction function reflects a trade-off
between losing marginal customers at yA and inframarginal rents on
types below yA, and so the reaction function does not depend on the first-
period cut off y ; while decreasing y decreases B’s sales on A’s turf, and
so makes firm B price more aggressively, as shown in Fig. 2, where the
curves R are the reaction curves on firm A’s turf when it had the entire
first-period market (which is why they intersect on the 45 degree line), and
R^ is firm A’s reaction curve on its turf as a function of the first-period
cut-off y :
The next step is to work backwards and determine the equilibrium of
first-period prices. Before presenting the analysis, we can identify some
general considerations to keep in mind.
Fig. 2. Second-period reaction functions as a function of the first-period cut-off.

1) If consumers are forward looking (as assumed by FT) they realize that
they will be offered a ‘‘poaching’’ offer in the second period. FT show
that this can lead to a less elastic first-period demand, and hence higher
first-period prices.
2) Firms may distort first-period price to increase second period profit.
Specifically, each firm would rather that its opponent have less infor-
mation about consumer preferences, and is willing to distort first-
period prices for that purpose. Moreover, this preference is large
enough that firms do better when neither of them has any information
about consumer’s identities. The impact of this consideration depends
on the way that changes in price change what is learned about con-
sumer demand, which in turn depends on the distribution of types.
3) If customers buy from their preferred firm in the first period (as they
do in FT) then second-period switching lowers welfare.
To explore this second point in more detail, we present a more detailed
analysis of second-period competition in the uniform case than is given in
FT. Solving for the intersection of the second period reaction curves (cor-
responding to equation (6) and (7) in FT) shows that yA ¼ ðy þ yÞ=3; yB ¼
bb a ¼ ðȳ þ y Þ=3: In its home turf, firm A sells to types below yA; this has
mass ðy 2 yÞ=3ðȳ yÞ ¼ ðy þ 2ȳÞ=6ȳ: On B’s turf A sells to types
between y and yb; this has mass ðȳ 2y Þ=3ðȳ yÞ ¼ ðȳ 2y Þ=6ȳ:
So the second-period profit of firm A is
2 2
ðy 2 y Þ2 ðȳ 2y Þ2 5y þ 5ȳ
þ ¼ ,
18ȳ 18ȳ 18ȳ
provided that y 4y ¼ y =2; so that there is poaching in both markets. By
symmetry this is also the second-period profit of firm B.
Note that the symmetric outcome y ¼ 0 is the global minimum of firm
A’s second-period profits; it does better not only with a larger first-period
market share but also with2 a smaller one! Specifically, when y ¼ 0 the
second-period profit is 5ȳ =18ȳ ¼ 52ȳ=18: As y increases to ȳ=2 profit
increases to ðȳ=2 þ 2ȳÞ2 =18ȳ ¼ ð25ȳ =4Þ=18ȳ ¼ 25ȳ=72: From this point
onwards, there is no competition in firm B’s second-period market. Firm
A’s profit is ðy 2 y Þ2 =18ȳ; which converges to the static equilibrium value
of ȳ=2 as y goes to ȳ:
This shows that both firms do best when neither has first-period infor-
mation. When y is near the endpoints, firms have less precise information
in the larger market, and hence competition there is less intense.
Perhaps surprisingly, in the uniform case this second-period considera-
tion has no impact on first-period pricing. This is because the first-period
equilibrium will have equal market shares, i.e., y ¼ 0; and because y ¼ 0
leads to the lowest level of second-period profit, there is no first-order effect
when it changes. For this reason, the only reason that first-period prices
differ from the static equilibrium is that consumer demand differs. In the
static case, the cut-off y shifts one-for-one with prices, while in the dy-
^ ÞÞÞ=ð1 dC Þ; because type y
namic setting y ¼ ðb a þ dC ðâðy Þ bðy
must be indifferent between the different plans to switch and get the
‘‘poaching price’’ next period, and so must be indifferent between buying
good A now at price a and then buying B tomorrow at price b; ^ or buying B
now at price b and then buying A tomorrow at price a^ : In the uniform case,
this leads to a less elastic first period demand ðj@y =@ajo1Þ; and hence
higher prices; with zero production costs and consumer types distributed on
the unit interval, the first-period price is ð1 þ dÞ=3 and the second-period
prices (on the equilibrium path) are 2/3 to the firm’s old customers and 1/3
to the customers it is trying to ‘‘poach.’’
This finding for the uniform case leaves open the possibility that for other
distributions the second-period-profit effect could have an impact on first-
period pricing. However, it seems plausible that y ¼ 0 is the global mini-
mum of second-period profits for general symmetric distributions, so that
the effect of second-period profit on first-period decisions vanishes, pro-
vided that the first-order approach is valid. However, the fact that firms
would do better in the second period with a less symmetric first-period
outcome suggests a possible non-concavity in the problem. The MHR as-
sumption makes the static optimization problem concave, which implies
that the firms’ first-period objective functions are concave for discount
factors close to 0 and any distribution that satisfy MHR; FT show that they
are also concave under the uniform distribution for all (common) discount
factors (that is, df ¼ dc ¼ d). However, concavity does not seem to be im-
plied by the MHR condition, and when it fails there can be mixed-strategy
equilibria. To investigate this possibility it may be interesting to abandon
the first-order approach altogether, and work with discrete types, as in
Esteves and Chen and Zhang.
Discrete distributions
In Esteves’ model, whenever the difference in price is less than yByA
each firm buys from their preferred firm, while if the difference is larger
than this all consumers buy from the same firm and no information is
revealed, which corresponds to the case y ¼
ȳ in FT. Again as in FT, the
second-period profits are symmetric in the information: firms do better
when the first-period prices are very different, but as far as second-period
prices go they are indifferent between having a large turf or a small one. To
simplify the analysis, Esteves assumes that consumers are completely myo-
pic. The first-period equilibrium is in mixed strategies, and she shows that
the probability that both firms have positive first-period sales decreases as
they become more patient. Moreover, she shows that first period prices tend
to fall as the discount factor increases.
Chen and Zhang suppose that there are three types. A mass g is loyal to A
(they get 0 utility from B, so they buy A whenever the price is below their
reservation value v), a mass g is loyal to B, and a mass of 12g who are
exactly indifferent. Neither firm can hope to sell to the loyalists of the other,
so what each firm wants to do is distinguish its loyalists from the neutrals.
Starting from equal first-period prices, a small increase in firm A’s price
shifts all of the neutrals to firm B, and results in an asymmetric knowledge
about the consumers: firm A knows who its loyalists are, but firm B does
not. Thus, in contrast to the previous two papers, the firm with the smaller
first-period sales has strictly higher second-period profits. They show that
this leads to prices that are, on average, higher than in the static equilib-
rium, even when consumers are myopic.
We should point out some unusual features of the assumed demand dis-
tribution. Specifically, second period profits when consumer types are
known are exactly the same as in the static model, while in general we may
expect that known types could lead to fiercer competition and lower profit.
This suggests that competition in the static model is particularly fierce,
which observation may help explain why equilibrium profits here are higher
than when firms lack information on purchase history.
Welfare
Finally, we compare the welfare effects of price discrimination in the
three models. In FT, the first-period outcome is efficient, so all second-
period switching lowers welfare. In Esteves, both the static equilibrium and
the first-period equilibrium of the two-period price discrimination game are
not fully efficient, due to the randomized nature of the equilibrium. More-
over, when the first-period prices reveal the customers’ types, the second-
period outcome is efficient, and there is no switching, even though firms
offer lower second-period prices to their opponents’ customers. This stems
from the two-point distribution of demand, and would not extend to a
discrete model with more types. Combining these two observations, we see
that price discrimination can increase efficiency provided that it doesn’t
lower first-period efficiency too much, and she shows that this is indeed the
case. In the Chen and Zhang model, efficiency considerations are moot, as
the only consumers whose purchases change when price discrimination is
allowed are those who are completely neutral. There can, however, be effi-
ciency implications of price discrimination when the reservation value of the
neutrals is less than the other players, as price discrimination allows the
firms to offer the neutrals a second-period price that is below their reser-
vation value without losing sales to the loyalists.
3.2 Infinite lived firms, overlapping generations of consumers, and short-

term contracts
Villas-Boas (1999) extends the FT model to the case of two infinite-lived

firms facing overlapping generations of consumers. Each consumer lives for
two periods, and each generation has unit mass. Each firm knows the iden-
tity of its own past customers, but not those of its opponent, and it does not
observe the consumer’s ‘‘age,’’ so it cannot distinguish young consumers
from old ones who bought from the opponent last period. The basic setup of
the model, and the notation, are the same as in FT, with y uniform on
½1=2; 1=2 and zero production costs. The timing of the game is a bit
different, as in each period the firms first simultaneously set the price for new
customers, and then set the prices to existing customers after observing the
price the competitor charges to new ones.
In order to focus on the dynamics of price discrimination, and abstract
from (possibly important) repeated game aspects, the paper restricts
attention to the state-space or MPE of the game. Given the linear-quadratic
nature of the model, there are MPE in which the strategies are piecewise
affined in the state variable, and these are the ones considered in the
paper.21 As a benchmark case, note that the MPE here would be exactly
the outcome in FT if, as in FT, firms can recognize both their own and the
opponent’s customers, and all prices are set simultaneously. If firms can
recognize both types of old customers, but prices are set sequentially
as specified above, the prices will be 1 þ dC ðdF =4Þ to new customers, and
the prices will be 3/4 and 1/2 to the firm’ and the competitor’s old
customers, as opposed to 2/3 and 1/3 with simultaneous price setting.
(Prices are higher with sequential moves because the reaction curves
slope up, this is a form of the ‘‘puppy dog effect’’ (Fudenberg and Tirole,
1984).)
We now turn to the MPE of the game where firms only recognize their
own customers. If the reservation value is high enough that all consumers
purchase every period, Villas-Boas shows that the equilibrium is again
characterized by cut-offs yt such that each new consumer arriving in period
t purchases from firm A if their type yoyt : Thus, the payoff-relevant state
in each period is simply the previous period’s cutoff.
The easiest part of the model to solve is the prices firms charge to their old
customers. Since these consumers will leave the market at the end of the
period, neither they nor the firm need to consider future periods in making
their decision, and since prices are set after observing the rival’s poaching
price, the firm faces a simple static maximization. In contrast, the price set
to unrecognized consumers must take into account that some of these are
new consumers who will purchase again in the next period, and the demand
of new customers must also take the future into account.
Neither of these complications is present in the case of complete myopia,
dF ¼ dC ¼ 0. Here, the cutoff converges to the steady state with equal-mar-
ket shares. Except possibly in the first period, the convergence is monotone,
and customers sort themselves as in FT: those with strong preference for one
21
The reason to consider piecewise affine strategies instead of affine ones is that there are ‘‘kinks’’ in
the value functions corresponding to the states where a firm completely retains all of its clientele; these
kinks are roughly analogous to the points
y in FT.
firm buy from that firm in each period, while those with more intermediate
preferences switch. As in FT, prices to the recognized consumers are lower
than in the static case. Prices to the unidentified consumers are also lower
than the static prices, while in FT the first-period price equals the static price
when firms are myopic; this is because the pool of unidentified consumers
here contains both new consumers (as in the first period of FT) and old
consumers who prefer the other firm.
Villas-Boas then considers the case of myopic firms but patient consum-
ers; this differs from the previous analysis where consumers take into ac-
count the prices they will be charged next period; it differs from FT because
a consumer who buys A in the first period is offered a second-period price
for B that is tailored to a mixture of ‘‘A-preferrers’’ (i.e. yoyt ) and new
consumers, as opposed to a ‘‘poaching price’’ for A-preferrers alone. This
mixed price will in general be less responsive to changes in y than is the
poaching price, which makes the marginal new customers more responsive
to changes in price. For this reason, the price to new consumers is lower
than in FT, and in fact it goes to 0 as dC-1.
Finally, Villas-Boas considers the case where dF and dC are both non-
zero. As in Esteves and Chen and Zhang, patient firms have an incentive to
shift their prices in a way that softens future competition, which here leads
to higher prices. In the case dC ¼ dF ¼ d-1, the price charged to new
consumers converges to 0 with d, while the price charged to old ones con-
verges to 1/2. Thus firms are worse off than when they could credibly share
their information. We discuss the issue of information sharing in Section 5.2
on credit markets.
3.3 Long-term contracts
As we remarked in Section 2, long-term contracts are used in a variety of

consumer markets. This section considers the impact of competition in
simple long-term contracts in the setting of the two-period FT model. Spe-
cifically, we suppose that in the first period firms A and B offer to sell their
goods this period at spot prices a and b, and that they also offer long-term
contracts to supply the goods in both periods for A and B. In the second
period, firms know the first-period prices announced by their rival, and they
also know from whom each consumer purchased, but do not observe the
contracts chosen by their rivals’ customers.
If a firm chooses to only sell long-term contracts, it would prevent
poaching by its rival; but the fact that a monopolist with commitment
power induces switching suggests that the complete lock-in will not
be optimal here either. And indeed, Fudenberg and Tirole show that the
equilibrium has the form depicted in Fig. 3: consumers who most prefer
A buy a long-term contract from A; this is the interval ½y; yA : The
next interval ½yA ; yA purchases A in each period on the spot market,
Fig. 3. Equilibrium long- vs. short-term contracts as a function of the consumer type.
interval ½yA ; y buys from A in the first period and B in the second, and so
on.22 Thus, as in the case of short-term contracts, there is inefficient
switching.
A key fact in determining the equilibrium outcome is that when firm A
locks in more of its customers with long-term contracts (increases yA ), it
becomes more aggressive on its turf in the second period, as cuts in its
second-period price a do not reduce revenue from locked-in consumers.23
Since changes in yA do not change firm B’s prices on firm A’s turf, increases
in yA lead both firms to set lower prices. Moreover, the monotone hazard
rate condition implies that the slopes of the second-period reaction curves
are less than 1, so increases in yA move the switchpoint yA to the right,
which means fewer agents switch. Hence, if the firms use any long-term
contracts at all, there will be less switching than with short-term contacts.
Fudenberg and Tirole show that on the path of a symmetric equilibrium,
firms do use some long-term contracts, so there is less switching (and more
efficiency) than with short-term contracts. The intuition for this is as fol-
lows: by locking in some of its customers, a firm can commit itself to more
aggressive second-period pricing on its own turf, which induces a lower
second-period poaching price from Firm B. The marginal first-period A
purchaser plans to switch in the second period, so lowering B’s poaching
price lets Firm A charge a higher first-period price, which raises its profit.
22
Because this is a deterministic model, equilibrium prices must satisfy the no-arbitrage condition
A ¼ a+da, so that all consumes who plan to purchase from A in both periods are indifferent between
purchasing the long-term contract or a sequence of short-term ones. The results reported here rely on the
tie-breaking assumption that when the no-arbitrage condition holds, it is the customers who prefer A
most choose the long-term contract. Intuitively, there is an option value to the sequence of short-term
contracts, and this value is increasing in the probability that the customer decides to purchase B instead
of A in the first period. It seems plausible that this option value is higher for consumers with higher
values of y, and indeed this tie-breaking rule corresponds to taking the limit of models where the second-
period valuation is imperfectly correlated with first-period value, and the distributions are ranked by
first-order stochastic dominance in the first-period valuation. Some sort of tie-breaking rule is needed in
any deterministic model where there are multiple ways of purchasing the same consumption stream.
23
Note that firm A does not directly set yA ; instead, this switchpoint is determined by the condition
that equilibrium prices satisfy the no-arbitrage conditions A ¼ a+da and B ¼ b+db.
Conversely, a firm always uses some short-term contracts. Indeed, using

only short-term contracts dominates using only long-term ones whenever
first-period sales exceed the isoelastic point y : To see why, suppose that all
customers in the interval ½y; y buy a long-term contract from Firm A, and
that y 4y : Now suppose that Firm A deviates and offers only a short-term
contract in the first period, where the price a is set so that a ¼ A dbðy^ Þ
^
þdy ; where bðy Þ is Firm B’s poaching price when none of Firm A’s cus-
tomers have a short-term contract. This price has been chosen so that a
consumer of type y gets exactly the same utility from purchasing A in the
first period at price a and then buying B at the poaching price as it received
from purchasing the long-term contract from A, and since the change does
not affect competition on Firm B’s turf it leads to the same first-period
cutoff.24 Moreover, firm A would receive exactly the same payoff as with the
long-term contract by offering a second-period price on its turf of a00 ¼
^ Þ y ; as this price will induce all of its first period customers to pur-
bðy
chase from it again. However, when y 4y ; this pricing is more aggressive
than is optimal, and firm A does strictly better by raising its second-period
price, even though this leads some customers to switch.
Fudenberg and Tirole go on to show that the equilibrium they construct
remains an equilibrium when more general contracts are allowed, but they
do not discuss uniqueness, and it is an open question whether more general
contacts can lead to qualitatively different outcomes. Moreover, as with the
analysis of short-term contracts, the MHR condition does involve a loss of
generality; the effect of long-term contracts with the sorts of distributions
studied by Esteves (2004) and Chen and Zhang (2004) is open as well.
3.4 Switching costs
To conclude this section, we return to the case of short-term contracts to

compare the impact of purely information-based duopoly poaching with
price discrimination in the presence of switching costs. These costs are real
social costs in, e.g., complementary equipment or in learning how to use the
product; as such they differ from ‘‘pecuniary’’ switching costs such as can-
cellation fees.
Before addressing price discrimination, we briefly discuss the forces
present in models of switching cost without price discrimination. (For ex-
tended surveys of the switching costs literature, see Klemperer, 1995; Farr-
ell and Klemperer, 2004.25) In two-period models such as Beggs (1989) and
Klemperer (1987a), all consumers are locked-in in the second period, while
24
It is easy to see that there is an equilibrium with the same cutoff. Fudenberg and Tirole prove that
(under their tie-breaking rule) any profile of first-period contracts leads to a unique first-period cutoff.
25
For early papers on switching costs see also, for example, von Weizsacker (1984), Klemperer
(1987b), Farrell and Shapiro (1988, 1989). For the case of endogenous switching costs see, for example,
Caminal and Matutes (1990) and Kim et al. (2001). For a recent survey on information technology and
switching costs see Chen and Hitt (2005).
none are in the first. Second-period lock-in leads second-period prices to be

higher than without switching costs, while first-period prices are lower, as
firms compete for the rents from locked-in customers. Finally, consumers in
the first period foresee being locked-in in the second period, and become
less price sensitive, which is a force toward higher prices in the first period.
To illustrate these forces, we will use a simple two-period model. Each
firm sells a fixed and given product in each of the two periods. Each con-
sumer buys at most one unit in each period. Consumers are uniformly
distributed along a Hotelling segment, whereas firms are located at the
extremes of the segment (as in the previous subsection). Transportation
costs are t per unit and production costs are zero. A fraction s of the
consumers who buy from Firm i in the first period incur a high-switching
cost if buying from Firm j6¼i in the second period (so that they never switch
firms in the second period). The parameter s can then be seen as an index of
switching costs. The remaining consumers, in fraction 1s, have zero-
switching costs.
Given these assumptions, we can start by determining second-period de-
mand for each firm. Let qi1 be the distance to Firm i of a consumer with
switching costs that is indifferent in the first period between the two firms
(note qi1 ¼ 1 qj1 ). Then Firm i is guaranteed a demand of sqi1 in the second
period from the consumers that have switching costs. The total demand in
the second period for Firm i is then sqi1 þ ð1 sÞððt þ pj2 pi2 Þ=2tÞ; the
unique second-period equilibrium prices are pi2 ðqi1 Þ ¼ t½1 þ sð2qi1 1Þ
=3=ð1 sÞ; and the second period equilibrium profit for Firm i as a func-
tion of qi1 is pi2 ðqi1 Þ ¼ t=ð2ð1 sÞÞð1 þ ðsð2qi1 1Þ=3ÞÞ2 : This illustrates a first
effect of switching costs. Consumers that bought initially from one firm
would continue to prefer that firm, and in addition have now a more intense
preference due to the switching costs. This would then decrease the demand
own-price sensitivity in the second period, which would lead to greater
prices and profits in the second period.
Working backwards to the first period, consumers without switching
costs behave exactly as in the static case, because their decisions do not
affect what happens in the second period. Consider now the decisions of the
consumers that have switching costs. For the marginal consumer buying
product i, denoted by qi1, the total cost of buying product i is pi1 þ tqi1 þ
dðpi2 ðqi1 Þ þ tqi1 Þ; while the total cost of buying product j is pj1 þ tð1 qi1 Þ þ
dðpj2 ðqi1 Þ þ tð1 qi1 ÞÞ: Indifference between buying product i and j leads then
to qi1 ¼ ð1=2Þ þ 3ð1 sÞðpj1 pi1 Þ=ð2t½3ð1 þ d sÞ dsÞ and a total demand
in the first period of
1 3ð1 sÞðpj1 pi1 Þ pj1 pi1

q̄i1 ¼ þs þ ð1 sÞ . (8)
2 2t½3ð1 þ d sÞ ds 2t
If consumers are myopic, or there are no switching costs, this reduces to
the static Hotelling demands. Equation (8) illustrates a second effect of
competition with switching costs: Switching costs and forward-looking

consumers make the first-period demands less price sensitive because the
marginal consumers realize that by buying one product they will be locked-
in and pay a higher price in the next period. This is a force toward higher
equilibrium prices in the first period.
The firms set first-period prices to maximize the total value of profits,
pi1 q̄i1 þ dpi2 ðqi1 Þ: This maximization illustrates a third effect of switching
costs. In order to get higher profits in the second period, firms charge lower
prices in the first period to increase qi1.
This is a force toward lower prices and lower profits. In this particular
problem this maximization by each firm yields unique first-period equilib-
rium prices pi1 ¼ pj1 ¼ tð1 þ dÞ=½1 þ dð1 s=3Þ:
In general, which effects dominate (for lower or higher profits) will de-
pend on the particular characteristics in the market. In the particular ex-
ample above, equilibrium profits are higher with switching costs. An
example where it goes the other way can be obtained if consumers have
small switching costs, change preferences from period to period, and are not
too patient. Beggs and Klemperer (1992) look at the impact of large
switching costs on the MPE of an infinite horizon duopoly model with
uniform pricing. Each period, a fraction u of new consumers enter the
market with horizontally differentiated preferences that are fixed over time.
Once a consumer purchases from a firm it is unable to purchase from its
rival in the future.26 In this model, firms use a single price both to exploit
locked-in consumers and to attract new ones, so the effects of switching
costs on prices are less obvious; Beggs and Klemperer show that switching
costs increase prices in symmetric equilibria of the affine MPE that they
consider.
In some markets, switching costs can be created endogenously by the
competing firms by putting incompatibility features in its products. This
possibility may end up making all firms worse off in equilibrium (e.g.,
Cabral and Villas-Boas, 2005). Nilssen (1992) distinguishes between switch-
ing costs that are incurred each time a consumer changes supplier, and
‘‘learning’’ costs that are incurred each time a consumer uses a supplier for
the first time. Nilssen argues that a greater relative size of switching to
‘‘learning’’ costs leads to higher prices for the loyal consumers, and lower
introductory prices.
Turning to our main interest of behavior-based pricing, we focus on the
model of Chen (1997), which is a two-period, two-firm model that is very
similar to that of Section 3.1, except that all consumers are identical in the
first period, and that after making their first-period purchases, each con-
sumer privately observes a switching cost s. As we will see, the main differ-
ence with the work discussed above is that second period prices on the two
26
As in Taylor (2003), discussed below, the model abstracts from the determination of initial market
shares, and takes these as exogenous.
‘‘turfs’’ are independent of the relative sizes of these two markets. We will
then discuss Taylor (2003) who extends Chen to oligopoly, multiple periods,
and switching costs that are correlated over time, and conclude with a brief
mention of some other related work.
Following Chen, assume that all consumers have common value v for each
of the two goods, and that their switching costs are distributed uniformly on
an interval ½0; s̄: In the second period, a consumer will switch from Firm
A to Firm B if the difference in prices a b^ is greater than his switching
cost, so sales on Firm A’s turf will be xð1 Gða bÞÞ ^ and xGða bÞ ^ for
Firms A and B, respectively, where x is the size of Firm A’s turf and G is the
cumulative distribution function for the switching costs. Since the size of
Firm A’s turf simply has a multiplicative effect on second-period profits, it
clearly has no impact on second-period pricing or sales, at least at interior
equilibria where both firms have sales on A’s turf.27 Intuitively, the fact that
a customer bought from Firm A last period tells us nothing at all about his
preferences, except that the customer must now pay the switching cost to use
B, so the size of Firm A’s turf has no bearing on second-period competition.
This is in contrast to the models of horizontal differentiation we considered
earlier, where if firm A has a larger first-period market share it knows that
the consumers in B’s turf have a stronger preference for B, and so Firm A is
more aggressive on Firm B’s turf as Firm B’s first-period sales decrease. For
this reason, we suspect that adding a small amount of horizontal differen-
tiation to the switching cost model would make the second-period prices
respond to market shares.
With the uniform distribution, each firm charges second-period prices c þ
2s̄=3 and c þ s̄=3 on its own and the rival’s turf respectively, where
v is assumed larger than c þ s̄; firms sell to 2/3 of their old consumers and 1/3
of their rivals, so second period profits are ð4s̄=9Þx þ ð1=9Þs̄ð1 xÞ ¼ s̄=3ðx þ
1=3Þ and ðs̄=3Þð1 x þ ð1=3ÞÞ; for Firms A and B, respectively. Because the
first period product is completely homogenous, and second-period profit is
increasing in market share, the first-period prices will be below cost: at the
profile where both firms charged marginal cost, and so have second-period
profit of 5s̄=18; either firm would gain by undercutting slightly, capturing the
whole market, and having second-period profit 4s̄=9: In fact, Chen shows
that the unique subgame perfect equilibrium has first-period prices of c
ds̄=3; at this point cutting price would incur a large enough first-period loss
to offset the second-period gain. Thus, the conclusion that prices rise over
time extends from switching-cost models without targeted pricing to switch-
ing-cost models with behavior-based pricing. This prediction is in contrast to
that of the FT model of short-term contracts, where prices rise over time.28
27
Chen shows that the equilibrium is interior; Taylor extends this finding to distributions G such that
both G and 1-G satisfy the MHR condition.
28
Of course the dynamics of prices are different in stationary infinite-horizon models such as Villas-
Boas (1999).
If firms cannot observe the consumers’ past purchases, then firms with
larger first-period sales will price less aggressively in the second period.
Chen shows that this would lead to less aggressive first-period pricing, so
that, as in FT, firms are made worse off when they can both engage in
targeted pricing.29 Moreover, consumers need to forecast first period sales
to know second-period prices, and the assumption of homogenous con-
sumers means that the model may have multiple equilibria.
As noted above, Taylor extends Chen’s analysis in several ways. To
simplify the analysis, he also assumes that consumers are already ‘‘as-
signed’’ to one of the firms at the start of the first period. For this reason,
first-period demand is very different than in Chen’s model, and maintaining
the rest of Chen’s set-up, first-period prices are now above marginal cost,
and second-period market shares depend on the initial conditions; prices in
the second period, being independent of market share, are the same as in
Chen.
Taylor extends this analysis to multiple periods, finding that prices in the
two markets are constant over time until the last period. This is intuitive:
only the most recent purchase matters for the evolution of switching costs,
so all periods before the last are strategically similar (given the assumption
that consumers enter the game already assigned to a firm). More surpris-
ingly, moving from two firms to three makes a substantial qualitative
difference: when there are at least three firms, at least three of them offer
marginal cost pricing to other firm’s customers. The reason that three is the
key number here is that with three firms, there are two firms competing to
get customers from each other firm, so that there is Bertrand competition
for the switchers. This insight suggests that it would be interesting to study
information-based price discrimination in models with three or more firms;
this will be complicated by the need to consider a richer specification of
preferences, with a two-dimensional taste parameter y. As usual with
differentiated products, we would not expect prices to be driven to marginal
cost, but new and interesting features could emerge.
Finally, Taylor considers a two-period model with two types of consum-
ers, those whose switching costs tend to be low and those whose costs
tend to be high. Here, a customer who ‘‘switches’’ in the first period is
thought on average to have lower switching costs, so that agents who switch
will be offered a lower price by their first-period supplier than agents who
buy from that supplier without switching. It would be interesting to extend
this analysis to more than two periods. In that case, consumers will be all
the more concerned about their ‘‘reputations,’’ and the impact of being
known as a low-cost switcher may be ambiguous, as firms may wish to
avoid ‘‘recruiting’’ consumers who are likely to soon move on to another
brand.
29
Chen analyzes one of the equilibria for the uniform-price model, we do not know whether there are
others.
In addition to these papers, we should mention the paper by Shaffer and

Zhang (2000) that looks at a static game corresponding to the last period of
the sort of two-period model studied above, with the additional feature that
switching may be more costly in one direction than in the other (see also
Shaffer and Zhang, 1995). With symmetric-switching costs, firms always
charge a lower price to their rival’s consumers, but this need not be true
when switching costs are sufficiently asymmetric. More recently, Dobos
(2004) analyzes a model that combines horizontal differentiation in the first
period, switching costs in the second, and network externalities in both; he
finds that profits are decreasing in the size of the network effect, as this
effect leads to more aggressive first-period pricing.30
4 Behavior-based pricing with multiple products, and product design
So far we have been assuming, for the most part, that each firm produces
a single good. We now consider cases where each firm may produce mul-
tiple versions of the same product. Even in the case where the set of goods is
fixed, this leads to interesting forms of behavior-based pricing, such as price
discounts for consumers who are upgrading as opposed to new purchasers.
In addition, there are the questions of how many different goods a firm will
choose to sell, and (assuming it has this choice) what their characteristics
will be.31
The literature on behavior-based pricing and multiple goods has studied
two rather different sorts of goods and demand structures. Fudenberg and
Tirole (1998) and Ellison and Fudenberg (2000) studies ‘‘upgrades’’ in
models of vertical differentiation, where all customers agree that one good
is better than the other; these models study only the monopoly cases. Thus,
these papers are most closely related to the literature we discussed in Sec-
tion 2. In contrast, Zhang (2005) studies endogenous product lines in a
Hotelling style duopoly model of horizontal differentiation that is similar to
the model of Fudenberg and Tirole (2000) except for the assumption of
quadratic ‘‘transportation costs.’’ We focus on these two sorts of models,
and do not discuss the related literature on the monopolist’s profit-max-
imizing menu of goods and prices in a static model.32 We do, however,
discuss the papers of Levinthal and Purohit (1989), Waldman (1996), and
Nahm (2004), which study the introduction of a new product in models
with anonymous consumers and a frictionless second-hand market.
Although behavior-based pricing is not considered in these papers, the
30
His model is an extension of Doganoglu and Grzybowski (2004), who consider the same preferences
but without price discrimination. Villanueva et al. (2004) investigates the strategic effects of firms
considering the life time value of customers.
31
This latter question can also be asked when each firm is only allowed to produce a single good, but
that question does not seem to have been explored in the literature on behavior-based pricing.
32
See Mussa and Rosen (1978) and Deneckere and McAfee (1996) for discussions of the way the
monopolist’s desire to extract surplus leads to distortions in the product line.
analysis of the anonymous case is an important benchmark for the effects of

behavior-based pricing.
4.1 Upgrades and buybacks with an anonymous second-handmarket
In this subsection and the next, we discuss the two-period model of

Fudenberg and Tirole (1998). We begin with the case of anonymous con-
sumers and a frictionless second-hand market, which corresponds to the
market for textbooks, and is also a useful benchmark for evaluating the
impact of behavior-based pricing. As noted above, behavior-based pricing
is impossible when consumers are anonymous, just as it is in the durable-
good models of Section 2. Indeed those models can be viewed as a special
case of this one, because whether or not there is a second-hand market that
makes no difference given that there is a single perfectly durable good and
all consumers enter the market at the beginning and remain until the end.
In period 1, the monopolist produces a low-quality version of a durable
good; this good is denoted L. In Period 2, the monopolist can produce both
L and an improved version H. These goods are produced under constant
returns to scale, with cost cL for L and cH ¼ cL+cD for good H, where
cDZ0.33 There is a continuum of consumers, indexed by yA[0,1]; a type-y
consumer has utility yV+I, where I is her net income, and V ¼ VL or
VH ¼ VL+VD, VD>0 depending on whether she consumes L or H. This is a
fairly standard demand structure, and it is easy to work with, but involves
some loss of generality, as can be seen from the fact that in a static model
the monopolist will not offer both goods if their costs are the same.34
Following the paper, we assume that VL>cL and VD>cD. To simplify, we
also assume that the distribution of types is uniform; the paper assumes that
the distribution has a continuous density that satisfies the monotone hazard
rate condition. The firm and the consumers use the common discount factor d.
Because the monopolist lacks commitment power, we solve the problem by
working backwards from the second period. The solution here depends on
the stock x1 of L that is already in the hands of the consumers, but the
assumptions of anonymity and a frictionless second-hand market mean that
we do not need to worry about which consumers bought the product, and
indeed we can suppose that all old units are sold in the second-hand market,
with some of them possibly repurchased by their original owners. The form
of the utility function implies that there will be three (not necessarily non-
empty) segments of consumers in the second period: types in the interval
[0, yL] do not consume; types in [yL, yH] consume good L, and types in [yH, 1]
consume good H. The market price of good L is then the value pL ¼ yLVL
that makes yL indifferent between purchasing L and not purchasing, while
33
The production cost of upgrading an L unit to H is the same as that of making H from scratch.
34
This can be seen by considering equation (9) when x1 ¼ 0.
the price of H makes yH indifferent between purchasing H or purchasing L,

so pH ¼ pL+yHVD.
If the mass of consumers consuming good L is greater than the existing
stock, i.e., if yHyL>x1, the monopolist is a net seller of L in Period 2;
when the reverse inequality holds, the monopolist is engaged in ‘‘buy-
backs,’’ and when yHyL ¼ x1, the monopolist is inactive on the L market.
Each of these regimes can arise for some values of the first-period stock;
moreover, each of these regimes can arise for an open set of parameters in
the full equilibrium, where x1 is determined by the monopolist’s first-period
sales.
When yHyL>x1, so there are net sales, the monopolist has second-
period profit
Ynetsales
2
¼ ðyH yL x1 ÞðyL V L cL Þ
þ ð1 yH ÞðyL V L þ yH V D cH Þ
¼ ð1 yL x1 ÞðyL V L cL Þ þ ð1 yH ÞðyH V D cD Þ. ð9Þ
Thus, it is as if the monopolist faces two separate, unlinked markets in
Period 2. All consumers above yL purchase L, with x1 of this coming from
the pre-existing supply. Separately, the monopolist supplies the ‘upgrade’ to
types above yH; this (fictitious) good has incremental cost cD and sells at
price yHVD. Thus, when the net-sales regime prevails, the monopolist sells
exactly the same amount of good L as it would if good H did not exist, and
sales of the old good follow the standard Coasian path discussed in Section
2. Similarly, price and sales in the upgraded market are not influenced by
x1. Thus, the first-order conditions for maximizing (9) are given by the
standard formulas:
yL V L cL 1 yL x1
¼
yL V L yL
and
yH V D cD 1 yH
¼ .
yH V D yH
When yHyLox1, so there are buy-backs, we suppose that the mon-
opolist has no use for Qrepurchased units. Thus, the payoff function in this
net sales
region is the same as 2 except that cL is replaced by 0. That is,
Ybuybacks
2
¼ ð1 yL x1 ÞðyL V L Þ þ ð1 yH ÞðyH V D cH Þ.
Note that once again the ‘‘upgrade market’’ decouples from the market
for L. However, the price for L (given x1 and the buy-back regime) is lower
than it would have been if H had not been introduced, for now the ‘‘effec-
tive cost’’ of L is zero. Thus, while the monopolist’s second-period payoff is
continuous at the boundary between net sales and buybacks, it has a kink
there, as the effective marginal cost changes from 0 to cL. For this reason,
the ‘‘inactive’’ regime is the equilibrium for a range of values of x1. In this
regime the constraint yHyL ¼ x1 is binding, and the markets do not
decouple.
Fudenberg and Tirole show (in Proposition 2) that there are numbers
0 x1 ox̄1 o1 such that when xox 1 the solution has net sales, and is ex-
Q net sales
actly the solution to maximizing ,2 while ignoring
Q net the net-sale con-
straint. For x1 ox1 ox̄1 the solution that maximizesQ buybacks2
sales
has negative
sales of L, while the solution that maximizes 2 has net sales; here,
the second-period equilibrium is at the kink. Finally, for x̄1 ox1 the solution
has buybacks. Moreover, pL is a continuous and weakly decreasing function
of x1, and yL+x1 is continuous and weakly increasing.
What we are really interested in is the full equilibrium of the two-period
game. Fudenberg and Tirole show that setting a first-period price of p1
leads to sales to all types above the cutoff value y(p1), so that the stock on
hand at the start of the second period is x1 ¼ 1y1. The monopolist’s
problem is thus to maximize the discounted sum of first- and second-period
profits, taking into account the way that first-period sales determine the
second-period regime. The following examples show that each regime can
arise for some parameter values, and give a flavor of when they might be
expected, but stop far short of a characterization of when each regime
prevails.
First, if cL Q
¼ cH ¼ 0, then there are always buybacks. To see this, note that
buybacks
in this case 2 simplifies to ð1 yL x1 ÞðyL V L Þ þ ð1 yH ÞðyH V D Þ;
so that the optimum in the H market is yH ¼ 1/2, which is the same as the
optimum in the L market when x1 ¼ 0.35 Thus, there are buybacks when x1 is
close to zero, and as x1 increases, yH is unchanged while yL+x1 increases, so
buybacks (which are x1+yLyH) increase as well.
Next, net sales occurs whenever cL ¼ 0 and cD is almost as large as VD, so
that the new good is sold to only the highest value consumers. This is true
for any value of the discount factor, but it is easiest to see for the case d ¼ 0,
as here first-period output is the amount sold by a static, zero-cost mon-
opolist, which is 1/2 for the uniform case considered here, while the first-
order condition for yL in the net sales regime simplifies to yL ðx1 Þ ¼
ð1 x1 Þ=2; so that yL ð1=2Þ ¼ 1=4o1=2 and the second-period solution fol-
lowing x1 ¼ 1/2 indeed has net sales.
Finally, the equilibrium will have neither sales nor buybacks if cD ¼ 0 and
cL ¼ cH is very close to VL. Intuitively, when cD ¼ 0 there will be no pro-
duction of the old good in Period 2, and because costs are close to VL, there
will be very little production of L in the first period, so x1 is small, which
makes buybacks less likely.36
35
This equality does not depend on the uniform distribution, but rather on the assumptions that costs
are zero and quality enters the demand function multiplicatively.
36
The formal argument uses continuity and monotonicity properties.
At this point, we should mention the related work of Levinthal and

Purohit (1989) and Lee and Lee (1994) on monopolists with an anonymous
second-hand market. Levinthal and Purohit consider a model with costless
production, where the second-period market is described by a pair of linear
demand curves, and the rental prices of each generation are equally affected
by an increase in the output of the new generation.37 In their model, buy-
backs are only optimal when the firm is sufficiently patient, and otherwise
there are net sales.38 Lee and Lee suppose that the monopolist is unable to
sell or buy units of the old product in period two.
4.2 Upgrades and buybacks with non-anonymous consumers
Fudenberg and Tirole go on to consider two other sorts of information

structures: ‘‘identified consumers,’’ where the firms know which consumers
purchased at Date 1, and ‘‘semi-anonymous consumers,’’ where consumers
can prove that they purchased if they wish to do so, but can also pretend
not to have purchased, which constrains the price to new customers to be
no lower than the ‘‘upgrade price’’ offered to old ones. Following the paper,
we now assume that V L 4dV H ; which implies that any first-period price
induces a cut-off y1(p1) such that the consumer of type y purchases when
y4y1 ðp1 Þ: This assumption is stronger than one would like, but we are not
aware of a weaker condition that guarantees a first-period cutoff, nor of
related analyzes that allow for disjoint sets of consumers to purchase in the
first period. We also assume that cL ¼ 0, and that cH ¼ 0 as well; the paper
does not make this last assumption.
We begin with the case of identified consumers. Here, the monopolist
faces two distinct second-period markets, patrons, and non-patrons. On the
patron’s market the monopolist maximizes ð1 yu Þ yu V D subject to yuZy1,
so yu ¼ max ½1=2; y1 ; and pu ¼ yuVD. On the non-patron’s market, the
monopolist will sell good H to consumers with values between yH and y1,
where yH is chosen to maximize (y1yH)yHVH; the solution to this is
yH ¼ y1/2, with price pH ¼ VHy1/2. Comparing the objective functions in
the two markets let us identify two competing effects. First, non-patrons of
any given type have more to gain from purchasing because they have a
lower payoff without a purchase; this ‘‘reservation utility effect’’ pushes the
upgrade price to be lower than the price to new consumers. On the other
hand, former customers have higher types; this ‘‘ratchet effect’’ means that
non-patrons should get lower prices. These effects will help us understand
when the identified and semi-anonymous cases coincide.
37
Note that in the Fudenberg and Tirole model, the price of L is yLVL and so depends only on the
supply of good L.
38
Their results imply that the inactive region never occurs without pre-commitment. They show that in
some cases the monopolist can gain from a first-period commitment not to produce L in the second
period, just as it could if good H did not exist.
Fudenberg and Tirole show that in equilibrium the monopolist chooses

y1>1/2, so that all old patrons upgrade, and there is no ‘‘leapfrogging’’ of
lower-value consumers past higher-value ones.39 Moreover in this case the
second-period upgrade price is y1VD.
In the semi-anonymous case, the payoff functions in the two markets are
the same as with identified consumers, but the markets are linked by the
customers’ incentive compatibility constraint, which requires that purpH,
The calculations above show that this constraint is slack, and the two
solutions coincide, if and only if VDrVH/2, or equivalently, if VDrVL, i.e.,
if the size of the innovation is not too large. The intuition for this is that for
large innovations, upgrading is very attractive to high-value types, so the
‘‘ratchet effect’’ dominates the reservation utility effect; this is true for
general distributions and not just the uniform.
Finally, Fudenberg and Tirole show that with costless production the
monopolist’s profits are higher under anonymity than with identified con-
sumers. With costless production, when y1Z1/2 (which is the relevant
range) the anonymous-market solution is for customers between y1/2 and
1/2 to consume L, and customers from 1/2 up to consume H; with identified
consumers, the monopolist sells H in the second period to all types above
y1/2. The commitment solution is to sell H to consumers above 1/2, and
nothing at all to the others; the anonymous solution is closer to this out-
come, and so yields higher payoffs. The point is that the presence of the
second-hand market leads the monopolist to sell less of H in Period 2,
which helps alleviate the commitment problem in Period 1. (Note that this
finding does not immediately extend to the semi-anonymous case, except
for parameters where it coincides with the solution with identified consum-
ers: The no-arbitrage constraint cannot help the monopolist in the second
period, for any given first-period outcome, but the constraint could have an
impact on first-period play.)
4.3 Endogenous innovation
Waldman (1996) and Nahm (2004) analyze endogenous innovation in the

anonymous case. Waldman supposes that there are only two types, yL and
yH, with yLVLocL. This means that the firm would not sell to the low types
in a one-period model, and moreover in the absence of the new good the
firm would not produce in Period 2. That is, the assumed demand structure
means that the firm would not face the usual Coasian commitment prob-
lem. However, the sale price of the low good in Period 1 is decreasing in the
probability that the firm will introduce an improved good H in the second
period, and Waldman shows that the firm does face a commitment problem
with respect to introducing the improved good.
39
Leapfrogging can occur when cH>0, as here the monopolist will not induce all old patrons to
upgrade but it will sell H to non-patrons so long as cH is not too high.
Nahm points out that this conclusion relies on the assumed demand
structure. In a two-type model with yLVL>cL, the price of good L will fall
over to yLVL in the second period whether or not the new good is intro-
duced, and the firm does not face a commitment problem with respect to
introducing the new good. Nahm goes on to investigate the incentives
for introducing the new product in a model of Section 4.1, where in between
Period 1 and Period 2 the firm spends resources on R&D, which in
turn determines the probability that the high quality good is available in
Period 2.
As we saw above, in the net-sales case, the second-period price of good L
is the same whether or not H is introduced, and investment in R&D only
influences payoff in the ‘‘upgrade’’ market. Hence, the monopolist does not
face a time-inconsistency problem with respect to R&D, and it chooses the
same level of investment that it would chose if it could commit to the choice
in Period 1. However, in the inactive and buy-back regimes, the second
period price of L is lower if H is introduced than if it is not. Hence, to
maximize first-period sales and overall profit, the monopolist would benefit
from a commitment that limited its R&D.
Ellison and Fudenberg (2000) analyze the semi-anonymous, costless-pro-
duction case in a model intended to correspond to markets for software. It
is very similar to that discussed above, with one good in Period 1 and the
possibility of producing an improved version in Period 2; the main differ-
ence is that their model includes (positive) network externalities. In their
model, consumers incur set-up or training costs each time they adopt or
upgrade their software, and differing versions of software are backwards
but not forwards compatible, so that users of the newest version of the
software enjoy the largest network benefits. In their dynamic model,40
consumers are ex-ante identical, but not all of them are present in the first
period. They show that the monopolist suffers from a commitment problem
that can lead it to introduce upgrades that are not only welfare-decreasing
but also lower its own overall present value. The idea of this result is simple:
in the second period the monopolist may prefer to sell upgraded, higher-
value software to new consumers, but this forces the old consumers to either
incur the costs of learning to use the new version or settle for smaller
network benefits due to incompatibility with new consumers. This can lead
to a loss of first-period profits that outweighs the second-period gain.
As it is common in models of network externalities, consumers’ pur-
chasing decisions have the flavor of a coordination game, and can have
multiple equilibria. Ellison and Fudenberg assume that in the second pe-
riod, new consumers coordinate on the equilibrium that is best for them,
and consider two different equilibrium-selection rules for the old consumers
40
The paper also considers a static model with a continuum of types, and shows that even with
commitment the monopolist may introduce socially inefficient upgrades to help it price discriminate.
That model is less closely related to the themes of this survey.
who are deciding whether to upgrade; in either case there is a region of the
parameter space where the monopolist introduces the upgrade when the
social optimum would be to sell only the old good in both periods.
4.4 Endogenous location choice in duopoly
Waldman, Nahm, and Ellison and Fudenberg consider a monopolist

whose innovation decision is whether to introduce or research an improved
version whose characteristics are fixed. Zhang (2005) considers endogenous
location choice in a two-period poaching model. The idea is that the rise of
flexible manufacturing makes it cheaper for firms to customize products to
various clienteles, and since purchase decisions convey information, firms
might want to design one product for its established customers and another
for those they are trying to poach from a competitor.
The information structure and institutional assumptions are the same as
in the short-term contracts section of Fudenberg and Tirole (2000), but the
payoff functions are different: Consumers are uniformly distributed on the
interval [0,1], while firm locations are endogenous, and transportation cost
is quadratic in distance: the utility of for type y of consuming a good at
Location a is v tðy aÞ2 ; where the reservation utility is assumed high
enough that in equilibrium all consumers purchase. At the start of the first
period, the two firms simultaneously choose Locations a and b, respectively,
and in the second period, each firm can produce products at two (or more)
locations, and offer different prices and goods to consumers whose first-
period actions were different.
In the base model, designing new products is costless.41 If the firms and
consumers have the same discount factor, or more generally, if the con-
sumers are sufficiently patient compared to the firms, the equilibrium is for
the firms to split the market in the first period, and for each firm to offer
two new and distinct models in the second period, with Firm A choosing ao,
an and firm B choosing bo, bn, where ‘‘o’’ and ‘‘n’’ are for old and new
consumers, respectively. However, as in the poaching models discussed in
Section 3, firms do better when they have less first-period information, and
if firms are sufficiently patient compared to consumers then the first-period
pure-strategy equilibria are asymmetric, with one firm capturing all of the
market, so that first-period purchases reveal no information.
To understand these results, we explain the outcome in the second-period
markets for types that have been revealed to lie in an interval [Z, Z+L],
which is the same as in a static model with these types as the single market.
It is interesting to note that although introducing varieties is costless, and
firms are allowed to introduce as many as they wish, in equilibrium each
firm only sells a single product. This fact is closely related to the fact that if
41
The paper speculates briefly about the case where innovation costs are such that firms introduce a
single new product in Period 2.
each firm can only introduce a single product, they will choose locations
outside the support [Z, Z+L] of the distribution of consumer types if such
locations are allowed, and at the boundaries of the distribution if it is not.
Intuitively, firms face a trade-off between locating near the center of the
distribution, which increases profits holding the opponent’s price fixed, and
locating toward the edges, which lessens price competition and raises the
opponent’s equilibrium price. With quadratic transportation costs and the
uniform distribution, the strategic effect dominates until the locations are
well outside the support of the distribution of types.42 The fact that the
optimal locations for a single product are outside of the support provides
an intuition for why introducing a second variety would not be helpful: if
the new variety is to provide an increase in efficiency, it must be closer to
the opponent’s location, but this would provoke the price competition that
the first location was chosen to avoid.
Now, consider firms simultaneously choosing locations and prices in two
different second-period markets, corresponding to the first-period purchase
of the consumers. The previous paragraph explains why each firm will
choose a single product for each market; in general, these products will
be different, and a better match for the tastes of the market they are de-
signed for.
Now we turn to the consumer’s decision in the first period. As in Section
3.1, the first-period decisions of consumers will generate a cut-off rule, so
that first-period sales identify two intervals of consumers, corresponding to
each firm’s turf. Also as in that model, the consumers who are near the
cutoff in the first period switch suppliers in the second period, and increased
consumer patience makes first-period demand less elastic. Consumers ben-
efit most when they are identified as being in a small interval, as this leads to
intense price competition; the firms’ second-period profit is highest when all
consumers purchase from the same firm in the first period, so that the
purchases reveal no information.
Working backwards to the firm’s first-period decisions, Zhang shows that
when consumers and firms are equally patient, and more generally if the
consumers are sufficiently patient compared to the firms, the first-period
outcome is symmetric, with Firms A and B located at equal distance from
the market center, and each taking half the market. In the second period,
each firm introduces two new products, one for each segment of the first-
period market. On the other hand, if firms are patient and consumers are
myopic, the firms are able to avoid segmenting the first-period market, and
their combined profits increase.
42
Economides (1986) studies the Hotelling location-price game where duopolists each offer one
product, with a uniform distribution on types, and transportation costs proportional to ta. He shows
that for aA[1.26, 1.67] the firms locate within the distribution of types, while for aA[1.67, 2] they locate
at the endpoints. (He constrains them not to locate outside of it.) For aA[1, 1.26] there is no pure-
strategy equilibrium; see D’Aspremont et al. (1979) for the linear and quadratic cases.
Zhang’s results on product design seem to reinforce the idea that cus-
tomer recognition leads to more intense competition, and lower profits. It
would be interesting to understand what happens if we have a longer time
horizon (possibly with changing consumer tastes), and what would happen
under product choice and monopoly, with customized product adverti-
sing (and where this customized advertising could also depend on past
behavior).
5 Related topics: privacy, credit markets, and customized pricing
This section briefly discusses the issues of consumer privacy protection,

pricing in credit markets, and standard third-degree price discrimination
that is based on exogenous characteristics. We focus on the work of Thisse
and Vives (1988), Dell’Ariccia et al. (1999), Dell’Ariccia and Marquez
(2004), Taylor (2004a), and Calzolari and Pavan (2005) and also discuss
Pagano and Jappelli (1993), Padilla and Pagano (1997, 2000), and Taylor
(2004b).
5.1 Privacy
As we have seen, the efficiency consequences of BBPD are ambiguous, so

there is some reason to consider the impact of various regulations and firm-
based initiatives that protect consumer privacy.43 One interpretation of
consumer privacy is that firms cannot track consumers’ past behavior.44
Consumers that buy early may be recognized as consumers that value the
product highly, and then be charged a higher price in subsequent periods. In
this sense consumers are hurt by losing their privacy, they are charged
higher prices. As discussed above, consumers, if they are aware of this loss
of privacy, may be strategic in the earlier periods, and refrain from pur-
chasing the product, not to reveal their high valuation. This may give firms
an incentive to commit to privacy protection.
Taylor (2004a) uses a variation of the two-period model of Section 2 to
focus on the privacy issue. Consumers interact sequentially with each of two
firms, and each consumer’s valuations for the products of the two firms are
positively correlated, so that, if the second firm is able to observe that a
43
This ambiguity should not be a surprise in view of previous results on related issues. Hirshleifer
(1971) noted that the efficiency impact of information acquisition is ambiguous when markets are
incomplete. This holds in particular for firms acquiring more information about the characteristics of
each consumer. For example, Hermalin and Katz (2006) show that third-degree price discrimination
may be better or worse from a social point of view than second-degree price discrimination. Wathieu
(2004) argues that information about consumers may lead to inefficiently many products being pro-
duced, each at too low a scale. For a recent survey on the economics of privacy see Hui and Png (2005).
44
Upon realizing that Amazon was charging different prices for the same item, possibly based on
different purchase histories, some consumers showed concern about shopping there (Customers Balk at
Variable DVD Pricing, Computerworld, September 11, 2000, p. 4).
consumer bought from the first firm, then the second firm’s beliefs about the
valuation of that consumer for its product is higher than if the consumer
declined to purchase. Taylor assumes that the second firm is unable to
commit to its prices until after consumers interact with the first one. Privacy
is the case in which the second firm is not able to observe whether a con-
sumer bought or did not buy in the first period. Without privacy, the first
firm can sell the list of its customers, and allow the second firm to price
discriminate between the consumers that bought and did not buy from the
first firm.
If there is no privacy, the first firm sells the customer data to the second,
and consumers do not foresee that sale (in the context of Section 2 this is the
case when the consumers are myopic), then the first firm has a greater
incentive to charge higher prices in order to make the customer data more
valuable. If consumers foresee that the first firm is going to sell the customer
data to the second firm, then they strategically refrain from buying, which
makes the customer data being sold less valuable, and gives incentives for
the first firm to lower prices. Firms prefer the no-privacy case when con-
sumers are myopic, but prefer consumer privacy if consumers are able to
foresee that under no privacy their purchase information is going to be sold.
Taylor shows that welfare can be higher or lower under consumer privacy
depending on the demand elasticity.
Calzolari and Pavan (2005) consider the case where two principals
sequentially contract with a common agent, and where the upstream prin-
cipal can sell its information to the downstream principal. They assume that
the agent’s valuations with the two sellers are perfectly correlated, which is
more restrictive than Taylor’s assumption of imperfect correlation, but
otherwise their model is more general.45 As in Taylor, the second principal
posts its contract after the consumer has already decided whether to accept
the contract of the first firm. By selling information to the downstream
principal, the upstream principal may get some payment from the down-
stream principal (possibly due to greater efficiency, or less information rents
provided to the agent, in the downstream relationship), or appropriate any
rents of the agent in the downstream relationship that are generated by this
sale of information. Calzolari and Pavan identify three conditions under
which, if the upstream principal can commit not to disclose any information
(commitment to privacy) she will choose to do so. The first condition is that
the upstream principal is not personally interested in the decisions taken by
the downstream principal. In the context of Taylor (2004a) this is just that
the profit of the first firm is independent on the decisions taken by the
second firm. The second condition is that the agent’s exogenous private
information is such that the sign of the single crossing condition is the same
for both the upstream and downstream decisions. In the context of Section
45
Ben-Shoham (2005) extends the Calzolari and Pavan analysis to allow for imperfect correlation, and
also for imperfect (i.e., noisy or partial) revelation of information from the first principal to the second.
2, this condition is just that the valuation of a consumer type is the same
across products. In Taylor (2004a) the valuation for the product of the first
firm is positively correlated with the valuation for the product of the second
firm. Finally, the third condition is that the preferences in the downstream
relationship are additively separable in the two contractual decisions. In the
context of Section 2, or Taylor (2004a), this is immediately obtained be-
cause the second-period profit or utility is independent of whether there was
a purchase in the first period.
It is interesting to try to informally relate the first condition with the
two-period model in Section 2. Denote the first-period profit under disclo-
sure of information as a function of the first-period action a1 as pd1 ða1 Þ; the
first-period profit under privacy as a function of the first-period action as
pp1 ða1 Þ; the second-period profit under disclosure of information as a func-
tion of the first-period action as pd2 ða1 Þ; and the second-period profit under
privacy as pp : Note that in the model of Section 2, the second-period profit
under privacy is independent of the first-period action.46 In the context of
Section 2, the firm chooses its first-period action under disclosure of in-
formation to maximize pd1 ða1 Þþpd2 ða1 Þ (where the discount factor was set to
one). In Calzolari and Pavan, the upstream principal is able to receive a
payment for the disclosure of information from the downstream principal
in the amount of pd2 ða1 Þ pp2 : The upstream principal chooses then its ac-
tion under disclosure of information to maximize pd1 ða1 Þþ½pd2 ða1 Þ pp2 ;
which results in the same optimal action as in the model of Section 2.
Finally, note that in the model of Section 2 the firm chooses privacy if
and only if maxa1 pp1 ða1 Þþpp2 maxa1 pd1 ða1 Þþpd2 ða1 Þ; while in the context of
Calzolari and Pavan the upstream principal chooses privacy if and
only if maxa1 pp1 ða1 Þ maxa1 pd1 ða1 Þþ½pd2 ða1 Þ pp2 : It is immediate that pri-
vacy is chosen in both models in exactly the same conditions (no customer
recognition in the model of Section 2). So, even though in Calzolari
and Pavan there are two principals, in the case where the upstream
principal expropriates the informational rent from the downstream
principal, the model corresponds to single-principal models discussed in
Section 2.
Calzolari and Pavan (2005) also show that under the second condition, if
the upstream principal discloses information to the downstream principal,
the increase in the rent that has to be given to the agent always offsets any
potential benefit from the sale of information, or from a greater rent of the
agent in the downstream relationship. This is because, if information is
disclosed, the agent becomes more protective of his type and the upstream
principal does not have the possibility of using any distortion of the down-
46
Calzolari and Pavan allow for the second-period profit to be additively separable in the two con-
tractual decisions, and therefore to be also a function of the first-period actions even under privacy. This
possibility does not affect the argument above.
stream relationship contractual variable to help the agent reveal his type.
This then implies that when the upstream principal is not personally in-
terested in downstream decisions (the first condition), then there is no ad-
vantage in disclosing information and the optimal policy is committing
to full privacy. The paper then argues that each of these conditions is
necessary for the full privacy result, and that if one of the conditions does
not hold it may be optimal for the upstream principal to disclose infor-
mation to the downstream principal. In such cases, there are also situations
in which disclosure of information benefits all three players.
Taylor (2004b) considers a market situation in which firms first post
prices, and then decide on how much to screen the customers that demand
their product. The profit that a firm derives from a customer depends not
only on the price charged, but also on the cost of servicing that customer
which varies in the population (and that is also not known by the cus-
tomer). The amount of screening chosen by a firm allows that firm to
receive a noisy signal about the cost of servicing a customer. More screening
reduces the noise of the signal. In relation to the papers above, this paper
can be seen as looking at quantity discrimination, while the papers above
looked at price discrimination. Given that the cost of servicing a certain
customer depends on the a priori unknown characteristics of the customer,
this model matches well the market features of credit markets, discussed
below.
Consider the case in which the screening device sometimes misses ‘‘bad
news’’, that is, good news are always identified appropriately as good news,
but bad news are only identified as bad news with some probability less
than one. Then, one can obtain that competitive firms screen customers too
much. A firm’s incentive to screen customers is given by the difference
between the cost of servicing the costly customers and the price it is getting
as revenue from those customers, while the social incentive is the difference
between the cost of servicing the costly customers and the consumers’ val-
uation. As in a competitive market the price is below the consumers’ val-
uation, a firm’s incentive to screen customers is greater than the social
incentive. If the screening device is not very good, or the social cost of
servicing the costly customers is small, then it may be better not to allow
firms to screen (customers have privacy) and for firms to service all cus-
tomers. If rejected customers stay in the market and apply for the other
firms, the situation may become worse, with even more equilibrium screen-
ing, so that no screening (privacy) is even better from a welfare point of
view. Consumers can improve their situation (of too much screening) by
reducing the quantity that they purchase.
Another possibility is for firms to offer consumers the option of disclosing
their valuation or keeping it private. McAdams (2005) considers this case,
in which consumers who do not disclose their valuation pay a ‘‘sticker
price,’’ while consumers who allow the firm to learn their valuation pay a
fee to get a ‘‘customized price,’’ and where learning a consumer’s valuation
is costly to the firms. McAdams shows that there are parameter values such
that welfare can increase if the firms are required to offer the same prices to
all consumers (and consumers are forbidden to reveal their valuation/give
up their privacy).
5.2 Credit markets
In credit markets, lenders may learn about the ability of their borrowers,
their customers, to repay loans; this information can then be used by the
firms in the future loans to those customers. In this case what a firm learns
about its previous customers relates to the cost of providing the customer
with a given contact, as opposed to the customer’s willingness to pay, which
has been the focus of the work we have discussed so far. This feature is also
present in other markets, such as labor markets (information about em-
ployees), rental markets (information about tenants), insurance markets
(information about policy holders), and some forms of service contracts
(fussy customers take more time to service). Our presentation here is cast in
terms of credit markets because the existing literature has used this type of
markets as main motivation.
We start by discussing what happens in credit markets when lenders have
private information about their own previous borrowers, and then consider
the possibility and effects of lenders sharing their information. The presen-
tation is based in large part on Pagano and Jappelli (1993), Padilla and
Pagano (1997, 2000), Dell’Ariccia et al. (1999), and Dell’Ariccia and
Marquez (2004).47 Some of the discussion is also related to some of the
material presented in the privacy section above, in particular, Taylor (2004b).
Following Dell’Ariccia et al. (1999), consider a market with two com-
peting Lenders 1 and 2. Borrowers have to borrow $1 to invest in a project
that pays R with probability y, and zero with probability 1y. Borrowers
are heterogeneous on the success probability y, with cumulative distribution
function G(y) (density g(y)) on [0,1]. Furthermore, the borrowers are, in-
dependent of y, in one of three groups: either they are ‘‘new’’ borrowers,
and so no lender knows about the borrower’s y; or they are ‘‘old’’ bor-
rowers from Lender 1, so that Lender 1 knows y, but this is not known by
Lender 2; or they are ‘‘old’’ borrowers from Lender 2, so that Lender 2
knows y, but this is not known by Lender 1. Let l be the proportion of
‘‘new’’ borrowers (1l of ‘‘old’’ borrowers), and let ai be the proportion of
‘‘old’’ borrowers from Lender i. Dell’Ariccia et al. (1999) assume that a
lender is not able to distinguish between ‘‘new’’ borrowers and ‘‘old’’ bor-
rowers from the other lender, and that, lenders first simultaneously set
interest rates ri for the borrowers for whom they do not know y, and then
they set, also simultaneously, the interest rates ri y for the borrowers for
47
For a discussion of informed versus arm’s-length debt see also Rajan (1992).
whom they know y.48 The paper focuses on the analysis of this market
interaction, which can be seen as the second period of a two-period
model.49
Consider first the behavior of a Lender i with respect to its previous ‘‘old’’
borrowers. These borrowers have access to an offer from the other lender at
an interest rate rj. In order for Lender i to attract them, it has to offer them
at least an interest rate rj. The expected profitability of a borrower of type y
is then yrj1. Lender i then wants only to extend credit to the borrowers
that will generate positive expected profit, that is, for the borrowers with
y 1=rj : Lender i expected profits from its previous ‘‘old’’ borrowers is then
Z 1
ai ð1 lÞ ðyrj 1ÞgðyÞdy.
1=rj
Note that these expected profits from the lender’s previous borrowers are
independent of the lender’s interest rate to the ‘‘new’’ borrowers.
Consider now the profit of a Lender i from the borrowers that borrow
from that lender for the first time, given interest rates (ri, rj). Lender i gets
an expected profit from the ‘‘new’’ borrowers of lðri EðyÞ 1Þ if riorj, of
ð1=2Þlðri EðyÞ 1Þ if ri ¼ rj, and of zero if ri>rj. The expected profits for
Lender i of the ‘‘old’’ borrowers of the other lender, due to the poor quality
borrowers that are denied credit by the other lender, are aj ð1 lÞ
Gð1=ri Þ½ri Eðy=y 1=ri Þ 1:
Because of the discontinuity of the expected profits from the ‘‘new’’ bor-
rowers at ri ¼ rj, by standard arguments (for example, related to Varian
1980), one can show that the market equilibrium involves mixed strategies
in the interest rates ri and rj. One can also show that the lender with a
smaller share of the ‘‘old’’ borrowers, makes zero expected profits from its
new customers, while the lender with a greater share makes positive ex-
pected profits from this type of customers. This is because the lender with a
greater market share of ‘‘old’’ borrowers suffers less asymmetric informa-
tion, and lends to less poor quality ‘‘old’’ borrowers than the lender with a
smaller market share of the ‘‘old’’ borrowers. Dell’Ariccia et al. (1999) go
on to show that this equilibrium with two lenders is exactly the same as the
equilibrium with a third lender potentially entering the market, as this new
lender would prefer to stay out. This is because this potential entrant can-
not protect itself from the lower quality ‘‘old’’ borrowers from both firms.
As the incumbent smaller market share lender makes zero-expected profits,
48
These two assumptions are as in Villas-Boas (1999), discussed in Section 3. Sharpe (1990), in the
context of credit markets, and with borrowers choosing investment levels, makes the assumption that
lenders make first the offers to the borrowers that they know, and then, after observing the offer policies
(but not the actual offers), make offers to the borrowers that they do not know.
49
The appendix of the paper presents some analysis on the two-period model (without discussing if
forward-looking borrowers would play a role), and argues, as in Sharpe (1990), that the first period
competition is more intense because of the informational advantages the lenders enjoy in the second
period.
the new entrant would make negative profits if entering the market (have
a positive market share), and prefers to stay out. We have then that the
ability to recognize previous customers in credit markets leads to blockaded
entry.50
Dell’Ariccia and Marquez (2004) considers a variation of the model
above where only one lender has previous ‘‘old’’ borrowers, this informed
lender has higher costs of funds than the competitor, and y is uniformly
distributed on the segment [0,1]. The paper fully characterizes the mixed-
strategy equilibrium, and analyzes how the existence of this informed lender
affects the loan portfolio allocation. Greater information asymmetry leads
to higher interest rates as the informed lender takes advantage of its in-
formation advantage. Furthermore, as the competitor has lower costs of
funds, the informed lender concentrates more on its previous borrowers, as
competing for the ‘‘new’’ borrowers requires now lower interest rates.
This problem of a new firm trying to poach some of the ‘‘old’’ customers
of an incumbent firm, and having to be aware of the lemons problem
associated with it, is also related to auction problems when one of the
bidders is better informed (as in e.g., Engelbrecht-Wiggans et al., 1983), and
to competition for auditing business, when the incumbent auditor is better
informed about the business risk of a client compared to a rival entrant
(e.g., Morgan and Stocken, 1998).
One issue that is particularly important in credit markets is what happens
if the lenders exchange information about the borrowers. Pagano and Jap-
pelli (1993) investigate this issue with two types of borrower quality, where
each lender is in a different ‘‘town,’’ and learns about the credit quality of
the borrowers in that town in the previous period. Some of the borrowers
change towns from period to period, and there is heterogeneity on the
return from the borrowers’ projects if successful. Lenders can price dis-
criminate across three types of borrowers: the safe ‘‘old’’ borrowers, the
risky ‘‘old’’ borrowers, and the ‘‘new’’ borrowers. If the interest rate to the
‘‘new’’ borrowers is too high, only the risky ‘‘new’’ borrowers apply for
credit. Consider first the case in which lenders are local monopolies in their
own towns. In this case, profits are decreasing in the proportion of ‘‘new’’
borrowers as the lenders have less ability to price discriminate between the
types of borrowers. If there is information sharing across towns, then
lenders can distinguish the types of all borrowers, and profits increase.
However, the lending volume increases with information sharing if the safe
‘‘new’’ borrowers were not served in the case without information sharing,
and decreases otherwise.
Consider now the case of competition where lenders can offer credit to
borrowers in neighboring towns, although at a cost disadvantage. ‘‘New’’
50
Baye et al. (1992) show the existence of a continuity of asymmetric equilibria in the symmetric
Varian (1980) model. It would be interesting to investigate the implications of those results for the model
above when there are more than two incumbents.
borrowers are assumed to come from far away towns. In order to simplify
the analysis (to get away from mixed strategy equilibria), Pagano and Jap-
pelli (1993) assume that outside lenders make offers after the offers made by
the local lenders. The paper finds that, as above, lenders are able to deter
entry if given their informational advantages, and that information sharing
leads to lower profits, with the greater threat from the potential entrants.
The incentives for lenders to share information depend then on the mo-
nopoly effects above for information sharing, and on the competition
effects against information sharing. Which effect dominates depends on
their relative strength.
Another potential important issue in credit markets is the possibility of
borrowers exerting effort to increase the probability of success of their
project. This issue is addressed in Padilla and Pagano (1997). In this case,
borrowers may be concerned about exerting effort and then being taken
advantage of by high interest rates from the informed lenders (hold-up
problem). Padilla and Pagano suggest that lenders may be able to correct
this incentive problem by committing to share their information about the
borrowers with other lenders, such that the borrowers can benefit from
interest rate competition. In another paper, Padilla and Pagano (2000)
consider the case in which lenders cannot take advantage of their infor-
mation about the borrowers because they compete away ex-ante any gains
from future private information. In this case the paper argues that the
lenders may still want to commit to share the borrowers default rate with
other lenders as an incentive device for the borrowers to exert more effort to
increase the probability of the project success. However, if the lenders share
the information about the type of the borrower, the incentives to exert
effort are lower than if only defaults are shared, and the borrowers exert the
same level of effort as if no information were shared.
5.3 Customized pricing
In some markets, competing firms may have information about the con-
sumer preferences and price discriminate based on consumer preferences.
Competition in such a setting may end up being more intense, if this leads
to less differentiation in the competition for each consumer.
Thisse and Vives (1988) consider this effect in the Hotelling line with two
firms located at the extremes of the segment [0,1]. Suppose that consumers
are uniformly distributed on this segment, and that a consumer located at x
pays ‘‘transportation costs’’ tx, if buying from the firm located at 0, Firm 0,
and ‘‘transportation costs’’ t(1x), if buying from the firm located at 1,
Firm 1.
If firms do not know the location of the consumers they have to charge a
uniform price for all consumers. Let the price charged by Firm 0 be p0, and
the price charged by Firm 1 be p1. Then, it is well known that the demand
for Firm 0 is D0(p0, p1) ¼ (t+p1p0)/2t, and that the demand for Firm 1 is
D1(p0, p1) ¼ 1D0(p0, p1). The equilibrium prices are then p0 ¼ p1 ¼ c+t
(assume constant marginal costs c), and the equilibrium profit for each firm
is t/2.
Consider now that the firms know the location of each consumer. Then,
each firm can charge a price per location x, pi(x). The price competition in
each location x is like competition with a homogeneous good, where the
consumer has different valuations for the product. For xr1/2 (the case of
x>1/2 is symmetric) we have in equilibrium p0 ðxÞ ¼ c þ tð1 2xÞ; p1 ðxÞ
¼ c; and the consumers choose Firm 0’s product. The average price received
as revenue by a firm is then c+t/2, and each firm has a profit of t/4, one half
of the profit when customized prices were not possible. This result points to a
general effect that competition with customized prices is more intense than
competition without customized prices, if customization leads to less differ-
entiation in the competition for each consumer. That is, competition with
customized prices becomes like competition with no differentiation, in which
at the equilibrium prices, an infinitesimal small price cut attracts all the
demand. Variations of this result can be seen in Borenstein (1985), Holmes
(1989), Corts (1998).51 For the case of competition with second-degree price
discrimination see, for example, Stole (1995), Villas-Boas and Schmidt-Mohr
(1999), Armstrong and Vickers (2001), Desai (2001). For a recent survey of
competition with price discrimination, see Stole (2004).52
However, as noted by Armstrong (2005), more information about the
consumer preferences may not necessarily lead to less differentiation and
lower profits. Armstrong notes that if the additional information is about
the ‘‘transportation costs’’ parameter in the traditional Hotelling model,
additional information leads to significantly higher prices for the consumers
with the higher transport costs; this may lead to higher equilibrium profits.
One interesting extension of the variation of the Thisse and Vives model
above is the case in which we allow firms to only know the locations of
some of the consumers in the line (the firm’s database), and therefore, can
only offer customized prices to those consumers. This case is considered in
Chen and Iyer (2002). We then have that at each location some consumers
are in the database of both firms, some consumers are in the database of
only one of the firms, and some consumers are not in any database. The
databases can be available from the firms’ internal sources or from external
sources such as syndicated vendors of information.53
Chen and Iyer show that firms may choose to have not all consumers in
their database as this alleviates price competition. However, it turns out
that allowing firms to offer some degree of customized prices leads to higher
profits than no customization at all. That is, there is an intermediate level
51
See also Katz (1984) for the case of price discrimination in monopolistic competition.
52
See also Armstrong (2005) for a recent survey on economic models of price discrimination.
53
This can then be seen as a later period of some dynamic interaction where firms learn the complete
preferences of some consumers (the consumers in the firm’s database).
of price customization that leads to higher profits. The intuition for why
having limited databases may alleviate price competition is related to
Grossman and Shapiro’s (1984), who show, in the context of uniform
prices, that decreased advertising costs may reduce profits because it leads
firms to increase their advertising. This increased advertising leads to more
consumers that can compare prices, which leads to a greater benefit for a
firm of cutting prices, and thus to lower equilibrium prices and profits. In
Chen and Iyer, larger databases allow firms to do more customized pricing,
which we know from Thisse and Vives, may lead to greater price compe-
tition. Ulph and Vulkan (2000) consider the incentives for firms to invest in
customization capabilities under different transportation cost functions.
Ulph and Vulkan (2001) discuss what happens when customization may
allow a firm to offer customized products. Iyer et al. (2005) consider the
effects of customized advertising (in a model similar to Grossman and
Shapiro, 1984, for uniform advertising), and show that customized adver-
tising decreases price competition.54
A related but different form of competition with price discrimination
is when firms with capacity constraints advance-sell their products,
possibly at a discount. Dana (1998) considers this case, and finds that in
equilibrium we may have advance-selling discounts that are bought by
consumers with lower valuation for the product, but that have a more
certain demand.
6 Conclusion
This paper presents a summary of existing research on the effects of firms

being able to recognize their previous customers, and behave differently
toward them. The importance of understanding the effects of this market
practice has increased in the recent past given the development of infor-
mation technologies and the Internet (e.g., web–browser cookies) that allow
firms to keep, gather, and process more information about their past cus-
tomers.55 This increase in information has led to the proliferation of cus-
tomer relationship management practices in most industries. As of now, it
seems that many firms collect more information about their customers’
behavior than they are able to process. As firms get better at processing this
large amount of information, the effects of customer recognition are going
to become more and more important. In fact, the Internet allows also firms
54
See also Stegeman (1991) and Roy (2000). For the case of imperfect targetability see Chen et al.
(2001).
55
See Rossi et al. (1996) for a discussion of available databases of purchase histories and their possible
use in direct marketing. Pancras and Sudhir (2005) present an empirical application of personalization
activities (for example, offering of coupons) in grocery retailing. Lewis (2005) presents an application to
subscriber data of a large metropolitan newspaper of the dynamic issues in pricing using the past
consumer purchase behavior.
to interact more directly with their customers, and better respond to this
increase in information.
Most of the work until now has been on the firms’ pricing decisions
(with the exception of the limited work discussed in Section 4). Firms use
consumer behavior to target many other sorts of decisions, including their
product offerings and communication policies. As of now we have still very
little understanding of how these activities can interact with the ability of
firms to recognize customers. This means that research on this problem has
so far just uncovered the ‘‘tip of the iceberg,’’ and that there is much work
to be done on behavior-based targeting in the future. It would also be
interesting to see more empirical work testing for the results presented in
this literature.56
Research to date has identified several pricing effects in both monopoly
and competition. As discussed in Section 2, in monopoly, we have to
account for both behavior of the firm anticipating the future gain of having
more information, and the strategic behavior of consumers anticipating
what firms will do in the future with their information. As discussed
there, we may end up having a ‘‘ratchet effect,’’ as consumers realize that
they would be hurt by revealing their information, so that they incur costs
(forgo utility) to conceal their preferences. Important factors in how these
forces play out include the relative discount factors of the firm and
the consumers, the feasibility of the firm offering long-term contracts, the
effect of new generations of consumers coming into the market, and the
effect of consumer preferences changing (with positive correlation) through
time.
In markets with multiple firms there is the additional effect of firms
poaching each other’s customers with special deals. This generates inter-
esting strategic effects, possibly inefficient switching, and effects on the
intensity of competition. In addition to the possibility of firms offering
long-term contracts, and the entry of new customers (or customers chang-
ing preferences), another effect that can be important in several markets is
the presence of switching costs or network externalities.
Allowing firms to recognize customers raises the question of what can
firms do with such information, and whether consumers should have the
right to privacy in their market interactions. Furthermore, in some markets,
the characteristics of consumers may affect profits directly and this may
have additional effects on the functioning of the market as discussed in
Section 5 in the context of credit markets.
Finally, the possibility of firms recognizing their past customers interact
with several market aspects that have been substantially studied in the past
56
There is already some related empirical work. See, for example, Goldberg (1996) and Guha and
Wittink (1996), who show that empirical dealer discounts for new cars are a function of whether it is a
first-time purchase and whether there is a trade-in.
such as customized pricing, switching costs, durable-goods markets, and

bargaining.
Acknowledgements
We thank Mark Armstrong, Michael Baye, Assaf Ben-Shoham, Yongmin

Chen, Terry Hendershott, John Morgan, Jae Nahm, Jean Tirole, and
Juanjuan Zhang for insightful comments on an earlier draft of this paper,
and for helpful conversations.
References
Acquisti, A., H.R. Varian (2005). Conditioning prices on purchase history. Marketing Science 24,
367–381.
Armstrong, M. (2005). Economics models of price discrimination. Working Paper, University College
London.
Armstrong, M., J. Vickers (2001). Competitive price discrimination. RAND Journal of Economics 32,
579–605.
Ausubel, L., R. Deneckere (1992). Durable goods monopoly with incomplete information. Review of
Baye, M.R., D. Kovenock, C.G. de Vries (1992). It takes two-to-tango: Equilibria in a Model of Sales.
Games and Economic Behavior 4, 493–510.
Baron, D., D. Besanko (1984). Regulation and information in a continuing relationship. Information
Economics and Policy 1, 447–470.
Battaglini, M. (2005). Long-term contracting with Markovian consumers. American Economic Review
95, 637–658.
Beggs, A.W. (1989). A note on switching costs and technology choice. Journal of Industrial Economics
37, 437–440.
Beggs, A., P. Klemperer (1992). Multi-period competition with switching costs. Econometrica 60,
651–666.
Ben-Shoham, A. (2005). Information and order in sequential trade. Working Paper, Harvard
University.
Borenstein, S. (1985). Price discrimination in free-entry markets. RAND Journal of Economics 16,
380–397.
Bulow, J.I. (1982). Durable-goods monopolists. Journal of Political Economy 90, 314–332.
Cabral, L., J.M. Villas-Boas (2005). Bertrand supertraps. Management Science 51, 599–613.
Calzolari, G., A. Pavan (2005). On the optimality of privacy in sequential contracting. Journal of
Economic Theory forthcoming.
Caminal, R., C. Matutes (1990). Endogenous switching costs in a Duopoly model. International Journal
of Industrial Organization 8, 353–373.
Chen, Y. (1997). Paying customers to switch. Journal of Economics and Management Strategy 6,
877–897.
Chen, P.-Y., L.M. Hitt (2005). Information Technology and Switching Costs. (this handbook).
Chen, Y., G. Iyer (2002). Consumer addressability and customized pricing. Marketing Science 21,
197–208.
Chen, Y., C. Narasimhan, Z.J. Zhang (2001). Individual marketing and imperfect targetability. Mar-
keting Science 20, 23–41.
Chen, Y., Z.J. Zhang (2004). The benefit of behavior-based Targeted pricing with strategic consumers.
Working Paper, New York University and University of Pennsylvania.
Conlisk, J., E. Gerstner, J. Sobel (1984). Cyclic pricing by a durable goods monopolist. Quarterly
Corts, K.S. (1998). Third-degree price discrimination in oligopoly: all-out competition and strategic
commitment. RAND Journal of Economics 29, 306–323.
D’Aspremont, C., J. Gabszewicz, J.-F. Thisse (1979). On Hotelling’s stability in competition. Econo-
metrica 17, 1145–1151.
Dana, J. (1998). Advance-purchase discounts and price discrimination in competitive markets. Journal
of Political Economy 106, 395–422.
Dell’Ariccia, G., E. Friedman, R. Marquez (1999). Adverse selection as a barrier to entry in the banking
industry. RAND Journal of Economics 30, 515–534.
Dell’Ariccia, G., R. Marquez (2004). Information and bank credit allocation. Journal of Financial
Deneckere, R., P. McAfee (1996). Damaged goods. Journal of Economics and Management Strategy 5,
149–174.
Desai, P. (2001). Quality segmentation in spatial markets: when does cannibalization affect product line
design?. Marketing Science 20, 265–283.
Dobos, G. (2004). Poaching in the presence of switching costs and network externalities. Working
Paper, University of Toulouse.
Doganoglu, T., L. Grzybowski (2004). Dynamic duopoly competition with switching costs and network
externalities. Working Paper, University of Munich.
Economides, N. (1986). Minimal and maximal product differentiation in Hotelling’s duopoly. Eco-
nomics Letters 21, 67–71.
Ellison, G., D. Fudenberg (2000). The neo-luddite’s lament: excessive upgrades in the software industry.
RAND Journal of Economics 31, 253–272.
Engelbrecht-Wiggans, R., P. Milgrom, R. Weber (1983). Competitive bidding and proprietary infor-
mation. Journal of Mathematical Economics 11, 161–169.
Esteves, R.B. (2004). Pricing under customer recognition. Working Paper, Oxford University.
Farrell, J., C. Shapiro (1988). Dynamic competition with switching costs. RAND Journal of Economics
19, 123–137.
Farrell, J., C. Shapiro (1989). Optimal contracts with lock-in. American Economic Review 79, 51–68.
Farrell, J., P. Klemperer (2004). Coordination and lock-in: competition with switching costs and net-
work effects. Handbook of Industrial Organization, forthcoming.
Freixas, X., R. Guesnerie, J. Tirole (1985). Planning under incomplete information and the Ratchet
effect. Review of Economic Studies 52, 173–191.
Fudenberg, D., D.K. Levine (1989). Reputation and equilibrium selection in games with a patient
player. Econometrica 57, 759–778.
Fudenberg, D., J. Tirole (1983). Sequential bargaining with incomplete information. Review of Economic
Studies 50, 221–247.
Fudenberg, D., J. Tirole (1984). The fat cat effect, the puppy dog ploy and the lean and hungry look.
American Economic Review, Papers and Proceedings 74, 361–368.
Fudenberg, D., J. Tirole (1991). Game Theory. The MIT Press, Cambridge, Massachusetts.
Fudenberg, D., J. Tirole (1998). Upgrades, trade-ins, and buybacks. RAND Journal of Economics 29,
238–258.
Fudenberg, D., J. Tirole (2000). Customer poaching and brand switching. RAND Journal of Economics
31, 634–657.
Goldberg, P.K. (1996). Dealer price discrimination in new car purchases: evidence from the consumer
expenditure survey. Journal of Political Economy 104, 622–654.
Grossman, G.M., C. Shapiro (1984). Informative advertising with differentiated products. Review of
Guha, R., D. Wittink (1996). The relationship between replacement time, brand loyalty and transaction
price for automobiles. Working Paper, Cornell University.
Hart, O.D., J. Tirole (1988). Contract renegotiation and coasian dynamics. Review of Economic Studies
55, 509–540.
Hermalin, B.E., M.L. Katz (2006). Privacy, property rights & efficiency: the economics of privacy as
secrecy. Quantitative Marketing and Economics 4, 209–239.
Hirshleifer, J. (1971). The private and social value of information and the reward to inventive activity.
Holmes, T.J. (1989). The effects of third-degree price discrimination in oligopoly. American Economic
Review 79, 244–250.
Hui, K.-L., I.P.L. Png (2005). The Economics of Privacy. this handbook.
Iyer, G., D. Soberman, J.M. Villas-Boas (2005). The targeting of advertising. Marketing Science 24,
461–476.
Katz, M.L. (1984). Price discrimination and monopolistic competition. Econometrica 52, 1453–1472.
Kennan, J. (2001). Repeated bargaining with persistent private information. Review of Economic Studies
68, 719–755.
Kim, B.D., M. Shi, K. Srinivasan (2001). Reward programs and tacit collusion. Marketing Science 20,
99–120.
Klemperer, P. (1987a). The competitiveness of markets with switching costs. RAND Journal of Eco-
nomics 18, 138–150.
Klemperer, P. (1987b). Markets with consumer switching costs. Quarterly Journal of Economics 102,
375–394.
Klemperer, P. (1995). Competition when consumers have switching costs: an overview with applications
to industrial organization, macroeconomics and international trade. Review of Economic Studies 62,
515–539.
Kreps, D., P. Milgrom, J. Roberts, R. Wilson (1982). Rational cooperation in the finitely repeated
prisoner’s dilemma. Journal of Economic Theory 27, 245–252.
Laffont, J.-J., J. Tirole (1990). Adverse selection and renegotiation in procurement. Review of Economic
Studies 57, 597–625.
Lee, I.H., J. Lee (1994). Durable goods monopoly under technological innovation. Working Paper,
University of Southampton.
Levinthal, D.A., D. Purohit (1989). Durable goods and product obsolescence. Marketing Science 8,
35–56.
Lewis, M. (2005). Research note: a dynamic programming approach to customer relationship pricing.
McAdams, D. (2005). A theory of sticker prices. Working Paper, M.I.T.
Morgan, J., P. Stocken (1998). The effects of business risk on audit pricing. Review of Accounting Studies
3, 365–385.
Mussa, M., S. Rosen (1978). Monopoly and product quality. Journal of Economic Theory 18, 301–317.
Nahm, J. (2004). Durable-goods monopoly with endogenous innovation. Journal of Economics and
Management Strategy 13, 303–320.
Nilssen, T. (1992). Two kinds of consumer switching costs. RAND Journal of Economics 23, 579–589.
Padilla, J., M. Pagano (1997). Endogenous communications among lenders and entrepreneurial incen-
tives. Review of Financial Studies 10, 205–236.
Padilla, J., M. Pagano (2000). Sharing default information as a borrower incentive device. European
Economic Review 44, 1951–1980.
Pagano, M., T. Jappelli (1993). Information sharing in credit markets. Journal of Finance 48, 1693–1718.
Pancras, J., K. Sudhir (2005). The personalization services firm: what to sell, whom to sell to and for
how much? Working Paper, University of Connecticut.
Rajan, R.G. (1992). Insiders and outsiders: the choice between informed and arm’s-length debt. Journal
of Finance 47, 1367–1400.
Rossi, P.E., R.E. McCulloch, G.M. Allenby (1996). The value of purchase history data in target mar-
keting. Marketing Science 15, 321–340.
Roy, S. (2000). Strategic segmentation of a market. International Journal of Industrial Organization 18,
1279–1290.
Schmidt, K. (1993). Commitment through incomplete information in a simple repeated bargaining
game. Journal of Economic Theory 60, 114–139.
Schwartz, M. (1997). Telecommunications reform and the United States: promises and pitfalls, in: P.J.J.
Welfens, G. Yarrow (eds.), Telecommunications and Energy in Systemic Transformation, Springer,
New York.
Shaffer, G., Z.J. Zhang (1995). Competitive coupon targeting. Marketing Science 14, 395–416.
Shaffer, G., Z.J. Zhang (2000). Pay to switch or pay to stay: preference-based price discrimination in
markets with switching costs. Journal of Economics and Management Strategy 9, 397–424.
Sharpe, S. (1990). Asymmetric information, bank lending and implicit contracts: a stylized model of
customer relationships. Journal of Finance 45, 1069–1087.
Sobel, J. (1984). The timing of sales. Review of Economic Studies 51, 353–368.
Sobel, J. (1991). Durable goods monopoly with entry of new consumers. Econometrica 59, 1455–1485.
Stegeman, M. (1991). Advertising in competitive markets. American Economic Review 81, 210–223.
Stokey, N. (1979). Intertemporal price discrimination. Quarterly Journal of Economics 93, 355–371.
Stokey, N. (1981). Rational expectations and durable goods pricing. Bell Journal of Economics 12,
112–128.
Stole, L. (1995). Nonlinear pricing and oligopoly. Journal of Economics and Management Strategy 4,
529–563.
Stole, L. (2004). Price discrimination in competitive environments. Handbook of Industrial Organization,
forthcoming.
Taylor, C.R. (2003). Supplier surfing: price-discrimination in markets with repeat purchases. RAND
Taylor, C.R. (2004a). Consumer privacy and the market for customer information. RAND Journal of
Taylor, C.R. (2004b). Privacy and information acquisition in competitive markets. Working Paper,
Duke University.
Thisse, J.-F., X. Vives (1988). On the strategic choice of spatial price policy. American Economic Review
78, 122–137.
Tirole, J. (1988). The Theory of Industrial Organization. The MIT Press, Cambridge, Massachusetts.
Ulph, D., N. Vulkan (2000). Electronic commerce and competitive first-degree price discrimination.
Technical Report, University College, London.
Ulph, D., N. Vulkan (2001). E-commerce, mass customization and price discrimination. Technical
Report, University College, London.
Varian, H. (1980). A model of sales. American Economic Review 70, 651–659.
Villanueva, J., P. Bhardwaj, Y. Chen, S. Balasubramanian (2004). Managing customer relationships:
should managers really focus on the long term? Working Paper, University of California,
Los Angeles.
Villas-Boas, J.M. (1999). Dynamic competition with customer recognition. RAND Journal of Economics
30, 604–631.
Villas-Boas, J.M., U. Schmidt-Mohr (1999). Oligopoly with asymmetric information: differentiation in
credit markets. RAND Journal of Economics 30, 375–396.
Villas-Boas, J.M. (2004). Price cycles in markets with customer recognition. RAND Journal of Eco-
nomics 35, 486–501.
von Weizsäcker, C.C. (1984). The cost of substitution. Econometrica 52, 1085–1116.
Waldman, M. (1996). Planned obsolescence and the R&D decision. RAND Journal of Economics 27,
583–595.
Wathieu, L. (2004). Marketing and the privacy concern. Working Paper, Harvard University.
Zhang, J. (2005). Dynamic segmentation and product design. Working Paper, University of California,
Berkeley.
Chapter 8
Information Technology and Switching Costs
Pei-yu Chen
Tepper School of Business, Carnegie Mellon University, 5000 Forbes Ave, Pittsburgh, PA 15213,
USA
Lorin M. Hitt
University of Pennsylvania, Wharton School, 571 Jon M. Huntsman Hall, Philadelphia, PA 19104,
USA
Abstract
Switching costs, the real or perceived cost of changing product providers, has
become increasingly important in information-intensive businesses. Reduced
search and transactions costs have made it possible for customers to more
readily evaluate alternative suppliers, increasing switching. At the same time,
the high fixed cost structure of customer acquisition and operations in in-
formation-intensive businesses increases the impact of customer switching on
profitability. Fortunately, the increasing availability of information and bet-
ter information technology has enabled firms to have greater understanding
and control over switching costs, raising the ability of firms to manage
switching costs. This paper presents a formal definition of switching costs,
surveys the literature on information technology (IT) and switching costs,
and discusses various ways firms may use to influence consumer switching
costs. A framework for managing switching costs and customer retention is
also provided along with a measurement framework for estimating the extent
of customer lock-in.
437
438 P.-y. Chen and L. M. Hitt
1 Introduction
Switching costs, the perceived disutility a customer would experience

from changing product or service providers, play a large and increasing role
in competition and strategy in IT markets and other information-intensive
businesses. Many of these markets are characterized by a high fixed cost
and a high cost of customer acquisition, but a low marginal cost of product
delivery. Profitability in these markets is therefore driven by the size of the
customer base and customer retention, which is at least partly determined
by switching costs. As noted by Shapiro and Varian (1999)
‘‘You just cannot compete effectively in the information economy unless you know how to
identify, measure, and understand switching costs and map strategy accordingly’’. (p. 133)
Switching costs can arise from a number of sources including contractual

provisions, product characteristics, and the amount of information cus-
tomers have about product alternatives. In addition, many information-
based products exhibit network effects, where the value of adopting a
technology increases with the number of other users adopting the same
technology or the number of complementary products available. These
network effects can also create significant barriers from switching to an
alternative technology or network (Shapiro and Varian, 1999). To the ex-
tent that networks can exist within firms, such as when there are benefits
from product compatibility among investments in the same or comple-
mentary technologies within the firm (Kat and Shapiro, 1985, 1994), these
within-firm network effects can also act as switching costs (e.g., Greenstein,
1993).
While all markets can be affected by switching costs, there are several
unique issues regarding customer retention in information or IT-based
products, which we will collectively refer to as ‘‘information-intensive’’
markets. First, information-intensive markets often have fewer market in-
efficiencies that often represent the principal obstacle for customer switch-
ing. For instance, in markets with high search costs, consumers become
‘‘locked-in’’ because it is not efficient to search for alternatives. This is less
likely to be relevant in markets where product information is readily avail-
able. Reduced information or transaction costs can generally lower the cost
of terminating or initiating a customer relationship, also affecting switching
costs. Second, information-intensive products often have significant com-
patibility issues. On the one hand, interoperability of technology products
can be extremely important, especially for complex technologies such as
telecommunications gear or for enabling technologies such as operating
systems or computer languages. However, the market has recognized this
leading to a strong trend toward standardization, allowing users to mix-
and-match products from different vendors. Thus, standards and the de-
cisions associated with determining standardization strategies are often of
special importance in information intensive industries. Finally, technologies
Ch. 8. Information Technology and Switching Costs 439
employed in many information-intensive businesses often enable firms to

obtain better information about their customers which is essential to any
attempt to manage customer retention. For instance, ‘‘clickstream’’ data
can enable the measurement of switching costs (see Section 5), and cus-
tomer relationship management systems (CRM) can allow firms to tailor
product offerings to enhance retention.
The literature on switching cost has been most extensively developed in
economics and marketing. In economics, the emphasis has been on under-
standing the wide range of potential competitive implications of market-
wide switching costs, principally from a theoretical perspective. For
instance, switching costs have been tied to pricing, profitability, entry de-
cisions, the formation of standards, and other issues considered in the eco-
nomics of industrial organization (see e.g. Klemperer, 1995; Farrell and
Klemperer, 2004). Some of the economic literature has made this more
specific, such as switching costs due to product compatibility or network
externalities (Katz and Shapiro, 1985) that are common in software markets
(Bresnahan, 2001). There also have been a few empirical studies examining
the market-wide implications of switching costs for credit cards (Ausubel,
1991; Calem and Mester, 1995; Stango, 2002), bank loans (Kim et al., 2003),
bank accounts (Shy, 2002), cigarettes (Elzinga and Mills, 1998), breakfast
cereal (Shum, 2004), and toll-free numbers (Viard, 2003). However, this
literature has principally focused on switching costs that affect all providers
in a market similarly (what we term ‘‘exogneous switching costs’’) rather
than placing emphasis on switching cost variations that arise from specific
actions of individual firms (‘‘endogenous switching costs’’).
While not typically labeled switching costs as such, marketing scholars
have also worked to understand the implications of switching cost from a
customer behavior standpoint, emphasizing issues such as brand loyalty,
repeat purchase behavior, and customer retention. Much of the research
here treats loyalty as an intrinsic characteristic of consumers and examines
the implications of variations across consumers in loyalty. For example,
Raju et al. (1990) study how exogenous brand loyalty affects firms’ pricing
(or promotional) strategies. Recent research has also examined how loyalty
impacts equilibrium prices in the setting where an intermediary (‘‘clearing-
house’’) provides competitive product and price information and firms are
faced with a strategic decision on whether or not to target price sensitive
‘‘shoppers’’ or ‘‘switchers’’ (Narasimhan, 1988; Baye, et al., 2004; Baye and
Morgan, 2004). Another strand of the literature, mainly empirical, focuses
on the identification of loyal customers and switchers by constructing em-
pirical models that can predict customer behavior (see e.g., Grover and
Srinivasan, 1987).
Research in information systems has mostly followed the economic ap-
proach, but tends to be more application focused. For instance, researchers
have considered switching cost measurement for online businesses (Chen
and Hitt, 2002b), the role of product compatibility in creating switching
costs in the market for telecommunications equipment (Forman and Chen,

2005; Chen and Forman, 2006), and the impact of switching costs and
network effects on competition between online, traditional, and hybrid
firms (Viswanathan, 2005).
Since comprehensive reviews of the switching costs literature already exist
from both the economics (see Klemperer, 1995; Farrell and Klemperer,
2004) and marketing perspectives (Jacoby and Chestnut, 1978; Seetharam
et al., 1999), our emphasis in this review will be on applications to infor-
mation systems and technology issues. However, since these applications
involve a blend of both economic and marketing approaches, we will also
present a formal approach integrating these perspectives (drawn heavily
from Chen, 2002).
The remainder of this review is organized as follows. Section 2 presents a
formal definition of switching costs in an economic choice framework,
which helps clarify some general points about switching costs. We review
some major theoretical and empirical studies of IT and switching cost as
well as discuss related issues in Section 3. Section 4 points out the impor-
tance of endogenous switching costs in high-tech and information-intensive
markets and discusses some instruments for influencing switching costs. We
then provide a general framework for managing customer retention and
consumer switching costs and a measurement framework for estimating the
magnitudes of switching costs in Section 5, followed by general conclusions
and a brief discussion on future research issues in Section 6.
2 Switching cost: definition and measurement issues
Switching cost is typically defined as the disutility a customer experiences

in switching products or product providers.1 These costs can include
the cost to terminate an existing relationship, start a new relationship, and
any other costs, either explicit (e.g., contractual penalties) or implicit
(e.g., risk aversion and uncertainty about new products) that makes devi-
ating from past product choice more costly than staying with a previous
choice.
To formally state this definition and explore its limitations we must in-
troduce some notation. For this discussion we assume that consumers’
preferences can be represented by a utility function that depends on the
consumer and attributes of products. Let uji to refer to the utility customer i
receives from purchasing product j and use the notation Sjki to capture the
switching cost of customer i switching from product j to product k. Here
and throughout, we will assume that switching costs are small relative to
1
To simply exposition we will refer to ‘‘products’’ as a product purchased from a particular product
provider. Unless necessary, we will not make a separate distinction between the product and the product
provider or between products and services.
wealth so we can treat switching cost and utility as additively separable, and
will focus on situation where there is unit demand.2
In a specific interval of time in a market, there are always new customers
who enter the market for the first time; consumers who have adopted one or
more existing products, and consumers who choose to abandon a product.
We allow for the existence of costs for first adoption or market exit in our
notation by defining a null product (+), which is also in the choice set.
Although entrance and exist costs are not typically defined as switching cost
in most models, these costs can also behave as switching costs. Following
this setup, S+k can be interpreted as the adoption (or entrance) cost of
product k, while Sk+ are the exit costs. If a consumer does not purchase a
product, we normalize utility to zero (u+i ¼ 0). We are interested in the
implications of switching costs for customer behavior, as this is the foun-
dation for all the other competitive implications.
The simplest illustration of how switching costs affect choices requires a
market with two consumers (iA[A,B]), two goods (jA[X,Y]), and two pe-
riods (t ¼ 0,1). Let customer A represent a customer who has previously
acquired product X, while customer B has not purchased any product in
this market previously (period 0). For each customer, there are three
choices in period 1, purchase X, purchase Y or exit the market (represented
as purchasing the null product +). Utilities for different choices in period
1 are given by3
Action Consumer A Consumer B
Purchase X uA
X uBX SB+X
Purchase Y uA A A
Y S XY S +Y uBY SB+Y
No purchase SA
X+
0
While this formulation appears somewhat complex, the literature has

adopted a convention where initial adoption and abandonment costs (S+ki
and Sj+i) respectively are ignored and switching costs as assumed constant
across products and consumers (Sijk ¼ S 8i; j; k; j; ka+). While ignoring
adoption and abandonment costs is probably reasonable in a wide variety of
settings, the assumption of constant switching costs across all choices can be
restrictive. In particular, a key component of ‘‘customer retention strategy’’ is
to create firm-specific switching costs or to identify customers who have
2
These assumptions are common in switching cost models. With unit demand, we avoid issues of
purchase quantity for a single product as well as the possibility that multiple products are purchased.
The same insights for the unit demand case generally apply to the more general case, but the notation
becomes considerably more cumbersome.
3
The separation of utility and switching cost is reasonably provided there are no significant wealth
effects (i.e., consumers disutility of switching does not depend on their existing wealth).
unusually high switching costs. For instance, in online brokerage the firm can
charge a fee to close an account or transfer assets, offer a subsidy for new
customers transferring in from specifically named competitors, offer special
programs to target ‘‘buy and hold’’ investors who have large portfolios that
are cumbersome to move among brokers.4 Thus, the assumption of constant
switching cost rules out strategic choices that enable firms to modify their
own switching costs, an issue we will return to later in this review. None-
theless, imposing these assumptions yields the classic switching cost analysis:
Action Consumer A Consumer B
Purchase X uAX uBX

Purchase Y uAYS
uBY
No purchase 0 0
For additional insights into this framework it is useful to consider prod-

uct choice behavior in a discrete choice framework (McFadden, 1974).
Without introducing extraneous technical details at this point, this formu-
lation argues that the probability of choosing product j is proportional to
utility.5 Take consumer A, who is an existing customer of product X. The
utilities she derives from purchasing X, Y and nothing are uA A
X, uY-S and 0,
respectively. If she has not purchased any product before, the utilities she
gets from X, Y and nothing would be uA A
X , uY and 0. Given these two utility
sets, the alternative product (Y) is relatively unattractive to consumer A in
period 1 because of their prior purchase of product X, even though the
underlying products are the same before and after consumer A adopts X. In
another words, customers are more likely to stay with an existing product
since it provides greater utility relative to other non-adopted alternatives,
even if the alternative product is identical in terms of non-switching cost
utility (u). This is why switching cost is typically associated with other
informal notions such as ‘‘stickiness’’, ‘‘lock-in’’, ‘‘customer retention’’ or
increasing ‘‘repeat purchase’’ propensity.
It should be noted that even the more general framework embeds an
assumption about customer behavior that can be important: it assumes that
customers do not make choices based on perceived future switching costs
(i.e., consumers are myopic regarding future switching events). Most insights
about switching cost and competition continue to hold if customers have
4
Another example of this behavior is a ‘‘competitive upgrade’’ that is common in software markets
where a firm will subsidize customers of other firms products to switch.
5
The simplest case is the logit model where the error term is independently and identically distributed

across products and consumers with the ‘‘extreme value’’ distribution (i.e., prob. ð Þ ¼ ee ; where
NoeoN). J The choice probability of product j out of a larger set of size J is then given by:
uj
P ul
pj ¼ e e :
l¼1
rational expectations about future switching behavior (Farrell and Klemp-

erer, 2004) or share a common expectation of the probability of future
switching (see, Chen, 2002, Chapter 2). However, it can become important if
we are interested in tradeoffs between customer acquisition and retention.
High switching cost might favor customer retention but deter initial pur-
chase if consumers are fully rational; there is no such tradeoff if customers
are myopic and ignore future switching. Since firms can benefit from en-
couraging customers to be myopic about future switching, this may explain
why contractual penalties for switching are often written ‘‘in small print’’ on
new purchase contracts. The general tradeoff between acquisition, retention,
and attrition has been discussed in the literature (Chen and Hitt, 2002b),
although the specific issue of customers’ perceptions of switching cost has
not been fully explored and is an opportunity for future research.
This simple framework illustrates why it can be difficult to separate out
the effects of product quality (increasing utility) or creating lock-in (in-
creasing switching costs) in empirical data since they can behave in similar
ways. Improving the quality of your product increases the utility of your
product relative to others. Raising switching costs lowers the utility of other
products for your existing customers. Thus, a firm can be successful at
retaining customers either because they offer a superior product (at least for
a specific set of consumers), or because they have high switching costs.
Much of the managerial literature on customer retention uses retention rate
as a measure of customer loyalty or consumer switching costs and does not
make a distinction between switching cost and quality in driving retention
(see. e.g., Reichheld and Schefter, 2000). This is perfectly reasonable if one
is trying to predict future customer retention or compare retention across
firms; it becomes problematic when one in interested in the drivers of cus-
tomer retention or how investments in specific practices can improve (or
decrease) retention. An excellent product can have poor sales if customers
face high switching costs and most customers have already adopted alter-
natives. Similarly, a poor product can show excellent customer retention if
switching costs are high and switching costs may not matter at all during
the initial introduction of a completely new product since there are no
existing customers.
The difficulty in separating out switching cost from other drivers of
product choice was highlighted in economic analysis of the Microsoft anti-
trust case. The key question was whether Microsoft’s dominant position is
due to superior quality, switching costs, or anti-competitive behavior (see
Liebowitz and Margolis, 1999). While being successful due to quality or
switching costs are both legal, the quality explanation is welfare promoting,
while the switching cost explanation may not be and might warrant some
sort of market intervention if it were acquired through or leveraged by
other anti-competitive behavior. We will discuss the distinction between
product utility and switching in more detail in Section 5 and provide a
framework to distinguish the two effects.
3 Switching costs, competition, and firm strategy
3.1 Switching costs and competition
Most of the early literature on switching costs considers the impact or the
consequences of (exogenous) switching costs. The usual analysis proceeds
with exogenously determined (and usually symmetric) switching and pro-
ceeds to examine how switching costs affect pricing, entry, and firm profit-
ability under different market conditions. Klemperer (1995) and Farrell and
Klemperer (2004) provide extensive surveys on these issues from economic
perspective, so we would only briefly summarize the general results
pertaining to switching costs and competition in this section.
Consumer switching costs provide firms’ market power over their existing
customers enabling them to charge a price premium (as high as S) or stra-
tegically price to deter switching, even to superior alternatives. This result
suggests that it makes sense for firms to invest to acquire new customers,
even incurring losses in the process, because they may charge higher prices
over their existing customers for repeat purchases (this has been termed
‘‘bargains-then-ripoffs’’). More broadly, in markets with switching costs,
firms face tradeoffs between harvesting profits by charging higher prices
over their existing customers and investing in market share by charging
lower prices to acquire new customers who will be valuable repeat pur-
chasers in the future (termed harvesting vs. investing) (Klemperer, 1995).
The general observation is that in otherwise competitive markets with
switching costs, firms have additional strategic options that can lead to
higher prices and greater profits.
Switching costs can also act as a deterrent to entry, or alter the adoption
rate of new technologies because they decrease the value of new products in
markets where most customers already use an existing product. This result
suggests that, when consumers’ switching costs are high, a vendor with
initial advantages in a market may be able to parlay its advantage into a
larger, lasting one (Katz and Shapiro, 1994). In addition, some authors
have noted that some results about single-product competition over many
periods with switching costs can carry over to multi-product competition in
a single period when there exist consumer shopping costs or cross-product
switching costs (Klemperer and Padilla, 1997; Farrell and Klemperer, 2004;
Forman and Chen, 2005). These results suggest that a firm with larger
product line has a strategic advantage over a firm offering fewer products in
the presence of shopping costs and cross-product switching costs.
3.2 Endogenous switching costs
An important feature of much of the literature on switching costs is the

dual assumptions of symmetry (all firms face the same switching costs) and
exogeneity (firms cannot through their own actions affect switching costs),
which we collectively refer to as ‘‘exogenous switching costs’’. However, the

important role switching cost can play in competition strongly suggests that
firms can benefit from investments or actions that affect customer switching
costs (‘‘endogenous switching costs’’). For instance, a firm will prefer
switching costs from but not to its product if possible.
The role of endogenous switching costs is gaining increasing interest in
the literature. Caminal and Matutes (1990) show that firms may wish to
commit to offering repeat-purchase coupons before competing in a two-
period duopoly in order to soften competition. An incumbent firm will also
have an incentive to create contractual switching costs through the use of
long-term contracts to lock-in their customers (Fudenberg and Tirole,
2000). In addition, a firm with large installed base may also prefer a tech-
nology design that is incompatible with other firms, while a smaller firm will
prefer compatibility with the incumbent firm (Katz and Shapiro, 1994).
Furthermore, given that a broad product line offers an advantage when
there are consumers’ shopping costs or cross-product switching costs, a firm
may strategically invest to increase its product breadth or to manipulate
cross-product switching costs so that consumers prefer to buy multiple
products from the same vendor (Farrell and Klemperer, 2004). Similarly,
Baye and Morgan (2004) show that each firm has a unilateral incentive to
implement costly loyalty programs, even though it results in lower equi-
librium profits for all firms. Demirhan et al. (2004) also show that in the
presence of rapid technological advances and declining IT costs, which may
provide the late entrant a cost advantage, an early entrant will have
stronger incentive to offer products that impose significant switching costs
on consumers to mitigate the damage from a decline in IT cost and preserve
its first mover advantage.
3.3 Switching costs in information-intensive markets
As strategic investments in switching costs become more important, a first

step in understanding how firms might be able to incorporate switching
costs into competitive strategy is to understand the sources of switching
costs in actual markets. Most of the theory literature in economics do not
distinguish between different sources of switching costs in their models or
just focus on a specific form of switching costs. However, Klemperer (1995)
in his classic review of switching costs lists some of the major sources of
switching cost identified in economic models including: the need for com-
patibility with existing equipment, transaction costs of switching suppliers,
the cost of learning to use new brands, uncertainty about the quality of
untested brands (e.g., search costs), discount coupons and similar devices
(e.g., loyalty programs), psychological costs of switching, or non-economic
‘‘brand loyalty’’. Several forms of these switching costs are especially rel-
evant to information products and the digital marketplace and warrant
further discussion. We build upon Klemperer’s categorization of switching
costs and discuss these switching costs in the context of high-tech and
information-intensive markets.
Search costs. These are the costs consumers must incur to locate an
alternative seller (Stiglitz, 1989). While search costs affect even consumers’
initial purchases, it can act just like consumer switching costs when a con-
sumer has better information on her current supplier, but not other sup-
pliers (Schmalensee, 1982; Moshkin and Shacher, 2000; Villas-Boas, 2004).
When product or price information is costly to acquire and consumers have
imperfect information about alternative products and prices, it may be
optimal for consumer to engage in limited search or not to search at all,
remaining with a current supplier. This is particularly problematic for
differentiated product markets where the costs of acquiring sufficient in-
formation about product alternatives and the cost of consumers processing
of this information can be high. Much of the literature on competition and
the Internet suggest that the Internet lowers search costs and thus poten-
tially increases switching (Bakos, 1997). However, there is some contrary
evidence to this notion. Numerous empirical studies suggest that consumer
search costs remain high in electronic markets or at least customers behave
as if search costs are high (see the survey by Baye et al., 2005).
Transactions costs. These are the costs required to initiate a new rela-
tionship, and in some cases to terminate an existing relationship. These
costs may include shopping costs, transportation costs, or costs associated
with opening a new account or closing an existing account. These costs
contribute to S+k (adoption costs) or Sk+ (exit costs) introduced in Sec-
tion 2. Transactions costs reduce switching even when information is cost-
less. For example, when transaction costs are high, consumers may prefer to
have all their transactions done with the same vendor, in order to save on
shopping costs. The success of the ‘‘one-stop shopping’’ strategy adopted by
Walmart and Amazon.com can be partially attributed to minimizing trans-
action costs. It is generally believed that the Internet and other electronic
technologies have reduced these costs since much of the process of starting
or terminating a relationship can be done without travel or conducted with
electronic support (e.g., using identity management software to communi-
cate user information to a new supplier). Technology cannot only lower
transactions costs for consumers. The same technologies can be used by
suppliers to assist consumers in switching providers as is becoming in-
creasingly common for technology-driven financial services products such
as mortgages, credit cards, and retail banking transaction accounts.
Learning costs. When products or sellers are different, consumers may
incur costs (time, money, or other effort) to learn to work with a new
product or a new seller. When these costs are seller or brand specific, they
are sunk costs and non-transferable from one relationship to another. These
costs can be especially high for differentiated information services such as
online brokerage (Chen and Hitt, 2002b). Indeed, Chen and Hitt (2002b)
show empirically that making an information service ‘‘easy to use’’ may
actually increase switching (although it increases acquisition), which they

attribute to the reduced need for sunk investments in learning. It has also
been noted that even using a web site has a learning curve—Johnson et al.
(2003) found that visit duration declines the more often a site is visited,
indicating a learning effect. In addition, having learned to use a site raises
its attractiveness relative to competing sites for the consumers. Thus, the
site will be more likely to be used than its competing alternatives, all other
things being equal. Thus, learning costs can represent a significant barrier to
product switching in many information-intensive markets.
Complementary investments. Many information products are such that
they require complements to be useful. For instance, digital music players
require purchase of compatibly formatted digital music. Switching to an-
other player may make this content unusable. For many digital products
and services, complementary assets are created by use. For instance, Micro-
soft Word and word-compatible files, eBay, and user ratings, online net-
working web sites and ‘‘friend lists’’ all represent complements created by
use. In some cases, such costs can be mitigated by the presence of ‘‘con-
verters’’ (e.g., Farrell and Saloner, 1992; Choi, 1997), but many service
providers deliberately make it difficult to utilize complementary assets when
the relationship is terminated.
In commercial relationships, firms often encourage sunk investments in
coordination technologies (Clemons et al., 1993). Technologies such as the
American Airlines SABRE system (airline ticket distribution) or the Baxter
Healthcare ASAP system (hospital supply ordering) owe some of their
success to the fact that firms that invested in these technologies not only
made sunk expenditures on the systems and associated training, but rede-
signed business operations around these technologies. These effects may be
especially strong in enterprise software markets such as Enterprise Resource
Planning (ERP) systems where the complementary investments may be
several times the size of the actual technology expenditure (Brynjolfsson
et al., 2005). However, not all technology investments have these charac-
teristics—the McKesson Economost system (inventory management for
drugstores) was easy enough to duplicate for other suppliers that it con-
veyed limited advantage. Moreover, the initial investments in redesigning
business processes to accommodate Economost gave firms the needed
knowledge to efficiently switch suppliers in the future.
Network effects and compatibility. Closely related to complementary in-
vestments, some products exhibit network effects that arise when a user
desires compatibility with other users or where increased consumption of
addition units of the same good creates additional value. In the presence of
network effects, users benefit from adopting products (and staying with
adopted products) with the most users. Even in the presence of a superior
technology, the coordination costs of changing all users to a new technol-
ogy may outweigh the advantage of a new technology (this is referred to as
‘‘excess inertia’’ in the network economics literature).
Network effects can create switching costs when they exist across the
choices of a single economic actor (either an individual or a firm). An
example is a firm’s investment in network switches used to manage com-
munications in local area networks. There are significant interoperability
benefits to having all switches in a firm purchased from the same manu-
facturer (see discussions in Chen and Forman, 2006, and Forman and
Chen, 2005). Since converting a switch to another network provider
changes the value of all other switches that had been purchased by the firm
previously, this behaves as a switching cost. Moreover, switch manufac-
turers have strong incentives to preserve these differences and do so by
creating proprietary ‘‘extensions’’ to the technology, even when the base
technology is trending toward standardization (Chen and Forman, 2006).
However, not all network effects are truly switching costs. For instance,
general network externalities present in the market as a whole as a result of
standardization, such as the adoption of TCP/IP networking standards on
the Internet, provide a benefit but not switching costs because they affect
overall utility by raising the value of the product. The critical difference is
whether an individual decision-making unit (e.g., a firm) has control over
the size of the ‘‘network’’ through their own choices.
Contractual switching costs. These are pecuniary incentives provided for
customers to make repeat purchases from the same provider, be it a store, a
service, or a manufacturer, for a certain period of time. These can be pos-
itive (e.g., ‘‘reward points’’ or ‘‘frequent flier programs’’) or negative
(‘‘penalty for early withdrawal’’ for deposit banking or an ‘‘early termina-
tion fee’’ for a wireless phone contract). While these costs apply to a wide
range of products, they appear to be unusually common for certain types of
information products or services—especially those that involve some sort
of new user subsidy for adoption. Contractual switching costs have some
unusual properties such as the fact that they are often time limited (e.g.,
frequent flier miles expire, contracts exist for a specified duration). How-
ever, most importantly, they represent a source of switching costs, which
can be (more) easily varied by firm and thus can represent a major source of
cross-firm variation in switching costs. Kim et al. (2001) have studied in-
centives to offer reward programs that create pecuniary switching costs.
3.4 Empirical evidence
The empirical literature on switching costs is much smaller and more

recent than the theory literature, due primarily to the difficulty in obtaining
detailed data on individual- or firm-specific decisions required to test hypo-
theses related to switching costs. For instance, conventional market share
data does not enable product quality and switching cost to be distinguished.
For market share data to be useful in retention studies, it would have to
distinguish between share of new customers and share of customers ac-
quired through switching which is rarely available for more than a single
firm. However, the increased availability of micro-level consumer behavior

data, especially for technology products and services, has greatly expanded
the ability to do research in this area. In this section, we review major
empirical studies on IT and switching costs.
Switching costs in software and other ‘‘high-tech’’ markets

One of the first empirical studies on IT and switching costs is Greenstein
(1993), who studied mainframe procurement decisions in government agen-
cies. He found that users of IBM mainframes are more likely to buy main-
frames from IBM in the future than are users of other vendors’ products,
even when controlling for buyer characteristics that might influence vendor
choice. This result suggests that buyers face significant switching costs from
past investments due to the need of compatibility within the firm, and
points out the importance of switching costs in influencing users’ purchase
decisions.
Forman and Chen (2005) consider a similar question of vendor choice in
the context of networking equipment such as routers and switches. Their
research strategy further enables them to decompose switching costs into
those created by the costs of learning new products and those created by the
need of compatibility with past investment. Using detailed data they show
that switching costs may arise from prior investments made at the same
establishment as well as investments made at other establishments within the
same firm. In addition, they show the installed base advantage of one
product (network routers) can spillover to other products (network switches)
as technology evolves. Chen and Forman (2006) further extend this research
by examining possible vendor actions in creating switching costs, including
manipulating ‘‘horizontal’’ compatibility between comparable rival products
and ‘‘vertical’’ compatibility between complementary products, maintaining
a broader product line, creating product suites, and targeting specific market
segments. Kauffman et al. (2000) examine how prior investments in pro-
prietary networking technology influence incentives to adopt a multi-bank
electronic network. Similarly, Zhu et al. (2006) investigate firms’ migration
from proprietary or less open inter-organizational systems to open-standard
inter-organizational system (i.e., the Internet) and show that experience with
older standards may create switching costs and make it difficult to shift to
open and potentially better standards.
Research on technology prices also supports a notion that switching costs
arising from learning and compatibility can create market-wide effects on
pricing. For instance, Brynjolfsson and Kemerer (1996) find that products
that adhered to the dominant standard for spreadsheet software com-
manded prices which were higher by an average of 46%. Gandal (1994)
found a similar result using different data. Overall, this literature has shown
that there exist significant switching costs in high-tech and software mar-
kets, and that a purchase decision today can have a far-reaching impact in
the future.
However, new IT innovations may decrease switching costs arising from

prior investments when buyers of the innovation must make new sunk cost
investments to take advantage of the new innovations (Brynjolfsson and
Kemerer, 1996; Bresnahan and Greenstein, 1999). For instance, Breuhan
(1997) demonstrates that firms switching from the DOS to Windows op-
erating system had lower switching costs of changing word processing and
spreadsheet vendors than firms that retained the same operating system.
Chen and Forman (2006) also find that the introduction of new products
(network switches) lead to a temporary ‘‘window’’ of lower switching costs
away from network routers. The introduction of the new product forced
firms to redesign and rebuild their network infrastructure, effectively ‘‘free-
ing’’ them from the switching costs arising from their installed base. How-
ever, even though switching costs may be reduced with new IT innovation,
they find that there still remain significant switching costs.
Switching costs in online markets

Although electronic markets are believed to have low switching costs since
a competing firm is ‘‘just a click away’’ (Friedman, 1999), recent research
suggests that there is significant evidence of switching costs and brand loy-
alty in online markets. For example, Smith and Brynjolfsson (2001) found,
using data from a price comparison service (the DealTime ‘‘shopbot’’), that
a consumer’s past purchase experience has significant predictive power of
her future store choice and that customers are willing to pay premium prices
for books from the retailers they had dealt with previously. Similarly, Lee
and Png (2004) showed that consumers bear significant shopping costs,
which represent another source of switching costs, for shopping books on-
line. Johnson et al. (2004) showed that customers engaged in limited across-
site search and tended to search fewer sites as they become more experienced
with online shopping. Moe and Fader (2004) have also explored the rela-
tionship between visit frequency and website experience and found evidence
supporting the fact that people who visit a store more frequently (indicating
greater lock-in) are more likely to buy. Bucklin and Sismeiro (2003) also
found evidence in support of the ‘‘lock-in’’ phenomenon.
Several marketing papers further compare brand loyalty in online shop-
ping environments and their offline counterparts and conclude that con-
sumer brand loyalty is not necessarily lower online, and in many cases,
brand loyalty is actually stronger online than offline. Danaher et al. (2003)
used data on 100 grocery brands and found that higher share (and therefore
better-known) brands have greater-than-expected loyalty when bought on-
line compared with an offline environment. Degeratu et al. (2000) also note
that there is less brand switching online, especially when online consumers
use a pre-set personal list. In addition, Andrews and Currim (2004) report
that online grocery shoppers consider only brands they have purchased
before and form a smaller consideration set, thereby remaining loyal to a
smaller number of brands.
There is also evidence of firms making firm-specific switching costs in-

vestments. Chen and Hitt (2002b) find that switching costs vary signifi-
cantly across different online brokerages even after controlling for customer
heterogeneity, suggesting that firms have considerable control over switch-
ing costs through a number of firm practices. Similarly, Goldfarb (2006)
reports evidence of switching costs at Internet portals after controlling for
unobserved heterogeneity (i.e., ‘‘spurious’’ state dependence).6 All these
studies suggest that switching costs and brand loyalty remain high in digital
markets, even though search costs are reduced and information can be
easily gathered in these markets, and that a significant source of this var-
iation is due to firm practices.
4 Endogenous switching costs and firm strategy in information-intensive

markets
As discussed above, IT is often associated with a reduction in market

‘‘frictions’’, frictions that can represent a significant portion of switching
costs. Search technology has enabled firms to identify and to evaluate trade
counterparties more efficiently. IOS enable firms to minimize the cost of
interacting with a broader group of suppliers. Increased standardization of
both the communications protocols (e.g., TCP/IP) and application inter-
faces (e.g., XML/SOAP) make these and other ‘‘coordination’’ technologies
less costly and more broadly available to the mass market. Collectively,
these arguments suggest that for high-technology products such as infor-
mation products or services, the inherent level of market wide (exogenous)
switching cost is lower. The removal of market inefficiencies makes the
market more competitive and vendors more vulnerable to competition,
creating incentives for firms to deliberately create firm-specific switching
costs (endogenous switching costs) to restore their market power.
In addition to lowering search and coordination costs, some have further
argued that information systems reduce the degree of product differenti-
ation among firms (Porter, 2001). He argues that that technology enables
firms to more rapidly imitate each other, either due to flexible manufac-
turing or imitation of software innovations, or by enabling entrants to
source from the same (or similar) pool of global suppliers with reduced
transaction risk. Of course, the counterargument to this observation is that
these same technologies enable firms to innovate more rapidly and to better
meet customer needs enabling greater product differentiation (see e.g.,
Tapscott, 2002; Clemons et al., 2003). Regardless of which perspective
prevails, it is clear that imitation of technology-enabled innovations is much
more rapid.
6
Although these switching costs estimates drive only 11–15% of market share on a choice-to-choice
basis, which is much smaller than the effects found by Shum (2004) in breakfast cereals and that by
Keane (1997) found in Ketchup.
When firms enjoy a smaller period of near-monopoly position on a new

product, creating switching costs becomes an important part of strategy.
For instance, Capital One Financial initially gained considerable advantage
by novel product designs that help them targeting ‘‘profitable’’ customers in
the credit card industry. However, it became apparent that many of these
designs were easily imitated. This encouraged them to make large invest-
ments in ‘‘retention specialists’’ who worked to retain customer accounts
and reduce switching. While the product designs were easily copied, the
organizational and human assets involved in customer retention proved
hard to duplicate, and may have been a key source of competitive advan-
tage for Capital One (Clemons and Thatcher, 1998). While attracting
profitable customers is always an important strategy, in markets with rapid
innovation and imitation, how to keep these profitable customers will be the
key to a firm’s long-term profitability. To better manage customer acqui-
sition and retention, Blattberg and Deighton (1996) have introduced the
concept of ‘‘customer equity’’ that balances spending on obtaining and
keeping customers.
Switching costs have also become important in some technology-enabled
markets, especially Internet-based products and services, simply because of
the high cost of customer acquisition. At the peak of the ‘‘Internet bubble’’
firms were actively spending up to $1,000 per new customer with the ex-
pectation of converting this up-front investment into an offsetting stream of
revenue with inherently high margin. While this type of activity has con-
siderably decreased, referral rates on the order of $20–$80 per customer7 are
still common. Given the competitiveness of online markets and the attend-
ant low margins, this leaves length of relationship as a key driver of overall
customer profitability. Using this type of analysis, Reichheld and Schefter
(2000) calculate that by retaining 5% more customers, online companies
can boost profits by 25–95%.
As exogenous, market-wide switching costs are reduced, endogenous
switching costs have the potential to play a much greater role in compe-
tition. While the literature in this area is still developing, a number of
studies have identified mechanisms for how firms can either restore switch-
ing costs eroded by external forces, or alter their own switching costs.
Strategic restoration of search costs. As search costs for prices and product
information are reduced with the use of search and shopping engines, firms
may have additional incentives to differentiate their products in order to
induce search costs (Kuksov, 2004). Indeed, we do observe many retailers
strategically introduce myriad variations of a product, which in turn lead to
consumer search costs to locate a product of their interest (Bergen et al.,
7
A common example in 2005 is the existence of services that provide ‘‘free’’ products such as iPods,
computers, designer handbags, or digital cameras when a customer signs up for one or more trial
services and encourages other customers to do the same. This entire business is based on the presence of
referral fees as high as $80 per new customer to offset the cost of these products.
1996). Ellison and Ellison (2001) have also noted that firms try to adopt a
number of strategies that make search more difficult. Many firms also use
rapid and unpredictable price changes to prevent consumers from consist-
ently learning about the identity of the low-price provider and also to
prevent competitors from systematic undercutting (Baye et al., 2005). For
example, airlines have responded to increased price transparency provided
by computerized reservation systems (CRS) but creating large numbers of
constantly changing fares, so many that finding best fare for a customer can
be problematic even with technological support (Clemons et al., 2002).
Studies of online price show considerable price dispersion, perhaps
even more so than offline stores (Brynjolfsson and Smith, 2002; Baye
et al., 2005).
Product design. Firms may also induce switching costs through product
design, such as adopting proprietary standards. For example, Sony has
adopted a proprietary technology (e.g., the memory stick) for its digital
camera. An incompatible or different product design may also lead to
learning costs that are brand-specific and cannot be transferred. In the case
where complementary products are also needed, a customer is essentially
locked-in to the vendor for all other complementary products when the
vendor adopts an incompatible product design for all its products. How-
ever, it is important to note that the decision of proprietary or incompatible
product design can only be profitable when enough demand can be sus-
tained, which is often difficult to predict ex-ante. For instance, there is
considerable debate about whether Apple’s strategy of closed standards has
helped or hurt their market position in personal computers.
Even in markets where standardization has eliminated switching costs
from incompatibility, it has been a common practice for vendors to ‘‘ex-
tend’’ standards by adding proprietary enhancements to their products.
Although these proprietary enhancements may be added to improve func-
tionality and add value to customers, these changes also make compatibility
and interoperability among competing products more difficult to achieve
(Farrell and Saloner, 1992; Wickre, 1996). In addition, vendors may also
extend standards in ways that affect the ‘‘vertical’’ compatibility (or inter-
operability) between complementary products, thereby discouraging con-
sumers from ‘‘mix-and-match’’ purchases. For example, even though
Cisco’s Internetworking Operating Systems (IOS) software has published
standards, Cisco itself has claimed that there is enough proprietary code
within IOS to allow Cisco products to ‘‘work better when they talk to each
other, rather than machines made by rivals’’ (Thurm, 2000).
Product line design and bundling strategy. Klemperer and Padilla (1997)
have demonstrated that selling an additional product can provide strategic
benefits for a firm in the market for its current products if consumers have
shopping costs. The strategy of maintaining a broader product line has also
been empirically found to be associated with reduced consumer switching
(Chen and Hitt, 2002b; Chen and Forman, 2006). Another advantage of
multi-product firms is that they may ‘‘bundle’’ products, which lead to

contractual shopping costs between products (Farrell and Klemperer,
2004), which can be an effective strategy to deter new entry (Nalebuff,
1999). Microsoft has pioneered this strategy in both their operating system
and office software product lines. In markets for communications equip-
ment, vendors commonly include heterogeneous products in a single
‘‘product suite’’ designed to serve many buyer needs (e.g., 3Com’s Office-
Connect or Bay Networks BayStack product lines). These product suites
lower buyers’ costs of identifying and adopting complementary products
while simultaneously creating switching costs.
Customer profiling and personalized services. The use of advanced IT, such
as the use of cookies, log files, data mining technologies, customer profiling
techniques, collaborative filtering technologies (e.g., recommendation sys-
tems), and other personalization technology, also allows firms to ‘‘identify’’
their customers and their needs and to act accordingly. Many of these
efforts to restore switching costs also add value to the customer even if they
were intended principally to reduce switching. Personalized services and
products tailored to customer needs may improve customer satisfaction,
which in turn, lead to consumer switching costs if another firm cannot offer
similar services or products (perhaps due to lack of relevant data). The
benefits of personalized services or product offerings may increase over time
as the firm has more information about customers. For instance, recom-
mendation systems become more precise as the consumer visits the site
more often, makes more purchases and provides more product reviews.
Thus, customers concentrating their purchases at a single online retailer
may receive greater value and are more likely to return. In addition, the
ability of firms to track customers’ usage patterns also make it easier to
identify different customer types, and allow firms to identify potential
switchers and act accordingly before they actually switch (Chen and Hitt,
2002b). Customer recognition also allows firms to engage in ‘‘behavior-
based price discrimination’’ (Acquisti and Varian, 2005; Fudenberg and
Villas-Boas, 2005), although the customer uproar over Amazon’s attempt
to offer higher prices to loyal customers has made firms cautious about
implementing highly transparent and purely redistributive strategies for
‘‘exploiting’’ loyalty.
Virtual network effects and brand-specific investments. Switching costs
arising from network effects can arise even for products without technical
compatibility issues if past investment or frequency of use influences future
value. For example, recommendation systems can become more precise not
only when a user visits a site more often (as discussed earlier), but also when
more users visit or make purchases at the site because with a large installed
base there are likely to be more people with similar tastes who have rated
other products. Rating systems in consumer or commercial electronic mar-
kets (e.g., eBay’s ‘‘positive’’, ‘‘negative’’, ‘‘neutral’’ system) also become
more trustworthy as more transactions are recorded. Moreover, ratings are
generally not portable across services or sites, thus encouraging users to

concentrate their transactions at a single site. This is a type of lock-in that is
analogous to own-choice network effects, similar to the mainframe com-
puter and network switchgear settings discussed earlier.
Market segmentation and loyalty programs. While firms may invest to
increase firm-specific switching costs, switching costs may differ across
buyers for reasons unrelated to vendor product strategy. The fact that much
of customer loyalty is due to latent customer traits is a central tenet of most
marketing models of retention and repeat purchase. For example, previous
literature has demonstrated that larger firms (Kauffman et al., 2000) and
those with more complicated infrastructure (Bresnahan and Greenstein,
1996; Forman, 2005; Forman and Goldfarb, 2005) will be slower to adopt
new IT innovations and face higher switching costs because of the difficulty
of integrating new systems. This is true even though these firms are more
technologically sophisticated (Raymond and Paré, 1992). The marketing
literature has shown for a wide range of consumer products that customers
tend to show considerable brand loyalty. Technologies such as CRM sys-
tems and other data analysis technologies can help reveal these differences
among customers which can be coupled to retention investments. One
commonly used customer retention strategy is to offer incentives and re-
wards to customers to make it worthwhile to return. Bell and Lal (2002)
note that a successful implementation of a loyalty program should (1)
reduce price competition and increase their profits due to switching costs
and (2) reduce marketing expenses by focusing attention on retaining the
loyal customers and capturing an increasing share of their wallet.
5 A framework for managing switching costs
5.1 Introduction
In order to effectively manage customer retention, it is important to have

a framework for managing customer retention and methods of measuring
switching costs and understand the factors that influence them. In this
section, we propose a conceptual framework for understanding the role of
product strategy and switching cost, and integrating the switching cost
(economic), customer loyalty (marketing), and systems perspectives of
switching cost in electronic markets, as depicted in Fig. 1.
In this conceptual model, a firm’s customer base is driven by two major
factors: (1) the share of new customers which is determined by the overall
attractiveness of the product relative to other competing products (facto-
ring influencing customer acquisition may include price and other non-price
attributes, as well as other marketing variables such as advertising, place,
and promotion) and (2) the degree of retention of existing customers, which
is affected by both product utility and switching costs. Switching costs, in
Marketing
efforts
Product Relative
Utility Acquisition
attributes
Firm
practices
Customer Switching Customer

loyalty Retention
costs base
Product
nature
Market factors
• Search costs
• Regulation
Fig. 1. Integrated view of switching cost strategy.
turn, can be driven by firm practices (which result from retention invest-
ments such as loyalty programs), customer loyalty (which result from in-
trinsic customer characteristics and firms’ success in targeting/segmentation
of loyal customers), and the fundamental nature of the product (such as
purchase frequency, learning, or customer co-investment). In addition,
product design may also influence consumer switching costs.
Regardless of firms investments in creating switching costs, there are large
and significant market-wide switching costs in most markets. While these
factors do not enable firms to differentiate themselves, they provide a
source of market inefficiency that reduces competition overall and enables
strategies (such as price discrimination) to be employed that can only be
profitably used when markets are imperfectly competitive. Search costs play
an important role in determining market-wide switching costs—when
search costs are high, consumers do not evaluate alternatives and are
therefore much less likely to switch product providers.
This framework offers an integrated model for studying retention ability
and switching cost of firms. It differs from traditional approaches by the
incorporation of customer and firm-specific switching components into the
same model, and separating switching costs and product utility in deter-
mining retention level. We now develop a simple analytical model based on
this conceptual framework to explore some competitive implications of
switching cost, especially how market characteristics affect firms’ incentives
to make retention investments and how these investments affect equilibrium
market outcomes.
5.2 A model of customer retention
Consider a market with multiple customers (who may have different

preferences) and multiple firms (who may offer different products and
have different levels of switching cost). We assume that customers choose
among firms with a stochastic process based on the random utility frame-
work. Customers’ utilities (u) are comprised of two parts: a systematic
component related to the observable and unobservable characteristics of
the good (n), and a random component which is idiosyncratic to an indi-
vidual customer and arises due to specific tastes or random error in selec-
tion (e). To simplify the mathematical structure of our model we adopt the
discrete choice, multinomial logit formulation (MNL) to describe customer
behavior, which will be described in more detail below. We allow firms to
change their switching cost level through investment and presume that
customers do not incorporate switching costs into their initial product
adoption choices.
Our model will focus on predicting short-run (two period) firm behavior
in transitioning from some initial state (market share, product quality,
prices and switching cost level for each firm) to a new state where allowable
transitions are defined by a technology. There is no requirement that the
initial state represent any sort of equilibrium as it could arise from a variety
of situations (different fixed cost of entry, previous period growth, nature of
competition, etc.), which could yield a wide variety of potential initial
equilibria. However, given an initial state we require that firms make op-
timal pricing and investment choices to maximize second period profits.
Extensions to multiple future states (including multi-period stable
equilibria) and a complete characterization of potential equilibria for the
initial states are beyond the scope of this analysis.
The MNL choice formulation suggests that for a set of N consumers
choosing among M firms,8 we can write the utility of a particular consumer
if she chooses firm j (jA[1,2,yM]) as:
uj ¼ vj þ j . (1)
If we observe a customer choosing firm j, we can infer that this choice
provides the consumer with the highest utility over the set of M firms. That
is, the probability that a consumer will choose firm j is determined by the
relative utility level:
pj ¼ prob:ðuj uk ; 8kÞ. (2)
Under MNL, the error term is assumed to be independent and identically
distributed

with a double exponential distribution (i.e., prob:ðj Þ
¼ ee ; where 1oo1). Assuming these error terms are independent
8
We assume the market is covered, and every customer will choose one and exactly one firm.
and identically distributed this yields a very simple expression for choice
probability:
evj
pj ¼ . (3)
P
M
v
el
l¼1
We have assumed that a firm could invest to build switching cost to pre-
vent its customers from switching. For a consumer who chose firm j in period
1, her utility from choosing to stay with firm j in period 2 remains the same
(except for the random component):
ujj ¼ vj þ j . (4)
(as before, the notation uab denotes the utility a customer gets if she chooses
firm a in period 1 and then switches to firm b in period 2; when a ¼ b, the
customer stays with her existing provider).
If the customer decides to replace product j with another (product k), she
incurs a disutility or switching cost. We allow for firm variation in switching
costs, but impose a simplifying assumption that switching cost is only de-
termined by your existing choice, not any new product choice.9 We use Sj to
indicate the costs of switching out of product (or provider) j. With this
setup, the utility of choosing a firm other than j, say k, will be reduced by Sj:
ujk ¼ vk Sj þ k . (5)
This yields a retention rate (r) for firm (j) of:
evj
rj ¼ pjj ¼ P . (6)
evj þ evl sj
laj
We also define a firm’s relative ability to attract switchers from other

firms as pjs, which may or may not be the same as pj. For example, ex-
perienced customers (who have been in the market before) may value
different attributes more or less than new customers. However, while pjs can
be different from pj they are determined by the same underlying attributes.
Consequently, improving an attribute will increase both choice probabilities
proportionally. Since a customer switching out of firm k will not include k
in the choice set, the acquisition rates for firm j, where j6¼k, is given by pjs/
1pks.
Given any initial market share distribution (a vector ms, with components
msj8j) for J firms (indexed j ¼ 1yj), the vector of market shares for next
period (ms0 ) is determined by the market growth rate (g, a percentage growth
from previous state), the rate of customer retention for firm j (rj from 6), and
9
As discussed earlier, programs such as competitive upgrades can create differing switching costs
between any two firms (which need not be symmetric). We will ignore the switching destination in this
analysis.
the rate of customer acquisition for firm j (pj from 3 and pjs). Note that
according to our framework pjs and pj are functions of product attributes
(v), and retention (r) is a function of product attributes and switching costs
(s) which can be altered by investment at the firm level. This yields the
following expression for future market share as a function of previous mar-
ket share:10
msj X msl psj

ms0j ¼ r j þ g pj þ ð1 rl Þ . (7a)
1þg laj
1þg 1 psl
Alternatively, we can express this in terms of the total number of cus-

tomers (Nj) and number of new customers (n):
X psj
N 0j ¼ N j rj þ n pj þ N l ð1 rl Þ . (7b)
laj
1 psl
One implication of this formulation is that a firm’s utility has two effects
on market share, a direct effect in that better products will be purchased by
more customers initially (from 3), and an indirect effect in that better prod-
ucts will have higher retention rates (from 6). Thus, a high retention rate does
not necessarily indicate high switching costs but maybe just a result of high
utility level (due to superior quality or lower price) from the product.
To model search cost we follow the approach of Salop and Stiglitz (1977)
in dividing the customer population into informed customers with zero
search costs (a proportion m) and uninformed customers with identical
positive search costs (a proportion 1m). Presumably, electronic markets
enable more consumers to become informed which leads to an increase in m.
For simplicity, we assume that the search costs for uninformed customers
are sufficiently large to deter search.11 This implies that customers with zero
search costs always buy the product with highest utility given the product
utility levels and idiosyncratic preferences (in the form of the residual in the
random utility framework), and reevaluate their preferences each period
based on utility levels and switching costs. Uninformed customers choose
products at random and do not switch (since they have no information
about alternatives). This yields the following variants of equation 7a and b
10
Note that we can incorporate attrition rates in this model. As long as attrition rates are the same
across firms, the market share of each firm after accounting for attrition rates will not change. In
addition, we have implicitly assumed that each sale is equally profitable. However, in cases where not
each sale is equally profitable, we can easily extend our model by incorporating a weight to indicate
different profitability or segment customers into different profitability gradient and apply the model for
each segment.
11
If we know that customers face a specific amount of search cost, then it becomes switching cost.
incorporating the two customer types:

" #
msj X msl psj
0
msj ¼ m rj þ g pj þ ð1 rl Þ
1þg laj
1þg 1 psl

msj 1
þ ð1 mÞ þ g , ð8aÞ
1þg J
" #
X psj 1
0
or N j ¼ m N j rj þ n pj þ N l ð1 rl Þ þ ð1 mÞ N j þ n .
laj
1 psl J
(8b)
This equation offers a framework for obtaining optimal retention level
for firms if given cost and price information. The effect of exogenous fac-
tors can also be straightforwardly determined by differentiating this
expression.
Several interesting results could be obtained from this equation (all re-
sults are stated without proof; the proofs are available from the authors
upon request):
‘‘Own-firm’’ effects:
1. Without superior product or high switching cost, reductions in search
cost have a negative impact on firms with a large market share.12
This result suggests that it becomes more difficult to sustain market share
advantage as search cost goes down, and the market may become
more unstable. In digital markets where transaction costs and search costs
are greatly reduced, it implies that firms with large market share in
the traditional market will not be able to guard their market share if they
do not improve their product value or invest in switching costs. This ar-
gument is supported by Clemons et al. (1996) and is consistent with much of
the discussion of competition between ‘‘bricks and mortar’’ and online
retailers.
2. The marginal value of both investments in quality and retention is in-
creasing as search cost goes down (i.e., m goes up) and as market share
goes up

@ð@ms0j =@vj Þ @ð@ms0j =@sj Þ
0; 0
@m @m
12
For example, assume that pi ¼ psi ¼ 1=J and si ¼ 0 8i; we can show that
ð@ms0j =@mÞ0 for ms0j 41=J:
and

@ð@ms0j =@vj Þ @ð@ms0j =@sj Þ
0; 0
@msj @msj
Moreover, retention investment becomes relatively more important when
growth is slow or market share is high.
3. All other factors being equal, the minimum retention level required to
sustain a firm’s market share is increasing in market share, number of
firms in the market and competitors’ investment in retention. The mini-
mum retention level required to sustain a firm’s market share is also
increasing in market growth when the firm’s market share is greater than
average.
Result 2 suggests that quality investments and retention investments are
more important in markets with lower search costs (e.g., online retailing).
Result 3 suggests that these investments are more important for firms with a
large market share to defend. This provides another reason why initially
dominant firms in online retailing are often also the most aggressive in
offering technologies such as personalization or recommendation systems
that encourage customer retention.
‘‘Cross-firm’’ effects:
4. The marginal value of product quality investments is a decreasing func-
tion of competitors’ switching costs ð@ð@ms0j =@vj Þ=@sk 0; 8kajÞ:
This result suggests that retention investments and product investments
are strategic substitutes. In specific, it suggests that investments in switching
costs construct not only entry barriers, but also make competitors less
aggressive in acquiring new customers or improving their product offerings.
Although we do not know the exact reasons, we do observe that Barne-
sandNoble.com was less aggressive in their investments in both website
design and recommendation systems, while Amazon.com has been the pio-
neer in technology investments for customer retention.
Collectively, these results suggest a close linkage between investments in
product utility and investments in retention. In high-growth markets, firms
have greater incentive to invest in product utility and customer acquisition
than retention. These effects tend to be more important for incumbents
than entrants who need not consider the impact of their actions on their
existing customer base.
5.3 Measuring switching costs
The choice framework proposed in the previous section and the associ-
ated discussion can also be utilized for empirical work where the underlying
goal is to estimate switching costs for multiple firms and identify the factors
that affect switching behavior. As noted in Section 2, using retention rate as
proxy for firm’s switching cost is misleading because product attributes and
marketing variables affect retention rates as well. Building upon the ran-
dom utility modeling framework, Chen and Hitt (2002a) develop a simple
strategy that allows us to quantify switching costs by filtering out the effect
from other types of investments, such as product improvement or price
reductions. We briefly describe the measurement strategy here.
As before, we model the buyer i for vendor choice j as uij ¼ vij þ ij ; which
is comprised of two parts: it contains a component (vji) which captures the
measured preference of buyer i for a particular vendor j and a random
component (eji) which summarizes the contribution of unobserved varia-
bles. We extend the earlier discussion by expressing vji, the systematic utility
component, as a function of customer characteristics and product at-
tributes. Depending on customers past behavior (using notation introduced
in Section 2) yields a set of utilities for different choices:
ui+j ¼ gij þ X j bij þ Z i lij Si+j þ ij ,

uijj ¼ gij þ X j bij þ Z i lij þ ij ,
uikj ¼ gij þ X j bij þ Z i lij S i+j Sik+ Sikj þ ij ,
uik+ ¼ S ik+ ; and
ui++ ¼ 0,
gji captures buyer i’s unobserved inherent tastes for product j (or unob-
served buyer–seller match). Xj is a vector of firm or product attributes
including price or cost index and a collection of other relevant attributes,
which may include marketing variables. bji is a vector of relative weights
customer i puts on different firm or product attributes. Zi is a set of ob-
served customer characteristics and vector lji captures customer preference
parameters to justify that customers may have heterogeneous preference
over the choices she has. S+ji is the initial adoption cost of product j
(beyond the price paid). Note that adoption cost can be negative if a firm
subsidies new customers for adoption. Skii is the costs incurred by customer
i from switching from k to j, and Sk+i the exit cost. Finally, eji, the random
component summarizes the contribution of all other unobserved variables
as well as customer i’s idiosyncratic, specific tastes or random error in
selection.13 Each consumer will choose the product which maximizes her
13
Note, we can further decompose eji into two or more variables capturing effects from unobserved
factors (e.g., unobserved product attributes or demand shift) and customers’ idiosyncratic tastes. This
strategy is often adopted when there are suitable instruments to capture the effects of some unobserved
factors (see Berry et al., 1995; Nevo, 2000).
utility, that is, a customer (i) will choose product j if and only if
uij 4uik ; 8kaj; which implicitly defines the set of consumer attributes that
lead to the choice of good j.
However, the estimations of the most general framework require exten-
sive data on customer and firm characteristics and over time, so most em-
pirical models are simplified by putting a number of restrictions. For
example, many models examine buyer choices conditional on customers
making a purchase decision (i.e., do not consider customers who do not
make any purchases). Moreover, almost all, if not all, models assume
S i+j ¼ 0 and Sik+ ¼ 0 and put restrictions on Skji. The economics literature
typically assumes Sikj ¼ S 8i; j; k: Chen and Hitt (2002b) and Chen and
Forman (2006) assumeS ikj ¼ Sk 8i; j; where switching cost depends on only
where the customer is from but not where she switches to. In addition, it is
not always possible to estimate gij ; bij ; or lij (which collectively determines
consumer demand absent switching costs): the estimation of which usually
requires many observations from the same individual. As an alternative, it
is often assumed that unobserved individual attributes are random draws
from a known distribution (e.g., random-coefficients logit model).
The simplest case to estimate consumer demand for differentiated products
is the standard logit model, which assumes that unobserved individual het-
erogeneity enters the model only through the error term, eji, that is, vji ¼ vj (or
equivalently, gij ¼ gj ; bij ¼ bj ; and lij ¼ lj ), and the error term is independ-
ently and identically distributed across products and
consumers with the
‘‘extreme value’’ distribution (i.e., prob:ðj Þ ¼ ee ; where 1oo1).
The market share (or the choice probability)
M of product j in the absence of
P vl
switching costs is given by pj ¼ evj e :
l¼1
However, this type of error structure is governed by independence of
irrelevant alternatives (IIA)—that is, the ordinal ranking of any two prod-
ucts does not depend on the attributes of other alternatives or even the
presence or absence of an alternative choice. This assumption can produce
unreasonable substitution patterns. The MNL framework can be general-
ized using the McFadden (1978) generalized extreme value (GEV) model
that allows for richer patterns of substitution among alternatives. One such
example is the nested logit model, where choices are grouped into different
clusters and choices within a cluster are better substitutes for each other
than choices across clusters (this has been applied by Chen and Forman,
2006). Alternatively, mixed logit models can be used (McFadden and Train,
2000), which allow the parameter associated with each observed variable
(e.g., its coefficient) to vary randomly across customers, thus allowing
for more complex demand elasticities across products and possibly price
endogeneity (see Berry, 1994; Nevo, 2000, for applications of these
approaches). Forman and Chen (2005) apply the mixed logit model to
the network switchgear industry.
The framework introduced in this section can be used to measure switch-

ing costs using any data source that has multi-period customer choice data,
even when customer or product characteristics are not known for these
customers. Indeed, since the resulting choice probabilities in large samples
are equivalent to market shares, Chen and Hitt (2002a) have shown that,
when the preferences for new customers and existing customers are drawn
from the same distributions, switching cost estimates can be made at an
aggregate level, only knowing the starting market share of each product and
the breakdown of next period market share into new customers, switchers,
and existing customers for each firm.
When more detailed data on customers and firms is available, we can
estimate a more general demand model and also distinguish between
different sources of switching costs. Chen and Hitt (2002b) estimated a logit
model with data from online brokerage usage derived from Media Metrix, a
firm that tracked customers’ usage of the Internet. They found considerable
variation in switching costs across online brokers, and that switching cost
and product quality were not perfectly correlated. These two observations
suggest that at least in this industry, different strategies pursued by different
firms yield different marketplace outcomes in terms of switching cost and
the resultant impact on overall market share. Using more detailed data on
customer and firm characteristics, they found that switching was lower for
firms that had minimum investment requirements to open an account and
that offered a broader product line. Customers who were frequent users
were also more loyal, while customers who changed their usage pattern or
had adopted multiple brokers at one time were much more likely to switch.
Forman and Chen (2005) utilize this framework, adopting a mixed logit
model to estimate the impact of network effects and other sources of
switching costs on vendor choice in the market for routers and switches.
Their results show that the size of a firm’s network as well as learning costs
significantly increases the switching costs of changing vendors. They also
demonstrate that although new IT innovation (i.e., switches) did tempo-
rarily lead to lower switching costs, there still remained significant costs of
switching vendors.
6 Conclusion
Previous theoretical work has shown that the presence of switching costs
can have a substantial effect on profitability. In this paper, we survey prior
literature on IT and switching costs and argue that switching costs man-
agement becomes more important in high-tech and information-intensive
markets. However, the creation of switching costs requires substantial and
deliberate investments by the firm in customer retention. Only by under-
standing the sources and magnitude of these switching costs it is then pos-
sible to understand tradeoffs between investments in loyalty and retention
programs and other types of investments such as advertising (for building

new customer acquisition rates), technologies, and service-level improve-
ments, and price reductions which raise both the acquisition and retention
rates simultaneously.
In this paper, we describe several sources of switching costs, especially
those relevant in high-tech and information-intensive markets, and discuss
several strategies that have the potential to influence consumer switching
costs. We further provide an integrated framework for the management of
switching costs and customer retention. Based on the framework, we have
shown that in markets with low search costs and slowing growth (common
to most retail and some commercial internet-mediated markets) customer
retention investments become more important in determining the degree of
competition and overall market structure. Overall, we have concluded that
conditions exhibited in electronic markets like reductions in search cost and
entry barriers have strengthened firms’ incentives in strategic retention in-
vestments. Moreover, lowered barriers to entry introduce new competitors
and many asymmetries into the market; thus creating different incentives in
retention investment across firms and result in heterogeneous switching
costs. In addition, we provide a framework for measuring switching costs.
The ability to measure switching costs not only allows firms to understand
the effectiveness of existing retention investments but to measure the out-
comes of new customer retention initiatives.
In closing this survey, we would like to point out a few opportunities for
future research. First of all, the availability of more extensive data on both
customers and firms makes it possible to estimate the magnitudes and drivers
of consumer switching costs more directly and more precisely. As noted in
Section 2 and also from previous literature (e.g., Heckman, 1981), one major
challenge in identifying switching costs lies in the difficult to separate true
switching costs (or ‘‘true state dependence’’) from spurious state dependence.
Spurious state dependence occurs when a buyer continues to purchase the
same product for reasons unrelated to real switching costs (e.g., the product
has a better ‘‘fit’’). With more extensive data on products and customer
behavior, it becomes possible to estimate real switching costs with proper
controls for spurious state dependence. The ability to measure switching
costs and their sources has important strategic implications. It allows firms to
better evaluate their investments and various strategies. For example, it be-
comes possible for firms to estimate the amount of switching costs associated
with adopting a broad product line strategy or a particular loyalty program.
The ability to measure switching costs also offers great opportunities in
testing hypotheses from previous theory literature. For example, Nilseen
(1992) predicts that there is a difference between switching costs that are
incurred each time a consumer changes supplier (or transactional costs), and
‘‘learning’’ costs that are incurred each time a consumer uses a supplier for
the first time. Transactional switching costs give consumers less incentive to
switch than do learning switching costs and lead to lower prices for new
consumers and higher prices for loyal consumers. The ability to unpack
switching costs thus makes it possible to test this prediction. Moreover,
previous theory has shown that a firm with large installed base would prefer
to adopt an incompatible technology (i.e., choosing high switching costs),
while firms producing differentiated products would prefer compatibility
(i.e., choosing zero switching costs). It would also be interesting to see
whether this holds true empirically. In addition, how switching costs affect
entry and product differentiation may also be answered empirically.14 It
would also be interesting to investigate whether and when a firm is better off
charging lower or prices to loyal customers with real data.
On the other hand, several conditions that are present in traditional
markets have been changed as technology advances, which may render
many previous predictions invalid and therefore provide new research op-
portunities. In particular, the advances of the Internet and communication
technologies have also brought along new business opportunities and made
new strategies possible: new markets (e.g., Ebay.com, Priceline.com, search
markets) emerge; new ‘‘products’’ (such as search results and recommen-
dations, automatic agent services, customized offerings) are created, and
new business strategies (e.g., pay-per-use, customized bundling such as
creating your own CDs) are also made possible. The relationship of these
strategies to customer retention is largely unknown. In some cases, the
effects are theoretically ambiguous. For example, personalization technol-
ogies allow firms to serve individual customer needs better, which many
have argued could lead to higher consumer switching costs. However, as
technology allows firms to identify and serve customer needs more per-
fectly, it can also increase competition and undermine differentiation (Chen
et al., 2001; Wattal et al., 2004). Thus, much of the relationship between
new Internet-enabled strategies and customer retention must be examined
empirically. What is clear, however, is that customer retention is becoming
increasingly important component of strategy and that understanding and
managing retention is especially critical in information-intensive markets.
This also suggests that the design of loyalty programs under different mar-
ket conditions will be an important subject.
Acknowledgments
We would like to thank Michael Baye, Eric Bradlow, Eric Clemons, Chris
Forman, Avi Goldfarb, Terry Hendershott, Paul Kleindorfer, John Morgan,
Ivan Png, Sandy Slaughter, Detmar Straub, Hal Varian, Dennis Yao, Shin-yi
Wu, and seminar participants at Carnegie-Mellon University, Georgia
Tech, MIT, New York University, the University of British Columbia, the
14
Previous theory suggests that switching costs can either discourage or promote entry and that
switching costs may lower firm incentives to differentiate (see the survey by Farrell and Klemperer,
2004).
University of California at Berkeley, the University of California at Irvine,

the University of Maryland, the University of Rochester, the University of
Texas at Austin, the Wharton School, the Workshop on Information Sys-
tems and Economics, and the International Conference on Information
Systems (ICIS) for helpful comments and suggestions that contributed to this
paper. Funding for this research was provided by Wharton eBusiness In-
itiative, National Science Foundation (Grant IIS-9733877), and the Wharton
School.
References
Acquisti, A., H.R. Varian (2005). Conditioning prices on purchase history. Marketing Science 24(3), 1–15.
Andrews, R.L., I.S. Currim (2004). Behavioural differences between consumers attracted to shopping
online versus traditional supermarkets: implications for enterprise design and marketing strategy.
International Journal of Internet Marketing and Advertising 1, 38–61.
Ausubel, L. (1991). The failure of competition in the credit card market. American Economic Review 81,
50–81.
Bakos, Y. (1997). Reducing buyer search costs: Implications for electronic marketplaces. Management
Science 43, 1676–1692.
Baye, M.R., J. Morgan (2004). Brand and price advertising in online markets. Working Paper, Indiana
University and UC-Berkeley.
Baye, M.R., J. Morgan, P. Scholten (2004). Price dispersion in the small and in the large: Evidence from
an internet price comparison site. Journal of Industrial Economics 52, 463–496.
Baye, M.R., J. Morgan, P. Scholten (2005). Information, Search and Price Dispersion. This handbook.
Bell, D., R. Lal (2002). The impact of frequent shopper programs in grocery retailing. Review of
marketing science Working Papers, 2.
Bergen, M., S. Dutta, S.M. Shugan (1996). Branded variants: A retail perspective. Journal of Marketing
Research 33, 9–19.
Berry, S. (1994). Estimating discrete-choice models of product differentiation. The RAND Journal of
Economics 25(2), 242–262.
Berry, S.J., Levinsohn, A. Pakes (1995). Autombobile prices in market equilibrium. Econometrica 60(4),
889–917.
Blattberg, R.C., J. Deighton (1996), Managing marketing by the customer equity tests. Harvard Business
Review, July–August, 136–144.
Bresnahan, T. (2001). The economics of the Microsoft case. Presented at the MIT Industrial Organization
workshop.
Bresnahan, T., S. Greenstein (1996). Technical progress and co-invention in computing and the uses of
computers. Brookings Papers on Economic Activity: Microeconomics, 1–83.
Bresnahan, T., S. Greenstein (1999). Technological competition and the structure of the computer
industry. Journal of Industrial Economics 47, 1–39.
Breuhan, A. (1997). Innovation and the persistence of technological lock-in. Ph.D. dissertation, Stanford
University, Palo Alto.
Brynjolfsson, E., C. Kemerer (1996). Network externalities in Microcomputer Software: an econometric
analysis of the spreadsheet market. Management Science 42, 1627–1647.
Brynjolfsson, E., D. Fitoussi, L.M. Hitt (2005). How big is IT? MIT Working Paper.
Bucklin, R.E., C. Sismeiro (2003). A model of web site browsing behavior estimatedon clickstream data.
Journal of Marketing Research 40, 249–267.
Calem, P., L. Mester (1995). Consumer behavior and the stickiness of credit-card interest rates.
Caminal, R., C. Matutes (1990). Endogenous switching costs in a Duopoly model. International Journal
of Industrial Organization 8, 353–374.
Chen, P-Y. (2002). Essays on the economics of electronic markets: pricing, retention and competition.
Ph.D. dissertation, The Wharton School.
Chen, P-Y., C. Forman (2006). Can vendors influence switching costs and compatibility in an envi-
ronment with open standards. MIS Quarterly 30, 541–562.
Chen, P-Y., L. M. Hitt (2002a). The economics of retention and switching costs: measurement and
applications. Working Paper, Carnegie Mellon University and The Wharton School.
Chen, P-Y., L.M. Hitt (2002b). Measuring switching costs and the determinants of customer switching
in internet-enabled businesses: A study of the online brokerage industry. Information Systems
Research 13, 255–274.
Chen, Y., C. Narasimhan, Z.J. Zhang (2001). Individual marketing with imperfect targetability.
Marketing Science 20, 23–41.
Choi, J.P. (1997). The provision of (two-way) converters in the transition process to a new incompatible
technology. Journal of Industrial Economics 45, 167–182.
Clemons, E., I-H. Hann, L.M. Hitt (2002). The nature of competition in electronic markets: An em-
pirical investigation of on-line travel agent offerings. Management Science 48, 534–549.
Clemons, E.K., D.C. Croson, B.W. Weber (1996). Market dominance as a precursor of a firm’s failure:
emerging technologies and the competitive advantage of new entrants. Journal of Management In-
formation Systems 13, 59–75.
Clemons, E.K., M.C. Row, S.P. Reddi (1993). The impact of information technology on the organ-
ization of economic activity: the ‘move to the middle’ hypothesis. Journal of Management Infor-
mation Systems 10, 9–35.
Clemons, E.K., M. Thatcher (1998). Capital one: Exploiting an information-based strategy. Proceedings
of 31st Hawaii International Conference on System Sciences.
Clemons, E.K., B. Gu, R. Spitler (2003). Hyper-differentiation strategies: delivering value, retaining
profits. Proceedings of 35th Hawaii International Conference on System Sciences.
Danaher, P.J., W.W. Isaac, A.D. Robert (2003). A comparison of online and offline consumer brand
loyalty. Marketing Science 22, 461–476.
Degeratu, A., A. Rangaswamy, J. Wu (2000). Consumer choice behavior in online and traditional
supermarkets: the effects of brand name, price and other search attributes. International Journal of
Research in Marketing 17, 55–78.
Demirhan, D., V.S. Jacob, S. Raghunathan (2004). Strategic IT investments: impact of switching cost
and declining IT cost. Working Paper, University of Texas-Dallas.
Ellison, G., S.F. Ellison (2001). Search, obfuscation, and price elasticities on the internet. Working
Paper, MIT.
Elzinga, K.G., D.E. Mills (1998). Switching costs in the wholesale distribution of cigarettes. Southern
Economic Journal 65, 282–293.
Farrell, J., P. Klemperer (2004). Competition with switching costs and network effects. Draft for
Handbook of Industrial Organization, Vol 3. Available on www.paulklemperer.org.
Farrell, J., G. Saloner (1992). Converters, compatibility, and the control of interfaces. Journal of
Forman, C. (2005). The corporate digital divide: determinants of internet adoption. Management
Science 51, 641–654.
Forman, C., P.-Y. Chen (2005). Switching costs and network effects in the market for routers and
switches. Working Paper, Carnegie Mellon University.
Forman, C., A. Goldfarb (2005). ICT Diffusion to Businesses. This handbook.
Friedman, T.L. (1999). Amazon.you. New York Times, Feb. 26.
31, 634–657.
Fudenberg, D., J.M. Villas-Boas (2005). Behavior-Based Price Discrimination and Customer Recog-
nition. This handbook.
Gandal, N. (1994). Hedonic price indexes for spreadsheets and an empirical test for network exter-
nalities. RAND Journal of Economics 25, 160–170.
Greenstein, S.M. (1993). Did installed base give an incumbent any (measurable) advantages in federal
computer procurement?. RAND Journal of Economics 24, 19–39.
Goldfarb, A. (2006). State dependence at internet portals. Journal of Economics and Management
Strategy 15(2), 317–352.
Grover, R., V. Srinivasan (1987). A simultaneous approach to market segmentation and market struc-
turing. Journal of Marketing Research 24, 139–153.
Heckman, J. (1981). Statistical models for discrete panel data, in: C.F. Manski, D. McFadden (eds.),
Structural Analysis of Discrete Data with Econometric Applications, The MIT Press, Cambridge,
MA, pp. 179–195.
Jacoby, J., R.W. Chestnut (1978). Brand Loyalty Measurement and Management. Wiley, New York.
Johnson, E., W. Moe, P. Fader, S. Bellman, G. Lohse (2004). On the depth and dynamics of online
search behavior. Management Science 50, 299–308.
Johnson, E.J., S. Bellman, G.L. Lohse (2003). Cognitive lockin and the power law of practice. Journal of
Marketing 67, 62–75.
Katz, M., C. Shapiro (1994). Systems competition and network effects. Journal of Economic Perspectives
8, 93–115.
Katz, M.L., C. Shapiro (1985). Network externalities, competition, and compatibility. The American
Kauffman, R., J. McAndrews, Y.-M. Wang (2000). Opening the ‘‘black box’’ of network externalities in
network adoption. Information Systems Research 11, 61–82.
Keane, M.P. (1997). Modeling heterogeneity and state dependence in consumer choice behavior. Journal
of Business and Economic Statistics 15, 310–327.
Kim, B.-D., M. Shi, K. Srinivasan (2001). Reward programs and tacit collusion. Marketing Science 20,
99–120.
Kim, M., D. Kliger, B. Vale (2003). Estimating switching costs: the case of banking. Journal of Financial
Intermediation 12, 25–56.
Klemperer, P. (1995). Competition when consumers have switching costs: an overview with applications
to industrial organization, macroeconomics, and international trade. The Review of Economic Studies
62, 515–539.
Klemperer, P.D., A.J. Padilla (1997). Do firms’ product lines include too many varieties?. RAND
Kuksov, D. (2004). Buyer search costs and endogenous product design. Marketing Science 23, 490–499.
Lee, S.-Y., I. Png (2004). Buyer shopping costs and retail pricing: An indirect empirical test. Review of
Marketing Science 2, Article 6.
Liebowitz, S.J., S.E. Margolis (1999). Winners, Losers & Microsoft: Competition and Antitrust in High
Technology. Ind. Institute.
McFadden, D. (1974). Conditional logit analysis of qualitative choice behavior, in: P. Zarembka (ed.),
Frontiers in Econometrics, Academic Press, New York.
McFadden, D. (1978). Modelling the choice of residential location, in: A. Karlquist, L. Lundqvist, F.
Snickars, J.W. Weibull, (eds.), Spatial Interaction Theory and Residential Location. North Holland,
New York, pp. 75–96.
McFadden, D., K. Train (2000). Mixed MNL models for discrete response. Journal of Applied Econo-
metrics 15(5), 447–470.
Moe, W., P.S. Fader (2004). Capturing evolving visit behavior in clickstream data. Journal of Interactive
Marketing 18, 5–19.
Moshkin, N., R. Shacher (2000), Switching cost or search cost? Working Paper #3-2000, Foerder
Institute for Economic Research.
Nalebuff, B. (1999). Bundling. Working Paper, School of Management, Yale University.
Narasimhan, C. (1988). Competitive promotional strategies. Journal of Business 61(4), 427–449.
Nevo, A. (2000). A practitioner’s guide to estimation of random-coefficients logit models of demand.
Journal of Economics & Management Strategy 9(4), 513–548.
Porter, M. (2001). Strategy and the internet. Harvard Business Review 79, 62–78.
Raju, J., V. Srinivasan, R. Lal (1990). The effects of brand loyalty on competitive price promotional
strategies. Management Science 36(3), 276–304.
Raymond, L., G. Paré (1992). Measurement of information technology sophistication in small man-
ufacturing businesses. Information Resources Management Journal 5, 4–16.
Reichheld, F.F., P. Schefter (2000). E-loyalty: your secret weapon on the Web. Harvard Business Review
78, 105–113.
Salop, S., J. Stiglitz (1977). Bargains and ripoffs: a model of monopolistically competitive price
dispersion. Review of Economic Studies 44, 493–510.
Schmalensee, R. (1982). Product differentiation advantages of pioneering brands. American Economic
Review 72, 349–365.
Shapiro, C., H. Varian (1999). Information Rules: A Strategic Guide to the Network Economy. University
Press, Boston, Harvard.
Shum, M. (2004). Does advertising overcome brand loyalty? Evidence from breakfast cereals. Journal of
Economics and Management Strategy 13, 241–272.
Shy, O. (2002). A quick-and-easy method for estimating switching costs. International Journal of In-
dustrial Organization 20, 71–87.
Smith, M., E. Brynjolfsson (2001). Customer decision making at an internet shopbot: brand still matters.
The Journal of Industrial Economics 49, 541–558.
Stango, V. (2002). Pricing with consumer switching costs: evidence from the credit card market. Journal
of Industrial Economics 50, 475–492.
Stiglitz, J.E. (1989). Imperfect information in the product market, in: R. Schmalensee, Willig, R.D.
(eds.), Handbook of Industrial Organization, Vol. 1, Elsevier, North-Holland, pp. 769–847.
Thurm, S. (2000). Microsoft’s behavior is helping Cisco learn how to avoid antitrust trouble. Wall Street
Journal June 6, P.A1.
Viard, B.V. (2003). Do switching costs make markets more or less competitive? The case of 800-number
portability. Working Paper #1773, Graduate School of Business, Stanford University.
Villas-Boas, J.M. (2004). Consumer learning, brand loyalty, and competition. Marketing Science 23,
134–145.
Viswanathan, S. (2005). Competing across technology-differentiated channels: the impact of network
externalities and switching costs. Management Science 51, 483–496.
Wattal, S., P.-Y. Chen, R. Telang (2004). On personalization technology adoption and information
sharing in digital markets. Working Paper, Carnegie Mellon University.
Wickre P. (1996). Stick to true standards-based solutions. Network World Fusion July 22. http://
www.nwfusion.com/archive/1996/96-07-22stic.html
Zhu, K.K.L., V. Kraemer, V. Gurbaxani, S. Xu (2006). Migration to open-standard Inter-organiza-
tional systems: Network effects, switching costs, and path dependency. MIS Quarterly 30, 515–539.
Chapter 9
The Economics of Privacy
Kai-Lung Hui
Department of Information Systems, City University of Hong Kong, Hong Kong
I.P.L. Png
Department of Information Systems, National University of Singapore, Singapore
Abstract
This chapter reviews economic analyses of privacy. We begin by scrutinizing

the ‘‘free market’’ critique of privacy regulation. Welfare may be non-mono-
tone in the quantity of information; hence, there may be excessive incentive to
collect information. This result applies to both non-productive and produc-
tive information. Over-investment is exacerbated to the extent that personal
information is exploited across markets. Further, the ‘‘free market’’ critique
does not apply to overt and covert collection of information that directly
causes harm. We then review research on property rights and challenges in
determining their optimal allocation. We conclude with insights from recent
empirical research and directions for future research.
1 Introduction
Information privacy has been defined as the individual’s ability to control

the collection and use of personal information (Westin, 1967; Stigler, 1980).
The invention and development of computing technologies led to wide-
spread concern about collection of personal information in various con-
texts, including employment, finance, marketing, and government. In
response to these concerns, the US Congress passed the Privacy Act of
1974, the Organization for Economic Co-operation and Development
(OECD) published guidelines on privacy protection and transborder data
flow (OECD, 1980), and the European Union (EU) adopted Directive
471
472 K.-L. Hui and I.P.L. Png
95/46/EC on data protection. The EU directive prohibits transfer of infor-

mation to jurisdictions that do not accord adequate protection.
The development of the Internet and the advent of e-commerce have
amplified public concern about privacy. With every web site visit, a browser
leaves an electronic trace which can later be retrieved to analyze the con-
sumers’ online browsing and shopping behavior. Another technology—the
cookie—stores identifying information about consumers. Using clickstream
and identifying information, web sites can profile visitors. Such profiling
could benefit consumers by more precisely identifying their needs.1 How-
ever, it can also be used to effect price discrimination or exclude individuals
with less attractive characteristics.2 Some organizations even sell customer
information to third parties, which subject their customers to further pri-
vacy intrusion.3,4
Clearly, technology has significantly changed business practices, but new
opportunities present new concerns. Westin (2001) concludes ‘‘There has
been a well-documented transformation in consumer privacy attitudes over
the past decade, moving concerns from a modest matter for a minority of
consumers in the 1980s to an issue of high intensity expressed by more than
three-fourths of American consumers in 2001.’’
Within the United States, the Federal Trade Commission (FTC) oversees
personal information privacy in consumer transactions. In the 1990s, the
FTC emphasized fair information practices (FIPs) in its policy toward
consumer privacy. Subsequently, however, the FTC revised its thinking and
considered that the cost of obtaining consumers’ consent for information
sharing and use would far exceed the potential benefit (Muris, 2003). Ac-
cording to this view, the FIPs were inefficient and the FTC should follow
the approach under the Fair Credit Reporting Act of 1970. This approach
generally allows the use of personal information, while focusing on en-
forcement against misuses. Specifically, ‘‘the most important objective of a
privacy agenda should be stopping practices that can harm consumers’’
(Muris, 2003).
The other major privacy issue identified by Muris (2003) was spam:
‘‘Spam is one of the most daunting consumer protection problems that the
1
See, for instance, Moe and Fader (2001, 2004), Bucklin and Sismeiro (2003), Montgomery et al.
(2004), and Sismeiro and Bucklin (2004).
2
‘‘Giving the web a memory cost its users privacy,’’ New York Times, September 4, 2001. Ama-
zon.com’s application of dynamic pricing illustrates consumers’ privacy dilemma (‘‘On the web, price
tags blur; what you pay could depend on who you are,’’ Washington Post, September 27, 2000).
3
For instance, Amazon.com’s privacy policy states: ‘‘As we continue to develop our business, we
might sell or buy stores, subsidiaries, or business units. In such transactions, customer information
generally is one of the transferred business assets y in the unlikely event that Amazon.com, Inc., or
substantially all of its assets are acquired, customer information will of course be one of the transferred
assets.’’
4
The U.S. Federal Trade Commission (2005a) has taken enforcement action against an online shop-
ping cart provider that rented customer information to third-party marketers, in violation of the dis-
closure policies published to consumers using the shopping cart.
Ch. 9. The Economics of Privacy 473
Commission has ever faced.’’ Muris worried whether legislation and legal
sanctions could resolve the problem of spam.
Clearly, privacy is an important policy and business issue. What has been
the contribution of academic scholarship, and in particular, economics, to
the issue? Academic discourse on individual privacy dates back at least to
the seminal Harvard Law Review article of Warren and Brandeis (1890).
Privacy is a multi-disciplinary issue that has been and should be analyzed
from multiple perspectives—law, psychology, sociology, political science,
and economics.5 Economics is an especially appropriate discipline as it
provides a framework to appreciate the key trade-offs in policy toward
privacy.
The earliest economic analyses of privacy focused on the efficiency of
markets for personal information. Since the Privacy Act of 1974 regulated
only government records, the immediate issue was whether the collection
and use of personal information by private sector entities should be reg-
ulated. The ‘‘Chicago School’’ (Posner, 1978, 1979, 1981; Stigler, 1980)
contended that regulation is not needed—markets for personal information
would work as well as markets for conventional goods and services.
However, the Chicago School’s argument ignored the ways in which
personal information is collected. Realistically, accurate personal informa-
tion does not come from nowhere; resources must be expended to collect
the information, and the collection could have undesirable consequences on
consumer welfare.
For the most part, the Chicago School focused on just one dimension of
privacy, viz., secrecy, and overlooked two other dimensions—autonomy
and seclusion (Hirshleifer, 1980; Camp and Osorio, 2003). While secrecy
concerns privacy of information, autonomy concerns freedom from obser-
vation and seclusion concerns the right to be left alone. Besides markets for
secrecy, we are also interested to know whether markets for autonomy and
seclusion work well.6,7
From an economic standpoint, governments, businesses, and other
organizations use personal information about individuals in three ways.
First, they use personal information to customize goods and services, dis-
criminate more effectively between people with differing willingness to pay
5
See, for example, Culnan and Bies (2003), Eddy et al. (1999), Goodwin (1992), Hirshleifer (1980),
Laudon (1996), Petty (2000), Posner (1978, 1979, 1981), Schwartz (1968), Smith (2001), Stigler (1980),
Stone and Stone (1990), Tolchinsky et al. (1981), and Woodman et al. (1982).
6
Hirshleifer (1980) cited telemarketing as an example of violation of autonomy. Actually, telemar-
keting requires personal information, viz., a telephone number, and involves an intrusion into the right
to be left alone, hence it involves violation of secrecy and seclusion. An example that more clearly
exemplifies autonomy is nude sunbathing. A peep does not need the subject’s personal information to
intrude on the subject’s autonomy.
7
Posner (1981) did acknowledge the definition of privacy as peace and autonomy, but he dismissed
these aspects by saying ‘‘to affix the term privacy to human freedom and autonomy (as in Jack Hi-
rshleifer) is simply to relabel an old subject—not to identify a new area for economic research y the
range of economic applications in this area seems limited.’’ (p. 405)
or differing reservation wage, and sort more effectively among people with
different personal characteristics (Mussa and Rosen, 1978; Katz, 1984;
Moorthy, 1984; Varian, 1985; Hart and Tirole, 1988; Tirole, 1988, Chapter
3; Png, 1998, Chapter 9). The Chicago School posits that these uses of
personal information lead to socially efficient outcomes and require no
government regulation.
However, the use of personal information to profile individual persons
imposes an indirect or consequential externality as some suffer from paying
relatively higher price, receiving a relatively lower wage, or being excluded
from enjoying a particular good or service. Hence, the exploitation of per-
sonal information could lead to ex post inefficiencies. Hirshleifer’s (1971)
classic analysis shows that the result of such information might simply be
re-distribution, and so, from a social viewpoint, there might be over-
investment in information. Even if consumer information is costless,
the seller’s private incentive to maximize profit may be inconsistent with
maximizing social welfare. Some consumers may get priced out of the
market when more information is available to the seller, even though it is
socially efficient for them to consume the item (Varian, 1985; Hart and
Tirole, 1988; Thisse and Vives, 1988; Fudenberg and Villas-Boas, 2006).
Second, a seller may collect personal information in one market for use
by itself or others in another market. Then, the seller may have an excessive
incentive to collect consumer information, at the expense of some of its own
potential consumers (Taylor, 2004).8 That is, the option of selling consumer
information for extra revenue may further reduce social efficiency both
from benefit (loss in trades and increase in deadweight losses) and cost (the
effort in compiling the information) perspectives.
The third way in which organizations use personal information about
potential clients is to direct unsolicited promotions, in person, by mail,
telephone and fax, and electronically. These solicitations impose costs of
intrusion on recipients and are a direct externality. Unsolicited marketing is
one type of intrusion against seclusion (Camp and Osorio, 2003). A pref-
erence for seclusion is like a taste for privacy in that intrusions cause a
direct externality, unrelated to any effect on the terms of any transaction or
trading relationship (Laudon, 1996). Computing technologies have facili-
tated a flood of unsolicited promotions, which cause annoyance and affect
productivity. However, most privacy research has ignored the implications
of these uses of personal information.
Finally, opposing views on privacy and information use have led to dif-
ferent suggestions on whether property rights in personal information
should be established and how they should be assigned. The Chicago
School supports free collection and use of information; hence the issue of
8
European Union Directive 2001/29/EC grants copyright protection to compilers of databases even if
they did not create the information compiled. This right would further encourage sellers to develop
consumer databases.
property rights is moot. Hermalin and Katz (2006) suggest that individuals
might voluntarily reveal their personal information to trading partners
anyway. Therefore, it does not matter how property rights are assigned.
However, others argue that exclusive rights should be granted to individuals
so that they can control the collection and subsequent use of their infor-
mation (Noam, 1995b; Laudon, 1996). Marketers would then internalize
the privacy costs that they impose on consumers. We examine each of
these arguments and highlight some challenges in determining the optimal
allocation of property rights.
This chapter reviews economic analyses of privacy. Section 2 begins with
the free market approach. Sections 3 and 4 discuss the indirect consequen-
tial externality that arises from the use of personal information. Section 5
reviews direct externalities. Then, Sections 6 and 7 discuss the possible
resolution of privacy through property rights and regulations. Section 8
reports some empirical findings, while Section 9 concludes with directions
for future research.
2 ‘‘Free market’’ approach
The Chicago School (Posner, 1978, 1979, 1981; Stigler, 1980) resolutely
affirms that markets for personal information would work as well as mar-
kets for conventional goods and services. Government regulation would
impede economic efficiency. For instance, unskilled workers would suffer
relatively more than skilled workers from restrictions on employers in the
collection and use of personal information about workers. Likewise, low-
income borrowers would suffer relatively more than wealthy borrowers
from restrictions on lenders in the collection and use of personal informa-
tion about borrowers.
The ‘‘free market’’ approach to privacy may not work efficiently, how-
ever, for several reasons. First, the Chicago School focuses on ex post
efficiency, but overlooks that open and perfect information may destroy the
basis for some markets with risk and asymmetric information (Hermalin
and Katz, 2005). Take the insurance market as an example. If an insurer
cannot distinguish persons with different health, it may offer medical in-
surance to healthy and unhealthy persons at the same premium. Then, what
the Chicago School views as an inefficient cross-subsidy from healthy to
unhealthy persons in an ex post sense could also be viewed as insurance
against bad health in an ex ante sense. However, if the insurer can use
personal information to distinguish persons by health level, then it would
differentiate policies according to the person’s health. Then, information
collection would have undermined the market for insurance against
bad health. The same argument applies to other markets where the ‘‘qual-
ity’’ on one side is private information. Examples include human resources,
investments, and betting on sports.
Second, and more fundamentally, within the context of ex post efficiency,

the Chicago School’s argument works only when sellers have perfect in-
formation about consumers. However, welfare may not be monotone in the
quantity of personal information (Hermalin and Katz, 2005). In a setting of
‘‘second-best,’’ an increase in the quantity of personal information might
reduce welfare, and accordingly, protection of privacy might raise welfare.
This issue is further complicated when personal information is collected in
one market for use in another.
Third, the Chicago School critique overlooks various direct externalities
associated with the collection and use of personal information. These in-
clude direct marketing solicitations that overtly intrude into personal se-
clusion as well as covert intrusions into personal secrecy and autonomy.
The first issue (ex ante vs. ex post efficiency) is fairly trivial and we shall
not elaborate it here. The second and third issues concern non-trivial pro-
duction and exploitation of personal information, which are at the heart of
many ongoing privacy debates. We survey recent economic advances on
these two issues below.
Most economic analyses focus on overt collection and use of personal
information, where the subject is aware that her personal information is
being collected and used. Following the literature, our review will empha-
size overt collection and use. Where relevant, we will also discuss covert
collection and use.
3 Within-market consequential externalities
In this section, we consider how the collection and use of personal in-
formation within a single market affects the efficiency of market outcomes.
The collection and use impose a consequential (rather than direct) exter-
nality. For the most part, within-market consequential externalities apply
to the secrecy dimension of privacy.
Personal information is widely used to devise customized offers (prod-
ucts, prices, employment contracts, insurance, etc.) that better suit the
tastes or characteristics of particular individuals.9 To evaluate whether such
customization promotes exchange and hence market efficiency, many eco-
nomic analyses draw from the literature of asymmetric information
(Akerlof, 1970; Spence, 1973; Stiglitz, 1975) and product differentiation
(Mussa and Rosen, 1978; Katz, 1984; Moorthy, 1984).
In the following review, we adopt the classification of Hermalin and Katz
(2006) and distinguish two classes of situation where privacy might matter.
In one, personal information is not productive—the costs of the uninformed
9
See, for instance, Chen et al. (2001), Chen and Iyer (2002), Acquisti and Varian (2005), Calzolari and
Pavan (2005), Ghose and Chen (2003), Odlyzko (2003), Taylor (2004a) and (2004b), Wathieu (2002),
Chellappa and Sin (2005), and Wattal et al. (2004).
party do not depend on the personal characteristics of the informed party

as, for instance, in the case of pure price discrimination. In the other class,
personal information is productive—the costs of the uninformed party do
depend on the personal characteristics of the informed party as, for in-
stance, in the case of an employer recruiting workers of differing skill or an
insurer covering persons with differing health.
3.1 Non-productive information
Hermalin and Katz (2006) provide the simplest model of the issue. Con-
sider a monopoly that has asymmetric information about consumers, where
the consumers have either high or low valuation for some item. The mar-
ginal cost of the item is sufficiently low that it is efficient to provide to both
consumer types. Generally, the seller’s pricing strategy depends on its in-
formation about the consumer population. It provides a set of consumption
levels from which consumers choose and thereby self-select.
Suppose that, originally, the seller sold only to the high-type consumers.
Additional information would enable the seller to better sort between high
and low types. If it leads the seller to sell to both types, then welfare would
rise. However, suppose that, originally, the seller sold to a pool of both high
and low types. If the additional information leads the seller to reduce the
quantity provided to the low types, it would reduce welfare. Accordingly,
privacy regulation (which would reduce the availability of personal infor-
mation) might raise or reduce social welfare.
Bulow and Klemperer (2006) apply the auction theory concept of affil-
iation to analyze situations where competing sellers acquire different pieces
of information about a consumer. While a seller will raise price against
consumers with a relatively high willingness to pay for its product, it would
reduce price toward consumers with relatively low willingness to pay. Other
sellers would respond to the price cuts, and overall, the expected price to the
consumer would be lower.
The implications of privacy regulation are more complex in a setting
that unfolds over time, where consumers may make repeat purchases and
sellers can condition price on the consumer’s purchase history. Research
into this aspect overlaps quite closely with the economics of ‘‘behavior-
based price discrimination’’ (Fudenberg and Tirole, 2000), which is
reviewed by Fudenberg and Villas-Boas (2006) in this Handbook. The
pioneering analysis is due to Hart and Tirole (1988).
For simplicity, we present the analyses of Acquisti and Varian (2005) and
Taylor (2004). As in the static case, there are two types of consumer, with
the high type willing to pay more for the item than low-type consumers.
Also, the marginal cost of the item is sufficiently low that it is efficient to
provide it to both the types. The difference with the static case is that there
are two periods.
The seller’s pricing strategy depends on its information about the con-
sumer population. With privacy regulation that prevents collection of per-
sonal information, the seller would set the same price over time, which price
depends on the composition of the consumer population. In particular, if
the proportion of low-type consumers is high enough (or, more generally,
the demand is sufficiently elastic), the seller would set a price low enough so
that both consumer types buy the item, and such that the high-type con-
sumers enjoy a positive surplus. This equilibrium is efficient.
Now, suppose that the seller can infer the consumer types from their
purchase history. Specifically, in the first period, the seller can set a suf-
ficiently high price that only high types buy, and the remaining consumers
(who do not buy) are revealed to be low types. Then, the seller can con-
dition prices in subsequent periods on first-period purchase behavior, and
so, perfectly price discriminate (Acquisti and Varian, 2005 call this ‘‘price
conditioning’’).
Accordingly, if personal information collection is feasible, the seller faces
a trade-off: by charging a high price in the first period, it forgoes profit from
the low-type consumers, but it gains from identifying the high-type con-
sumers and price discriminating against them in subsequent periods. From
the viewpoint of social welfare, the low-type consumers suffer a deadweight
loss from not consuming in the first period.
It is easy to predict what increases the seller’s incentives to collect con-
sumer information. In the stylized example above, a wider gap between the
high- and low-type consumers’ valuations, a higher proportion of high-type
consumers, a longer time horizon (i.e., more future repurchases), and more
precise addressing of the consumer segments, would increase the seller’s in-
centive to use a high price to screen the consumer segments in the first period.
Note that the collection of personal information could also raise welfare.
This arises when, absent the ability to record transaction information (and
thereby discriminate), the seller chooses to sell only to high-type consumers.
By enabling discrimination, the collection of purchase history then leads the
seller to sell to low-type consumers as well, and so, raises welfare.10
Another consideration is that consumers might also act strategically.
Suppose again that, when unable to record transaction information, the
seller sells only to high-type consumers. If low-type consumers can credibly
reveal their personal characteristics,11 they would also produce information
and so persuade the seller to offer them the item at a lower price.12 The
efforts of the seller and low-type consumers to produce information are
strategic substitutes (Bulow et al., 1985).
10
Generally, price discrimination might raise or reduce welfare (Varian, 1985).
11
Students may produce school or university identity cards and seniors may show proof of age to
qualify for lower prices. In the employment context, job seekers may produce reference letters from past
employers, professional certificates, and school transcripts to prove their ability.
12
Hermalin and Katz (2006) also make this point in discussing property rights.
Now, if privacy regulation increased the seller’s cost of information col-

lection, then the seller would collect less information. In turn, this would
lead low-type consumers to produce more information. If the response of
the low-type consumers is sufficiently vigorous, the net result might be
paradoxical—the total amount of information produced and social welfare
could both increase (Gould, 1980). Similarly, relaxing privacy regulation,
which reduces the cost of information production, could lead to less in-
formation being produced and reduce welfare.
The implications of privacy regulation are subtler still in a setting with
competition among multiple sellers. As reviewed by Fudenberg and Villas-
Boas (2006) in this Handbook, even if each seller would gain individually by
being the only one to engage in price conditioning, if all sellers engage in the
practice, then it might intensify competition and thereby reduce the sellers’
combined profits. Further, as in the monopoly case, privacy regulation may
raise or lower social welfare. However, by contrast with the monopoly case,
the use of price conditioning among competitive sellers may raise consumer
surplus.13
Wathieu (2004) addresses a different issue—the impact of privacy reg-
ulation on the cost of product variety. Consider a setting where consumers
have specific tastes for different products. Ex ante, a seller cannot distin-
guish the consumer types, and it incurs an advertising cost to address each
individual consumer with a product. The advertising cost must be repeated
for each product that the seller markets to a particular consumer. If the
seller acquires and uses consumers’ personal information to segment the
demand, it can reduce advertising costs because the advertisements are
more accurately directed. In this context, by hindering segmentation, ‘‘pri-
vacy’’ may increase the sellers’ advertising costs.
However, despite the saving in advertising costs, when production is
characterized by economies of scale, allowing the seller to gain access to
consumer information could lead to excessive product variety. With the
information, the seller will have excessive incentive to price discriminate
and extract surplus from mainstream consumers. In this case, mainstream
consumers would prefer information privacy, and so avoid being identified
and hence avoid a higher price.
Generally, the collection and use of non-productive personal information
may redistribute surplus among sellers and consumers, but it does not nec-
essarily generate more exchange (Hirshleifer, 1971). Specifically, the col-
lection and use of customer purchase histories has private value to sellers,
but need not create social value. In fact, it may diminish social welfare by
reducing the consumption of the low-type consumers. In the monopoly
13
See also Choudhury et al. (2005) for a related analysis on competition between firms that employ
personalized pricing technologies, and Bouckaert and Degryse (2006) for an analysis of the differing
impacts of opt in and opt out privacy policies on firm entry and social welfare.
setting, consumer concerns about price discrimination seem to be well jus-

tified. On the other hand, in competitive settings, price conditioning may
benefit consumers. Further, sellers may use personal information to reduce
marketing costs (Wathieu, 2004). Finally, changes in privacy regulation to
adjust the cost of collecting or producing personal information may lead to
conflicting adjustments in the production of information. Clearly, the social
value of privacy regulation is ambiguous.
3.2 Productive information
Hermalin and Katz (2006) provide a simple model of the issue.14 Com-
petitive employers face a heterogeneous population of workers, some of
whom have high productivity while others have low productivity. Each
employer needs just one worker. In the economic efficient allocation, both
types of worker would be employed.
Suppose that the original equilibrium pools high and low types at a
common wage. Since both types of worker are employed, this equilibr-
ium is efficient. Now, divide the worker population into two pools.
With additional information, employers can more accurately identify
high-type workers. If the proportion of high types in the ‘‘good’’ pool is
sufficiently large (and that in the other ‘‘bad’’ pool is low), then in com-
petitive equilibrium, employers will employ all workers in the ‘‘good’’ pool
at a common wage, but pay a low wage to the bad pool. The low wage
would attract only low types; hence, the high-type workers in the bad pool
would be unemployed. This would reduce welfare relative to the original
equilibrium.
By contrast, suppose that the original equilibrium included only low
types. This adverse selection equilibrium is not efficient. Again, divide the
worker population into two pools, and suppose that additional information
enables employers to more accurately identify high-type workers. If the
proportion of high types in a ‘‘good’’ pool is sufficiently large, then in
competitive equilibrium, employers will employ all workers in the ‘‘good’’
pool at a common wage. This would raise welfare relative to the original
equilibrium.
Taylor (2005) also addresses the issue of over/under-investment in pro-
ductive personal information in a competitive setting, but using a somewhat
different setting. Each employer seeks a worker, who has either high or
low productivity. The worker does not know her own productivity. In
the economic efficient allocation, only the high-type worker would be
employed. The employer can invest in information about the worker.
When the information about high-type workers is perfect but information
about low-type workers is subject to error, the employer will over-invest in
14
Their setting is not quite the simplest possible, as it supposes there to be competition on the seller
side. An even simpler setting would have just a monopoly seller.
information. However, when the information about high-type workers is

subject to error but information about low-type workers is perfect, the
employer will under-invest in information.
The analyses of Hermalin and Katz (2006) and Taylor (2005) imply that
there is no simple rule: whether privacy of personal information raises or
reduces welfare depends on the circumstances.
A separate stream of research has considered the role of personal infor-
mation privacy in tax policy. In this setting, the less-informed party is the
government. The government uses income tax to re-distribute income from
high- to low-income earners. If the government sets tax rates after indi-
viduals have decided their investment in something that increases their fu-
ture earnings, say education, a time consistency problem arises. Fearing
that the government will set high tax rates in the future, taxpayers will
under-invest in education (Boadway et al., 1996).
In this context, a privacy policy is an effective way by which the gov-
ernment can commit to lower tax rates in the future (Konrad, 2001):
the privacy policy limits the government’s ability to collect information
and hence to levy high tax rates. Accordingly, the privacy policy serves to
encourage taxpayers’ investment in activities that increase their future
earnings.
Dodds (2003) considers a different setting, where individuals of two types
benefit from a public good. The socially efficient quantity of the public good
depends on the number of high-productivity persons. The high-productivity
persons are reluctant to reveal themselves as they must then contribute
relatively more toward the public good. The issue is closely related to that
of taxpayer compliance where taxpayers must report their income subject to
government auditing. As in the taxpayer compliance analyses (Mookherjee
and Png, 1989), Dodd’s key result is that 100% auditing does not maximize
welfare. He interprets this to mean that some degree of privacy is socially
efficient.
Generally, in competitive settings, an improvement in the accuracy of
productive personal information may lead the less informed party (seller or
employer) to include more or exclude some marginal persons (consumers or
workers). This is a consequential externality on some members of the bet-
ter-informed side of the market. The consequential externality might be
positive or negative. It is surprising that the grounds for privacy do not
seem to be weaker with respect to productive as contrasted with non-pro-
ductive personal information.
4 Cross-market consequential externalities
In this section, we consider how the collection of personal information

in one market for use in another market affects the efficiency of market
outcomes.15 The collection and use impose a consequential (rather than

direct) externality. For the most part, cross-market consequential external-
ities apply to the secrecy dimension of privacy.
Marketers may compile customer databases for sale to third parties. For
example, e-mail portals may pass personal details of account holders to
third parties who then use the information to promote their goods and
services. The policy implications with respect to the third-party ‘‘informa-
tion buyers’’ are similar to those in the cases that we have reviewed in the
preceding section. Hence, we consider only the actions of the ‘‘information
sellers.’’
The central theme is that the marketer may have even more incentive to
collect consumer information in a cross-market than in a within-market
setting. Recall the monopoly model of within-market collection and use of
non-productive information over time. As analyzed in Section 3.1, the seller
has an excessive incentive to price high in the first period and so identify the
high-type consumers.
This incentive is reinforced if the seller can sell the personal information
collected to third parties—the revenue from selling customer information
would raise the marginal return from the price experiment (Taylor, 2004).
Hence, the seller is more likely to set a high first-period price. When de-
mand is somewhat elastic (i.e., the seller would sell to all consumers absent
the opportunity to sell information), the option to sell consumer informa-
tion would lead the seller to restrict output, and hence, reduce welfare.
Addressing a similar problem, Calzolari and Pavan (2005) develop a very
sophisticated model that considers interaction between two different un-
informed parties, say sellers, with a single informed party, say a buyer
whose characteristics are private information, over time. They identify
conditions under which the early seller will transfer information about the
buyer to the later seller. In particular, when the early seller is not interested
in the exchange between the buyer and later seller, the buyer’s valuations
toward the two sellers’ products are positively correlated, and the buyer’s
preferences in the two sellers’ products are separable, then the early seller
may prefer to protect the buyer’s privacy.
By contrast, when any one of these conditions is not met, the early seller
can capture additional rents arising from information or contractual ex-
ternalities. The effect of privacy on welfare is ambiguous—privacy may
promote the exchange between the buyer and later seller, but it could also
introduce new distortions in the buyer’s exchange with the early seller.
Overall, it seems that the selling of personal information benefits ‘‘infor-
mation buyers’’ (secondary sellers). As for social welfare in secondary
markets, it could increase or decrease depending on the composition of the
15
Cross-market externalities imply ‘‘secondary use’’ of personal information. Secondary use can also
occur within the same market. For instance, a marketer might use a delivery address submitted for an
online purchase to promote related items.
consumer population (as discussed in Section 3). However, in the primary

market, where the personal information is collected, welfare may decrease
because sellers have greater incentive to raise price in order to classify
consumers (Taylor, 2004; Calzolari and Pavan, 2005). The primary sellers
can then compile customer information for sale to secondary sellers. There-
fore, a cross-market externality may emerge when sale of consumer infor-
mation is allowed.
In general, sale of consumer information is more likely to be beneficial
when the potential of such information is high, e.g., when the classification
of consumers can help to match seller offers and interested consumers. If
the information does not lead to more efficient exchange in secondary
markets, then it may be worthwhile to discourage its sale, which would in
turn discourage primary sellers from collecting the information.
5 Direct externalities
In this section, we consider direct externalities arising from the collection

and use of personal information within the same market and across mar-
kets. Direct externalities apply to the secrecy, autonomy, and seclusion
dimensions of privacy.
A major use of personal information is to direct unsolicited promotions
by mail, telephone, e-mail, and in person. To the extent that such solic-
itations impose costs on consumers that marketers ignore, they generate a
negative externality (Petty, 2000).
Van Zandt (2004) analyzes a setting where heterogeneous sellers send
messages to promote different items. Consumer attention is a scarce re-
source—each consumer can process a limited number of messages. Hence,
consumers incur costs to ‘‘open’’ marketing messages. They respond to
sellers and purchase if and only if they have read the messages and are
interested in the item. Sellers have private information on consumer inter-
ests, and they decide strategically how many messages to send, and which
consumers to target. Sellers may over-promote their products. Accordingly,
measures that inhibit solicitations (e.g., that increase communication cost
or a tax on solicitations) may help sellers to focus their marketing effort and
hence improve social welfare.16
In a similar setting, Anderson and de Palma (2005) show that over-
promotion by sellers could even lead to market failure—the quality of
messages may become so low that consumers choose not to read any mes-
sages. Such market failure is reminiscent of the well-known ‘‘lemons’’
16
However, the sellers’ profits would increase only if they have sufficiently accurate data on consumer
interest (Van Zandt, 2004). To this extent, consistent with the Chicago School’s view, privacy (or more
specifically, secrecy) of personal information may not be desirable.
problem (Akerlof, 1970). Evidently, increasing solicitation cost could well

raise welfare, and therefore regulation may play a positive role.17
Akcura and Srinivasan (2005) consider the cross-market collection and
use of information. Sellers may collect personal information about con-
sumers in a primary market and use it in a secondary market. In deciding
how much personal information to reveal, consumers balance the benefit
from consuming the primary item against direct privacy costs.18 The higher
the rate at which consumers expect sellers to cross-sell personal informa-
tion, the less information consumers would reveal. Accordingly, sellers may
choose to limit the extent to which they cross-sell personal information, and
so, persuade consumers to provide more information in the primary mar-
ket.
Hann et al. (2006a) analyze direct marketing in a setting with two types of
consumer—one with a high value for the item being marketed and the other
with a low value. Each direct marketing solicitation causes some harm, but
a consumer can get the item only through the solicitation. Consumers can
take actions to reduce the harm (‘‘marketing avoidance’’). For instance, to
avoid telemarketing solicitations, consumers can conceal (e.g., by register-
ing with do-not-call lists) or deflect (e.g., by screening incoming telephone
calls). Sellers cannot distinguish the consumer types ex ante. Ideally, they
would promote only to high-type consumers. Instead, they incur costs to
solicit the entire consumer population, and then discover consumer types ex
post.
Seller solicitations are a strategic complement (Bulow et al., 1985) with
concealment by low-type consumers. If the cost of concealment measures
were to fall, low-type consumers would raise concealment, and sellers would
increase marketing. Indeed, since 2003, the US enforcement of a nationwide
‘‘do not call list’’ may have led to an increase in the return on investment
from telemarketing (Direct Marketing Association, 2004). However, from a
welfare perspective, consumer concealment is less favorable than deflection,
because it concentrates seller solicitations on a smaller number of consum-
ers (Hann et al., 2006a). A consumer needs only one solicitation to enjoy
the product, and additional solicitations add to harm. Accordingly, con-
centrating the solicitations raises the expected harm relatively more than the
benefit.
Although the scenarios of hacking and eavesdropping appear to be quite
different from that of direct marketing, the formal analysis is quite closely
related. Consumers who provide information to vendors or use commu-
nication services may be subject to covert intrusions into their privacy,
17
Further, Gantman and Spiegel (2004) consider the trade-off in software that incorporates adver-
tising banners (‘‘adware’’) between the benefit to consumers of receiving targeted information which
improves their choice of product against the privacy cost.
18
Akcura and Srinivasan (2005) do not specify the nature of these costs, but they could presumably
encompass the inconvenience from receiving unsolicited direct marketing and the harm from possible
intrusion into the seller’s database.
which impose direct and indirect costs. In response, consumers could take
defensive actions like encoding and encryption, which are costly and might
also diminish the benefit from consumption (Noam, 1995a). The strategic
impact of such defensive actions could be analyzed in the same way as
marketing avoidance (Hann et al., 2006a) and, more generally, private se-
curity (Koo and Png, 1994).
The research just reviewed emphasizes externalities from one side of a
market to another. Another important class of direct externalities is that of
peer-to-peer externalities. August and Tunca (2004) study the incentives of
end users to patch security flaws in computer systems. Computer viruses
exploit flaws in one computer system to penetrate others, and are more
likely to succeed the fewer users patch flaws. The key policy implication is
that, where users differ in their value from use of the system, mandatory
patching is not optimal.
Although August and Tunca focus on computer viruses, their analysis
may apply more generally to applications in which consumers reveal
the personal information of others. Examples of such applications include
instant messaging services and online communities (e.g., friends.com),
where users are asked to refer their peers to service providers. In some
cases, service providers may even covertly traverse the e-mail boxes of users
to recruit new potential users (much like the way computer viruses infect
other systems). Despite the risks of such privacy invasion, August and
Tunca’s analysis suggests that mandating users to protect privacy need not
be optimal.
Overall, it is clear that, in circumstances involving direct externalities,
privacy of personal information would increase social welfare. However,
sweeping solutions, such as disallowing the collection and use of personal
information, would not be optimal—they would prevent interested con-
sumers from enjoying the items being promoted (Van Zandt, 2004; And-
erson and de Palma, 2005; Hann et al., 2006a) or cause consumers to forego
some implicit benefits (Akcura and Srinivasan, 2005).
6 Property rights
Will the appropriate assignment of property rights (self-regulation) re-

solve the issue of privacy? The Chicago School posits that a free market for
information yields social efficiency. Hence, an explicit allocation of prop-
erty rights may shift society away from a socially efficient equilibrium and
reduce welfare. For instance, granting workers property rights to their
personal information may cause an employer to reduce employment.
In their analyses of both non-productive and productive information,
Hermalin and Katz (2006) show that the market outcome is identical re-
gardless of how property rights over personal information are assigned.
Specifically, in the case of non-productive information, the monopoly seller
can compel customers to reveal their type. In the case of competition with
productive information, high-type workers will identify themselves, thus
revealing the low types. Similarly, Kahn et al. (2000) show that, if there is
sufficient flexibility in contracting, information would be revealed to an
efficient degree. The outcome obeys the Coase Theorem—it does not matter
whether or how property rights to personal information are initially as-
signed.19
However, the analyses of Hermalin and Katz (2006) and Kahn et al.
(2000) apply to situations where the collection and use of personal infor-
mation take place within the same (primary) market. What if the relatively
uninformed party uses the information in secondary contexts as, for in-
stance, when a marketer sells consumer information gathered at one web
site to third-party spammers? Then a cross-market externality will arise.
The parties with personal information will certainly consider the cross-
market externality when deciding how much personal information to reveal
(Akc- ura and Srinivasan, 2005).
The impact of the allocation of property rights to personal information in
the primary market may well depend on the relation between the party’s
positions in the primary and secondary markets. Will a high-type worker in
the primary market also be a high-type worker in the secondary market?
When the secondary use of the information is uncertain, property rights
may have a role.
Further, in the case of direct externalities, property rights would clearly
help to resolve the harms that sellers impose on consumers, and also peer-
to-peer harms among consumers.
Therefore, it may be worthwhile to attach a value to personal informa-
tion, at least in terms of restricting future uses of the information. The
challenge then lies in how such a value is determined.
The first issue is that the parties with property right over their personal
information may not fully take account of the potential benefit of the in-
formation to uninformed parties. For instance, a common regulatory rem-
edy for unsolicited promotions is the ‘‘do not contact’’ list. However,
potential consumers may ignore sellers’ profit when deciding to register
with ‘‘do not contact’’ lists, and hence may tend to over-register relative to
the welfare optimum (Anderson and de Palma, 2005).20
It is quite natural to expect that allowing consumers to set their own
values for personal information may lead them to over-value data
(Schwartz, 2004). Hence, the second issue is that consumers may attach
too high a price to their personal information, which might excessively raise
the barrier to potential buyers of the information. Specifically, economic
experiments have repeatedly shown that people demand a higher price for a
19
See also Chellappa and Shivendu (2003).
20
In 2003, the US government established a nationwide ‘‘do not call’’ registry. By August 18, 2005, the
registry has recorded 100 million entries (Federal Trade Commission, 2005b).
property when another person seeks to use it than the price that they would
offer to protect the property from being used (see, e.g., Boyce et al., 1992).
In the context of personal information, individuals’ ‘‘willingness to ac-
cept’’ (WTA) for use of their personal information (when they have explicit
property rights over the information) may exceed their ‘‘willingness to pay’’
(WTP) for protection of their information from exploitation (when no
property right is granted). Granting property rights to individuals and al-
lowing them to name their own price may lead to under-usage of infor-
mation, whereas allowing the free use of personal information could lead to
over-usage.
The difference between WTA and WTP for personal information
could help explain the disparate findings from opinion polls (e.g., Harris
Interactive, 2001, 2003) and behavioral experiments (e.g., Ackerman et al.,
1999; Hann et al., 2003; Hui, 2006; Hui et al., 2006). Specifically, when
polled for their opinions on or attitudes toward privacy, people may assume
they ‘‘own’’ their personal information and hence demand a high price for
use of their information. By contrast, when confronted with actual infor-
mation requests and when they realize that protecting their personal infor-
mation may be ‘‘costly’’ (e.g., they may not be able to use a web site or
complete a transaction if they do not supply the information), they demand
less compensation. The behavioral experiments cited above have shown
that people provide their information in exchange for even small rewards or
incentives.
Clearly, it would be misleading to judge the importance of privacy from
opinion polls alone. Rigorous experiments are necessary to gauge the actual
value that people attach to their personal information under various cir-
cumstances. Perhaps the Becker–DeGroot–Marschak (BDM) mechanism
(Becker et al., 1964) can be employed to elicit the incentive-compatible
reservation prices that people place on their personal information. It would
be important to recognize the likely gap between WTA and WTP, and
assess the benefits of allocating property rights accordingly.
7 Regulation
Assignment of property rights will resolve privacy issues only in contexts

where the collectors and users of personal information and their subjects of
the information can enter into contractual arrangements. But what about
contexts where such arrangements are inconvenient or even impractical, for
instance, widespread peer-to-peer externalities in the decision of computer
users whether to patch security vulnerabilities? In law, this is the domain of
tort law and regulation as contrasted with contract law.
Tang et al. (2005) consider a setting where intrusion of privacy imposes a
direct cost on consumers. Consumers differ in their sensitivity to intrusion,
while sellers differ in their cost of protecting privacy. When few consumers
are sensitive, welfare is maximized with a regime of ‘‘caveat emptor,’’ as

businesses avoid the cost of protecting privacy. By contrast, when many
consumers are sensitive, welfare is maximized with mandatory privacy reg-
ulation, as consumers avoid the cost of comprehending each business’ pri-
vacy policy. In the intermediate case, welfare is maximized with privacy
seals—the low-cost businesses choose to purchase the seal, while the high-
cost businesses do not.21
A key reason in favor of regulation is that it may be a more effective form
of commitment than contractual arrangements. Our review above (Sections
3–5) has pointed to various situations of both consequential and direct
externalities where commitment to protect privacy increases welfare. Spe-
cifically, analyses of behavior-based price discrimination in competitive
settings show that businesses may benefit from privacy of personal infor-
mation (Fudenberg and Villas-Boas, 2006).
Ironically, business interests oppose proposals to tighten privacy regu-
lation. The US national cost of complying with these legislative proposals
has been estimated to be US$9–36 billion (Hahn, 2001). For just catalog
and Internet clothing retailers, the Direct Marketing Association estimated
that opt-in restrictions to use of demographic information by third parties
would raise costs by US$1 billion (Turner, 2001).
The economic analysis of consequential externalities suggests that
whether and how privacy increases welfare depends on the particular cir-
cumstances. Consequently, there will be no magic ‘‘one size fits all’’ solu-
tion, but rather, regulation should be tailored to the circumstances. For
instance, communication between persons with a particular relationship,
including husband–wife, penitent–clergy, patient–doctor, attorney–client,
citizen–census taker is commonly protected by ‘‘privilege.’’ The patient–
doctor privilege encourages an uninhibited exchange of information and so,
enhances overall community health (Noam, 1995a).
Muris (2003) had proposed to generally allow free use of personal in-
formation, while focusing enforcement against misuse. The focus on infor-
mation use is consistent with consumer preferences (Wathieu and
Friedman, 2005). However, in the studies that we reviewed in Sections 3
and 4, welfare could be reduced by apparently legitimate uses of informa-
tion that did not cause direct harms. Hence, requirements for consumer
consent to collection and use of personal information (as stipulated in the
FIPs) could raise social welfare. Accordingly, the key issue is how to bal-
ance the interests of sellers and consumers, and obviously a sweeping ‘‘use’’
or ‘‘no use’’ solution would not work across all contexts. Wherever it is
feasible to ascertain the benefits and costs of information use, the obvious
solution is industry-specific regulation, as in the Fair Credit Reporting Act
of 1970.
21
Information providers could also commit to privacy protection through service-level agreements
with their users (Pau, 2005).
Regulation must be tailored even with regard to direct externalities, for

which it is unambiguous that privacy would raise welfare. As mentioned
earlier, a common regulatory remedy for unsolicited promotions is the ‘‘do
not contact’’ list. While a ‘‘do not call’’ list may resolve telemarketing, a
similar ‘‘do not spam’’ list might be counterproductive. Illicit spammers
account for the bulk of spam, and they might well spam addresses on the
‘‘do not spam’’ list (Hahn, 2003; Muris, 2003). With regard to spam, a tax
appears to be the most promising solution (Kraut et al., 2002; Van Zandt,
2004; Anderson and de Palma, 2005), and generally, deflection is to be
preferred over concealment (Hann et al., 2006a).
8 Empirical evidence
To gauge the economic significance of privacy as a public policy issue, it

is vital to know how much people value their privacy. Polls and surveys
have repeatedly shown that people are concerned about privacy (Westin,
2001). However, the key policy issue is not whether individuals value pri-
vacy. It is obvious that people value privacy. What is not known is how
much people value privacy and the extent to which people differ in their
valuations.
Despite tremendous debate and policy interest, there has, to date, been
little research into this question (Hahn, 2001). Indeed, it has been conjec-
tured that ‘‘measuring the value of consumer privacy may prove to be
intractable’’ (Ward, 2001).
Recent opinion surveys and experimental research provide some insights
into this question. In November 1998, among 381 US respondents to an
online survey, most were willing to reveal personal information but would
not reveal personal identifying information (Ackerman et al., 1999). For
instance, 58% would report income, investments, and investment goals to
obtain customized investment advice, but only 35% would also reveal their
name and address.
In May–June 2000, the Pew Internet and American Life Project found
that, among 1,017 American Internet users, 54% would provide personal
information in order to use a web site, whereas only 27% were hard-core
privacy protectionists who would never provide their personal information
to web sites (Fox et al., 2000). In February–March 2003, the Annenberg
Public Policy Center of the University of Pennsylvania found that, among a
sample of 1,200 respondents aged 18 years or older who used the Internet at
home, most who did not accept a web site’s data collection policy would
nevertheless disclose their real name and e-mail address if they valued the
web site (Turow, 2003).
More compelling than surveys are various experiments that gauged sub-
jects’ willingness to reveal personal information. Hui et al. (2006) conducted
a field experiment to measure the likelihood that individuals would provide
personal information to Internet businesses. By estimating a discrete choice

model using real online participation data, they found that people were
willing to disclose more personal information in exchange for small mon-
etary incentives. Similarly, in laboratory experiments, simple interface re-
design could induce consumers to disclose more personal information (Hui,
2006), or opt-in to receiving future newsletters (Lai and Hui, 2004, 2005).
Wattal et al. (2004) procured data from field experiments conducted by
an e-commerce vendor. The vendor contacted potential customers with
different dimensions of customization—some received customized product
information, while others received personalized greetings, e.g., ‘‘Dear Ms
ABC.’’ Consumers responded positively to customized product offerings,
but negatively to personalized greetings.
Wathieu and Friedman (2005) suggest that privacy concerns are sensitive
to indirect consequences of information transmission. In particular, they
argue that personal information may not have intrinsic value, and the flow
of personal information may not be the key privacy concern. Rather, it is
the concern about information use that affects consumer behavior. Their
argument was supported in an experiment that involved 647 subjects from a
US business school research pool.
An experiment at Humboldt University provides further indirect evidence
(Berendt et al., 2005). Two hundred and six volunteers interacted with an
anthropomorphic 3-D shopping bot to select a compact camera or winter
jacket. The bot engaged subjects in dialogue about product attributes and
also posed ‘soft’ questions typical of selling in conventional stores. The
experimental subjects willingly revealed personal identifying information to
the bot, specifically, 35–40% provided their home address.
Earp and Baumer (2003) conducted an online survey with 415 respond-
ents. Each respondent was randomly shown one of 30 web pages, from well-
or lesser-known retail, medical/health, and financial sites. Respondents
were most willing to reveal gender and age, and least willing to reveal their
social security numbers. Moreover, they were significantly less willing to
provide personally identifiable information (phone number, home address,
e-mail address, social security number, and credit card number) to lesser
known than well-known web sites.
The surveys and experiments clearly show that people value privacy, but
to an extent less than some privacy advocates have claimed. In particular,
many survey respondents indicated use of web sites as a sufficient moti-
vation to provide personal information. The results suggest that consumer
information can be directly solicited in exchange for simple monetary or
procedural measures. Further, they also suggest that governments should
evaluate practical implications for Internet businesses before introducing
stringent privacy regulations.
A question related to individuals’ value for privacy, in general, is how
they value the use of their personal information. A set of conjoint analyses
at Singapore and US universities show that people are willing to bear the
risks of improper access to or secondary use of their information in ex-

change for monetary incentives or increased convenience (Hann et al.,
2003). In particular, the U.S. and Singapore subjects valued improper ac-
cess to personal information at around US$11–20, whereas they valued
secondary use at around US$8–27.22 Hence, despite consumers’ protests
against price discrimination, sale of personal information to unauthorized
third parties, spam, etc., it may indeed not be that difficult to convince them
to agree to these information uses.
Hann et al. (2003) also identified three distinct segments in the consumer
population—privacy guardians (the majority), information sellers, and
convenience seekers. However, these segments were not significantly cor-
related with demographic characteristics. By contrast, using census data,
Varian et al. (2004) identified household characteristics of telephone num-
bers registered with the US national ‘‘do not call’’ list. Those with annual
incomes exceeding US$100,000 and college-level education were signifi-
cantly more likely to register, while those with a member in the 13–19 age
group were significantly less likely to register. It is intuitive that wealthier
households would suffer more annoyance from telemarketing calls. Why
households with teenagers suffer relatively less is more of a puzzle.
In the context of direct e-mail marketing, marketers do not bear the
privacy costs imposed on consumers. Since the cost of spam is very low
(e.g., Muris, 2003), do spammers broadcast their solicitations randomly? In
a field experiment, Hann et al. (2006b) find that spam is not random but
rather targeted. Specifically, the incidence of spam was higher among e-mail
accounts created with particular service providers, accounts with particular
declared interests, and accounts associated with persons more likely to
make online purchases (Americans rather than Singaporeans, adults rather
than teenagers).
Further, the spam arena provides evidence of the relative effectiveness of
regulation vis-à-vis self-regulation. Web sites do indeed comply with their
published privacy policies (Jamal et al., 2003). Hence, if self-regulation of
privacy were economically efficient, it could work. Further, mandatory
regulation tends to drive out self-regulation: web sites in the United King-
dom, which mandates privacy regulation, provide stronger privacy protec-
tion than those in the United States, which follows a self-regulatory
approach (Jamal et al., 2005).
To conclude, the evidence so far indicates that consumers are not truly
so sensitive about privacy. Economic solutions, such as the exchange of
personal information for monetary incentives, convenience, or special re-
sources, may suffice to regulate the market for personal information
(Noam, 1995b; Laudon, 1996). The contentious debate about privacy reg-
ulation may have been misdirected—the question does not lie in whether
22
See also Baumer et al. (2005) for the use of experimental economics to quantify the values that
consumers place on privacy and security.
tighter control should be placed on information collection and use, but in

setting the right ‘‘prices’’ for personal information.
9 Future directions
Clearly, a free market in personal information will not provide an eco-

nomically efficient outcome. With regard to consequential externalities
within and across markets, privacy over personal information may raise or
lower welfare depending on the circumstances. This should not be surpris-
ing, as, generally, the direction of welfare gain in ‘‘second-best’’ situations is
a priori ambiguous. Given that, it would be interesting to explore whether
privacy regulation is relatively more likely to increase welfare in the context
of non-productive as compared with productive information.
We see several other directions for future research. First, in all of the
various models that apply the asymmetric information approach, it is as-
sumed that the uninformed party knows of the existence of the parties with
private personal information and knows their distribution of personal
characteristics, but just does not know the characteristics of individual
persons. For instance, in the setting of Acquisti and Varian (2005), only
high types buy in the first period, so everyone else must be a low type. But
what if the uninformed party does not even know the distribution of per-
sonal characteristics? Would the results be the same if the analysis begins
from the uninformed party’s beliefs about the distribution of the other
party’s personal characteristics?
Second, personal information, like information in general, is a public
good (Stigler, 1980). Economists have given little attention to the public
good aspects of privacy, specifically, the conditions for the optimal pro-
duction and usage when the marginal cost of usage is zero. For instance, if
disclosure of AIDS test results were mandatory, individuals might forgo
testing, which would lead to unintended adverse consequences (Hermalin
and Katz, 2006).
Third, as our discussion of WTP vis-à-vis WTA makes clear, there is
substantial potential to apply behavioral economics for a better under-
standing of privacy. Personal information is such a sensitive thing that
individual behavior is relatively more likely to depart from the rational
model with respect to personal information than other things. Preliminary
research has shown that consumers may often not have well-defined pref-
erences on privacy—it is possible to influence their willingness to reveal or
consent to use of their personal information by varying data solicitation
procedures, even trivially (Hui, 2006; Lai and Hui, 2004, 2005).
Fourth, prior research and discussion has focused on privacy of personal
information. Do the same analyses and conclusions apply to privacy of
corporate information? Under what circumstances does protection of cor-
porate information raise social welfare? This question is the counterpart to
a key issue in accounting research, viz., disclosure. The issue of corporate

privacy also bears on two other concepts—trade secrets in intellectual
property and corporate reputation.23
Fifth, we should mention economics-oriented research into the technology
of privacy. Loder et al. (2006) apply the theory of mechanism design to
devise an incentive-compatible technology to screen out spam. Serjantov
and Clayton (2005) use a stylized model and a set of e-mail data to examine
the implications of various spam-blocking strategies. More generally, an
interesting direction for research is to apply economics to the technology of
privacy, and specifically, issues of system and software security.
Finally, to ensure the currency of this review, we have created a
complementary wiki at http://infoprivacy.pbwiki.com/. All scholars are
invited to contribute information and links.
Acknowledgments
We thank Jean Camp, Robert Hahn, Karim Jamal, Luc Wathieu, and the
editor, Terry Hendershott, for helpful comments.
References
Ackerman, M.S., L. Cranor, J. Reagle (1999). Privacy in e-commerce: examining user scenarios and
privacy preferences, in Proceedings of the ACM Conference in Electronic Commerce, ACM Press,
New York, pp. 1–8. Available at http://portal.acm.org/citation.cfm?id=336995&jmp=cit&coll=
portal&dl=ACM&CFID=439217&CFTOKEN=36198287#CIT
Acquisti, A., H.R. Varian (2005). Conditioning prices on purchase history. Marketing Science 24(3),
367–381.
Akc- ura, M.T., K. Srinivasan (2005). Research note: customer intimacy and cross-selling strategy.
Management Science 51(6), 1007–1012.
Akerlof, G.A. (1970). The market for ‘lemons’: quality uncertainty and the market mechanism. Quar-
terly Journal of Economics 84(3), 488–500.
Anderson, S.P., A. de Palma (2005). A theory of information overload. Unpublished manuscript,
Department of Economics, University of Virginia.
August, T., T. Tunca (2004). Network software security and user incentives. Unpublished manuscript,
Graduate School of Business, Stanford University.
Baumer, D.L., J.B. Earp, J.C. Poindexter (2005). Quantifying privacy choices with experimental eco-
nomics. Unpublished manuscript, College of Management, North Carolina State University.
Becker, G.M., M.H. DeGroot, J. Marschak (1964). Measuring utility by a single response sequential
method. Behavioural Science 9(July), 226–232.
Berendt, B., O. Günther, S. Spiekermann (2005). Privacy in e-commerce: stated preferences vs actual
behavior. Communications of the ACM 48(4), 101–106.
Boadway, R., N. Marceau, M. Marchand (1996). Investment in education and the time inconsistency of
redistributive tax policy. Economica 63(250), 171–189.
23
For a preliminary discussion on corporate privacy and the related regulatory considerations, see
Posner (1978, 1979).
Bouckaert, J., H. Degryse (2006). Opt in versus opt out: a free-entry analysis of privacy policies, in: 5th
Workshop on the Economics of Information Security (WEIS 2006), Robinson College, University of
Cambridge, England 26–28 June 2006.
Boyce, R.R., T.C. Brown, G.H. McClelland, G.L. Peterson, W.D. Schulze (1992). An experimental
examination of intrinsic values as a source of the WTA–WTP disparity. American Economic Review
82(5), 1366–1373.
Bucklin, R.E., C. Sismeiro (2003). A model of web site browsing behavior estimated on Clickstream
data. Journal of Marketing Research 40(3), 249–267.
Bulow, J., J. Geanakoplos, P. Klemperer (1985). Multimarket oligopoly: strategic substitutes and com-
plements. Journal of Political Economy 93(3), 488–511.
Bulow, J., P. Klemperer (2006). Privacy and prices. Discussion paper, Nuffield College, Oxford University.
Calzolari, G., A. Pavan (2005). On the optimality of privacy in sequential contracting. Journal of
Economic Theory 30(1), 168–204.
Camp, L.J., C. Osorio (2003). Privacy enhancing technologies for internet commerce, in: O. Petrovic, M.
Ksela, M. Fallenböck, C. Kittl (eds.), Trust in the Network Economy, Springer, Berlin, pp. 317–332.
Chellappa, R.K., S. Shivendu (2003). A property rights approach to consumer concerns of privacy in
online personalization: incentives and welfare implications. Available at http://ssrn.com/ab-
stract=457003. Accessed August 15, 2005.
Chellappa, R.K., R. Sin (2005). Personalization versus privacy: an empirical examination of the online
consumer’s dilemma. Information Technology and Management 6(2–3), 181–202.
Chen, Y., G. Iyer (2002). Consumer addressability and customized pricing. Marketing Science 21(2),
197–208.
Chen, Y., C. Narasimhan, Z.J. Zhang (2001). Individual marketing with imperfect targetability. Mar-
keting Science 20(1), 23–41.
Choudhury, V., A. Ghose, T. Mukhopadhyay, U. Rajan (2005). Personalized pricing and quality dif-
ferentiation. Management Science 51(7), 1120–1130.
Culnan, M.J., J.R. Bies (2003). Consumer privacy: balancing economic and justice considerations.
Journal of Social Issues 59(2), 323–342.
Direct Marketing Association. The DMA 2004 response rate report. New York, October 2004.
Dodds, S. (2003). Privacy and endogenous monitoring choice when private information is a public good.
Department of Economics, Carleton University, December.
Earp, J., D. Baumer (2003). Innovative web use to learn about consumer behavior and online privacy.
Communications of the ACM 46(4), 81–83.
Eddy, E.R., D.L. Stone, E.F. Stone-Romero (1999). The effects of information management policies on
reactions to human resource information systems: an integration of privacy and procedural justice
Perspectives. Personnel Psychology 52, 335–358.
Federal Trade Commission (2005a). Internet service provider settles FTC privacy charges. March 10.
Available at http://www.ftc.gov/opa/2005/03/cartmanager.htm.Accessed April 8, 2005.
Federal Trade Commission (2005b). Statement of Federal Trade Commission Chairman Deborah Platt
Majoras on the 100 millionth number on the national do not call registry. Available at http://
www.ftc.gov/opa/2005/08/dncstatment.htm. Accessed September 14, 2005.
Fox, S., L. Rainie, J. Horrigan, A. Lenhart, T. Spooner, C. Carter (2000). Trust and Privacy Online:
Why Americans Want to Rewrite the Rules. Pew Internet & American Life Project, Washington DC.
Available at http://www.pewinternet.org/reports/toc.asp?Report=19. Accessed March 31, 2005.
31(4), 634–657.
Fudenberg, D., J.M. Villas-Boas (2006). Behavior-based price discrimination and customer recognition.
Handbook in Information Systems, Chapter 7.
Gantman, N., Y. Spiegel (2004). Adware, shareware, and consumer privacy. Working Paper No. 04-02,
NET Institute, October. Available at http://ssrn.com/abstract=618741.
Ghose, A., P.Y. Chen (2003). Personalization vs. privacy: firm policies, business profits and social
welfare. Working Paper, GSIA, Carnegie Mellon University.
Goodwin, C. (1992). A conceptualization of motives to seek privacy for nondeviant consumption.

Journal of Consumer Psychology 1(3), 261–284.
Gould, J.P. (1980). Privacy and the economics of information. Journal of Legal Studies 9(4), 827–842.
Hahn, R.W. (2001). An assessment of the costs of proposed online privacy legislation. AEI-Brookings
Joint Center for Regulatory Studies, Washington, DC.
Hahn, R.W. (2003). Outlawing spam won’t work. Policy matters 03-36. AEI-Brookings Joint Center for
Regulatory Studies, Washington, DC, October.
Hann, I.H., K.L. Hui, T.S. Lee, I.P.L. Png (2003). The value of online information privacy: an empi-
rical investigation. AEI-Brookings Joint Center for Regulatory Studies, Related Publication 03-25, ,
October.
Hann, I.H., K.L. Hui, T.S. Lee, I.P.L. Png (2006a). Consumer privacy and marketing avoidance: a
static model. Unpublished manuscript, Department of Information Systems, National University of
Singapore.
Hann, I.H., K.L. Hui, Y.L. Lai, T.S. Lee, I.P.L. Png (2006b). Who gets spammed?. Communications of
the ACM 49(10), 83–87.
Harris Interactive (2001). Consumer privacy attitudes and behaviors survey wave II. Available at
http://www.bbbonline.org/UnderstandingPrivacy/library/harris2-execsum.pdf. Accessed August 15,
2005
Harris Interactive. (2003). Most people are ‘‘privacy pragmatists’’ who, while concerned about privacy,
will sometimes trade it off for other benefits. Available at http://www.harrisinteractive.com/ha-
rris_poll/index.asp?PID=365. Accessed August 15, 2005.
Hart, O.D., J. Tirole (1988). Contract renegotiation and Coasian dynamics. Review of Economic Studies
55, 509–540.
Hermalin, B., M. Katz (2006). Privacy, property rights & efficiency: the economics of privacy as secrecy.
Quantitative Marketing and Economics 4(3), 209–239.
Hirshleifer, J. (1971). The Private and social value of information and the reward to inventive activity.
American Economic Review 61(4), 561–574.
Hirshleifer, J. (1980). Privacy: its origin, function, and future. Journal of Legal Studies 9(4), 649–664.
Hui, K.L. (2006). Consumer disclosure: the effects of company information presentation and question
sequence. Unpublished manuscript, Department of Information Systems, National University of
Singapore.
Hui, K.L., H.H. Teo, T.S.Y. Lee (2006). The value of privacy assurance: an exploratory field experi-
ment. MIS Quarterly, forthcoming.
Jamal, K., M. Maier, S. Sunder (2003). Privacy in e-commerce: development of reporting standards,
disclosure and assurance services in an unregulated market. Journal of Accounting Research 41(2),
285–309.
Jamal, K., M. Maier, S. Sunder (2005). Enforced standards versus evolution by general acceptance: a
comparative study of e-commerce privacy disclosure and practice in the United States and the United
Kingdom. Journal of Accounting Research 43(1), 73–96.
Kahn, C.M., J. McAndrews, W. Roberds (2000). A theory of transactions privacy. Working Paper 2000-
22, Federal Reserve Bank of Atlanta.
Katz, M.L. (1984). Firm-specific differentiation and competition among multiproduct firms. Journal of
Business 57(1 part 2), S149–S166.
Konrad, K.A. (2001). Privacy and time-consistent optimal labor income taxation. Journal of Public
Koo, H.W., I.P.L. Png (1994). Private security: deterrent or diversion?. International Review of Law and
Economics 14(March), 87–101.
Kraut, R.E., S. Shyam, J. Morris, R. Telang, D. Filer, M. Cronin (2002). Markets for attention: will
postage for email help? CSCW 02, New Orleans, ACM, 206–215.
Lai, Y.L. and K.L. Hui (2004). Opting-in or opting-out on the internet: does it really matter? in:
R. Agarwal, L. Kirsch, J.I. DeGross (eds.). Proceedings of the International Conference on
Information Systems, Washington, DC, December.
Lai, Y.L., K.L. Hui (2005). Luring the prey: attracting online consumer participation through choice
configurations in: J. Langenderfer, D.L. Cook, J.D. Williams (eds.). Proceedings of the American
Market Association (AMA) Marketing and Public Policy Conference, Washington, DC, May.
Laudon, K.C. (1996). Markets and privacy. Communications of the ACM 39(9), 92–104.
Loder, T., M. Van Alstyne, R. Wash (2006). An economic response to unsolicited communication.
Advances in Economic Analysis & Policy 6(1), Article 2.
Moe, W.W., P.S. Fader (2001). Uncovering patterns in cybershopping. California Management Review
43(4), 106–117.
Moe, W.W., P.S. Fader (2004). Dynamic conversion behavior at e-commerce sites. Management Science
50(3), 326–335.
Montgomery, A.L., S. Li, K. Srinivasan, J.C. Liechty (2004). Modeling online browsing and path
analysis using clickstream data. Marketing Science 23(4), 579–595.
Mookherjee, D., I.P.L. Png (1989). Optimal auditing, insurance and redistribution. Quarterly Journal of
Economics 104(2), 399–415.
Moorthy, K.S. (1984). Market segmentation, self-selection, and product line design. Marketing Science
3(4), 288–307.
Muris, T.J. (2003). The federal trade commission and the future development of U.S. consumer pro-
tection policy, in: Aspen Summit: Cyberspace and the American Dream, The Progress and Freedom
Foundation, August 19. Available at http://www.ftc.gov/speeches/muris/030819aspen.htm
Mussa, M., S. Rosen (1978). Monopoly and product quality. Journal of Economic Theory 18(2),
301–317.
Noam, E. (1995a). Privacy in telecommunications, Part I. New Telecommunications Quarterly 3(2), 51–59.
Noam, E. (1995b). Privacy in telecommunications, Part III. New Telecommunications Quarterly 3(4),
51–60.
Odlyzko, A. (2003). Privacy, economics, and price discrimination on the internet, in: Proceedings, 5th
ACM International Conference on Electronic Commerce. Available at http://portal.acm.org/
citation.cfm?id=948051&coll=ACM&dl=ACM&CFID=15950150&CFTOKEN=71609284&ret
=1#Fulltext
OECD (1980). OECD guidelines on the protection of privacy and transborder flows of personal data. 23
September 1980. Available at http://www.oecd.org/document/18/0,2340,en_2649_34255_1815186_
1_1_1_1,00.htm. Accessed May 17, 2005.
Pau, L.F. (2005). Privacy management contracts and economics, using service level agreements (SLA).
Working Paper, Rotterdam School of Management, RSM Erasmus University, P.O. Box 1738,
Rotterdam 3000, Netherlands.
Petty, R.D. (2000). Marketing without consent: consumer choice and costs, privacy, and public policy.
Journal of Public Policy and Marketing 19(1), 42–53.
Png, I. (1998). Managerial Economics. Blackwell Publishers, Malden, MA.
Posner, R.A. (1978). An economic theory of privacy. Regulation (May/June) 9(3), 19–26.
Posner, R.A. (1979). Privacy, secrecy, and reputation. Buffalo Law Review 28, 1–55.
Posner, R.A. (1981). The economics of privacy. American Economic Review 71(2), 405–409.
Schwartz, B. (1968). The social psychology of privacy. American Journal of Sociology 73(6), 741–752.
Schwartz, P.M. (2004). Property, privacy, and personal data. Harvard Law Review 117, 2056–2127.
Serjantov, A., R. Clayton (2005). Modeling incentives for email blocking strategies. Workshop on the
Economics of Information Security (WEIS05).
Sismeiro, C., R.E. Bucklin (2004). Modeling purchase behavior at an e-commerce web site: a task
completion approach. Journal of Marketing Research 41(3), 306–323.
Smith, H.J. (2001). Information privacy and marketing: what the U.S. should (and shouldn’t) learn from
Europe. California Management Review 41(2), 8–33.
Spence, A.M. (1973). Job market signaling. The Quarterly Journal of Economics 87(3), 355–374.
Stigler, G.J. (1980). An introduction to privacy in economics and politics. Journal of Legal Studies 9(4),
623–644.
Stiglitz, J.E. (1975). The theory of screening, education and the distribution of income. American
Economic Review 65(3), 283–300.
Stone, E.F., L.D. Stone (1990). Privacy in organizations: theoretical issues, research findings, and
protection mechanisms. Research in Personnel and Human Resources Management 8, 349–411.
Tang, Z., J. Hu, M.D. Smith (2005). Protecting online privacy: self-regulation, mandatory standards, or
caveat emptor. Working Paper, Tepper School of Business, Carnegie Mellon University, April.
Taylor, C.R. (2004). Consumer privacy and the market for customer information. RAND Journal of
Economics 35(4), 631–650.
Taylor, C.R. (2005). Privacy and information acquisition in competitive markets. Working Paper,
Department of Economics, Duke University.
Thisse, J.F., X. Vives (1988). On the strategic choice of spatial price policy. American Economic Review
78(1), 122–137.
Tirole, J. (1988). The Theory of Industrial Organization. MIT Press, Cambridge, MA.
Tolchinsky, P.D., M.K. McCuddy, J. Adams, D.C. Ganster, R.W. Woodman, H.L. Fromkin (1981).
Employee perceptions of invasion of privacy: a field simulation experiment. Journal of Applied
Psychology 66(3), 308–313.
Turner, M.A. (2001). The Impact of Data Restrictions On Consumer Distance Shopping. Direct Mar-
keting Association. Available at http://www.the-dma.org/isec/9.pdf. Accessed August 15, 2003.
Turow, J., (2003). Americans & online privacy: the system is broken. Annenberg Public Policy Center,
University of Pennsylvania. Available at http://www.annenbergpublicpolicycenter.org/04_info_soci-
ety/2003_online_privacy_version_09.pdf. Accessed March 31, 2005.
Van Zandt, T. (2004). Information overload in a network of targeted communication. RAND Journal of
Economics 35(3), 542–560.
Varian, H.R. (1985). Price discrimination and social welfare. American Economic Review 75(4), 870–875.
Varian, H.R., F. Wallenberg, G. Woroch (2004). Who signed up for the do-not-call list? Third Workshop
on Economics and Information Security, University of Minnesota, Minneapolis, May.
Ward, M.R. (2001). The economics of online retail markets, in: G. Madden, S. Savage (eds.), The
International Handbook on Emerging Telecommunications Networks, Edward Elgar Publishers,
Cheltenham, UK.
Warren, S., L. Brandeis (1890). The right to privacy. Harvard Law Review 4(1), 193–220.
Wathieu, L. (2002). Privacy, exposure and price discrimination. Harvard Business School Marketing
Research Paper 02-03, October.
Wathieu, L. (2004). Marketing and the privacy concern. Unpublished manuscript, Harvard Business
School, March.
Wathieu, L., A. Friedman (2005). An empirical approach to understanding privacy valuation. Unpub-
lished manuscript, Harvard Business School.
Wattal, S., P.Y. Chen, R. Telang (2004). On personalization technology adoption and information
sharing in digital markets. Working Paper, Carnegie Mellon University.
Westin, A.F. (1967). Privacy and Freedom. Atheneum, New York, NY.
Westin, A. (2001). Testimony before U.S. House of Representatives, Committee on Energy and Com-
merce, Subcommittee on Commerce, Trade, and Consumer Protection, Hearing on ‘‘Opinion Sur-
veys: What Consumers Have To Say About Information Privacy,’’ May 8.
Woodman, R.W., D.C. Ganster, J. Adams, M.K. McCuddy, P.D. Tolchinsky, H. Fromkin (1982). A
survey of employee perceptions of information privacy in organizations. Academy of Management
Journal 25(3), 647–663.
Chapter 10
Product Bundling
Xianjun Geng
University of Washington Business School, Seattle, WA 98195, USA
Maxwell B. Stinchcombe
Department of Economics, The University of Texas at Austin, Austin, TX 78712, USA
Andrew B. Whinston
Department of Management Science & Information Systems, The University of Texas at Austin,
Austin, TX 78712, USA
Abstract
Product bundling refers to the business practice where a seller sells two or
more heterogeneous products or services in a package. To date the bundling
literature offers three major reasons for a seller to bundle products: cost-
saving, price discrimination, and bundling as a competition tool. This paper
surveys the bundling literature with a focus on understanding when and how
a seller can use bundling for price discrimination or as a competition tool.
1 Introduction
Product bundling refers to the business practice where a seller sells two or
more heterogeneous products or services in a package with a single price.
Examples of bundling include Microsoft Office, which contains several
stand-alone products such as Word and PowerPoint; Yahoo! Music, in
which for $4.99, a consumer can listen to any of 1 million music tracks
within a month timeframe; and, in a more traditional economy setting,
499
500 X. Geng et al.
Pepboys Auto’s maintenance service, in which a driver can have both oil
change and tire rotation done under a single discount price.
To date, the bundling literature offers three major reasons for a seller to
bundle products. First is cost saving, an efficiency argument. See, for in-
stance, the statement by Nalebuff (2004, p. 161):
In a larger sense, almost everything is a bundled product. A car is a bundle of seats, engine,
steering wheel, gas pedal, cup holders, and much more. An obvious explanation for many bun-
dles is that the company can integrate the products better than its customers can.
Other than reducing the integration cost, bundling may also reduce trans-
action and distribution costs involved in the selling process whenever seller-
side bundling simplifies the shopping and shipping processes (Salinger,
1995; Bakos and Brynjolfsson, 2000b). For instance, vacation packages
offered by major online travel sites can save consumers significant amount
of search time. Whenever cost saving via bundling is feasible, a bundle has
intrinsic cost advantage over individual products, thus bundling may give a
seller a higher margin than unbundling.
Although often mentioned in the literature, the cost-saving argument
received little attention in theoretical expositions due to its straightfor-
wardness. Instead, the literature has predominately focused on cases where
bundling itself does not offer any cost advantage over selling individual
products. Two other reasons for bundling are frequently mentioned: price
discrimination and bundling as a competition tool. The former is extensively
studied under monopoly setups, whereas the latter under duopoly setups.
In this paper, we survey the bundling literature with a focus on using bun-
dling for price discrimination and as a competition tool. Earlier research on
using bundling for price discrimination argues for the logic that, by bun-
dling products together, heterogeneity in buyer valuations can often be re-
duced, which in turn enables a monopolist to better capture consumer
surplus when marginal production cost is not high (Stigler, 1968; Adams
and Yellen, 1976). This logic is well illustrated in the case where buyer
valuations over two products are negatively correlated, such as in Adams
and Yellen (1976). Researchers soon discover that negative correlation is
not a necessary condition for bundling to be profitable for a monopolist
(Schmalensee, 1984; McAfee et al., 1989). Nevertheless, the intuitions that,
as far as pure bundling (if a bundle is offered, no sub-bundles or individual
products are offered) is concerned, bundling benefits a seller when it reduces
valuation heterogeneity and if marginal cost is low, hold to today even if
complementarity is also considered (Schmalensee, 1984; Lewbel, 1985;
Guiltinan, 1987; Eppen et al., 1991; Venkatesh and Kamakura, 2003).
When mixed bundling is considered, however, McAfee et al. (1989) show
that a seller may bundle in a much broader range of cases, even if marginal
costs are high, or if bundling does not reduce valuation heterogeneity. We
discuss these developments in Section 2.
Ch.10. Product Bundling 501
While earlier studies on bundling two products are more concerned with
when a monopolist should bundle, a number of recent papers shift the focus
to how to bundle—namely, what bundling strategy should a seller pick
(unbundling, pure bundling, or mixed bundling), and what price the seller
should choose for each sub-bundle or individual product. Answering these
questions requires researcher to deal with more than two products, which is
the norm, rather than the exception, in practice. This is discussed in Section
3. The literature takes two routes in dealing with the ‘‘how’’ question. One
stream of research tries to provide numerical solutions to optimal mixed-
bundling prices using integer-programming approaches (Hanson and Mar-
tin, 1990; Hitt and Chen, 2005). Computational complexity is a major
challenge to this stream of research. The other stream focuses on getting
analytical results for pure bundling or very simple forms of mixed bundling
(Armstrong, 1999; Bakos and Brynjolfsson, 1999, 2000a,b; Fang and Nor-
man, 2005; Geng et al., 2005).1 One important result for this stream of
research is that bundling can be surprisingly profitable, and at the same
time extremely simple.
Research on using bundling as a competition tool falls into two catego-
ries: entrance deterrence (also called the leverage theory) and product differ-
entiation. In entrance deterrence, bundling is also referred to as tying
(Schmalensee, 1982; Whinston, 1990; Choi and Stefanadis, 2001; Carlton
and Waldman, 2002; Gilbert and Riordan, 2003; Heeb, 2003; Nalebuff,
2004).2 The seminal paper by Whinston (1990) shows that, if a seller can
commit to pure bundling, it is then possible for pure bundling to leverage its
monopoly power in one market to another market where it competes with
other rivals.3 Intuitively, once committed to pure bundling, the seller will
have to compete aggressively with rivals, since losing the battle now means
losing sales in both product markets. Anticipating the aggressiveness of the
seller, rivals are expecting lower profits by entering the competition, and
thus may opt out of the market in the first place.
Subsequent papers expand the applicability of using bundling for en-
trance deterrence to other industrial scenarios. Choi and Stefanadis (2001)
discuss the case where products are complements, and rivals may fail to
enter a market even after incurring an entry cost. This is the case, for
example, when risky innovation is needed for entry. In this case Choi and
Stefanadis show that bundling is an effective entry barrier even if the seller
is facing entry threat from rivals in both product markets. Carlton and
1
Bakos and Brynjolfsson (1999) are among the earliest in this research stream. However, their analysis
and results regarding bundle pricing are significantly flawed as pointed out by Geng et al. (2005).
2
Though ‘‘bundling’’ and ‘‘tying’’ are interchangeable, ‘‘tying’’ does appear more frequently in papers
where the bundler is a monopoly in one market, and competes with other sellers in another market.
3
Though Whinston is the first to model tying’s leverage effect in a noncompletely competitive setup,
the debate on the so-called ‘‘leverage theory’’ started much earlier. See Whinston (1990) for a discussion
on this topic. Also see Whinston (2001) for an overview of the impacts of tying related to the web
browser industry.
502 X. Geng et al.
Waldman (2002) discuss a dynamic setting where a rival who competes in

the second market may evolve and eventually be able to enter the seller’s
monopoly market. They also show that bundling is an effective entry de-
terrence tool in this case. Recently Nalebuff (2004) shows that bundling is
valuable for the seller even if it does not engage in aggressive pricing, or
even if the rival has already entered the market (in the last case bundling no
longer serves entry deterrence—it simply maximizes the seller’s profit).
While the literature on using bundling for entry deterrence focuses on
how a seller can fend off all rivals and thus exclusively enjoys all markets it
is in, the literature on using bundling for product differentiation asks the
question of when two or more ex-ante homogeneous sellers can coexist and
both reap positive profits using bundling (Matutes and Regibeau, 1992;
Anderson and Leruth, 1993; Chen, 1997; Kopalle et al., 1999). In an in-
sightful paper, Chen (1997) analyzes the case where two sellers compete in a
first market, and both also sell another product in a second and competitive
market. Absent bundling, Bertrand competition drives both seller profits in
the first market to zero. If one seller uses bundling and the other does not,
however, both can earn positive profits since the bundle and the individual
first-market product are effectively differentiated products.
Using bundling as a competition tool is discussed in Section 4. Section 5
concludes it.
2 Bundling for price discrimination: the case of two products
To date the most studied stream in bundling literature is on using bun-

dling for price discrimination. We will also refer to ‘‘bundling for price
discrimination’’ as ‘‘monopoly bundling’’ as this stream of research always
considers a seller who is a monopolist in every market concerned. The focus
of this section is to answer the question of ‘‘when to bundle’’ in the simplest
setup—a two products setup.
We start by introducing a base model of monopoly bundling in Section
2.1, adapted from McAfee et al. (1989). Before jumping into any specific
research, we give a primer on several issues pertaining to bundling practice
and research in Section 2.2. This primer is important in setting up the strict
research context because confusion may arise on otherwise, for instance,
whether a specific research result is about pure or mixed bundling.
It turns out that the answers to ‘‘when to bundle’’ under pure and mixed
bundling are significantly different. Under pure bundling, as discussed in
Section 2.3, a seller should bundle when marginal costs are low and bun-
dling reduces heterogeneity among buyer valuations (Adams and Yellen,
1976; Schmalensee, 1984; Bakos and Brynjolfsson, 1999, 2000a,b; Geng
et al., 2005; Fang and Norman, 2005). Under mixed bundling (Section 2.4),
however, a seller may bundle in a much broader range of cases, even
if marginal costs are high, or if bundling does not reduce valuation
heterogeneity (McAfee et al., 1989). One special case is when buyer val-
uations of two products are independent. In this case, mixed bundling is
almost always the optimal choice for a monopolist.
At the end of this section, we briefly survey bundling of complements or
substitutes (Lewbel, 1985; Guiltinan, 1987; Eppen et al., 1991; Venkatesh
and Kamakura, 2003).
2.1 The base model
Consider a monopolist that offers two products, 1 and 2, to a buyer

population O of size 1. For product iA{1, 2}, each buyer oAO wants either
0 or 1 unit of each product with unit valuation vi (o)Z0. With respect to the
whole buyer population, non-negative random variables v1 and v2 follow a
joint distribution function F (v1, v2) and have finite means and variances,
which are common knowledge.4 There is no outside option, thus a buyer
will buy products whenever her surplus is non-negative. Moreover, resale is
not considered, as is the case in the literature.
The constant marginal cost for product is ci>0. We also assume that the
seller cannot refuse to sell product i to a buyer if the buyer has already
bought product j6¼i from it.5 In other words, a buyer can self-assemble the
bundle from individual products.
The monopolist’s objective is to maximize total profit from both prod-
ucts. Absent bundling, it can only optimize two variables: prices it charges
for products, p1 and p2, respectively. With the bundling option, it can opt-
imize a third variable (on top of the two individual prices): the price for the
bundle of both products, pB.
We define the following notations for ease of exposition. Let the marginal
distribution functions be derived from F (v1, v2) be F1 (v1) and F2 (v2),
respectively for products 1 and 2. Let the distribution function for a bundle,
that is, for vB ¼ v1,+v2, be derived from F (v1, v2) be FB (vB). Whenever
applicable, we use lower-case functions to denote the density functions,
such as having f (v1, v2) be F(v1, v2)’s density if the latter is differentiable.
2.2 Issues to be considered in monopoly bundling
Before we start discussing the literature, it is worthwhile for us to clarify a

few terms and issues relevant to bundling as follows:
4
Here, we ignore the issue of complementarity—it will be considered later in this section. Also, we
assume free disposal, so if a buyer does not want a product (i.e., has a negative valuation), she can
simply throw it away.
5
This is consistent with business practice where there are few examples where sellers restrict purchases
in this way. To our knowledge McAfee et al. (1989) is the only paper that discusses the case where the
seller can refuse sales of individual products.
504 X. Geng et al.
2.2.1 Pure bundling, mixed bundling, and the combinatorial issue

If the seller only offers the bundle for sale, it is called pure bundling. If,
instead, the seller offers both the bundle and individual products for sale, it
is called mixed bundling.
Both selling strategies are seen in business practice. As an example of
pure bundling, Yahoo! Music sells the access to its whole music library via
subscriptions to consumers, and a consumer cannot purchase individual
tracks or only part of the library. As an example of mixed bundling, Mi-
crosoft sells its office productivity software in a bundle called Microsoft
Office, and at the same time buyers can purchase individual productivity
software, such as Microsoft Word, separately if preferred.
The possibility of mixed bundling makes analyzing the seller’s pricing
strategy complex as there are three variables, p1, p2, and pB to be optimized
at the same time. Note that this is still under the two products case. When
there are more than two products, the complexity of mixed-bundling pric-
ing quickly explodes because of the existence of sub-bundles: bundles that
include some, but not necessarily all, of the products available. Generally, if
there are n products, the total number of sub-bundles is 2n1, which, except
for a few special cases that we will discuss in the next section, makes the
mixed-bundling pricing issue NP-hard.
2.2.2 Bundling products and bundling buyers

The vast majority of papers on bundling deal with bundling products. In
doing so, researchers usually assume buyer valuations as random valuables,
such as v1 in the base model. One exception is Bakos and Brynjolfsson
(2000b), where bundles of buyers, instead of bundles of products, are con-
sidered. One prominent example of bundling buyers is site licensing, which
is popularly used in the software industry.
Though dealing with seemingly different issues, bundling buyers and
bundling products are actually dual problems to each other: it is straight-
forward if one thinks of the similarity between the problem of ‘‘bundling
many products together and asking for a single bundle price’’ and the
problem of ‘‘grouping many buyers together and asking for a single group
price’’—mathematically speaking the terms ‘‘buyers’’ and ‘‘products’’ are
interchangeable.
2.2.3 Monitoring purchase

McAfee et al. (1989) mention the issue of the seller monitoring purchases.
If the seller is able to tell whether a buyer has bought a product from it
before, it then has the ability to refuse selling the other product to this same
buyer, and therefore be able to have a pricing strategy where p1+p2opB.
Lacking the monitoring ability the seller can only pick prices where
p1+p2ZpB, since otherwise nobody will buy the bundle. To our knowledge
only McAfee et al. (1989) discuss the monitoring issue. In this paper we
only consider the no-monitoring case.6
2.2.4 Posted price and auction

To our knowledge, until now the vast majority of papers on bundling
consider selling via posted prices. The exception is combinatorial auctions,
where different products can be sold through a menu of bundle options in
an auction (Avery and Hendershott, 2000; Xia et al., 2004). These papers on
combinatorial auctions study the complexity of bidding strategies, incentive
compatibility with multiple products, and how the relationship between
price discrimination and auctions differs between the single- and multiple-
product cases.7 In this paper we focus on posted-price bundling.
2.3 Pure bundling
Although mixed bundling is the most general form of bundling, a con-

siderable amount of research has focused on the case of comparing only
pure bundling and unbundling, thus omitting the possibility of sub-bundles.
Ease of exposition is certainly a reason frequently cited in the literature.
Moreover, many argue that in practice a seller is able to commit to pure
bundling by the so-called technological tying (Gilbert and Riordan, 2003),
where a seller makes it difficult to separate products in a bundle. One
example is Microsoft’s claim that now Internet Explorer is an inseparable
part of the Windows operating system.
In this subsection we only consider unbundling and pure bundling, where
whenever the seller offers the bundle, it will not offer individual products
for sale at the same time.8 For ease of exposition, we also limit our atten-
tion to the symmetric case, that is, F(v1,v2) ¼ F(v2, v1) and c1 ¼ c2 ¼ c. As a
result, we have marginal distribution functions F1() ¼ F2().
We start by considering a very simple, yet quit revealing, case that is
adapted from Adams and Yellen (1976). Let the marginal distribution
functions be F1() ¼ F2() ¼ U[0,1], where U[0,1] is a uniform distribution
on [0,1]. These marginal distribution functions are illustrated in Fig. 1a. Let
the marginal cost be co1. If the seller adopts unbundling, then for product
6
In practice, nevertheless, there are cases where a bundle is more expensive than the sum of individual
products. One example is vacation packages at Expedia.com, which are frequently priced high and one
can often find individual hotel, air ticket, and car rental deals to beat packages. Nevertheless, this case
can be explained using the ‘‘convenience’’ argument, since a packaged deal saved tourists time, and
Expedia.com can easily and automatically assemble these packages at little cost. Therefore, it falls into
the cost-saving argument that we mentioned in the introduction.
7
We thank Terrence Hendershott for comments on auctions.
8
The literature does not offer convincing arguments on why a seller has to stick to pure bundling (as
mixed bundling is more general) other than that pure bundling is seemingly more analytically man-
ageable.
506 X. Geng et al.
F1(v1), F2(v2)
1 1
FB(vB)
1 1
a v1, v2 b vB
Fig. 1 Bundling of perfectly negatively correlated products.
1 its profit is (p1c)(F1(p1)) ¼ (p1c)(1p1). Its optimal unbundling price

is then p1 ¼ p2 ¼ ð1 þ cÞ=2; and total profit from both products is (1c)2/2.
We now add one more assumption to this simple case. Let the valuations
of these two products be perfectly negatively correlated, that is,
vB ¼ v1 þ v2 ¼ 1. (1)
This is illustrated in Fig. 1b. It is then evident that, if the seller bundles both
products and charge a bundle price of pB ¼ 1; all consumers will buy and
the seller gets a profit of 12c.
Comparing
pffiffiffi the seller’s profit under unbundling and bundling, we have if
c 2 1; the seller bundles; otherwise it unbundles.9
There are two important intuitions in the simple case that, quite re-
markably, hold true for all existing research regarding using pure bundling
for price discrimination (Adams and Yellen, 1976; Schmalensee, 1982, 1984;
Armstrong, 1999; Bakos and Brynjolfsson, 1999, 2000a,b; Geng et al., 2005;
Fang and Norman, 2005). First, for pure bundling to be better than unbun-
dling for the seller, pure bundling should reduce heterogeneity in buyer val-
uations. In the simple case discussed above, under unbundling (Fig. 1a), for
each individual product there’s a continuum of buyer valuations. Under
pure bundling (Fig. 1b), however, heterogeneity is completely eliminated as
now every buyer has a bundle valuation of 1.10
Note that in this case valuation heterogeneity is eliminated because of the
negative correlation between the valuations of two products. Thus, one
might speculate that there is a connection between negative correlation of
valuations and pure bundling. However, later research found out that, even
without negative correlations, pure bundling can still reduce heterogeneity
9
Without loss of generality, we assume that, if the seller is indifferent between bundling and unbun-
dling, it bundles.
10
Though the intuition on bundling reducing heterogeneity in buyer valuations is evident from the
case in equation (1), it is, however, difficult to formally define reduced heterogeneity. Note that reduced
heterogeneity does not equal to lowered variance in valuations since the profitability of bundling also
depends on the shape of distribution functions. Schmalensee (1984) uses ‘‘buyer diversity’’ instead of
heterogeneity in buyer valuations.
in buyer valuations. One example is Geng et al. (2005) discussed in the next
section.
Second, for pure bundling to be better than unbundling for the seller, mar-
ginal product cost cannot be too p high.
ffiffiffi In the case above, the seller should
adopt pure bundling only if c 2 1: To understand this, consider that
c ¼ 3/4. Under pure bundling, the marginal cost of a bundle is 3/2, which is
larger than buyer valuation, thus the seller cannot make any profit. Under
unbundling, however, for each individual product there exist buyers who
have a valuation higher than 3/4—selling to them yields a positive profit for
the seller.
Though the simple case above illustrates the two intuitions well, its model
setup is too narrow to apply to practice. Follow-up research, such as Sch-
malensee (1984), Armstrong (1999), Bakos and Brynjolfsson (1999,
2000a,b), Geng et al. (2005), and Fang and Norman (2005), has greatly
expanded the applicability of the pure bundling stream by dealing with
more general setups.
The valuation distribution function for the bundle, FB(vB), that is derived
from a general joint valuation distribution function, F(v1, v2), is usually a
complex expression because the derivation involves convolution. This
makes analyzing pure bundling in a general setup difficult. Schmalensee
(1984) opts for a special form of the joint valuation distribution—the
bivariate normal distribution, which is analytically less daunting (yet nu-
merical analysis is still needed to find results even in this case). Schmalensee
also notes that ‘‘the Gaussian family (is) a plausible choice’’ because of ‘‘the
frequency with which normal distributions arise in the social sciences’’ (p.
S212).
Formally, consider the following cased adapted from Schmalensee (1984).
Let F(v1, v2) be a symmetric bivariate normal distribution with mean m,
standard deviation s, and correlation coefficient r. Under unbundling, the
profit the seller gets from selling product 1 at price p is
p1 ¼ ðp cÞð1 FðpÞÞ.
For notational convenience, define a ¼ (mc)/s, z ¼ (pc)/s, and let G(v)

be one minus a normal distribution function with mean 0 and standard
variation 1. Then we have
p1 ¼ szGððp mÞ=sÞ ¼ szGðz aÞ.
It can be shown that this profit is maximized at z ¼ z ðaÞ where

@p1 =@zjz¼z ¼ sðGðz aÞ þ zG0 ðz aÞÞ ¼ 0: Let the optimal unbundling
profit for product 1 be p1 ðaÞ ¼ sz ðaÞGðz ðaÞ aÞ ¼ suðaÞ; where for con-
venience we define
uðaÞ ¼ z ðaÞGðz ðaÞ aÞ.

508 X. Geng et al.
Since we have du=da ¼ @u=@a ¼ zG0 ðz aÞ ¼ Gðz aÞ40; and

d 2 u=da2 ¼ G 0 ðz aÞ ¼ Gðz aÞ=z40; we know u(a) is strictly increas-
ing and convex.
Similarly, under pure bundling the seller’s maximum profit is
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pB ðaB Þ ¼ sB zB ðaB ÞGðzB ðaB Þ aB Þ ¼ 2sð ð1 þ rÞ=2Þuða= ð1 þ rÞ=2Þ,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where mB ¼ 2m,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sB ¼ ð1 þ rÞ=2d2s; and aB ¼ ðmB 2cÞ=sB ¼
a= ð1 þ rÞ=2: Therefore, pure bundling is more profitable than unbun-
dling if and only if
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð ð1 þ rÞ=2Þuða= ð1 þ rÞ=2Þ uðaÞ40. (2)
Solving (2) analytically, however, is infeasible even in this special case. Note
that, since G() is a function with no variable parameters (the same applies
to its first-order derivative), one can numerically solve for z ðaÞ; and then
further for u(a). It then can be shown that there exists a threshold function
a^ ðrÞo1:3 such that, for any given ro1, (2) holds whenever a (i.e., (mc)/s)
is larger than a^ ðrÞ:
The results of this symmetric bivariate normal distribution case are con-
sistent with the two intuitions in the simpler, complete negative correlation
case. First, we can roughly view a, that is, the ratio between mean (shifted
by c) and standard variation, as a good measure p offfiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
valuation heteroge-
neity in the normal distribution case. Since aB ¼ a= ð1 þ rÞ=24a if p6¼1,
the bundle has less valuation heterogeneity than individual products except
in the perfectly positive correlation case. Second, for a given r, bundling is
better for the seller only if (mc)/s) is large enough—in other words, if c is
small enough.
Finally, notice that in this latter case pure bundling can be better than
unbundling for the seller even if the valuations of both goods are positively
(yet not perfectly) correlated. Again, we can get a larger aB than a even if r
is positive.
2.4 Mixed bundling
The most striking difference between mixed bundling and pure bundling is
that the two intuitions in the pure bundling research do not carry over to the
mixed bundling case. Indeed, under mixed bundling a seller may choose
bundling even if bundling itself does not make valuation distributions more
homogeneous, or if marginal costs are very high. The reason lies in the fact
that, under mixed bundling, it is possible for the seller to capture at least part
of the buyer population who do not buy the bundle by offering individual
products for sale among the bundle. As a result, analysis of mixed bundling is
quite different than that of pure bundling, and deriving optimal prices under
mixed bundling is generally more difficult than that under pure bundling.
On the other hand and surprisingly, though deriving optimal prices are
difficult, strong results are still available regarding when a seller should adopt
mixed bundling, compared to unbundling (McAfee et al., 1989). Unlike
Schmalensee (1984) where the author first derives optimal prices and profits
under bundling and unbundling scenarios, then compares them, McAfee
et al. take a different analytical approach. Instead of trying to derive the
optimal mixed bundling prices, they ask the question of whether the seller can
improve its profit by switching from unbundling to mixed bundling.
Formally, consider the case where the joint density function for buyer
valuations, f(v1, v2), exists but not necessarily always continuous (adapted
from McAfee et al., 1989). Let gi(vi9vj), i ¼ 1, 2 and j6¼i, denote the con-
ditional density derived from f(v1, v2). Let pi be the optimal price for
product i under unbundling, and assume that gi(pi9vj) is continuous in Pi at
pi for any vj.
Now, consider switching from unbundling to mixed bundling by intro-
ducing a bundle with price pB ¼ p1 þ p2 : Apparently, the seller’s profit will
not change after the introduction of this bundle.
We then further increase the price for product 2 to p^ 2 ¼ p2 þ ; where
e>0. This does not affect the profit from buyers with v1 4p1 since they
either buy the bundle or only product 1. From buyers with v1 p1 ; the
seller will now get a profit of
Z Z p
1 1
BðÞ ¼ p2 þ c2 f ðv1 ; v2 Þ dv1 dv2 ðproduct 2 onlyÞ
pn2 þ 0
Z p1 Z 1
þ ðp1 þ p2 c1 c2 Þ f ðv1 ; v2 Þ dv2 dv1 ðbundleÞ.
p1 p1 þp2 v1
To show that mixed bundling is more profitable than unbundling, we only

need to have B0 ðÞ40; that is,
Z p
1
1 G2 ðp2 jv1 Þ g2 ðp2 jv1 Þðp2 c2 Þ f 1 ðv1 Þdv1
0

þ ðp1 c1 Þ 1 G 2 ðp1 jp1 Þ f 1 ðp1 Þ40. ð3Þ
To see how strong result (3) is, consider the special case of independent
valuations, that is, gi (vi9vj) ¼ fi(vi) for any vj, i ¼ 1, 2 and j6¼i. Then (3)
becomes
½1 F 2 ðp2 Þ f 2 ðp2 Þðp2 c2 ÞgF 1 ðp1 Þ þ ðp1 c1 Þ½1 F 2 ðp2 Þf 1 ðp1 Þ40.
Notice that ½1 F 2 ðpn2 Þ f 2 ðp2 Þðp2 c2 Þ ¼ 0 because p2 is the optimal in-
dividual price for product 2 under unbundling, (3) is further simplified to
ðp1 c1 Þ½1 F 2 ðp2 Þf 1 ðp1 Þ40

510 X. Geng et al.
which is always true. Therefore, under independent valuations the seller

always gets a higher profit than under unbundling (conditional on the as-
sumption about continuity we had earlier).
Two points are worth emphasizing in the above case. First, this strong
result holds without requiring particular distribution functions, thus its
generality is much broader than the two cases we discussed under pure
bundling. Second and nonetheless, it does not answer the question of what
the optimal mixed bundling prices are. To our knowledge, to date this still
remains unsolved whenever there is an infinite number of buyer types (i.e.,
when f (v1, v2) is positive at infinite number of points).
2.5 Extension: bundling complements or substitutes
Until now, we have focused on the case of additive valuations, that is,
vB ¼ v1+v2. In practice, however, we frequently see cases where vB>v1+v2
(complement or superadditive case) or vBov1+v2 (substitute or subadditive
case). A Web browser is of no use without an operating system. An MP3
player is valuable only when it is filled with music/audio files. Or, possessing
two different MP3 players in a bundle does not warrant the buyer twice as
much joy since their functionalities largely overlap.
A number of papers discuss bundling complements and substitutes by a
monopolist (Lewbel, 1985; Guiltinan, 1987; Eppen et al., 1991; Venkatesh
and Kamakura, 2003). Lewbel (1985) shows through examples that the
existence of complementarity affects a seller’s bundling decision. Guiltinan
(1987) introduces marketing applications of bundling. Eppen et al. (1991) in
a non-analytical paper propose that bundling is attractive for a seller of
complements.
To date the more insightful analytical paper on bundling complements
and substitutes is Venkatesh and Kamakura (2003). Venkatesh and
Kamakura use both modeling and numerical analysis to study comple-
ments and substitutes in a two-dimensional uniform distribution case. Their
findings regarding pure bundling are consistent with the intuitions in the
case without complementarity, namely, that reduced valuation heterogene-
ity and low marginal costs favor pure bundling over unbundling. Moreover,
they find that the seller is more likely to bundle complements than to bundle
substitutes.
3 Bundling for price discrimination: the case of many products
In business, we frequently see cases where more than two products are
bundled together. Microsoft Office contains at least six pieces of stand-
alone software. By subscribing to a magazine a reader gets dozens of issues
per year. Once enrolled, a Web surfer gets the access to more than 1 million
songs at Yahoo! Music.11 Whether to adopt bundling, and how to set

bundling prices are important business questions software vendors, news
media, and entertainment companies face.
This section surveys bundling of many products. While the focus of re-
search on bundling two products is to answer the question of ‘‘when to
bundle,’’ a majority of research on bundling many products is to answer the
question of ‘‘how to bundle.’’ In doing so, one stream of research tries to
provide numerical solutions to optimal mixed-bundling prices using integer
programming (IP) approaches (Hanson and Martin, 1990; Hitt and Chen,
2005). We discuss this stream in Section 3.1. The biggest hurdle for this
approach is that the computational complexity of the IP quickly explodes
when the number of products increases. Another stream, as in Section 3.2,
focuses on getting analytical results for pure bundling or very simple forms
of mixed bundling (Armstrong, 1999; Bakos and Brynjolfsson, 1999,
2000a,b; Fang and Norman, 2005; Geng et al., 2005). One important result
for this stream of research is that bundling can be surprisingly profitable,
and at the same time extremely simple. A number of papers, including
Bakos and Brynjolfsson (1999, 2000a,b) and Geng et al. (2005), use the
context of information goods that incurs little product costs. As informa-
tion goods are one of the central components of this digital economy,
research results in this stream are heavily cited and discussed in areas such
as information systems (IS), marketing, and economics.12,13
It is also important to point out that ‘‘many products’’ does not always
imply ‘‘a very large number of products.’’ For instance, the analysis in
Armstrong (1999), Fang and Norman (2005), and Geng et al. (2005) applies
to any number of products.
Formally, in this section, we consider the case of N products labeled 1, 2,
y , N. We also abuse the notation a little and let N ¼ {1, 2, y, N}. If
NoN(N ¼ N), then it is the case of finite (infinite) number of products.
Buyer o‘s valuation of product i is vi, where oAO and iAN. For any i, vi is
non-negative,14 and has finite mean mi and finite variance si2. Unless spe-
cifically noted, we focus on the additive case, that is, the valuation of a
bundle equals to the sum of the valuations of all products in the bundle.
11
Having access to a product does not mean a buyer will use the product. This matters only when it
costs a seller to produce products. In this review, we do not discuss the difference between ‘‘can use a
product’’ and ‘‘do use a product.’’
12
A third stream deals with multiproduct non-linear pricing. See Armstrong (1996) for example. This
stream is related to multiproduct discrimination in industrial organization (Mirman and Sibley, 1980)
and multidimensional mechanism design (McAfee and McMillan, 1988). To date, this stream of research
depends on quite restrictive, and hard to verify assumptions for model trackability, such as condition
(18) in Armstrong (1996).
13
Also, Fang and Norman (2005) in a working paper consider bundling an arbitrary number of
products in the independent and identically distributed symmetric log-concave distribution case.
14
Again, we assume free disposal.
512 X. Geng et al.
3.1 The integer programming approach
One reason that finding optimal bundling prices is very difficult is that
prior research often assumes that there is an infinite number of buyer types,
that is, there are infinite possible values of the (v1, v2, y, vN) vector. In the
IP approach, researchers assume that there are only finite buyer types.
Given this assumption, a quite general IP model can be formulated for
solving the bundling pricing issue.
Formally, let N be a finite number, and let there be at most M buyer types
where M is finite (adapted from Hanson and Martin, 1990). Let bj denote
the percentage of type j buyers in the whole buyer population, where jAM.
Any type j buyer has a constant valuation vector (v1j, v2j, y, vNj). Let B be
the set of all possible sub-bundles under mixed bundling. The size of B is
then 2N1. A type j buyer’s valuation of sub-bundle bAB is then vbj ¼
P
i2b v
P ij : The marginal cost for the seller to provide sub-bundle bAB is
b
cb ¼ i2b ci ; where ci is the marginal cost for product i. Let xj ¼ 1 if a
15
buyer of type i buys sub-bundle b, and 0 otherwise. Now the optimal
bundling pricing problem can be formulated as the following IP problem:
PM P b b b
IP problem: max j¼1 b2B xj ðp c Þbj
fpb :b2Bg
Constraints:
No arbitrage pb1 þ pb2 pb3 ; where b1 ; b2 ; b3 2 B; b1 \ b2 ¼
F; and b1 [ b2 ¼ b3
Unit demand xjbA{0, 1}, where
P jAM, bAB
Buys at most one bundle b2B xbj 1; where jAM
^ ^
IR and IC xjb ¼ 0 if ðvb pb Þo maxf0; maxfðvb pb Þjb^ 2 Bgg;
j j
where jAM, bAB
Non-negative price pbX0, where bAB
Solving this IP problem gives the seller the optimal prices for every pos-
sible sub-bundle.
The main problem with this IP approach is the computational complexity
problem. Notice that there are M(2N1) IC constraints, the number of
which explodes when either the number of buyer types, M, or the number of
products, N, increases. Hanson and Martin (1990) and Hitt and Chen
(2005) try to reduce this computational complexity. Particularly, Hitt and
Chen consider the case where a buyer’s payment depends only on how
many products, not what products, she buys. Therefore the seller can only
have at most N different bundle prices. They further assume that the
number of buyer types, M, is much smaller than N.
15
For simplicity and without loss of generality, we ignore the case where, when two or more sub-
bundles give type j buyers the same surplus, different type j buyers buy different sub-bundles from this
selection.
Off the IP approach, other papers dealing with bundling many products
offer alternative ways to simplify the optimal bundling pricing issue. For
instance, Venkatesh and Mahajan (1993) consider the sport ticket market.
Noticing that sport tickets are frequently sold in only two forms—individ-
ual tickets and season tickets, Venkatesh and Mahajan focus on studying
only the complete bundle and individual products.
3.2 Pure or simple mixed bundling of many products
The IP approach is significantly limited by the complexity involved, and

thus is not practical for solving bundling problems when the number of
products is very large, say, a thousand. In recent years, nevertheless, re-
searchers find that pure bundling or some simple forms of mixed bundling
can be surprisingly profitable for a seller. That is, to offer a single bundle of
all products when there is no marginal cost, or to offer a simple two-part
tariff when there are positive marginal costs.
This approach is first proposed by Armstrong (1999) and Bakos and
Brynjolfsson (1999), and further developed by Geng et al. (2005). Bakos
and Brynjolfsson and Geng et al. deal with the case with no marginal cost,
while Armstrong considers positive marginal cost.
To start, consider the following two-part tariff (adapted from Armstrong,
1999): a buyer can choose any product she wants; if a buyer decides to buy at
least one product, the seller charges her a fixed fee, P, plus the sum of mar-
ginal costs of all products this buyer purchases. For example, if a buyer buys
products 1, 4, and 5, the price she pays is P+c1+c4+c5. Therefore this
two-part tariff is a specific pricing schedule for mixed bundling.
Given such a pricing plan and if a buyer decides to pay P, she will buy
product i if and only if vi>ci. For convenience define P vî ¼ maxfvi ci ; 0g:
Then,Pa buyer will pay the fixed fee P if and only if N i¼1 v^j P: Denote
Y¼ N v^
i¼1 j : Let the mean and variance of Y be m Y and sY2, respectively.
Note that mY is the upper bound of the profit the seller can get (i.e., profit
under first-degree price discrimination). If the seller sets P ¼ (1e)mY, then
it can get a profit of
P ¼ ð1 ÞmY ProbðY 4ð1 ÞmY Þ

ð1 ÞmY Prob jY mY jomY

ð1 ÞmY 1 ðsY =mY Þ2
2
½1 sY =mY mY .
Set ¼ ðsY =mY Þ2=3 ; we have
2=3
1 P=mY 1 2 sY =mY (4)
Therefore, the seller’s profit from this two-part tariff bundling can close to
mY if sY/mY is small enough. For any given number of products, N, the seller
514 X. Geng et al.
can calculate sY/mY to determine the effectiveness of this bundling ap-

proach.
This bundling approach is especially promising when the number of
products is very large. When No1; let m ¼ minfEð^vi Þji 2 Ng and s̄2 ¼
maxfVarð^vi Þji 2 Ng; where s̄40: For simplicity, we ignore the case of
N ¼ 1—see Geng et al. (2005) for a rigorous discussion. Now we have
ðsY =mY Þ2 ðN s̄2 Þ=ðN 2 m2 Þ ¼ ðs̄= m Þ2 =N: Then from (4) we have
1 P=mY 1 2ðs̄= m Þ2=3 =N 1=3 (5)
When N ! 1; (5) implies that P-mY. In other words, when the number
of products is very large, this simple two-part tariff enables the seller to get
approximately the first-degree price discrimination profit!
Though this stream of research on bundling a very large number of
products started only around 1999, businesses have long used similar two-
part tariff pricing schemes. For instance, warehouse clubs such as Sam’s
Club and Costco charge each member an annual membership fee, after
which a member can purchase products at very low prices.
Similar analytical techniques, that is, pooling a large number of inde-
pendent (or slightly correlated), random variables together to lower overall
uncertainty, are also used in non-product bundling areas. As pointed out by
Armstrong (1999), earlier Yarri (1976) and Rubinstein and Yaari (1983)
have shown that first-best insurance policy is approachable when agents
have intensively repeated interactions.
This two-part tariff, nevertheless, has one apparent shortcoming: seller
profit converges at the rate of N1/3, which is very slow. As a result, in
general it only applies to cases where the number of products for sale is very
large, such as Yahoo! Music that has one million songs.
A special case is when all products have zero marginal costs (Bakos and
Brynjolfsson, 1999, 2000a,b; Geng et al., 2005). This is a good approxi-
mation, for instance, when information goods such as online news are
considered where the cost of duplicating information is almost zero. In this
special case, the simple mixed-bundling pricing degenerates to even simpler
pure bundling pricing, and the two-part tariff degenerates to a single fixed
fee, P, which can be viewed as the bundle price for the bundle of all prod-
ucts.16
Though Bakos and Brynjolfsson (1999) started this research stream, their
analysis is significantly flawed as pointed out by Geng et al. (2005). Geng
et al. 2005 consider bundling of information goods with decreasing values,
such as when buyers consume products sequentially along timeline and
when a discount factor exists. Geng et al. point out that, when the number
of products, N, goes to infinity, it is conceivable that mY/N converges to
16
Strictly speaking, in the information, goods case P is the price for a buyer to be able to consume a
product. It not required, nor feasible in many circumstances, for a buyer to consume all products after
paying the bundle price.
zero—otherwise the bundle will be infinitely valuable and pricey, which is

never the case in reality. In this case, all average measures (variables divided
by the number of products) converge to zero and thus are not useful in
deriving pricing suggestions. Instead, Geng et al. argue that the correct
measures are the ones on the complete bundle, such as P and P in the
discussion above.
Before moving on to duopoly cases, we should note that there are a few
empirical papers on monopoly bundling (all in marketing).17 Venkatesh
and Mahajan (1993) apply bundling to event tickets and test the relative
performance of unbundling, pure bundling, and mixed bundling using col-
lected game ticket and season ticket data. They find that, subject to careful
choosing of bundle prices, mixed bundling outperforms unbundling and
pure bundling. Ansari et al. (1996) extend Venkatesh and Mahajan’s re-
search to the case where a seller can choose N before deciding on bundling
strategies. Yadav (1994), in an interesting paper, shows that buyers might
be quite subjective on evaluating products in a bundle. Specifically, Yadav
argues that sometimes a bundle contains a primary product (the anchor)
that buyers tend to evaluate first. One example is a bundle that consists of
one expensive item and several less-expensive ones, where buyers tend to
first look at the expensive item. Yadav finds that a buyer’s valuations of
non-primary products are positively affected by their valuation of the pri-
mary product. Finally, in a recent paper Jedidi et al. (2003) propose a
random sampling-based model for measuring buyer valuations for different
bundles.
4 Bundling as competition tools
In this section, we discuss another reason for a seller to bundle prod-

ucts—using bundling as a competition tool when rivals exist. Indeed, this role
of bundling is widely publicized in both business and academia following
the trials on US vs. Microsoft, where Microsoft is allegedly using its mo-
nopoly power (and later judged so in several cases) in personal computer
(PC) operation systems to unfairly beat rivals in other software markets,
such as the Web browser market (Whinston, 2001).
There are two major research streams on using bundling as a competition
tool. First is on the leverage theory, where there are two products and two
ex ante asymmetric sellers: one seller has monopoly power over one prod-
uct, and competes with the other seller in the second product market. The
research focus is on whether the first seller can leverage its monopoly power
in the first product market to gain competitive advantage over (or even
17
One recent theoretical paper by Fang and Norman (2005) discusses the case of log-concave val-
uation distribution functions and gives analytical results for any number of products. However, Fang
and Norman only deal with partition bundles (where no two bundles overlap).
516 X. Geng et al.
foreclose) its rival in the second product market. In this stream, the name
‘‘tying’’ is frequently used in places of ‘‘bundling.’’
Earlier research, such as Director and Levi (1956), Bowman (1957),
Posner (1976), Bork (1978), and Schmalensee (1982), focuses on the case of
one monopoly market plus one competitive market. With the exception of
Schmalensee (1982), they argue that the monopolist cannot leverage its
market power to the competitive market. The underlying logic is presented
in Section 4.1.
In recent years, we observe a dramatic shift in opinions on whether leve-
raging is effective for the first seller. Starting from the seminal paper by
Whinston (1990), however, there comes a number of papers that shift the
setup of the second product market from a competitive one to a duopoly
one (Choi and Stefanadis, 2001; Carlton and Waldman, 2002; Gilbert and
Riordan, 2003; Heeb, 2003; Nalebuff, 2004). As pointed out by Whinston
(1990), ‘‘y tying may be an effective (and profitable) means for a mon-
opolist to affect the market structure of the tied good market by making
continued operation unprofitable for tied product rivals’’ (p. 838). This
change of market structure does not apply to competitive markets where
price is fixed to marginal cost and profit is fixed to zero. They then show
that leveraging can be effective in a wide range of cases, such as when the
monopolist can pre-commit to bundling (Whinston, 1990), when the rival
has the potential to eventually enter the first-product market conditional on
success in the second-product market (Carlton and Waldman, 2002), when
the seller already faces immediate entry threat by the rival in both markets
it operates (Choi and Stefanadis, 2001), and even when the rival already
enters the second market (Nalebuff, 2004). Interestingly, a large proportion
of these papers discuss the complement products case (Whinston, 1990;
Choi and Stefanadis, 2001; Carlton and Waldman, 2002; Gilbert and
Riordan, 2003; Heeb, 2003), which is becoming increasingly important in
this digital economy where products increasingly depending on each other
to function.18 We discuss the additive valuations case in Section 4.2 and the
complements case in Section 4.3.
The second major research stream on using bundling as a competition
tool considers two ex ante symmetric sellers (Matutes and Regibeau, 1992;
Anderson and Leruth, 1993; Chen, 1997; Kopalle et al., 1999). Having one
seller beating the other one is not the research focus. Instead, this stream
focuses on two questions. First, will sellers adopt bundling in equilibrium?
So far the results are mixed. Second, can the option of bundling (compared
to unbundled selling only) increase both sellers’ profits? In an insightful
paper Chen (1997) gives an affirmative answer—he shows that bundling can
18
One prominent example is software, where almost all types of application software depend on an
operating system, and some also depend on database software, to work. Another example is UDDI-
based Web services (http://www.uddi.org), where Web-based services are assembled dynamically and in
real time by using software components from many independent vendors.
be an effective price differentiation mechanism, which helps to avoid price

wars. This stream is discussed in Section 4.4.
Formally, we inherit the two products setup in Section 3 with the fol-
lowing modifications. For convenience we call the market for product i
‘‘market i,’’ where i ¼ 1, 2. There are now two sellers, A and B. Except for
Section 4.4, we assume that A is a monopolist in market 1, and A and B
compete in market 2. Let a buyer’s valuation of product 2 from seller A be
v2A, and from seller B be v2B.
4.1 A monopoly market plus a competitive market
Early research on leverage theory (before Whinston’s paper in 1990) fo-

cuses predominantly on the case of a monopoly market plus a competitive
market, and argues that leverage does not work (except for Schmalensee,
1982). To see the reason, consider the following simple case.
First note that, since market 2 is competitive, there is a prevailing market
price for product 2 that equals the marginal cost, c2. Now, suppose in
equilibrium A uses pure bundling and some buyers do buy the bundle. Let
the equilibrium bundle price be pB.
Consider a buyer o who buys the bundle. For this buyer we have v1 ðoÞ þ
v2 ðoÞpB and v1 ðoÞ þ v2 ðoÞ pB 4v2 ðoÞ c2 : The seller’s profit is then
pBc1c2. If, instead, the seller unbundles and sets p1 ¼ PBc2, the same
buyer still buys since p1 ¼ pB c2 ov1 ðoÞ; and the seller can still get a profit
of pBc1c2. Therefore, pure bundling cannot get the seller more profit than
unbundling. In other words, a monopolist in one market cannot get twice as
much monopoly profit by bundling into another competitive market.
It should be noted, nevertheless, that the seller may get a higher profit by
bundling, compared to unbundling, if mixed bundling is allowed. Sch-
malensee (1982) shows in an example that, if mixed bundling is allowed, the
monopolist can use the bundle price pB and price for product 1 p1 to
effectively screen buyers into self-selection and therefore to get a higher
profit.
4.2 A monopoly market plus a potentially duopoly market: the additive

valuations case
In the case where market 2 is competitive, seller B is a passive player

whose business is not affected by seller A in any way: its price is always c2,
and its profit is always zero. In a criticism of the earlier literature, Whinston
(1990) argues that one should explicitly consider the impact of bundling on
the market structure of market 2, namely, whether seller B will be deterred
from entering the market. To do so a duopoly structure, instead of a com-
petitive structure, is needed for analyzing market 2.
We start with a setup adapted from Whinston (1990). Let v1 be a constant
so that buyers are differentiated only by (v2A, v2B). It can be shown that,
518 X. Geng et al.
under this setup and in equilibrium, any mixed bundling pricing schedule
for seller A is equivalent to some unbundling pricing schedule. Therefore,
seller A cannot earn more from mixed bundling than from unbundling,
seller B will not earn less, and therefore bundling is not effective in entry
deterrence.
On the other hand, if seller A is able to pre-commit to pure bundling, then
bundling can be effective in reducing seller B’s profit—sometimes to the
level that seller B feels entrance not worth the effort.19 To see the reason,
first consider the unbundling case: seller A charges v1 (remember now it is a
constant) in market 1 and everyone buys product 1, and charges the un-
bundling duopoly price, p2B ; in market 2 and gets partial of the market.
When seller 1 commits to pure bundling, it will offer a single bundle at
price pB. If it picks pB ¼ v1 þ p2B ; it will lose some sales in product 1 since
now buyers with low valuations of product 2 will not buy the bundle at all.
In fact, under pB ¼ v1 þ p2B only those buyers who buy product 2 in the
unbundling case will buy the bundle now. It is clear that seller A would like
to win back buyers in market 1 by reducing bundle price: at the neighbor-
hood of pB ¼ v1 þ p2B the marginal benefit from reducing bundle price in
market 1 is v1, while the marginal loss in market 2 is 0. Therefore in equi-
librium seller A will choose pB ov1 þ p2B : This in turn intensifies compe-
tition in market 2 and reduces seller B’s profit. If the profit reduction is
severe enough, seller B will not enter market 2 as soon as it observes seller A
pre-committing to pure bundling!
In summary, Whinston (1990) points out that a pre-commitment to pure
bundling is a credible threat to seller B that seller A will overproduce
product 2—in order to sell more of product 1 since they are now bundled
together—and thus will compete fiercely in market 2. This creditable threat
thus might change the structure of market 2 from a potentially duopoly
market to a monopoly market.
In a follow-up paper, Nalebuff (2004) extends the applicability of using
pure bundling for entry deterrence to cases where (v1, v2) follows a two-
dimensional uniform distribution.20 He shows that, by pure bundling, seller
A is able to effectively reduce seller B’s profit even if A does not discount
the price of the bundle (i.e., does not make it lower than the sum of optimal
unbundling prices). Moreover, if seller B’s entry cost is very low, and it
enters despite bundling by seller A, it is possible that, post-entry, seller A’s
loss under pure bundling is less than that under unbundling. This makes
pure bundling an attractive option for seller A no matter entry happens or
not.
19
A common used way to pre-commit to pure bundling is technological tying. See the discussion in
Section 2.3.
20
Nalebuff (2004) points out that the results also hold for ‘‘the entire class of symmetric quasi-concave
densities’’ (p. 164). Also note that Nalebuff does not discuss mixed-bundling as he states ‘‘Given the
limited potential for (mixed bundling) to increase profits, we do not pursue it further’’ (p. 174). There-
fore, how seller A can pre-commit to pure bundling is an unanswered issue.
Contrary to Whinston (1990) and Nalebuff (2004), where the focus is on

using bundling for short-term entry deterrence, Wilson et al. (1990) con-
siders a long-term setup where seller B’s product is technologically superior,
and the markets will grow faster with better products. Wilson et al. argue
that in this setup seller A may want to unbundle and encourage buyers to
assemble product 1 from A with product 2 from B, which in the long run
leads to a larger market that eventually benefits seller A.
4.3 A monopoly market plus a potentially duopoly market: the complements

case
One interesting variation in using bundling for entry deterrence is the case
of complements. Specifically, several papers discuss the case of complete
complements, where a product is useless without the other. Examples in-
clude computer hardware and software—a piece of hardware will not work
without software, and vice versa; VCR and videocassettes; electronic key-
boards and power units; and so on.
Studying bundling of complements in a duopoly setup is important espe-
cially as we enter this information economy, where many products are pro-
duced as ‘‘components,’’ and they need to be assembled together as a
‘‘system’’ for them to work. Unlike in the traditional economy where assem-
bling is difficult for buyers (such as assembling auto components into a car),
which calls for cost-saving-based argument for bundling, in this information
economy many components are produced with the intention for buyers to
assemble them with ease. A music fan can easily use iPod, iTunes, and pur-
chased songs to assemble a pleasant mobile entertainment experience. A Web
surfer can easily download Firefox Web browser and use it on a Windows PC.
Most recently, the invention of Web services technology makes it feasible (yet
mature solutions are yet to arrive) for a non-technical buyer to pull services
from various web sites and easily assemble them to fit her own needs. As a
result, the cost-saving based argument for bundling is less relevant.
One important insight in this case of complements is that seller A’s bundling
strategy are significantly different from that of additively separable products.
Consider the following simply setup (also adapted from Whinston, 1990, Sec-
tion 3) where the value of a bundle, vB, is now a directly defined non-negative
random variable (now v1 and v2 are not applicable). Absent entry threat, let
seller A’s optimal pure bundle price be pB : Seller A, nonetheless, can achieve
the same profit in the following way: seller A unbundles both products, and
charge pB b on product 1 and b on product 2, where 0pbpc2.
Now observe that this unbundled selling strategy will still keep seller A’s
monopoly pure bundling profit even if seller B exists at the start of the
game: since the price A charges for product 2, b, is lower than the pro-
duction cost, seller B will never enter the market. Therefore, one important
message in bundling complete complements is that seller A can accomplish
520 X. Geng et al.
entry deterrence even if it does not explicitly use bundling strategy. By beefing
up the price of product 1 and lowering that of product 2, seller A can
achieve the exact same pure bundling result. Moreover, this strategy also
works when product 1 is a stand-alone product, while product 2 is fully
dependent on 1. That is, v1>0, v2 ¼ 0, and vB>v1. Examples include a PC
operating system and a Web browser—the former is still valuable without
the latter, yet the latter is useless without the former. The above discussion
suggests that, in order to deter entry in the Web browser market, seller A
does not need to explicit bundle operating system and Web browser to-
gether. Instead, A can simply charge a high price on the operating system,
and give away the Web browser for free.
Several papers extended this stream started by Whinston (1990). Carlton
and Waldman (2002) consider the case where seller B has the potential to
eventually enter the first-product market conditional on success in the sec-
ond-product market. This is a valuable extension since, after all, we observe
business expansion by successful companies all the time. One example is
Google, which, after being successful in online search, is now leaping into
PC software market such as photoprocessing (Picasa) and PC file search
(Google Deskbar). Carlton and Waldman find bundling to be effective for
seller A to fend off seller B, even if seller B is technically superior (i.e., of
higher quality). In this case seller A bundles not only to defend current
profit in market 2, but also to protect future profit in market 1.
Until now we have assumed that, at the start of the game, seller B can
assail seller A only in market 2. It gives the impression that successful entry
deterrence depends on seller B’s inability to attack market 1 at the start of
the game. In an interesting paper, however, Choi and Stefanadis (2001)
show that entry deterrence may still work even if seller B is able to enter
both markets. They consider a case where seller B faces probabilistic entry,
that is, after incur an entry cost, seller B may either succeed or fail to enter a
market. This is the case when technological innovation is considered (such
as to invent a search engine), and when successful innovation is not guar-
anteed even if a considerable amount of money is spent on R&D. When
seller A unbundles, seller B can get some return as long as it can successfully
enter one market as a buyer can assemble B’s product with the according
complement from A. When seller A bundles, however, a buyer will not buy
from B unless B offers both products—this implies that B needs to be
successful in entering both markets before it can earn anything back. The
chance of successfully entering both markets, nevertheless, is much lower
than that of successfully entering at least one. As a result, seller B may opt
out of competition when seller A bundles.
4.4 Duopoly bundling
Besides the leverage theory, another stream of research in using bundling

as a competition tool considers two ex ante symmetric sellers (McAfee et al.,
1989; Matutes and Regibeau, 1992; Anderson and Leruth, 1993; Chen,
1997; Kopalle et al., 1999), and focuses on how both sellers can benefit from
the option of bundling. We call this case duopoly bundling, as both sellers
are able to bundle.
McAfee et al. (1989), based their strong results regarding monopoly bun-
dling, point out that under duopoly and independent valuations, it is impossible
for both sellers to unbundle in equilibrium: if one seller unbundles, then the
residual markets for the other seller still displays independent valuations, and
thus this other seller should adopt mixed bundling as discussed in Section 3.
Matutes and Regibeau (1992) introduce another decision dimension into
the duopoly case, where before the game starts, both sellers can decide on
whether to make their products compatible with those of their rival’s. They
show that the answer on the compatibility issue is mixed. Moreover, given
that both sellers decide to have their products compatible, they will adopt
mixed bundling in equilibrium. Finally, both sellers will do better if they can
pre-commit to unbundling. In other words, the option of mixed bundling
puts both sellers into a prisoner’s dilemma.
In an insightful paper Chen (1997) considers a market setup unique from
prior research. He considers the case where the market for product 1 is a
duopoly market with homogeneous valuations, and that for product 2 is a
competitive market with heterogeneous valuations. Absent the competitive
market 2, Bertrand competition will drive both sellers’ profit to zero. With
market 2, however, in equilibrium one seller bundles and the other does not,
and both sellers earn positive profits. To understand this result, consider the
following setup adapted from Chen (1997). Let v1 be a constant and v1>c1,
and keep v2 as a random variable. Since market 2 is competitive, p2 ¼ c2. It
is straightforward that, if both sellers unbundle, both will charge a price of
c1 in market 1, and both will earn a profit of zero.
Now consider a two-stage game. In stage 1, both sellers pick their bun-
dling strategy. They can pick either unbundling, or pure bundling.21 In
stage 1 they set price p1 for product 1 under unbundling, or price pB for the
bundle under pure bundling.
We already know that, if both sellers choose unbundling, each earns zero
profit. It is straightforward to see that both still earn zero profit if both
bundle: in this case both will charge a bundle price of c1+c2. Now assume
that seller A unbundles with equilibrium price p1 and seller B bundles with
equilibrium price pB :
To see why in equilibrium we have pB 4c1 þ c2 ; suppose that, instead,

pB ¼ c1 þ c2 : Now let seller A charge a price of p1 ¼ c1 þ ; where e>0 and
is small. A buyer will buy from seller A if and only if v1 c1 40 and
v1 c1 4v1 þ v2 c1 c2 ; or equivalently,
o minfv1 c1 ; c2 v2 g (6)
21
In an extension Chen shows that mixed bundling is always weakly dominated by pure bundling.
522 X. Geng et al.
When e is small, inequality (6) holds for buyers whose valuation of product
2 is below the marginal cost. As a result, seller A can get a positive profit by
charging p1 ¼ c1 þ 4c1 : However, when seller A charges c1+e, seller B
can charge a bundle price of c1+e+c2 and also get a positive profit, which
contradicts the optimality of pB ¼ c1 þ c2 : Therefore in any equilibrium we
have pB 4c1 þ c2 : As a result we have p1 4c1 :
To summarize, the insight from Chen (1997) is that bundling can be an effec-
tive product differentiation mechanism to avoid price wars.22
5 Concluding remarks
In this paper we review the literature on bundling. Emphasis is put on two

areas: bundling for price discrimination and bundling as a competition tool,
where there are significant developments in recent years. In both areas,
bundling is shown to be a powerful tool for sellers in a wide range of cases.
It is worthwhile to point out that many of these recent theoretical de-
velopments are closely related to the developments in electronic commerce
(e-commerce). This is evident from the fact that a major portion of papers
on bundling cite information goods, such as software, online service, and
streaming music, as their examples. Here are three possible reasons on why
information goods and e-commerce play such an important role in the
advancement of the bundling research. First, many information goods are
not designed for stand-alone usage. Instead, they are considered ‘‘compo-
nents’’ that need to be assembled into a ‘‘system’’ for usage (such as as-
sembling various Linux components into a complete business computing
environment). Bundling naturally arises in these ‘‘system’’ setups. Second, it
is much easier to bundle information goods, compared to bundling physical
products. Finally, advancements in computing technologies make it pos-
sible for a firm to engage in complicated mixed bundling strategies (that are
formerly infeasible) in order to better extract consumer surplus and to
compete with other firms.
With regard to price discrimination, it is also worthwhile to compare
bundling with another active research topic: behavior-based price discrim-
ination (see, for instance, Taylor, 2002; Fudenberg and Villas-Boas, 2005),
where a seller may infer valuable information using a customer’s past
behavior for the purpose of more-accurate price discrimination. Recent
research on behavior-based price discrimination is also closely related to
e-commerce, where technological advancements provide sellers with
unprecedented (though sometimes controversial) power in tracking and
22
In addition, Kopalle et al. (1999) also study duopoly bundling in models where buyer choice over
alternatives is modeled as a logit model, instead of based on a buyer’s highest surplus as we assumed
throughout this survey. Kopalle et al. find that, under the logit buyer choice model, sellers may engage in
either bundling or unbundling in equilibria.
collecting customer information. It is interesting, though, to see the differ-

ences between behavior-based price discrimination and bundling for price
discrimination. First, the former usually employs complicated signaling/
screening mechanisms for self-selection by consumers, which is accompa-
nied by complex pricing schemes. The latter, on the other hand and es-
pecially in the case of a large number of products discussed in Section 3.2,
can have a surprisingly simple pricing scheme. In Section 3.2, the pricing
scheme is simply a single pure bundle price in the information goods case
(yet still very powerful in extracting consumer surplus). Second, behavior-
based price discrimination may raise privacy concerns when customer-level
data are collected, such as the consumer uproar several years ago when
Amazon sold CDs at different prices to buyers of different purchase his-
tories (Hui and Png, 2005). Bundling does not depend on collecting cus-
tomer-level data, thus there is little privacy issue. Indeed, in the pure
bundling case not only no customer-level data is collected, the seller also
sells at a single price (for the bundle) to all buyers. It is still price discrim-
ination, yet in a seemingly fair format. Besides the differences, bundling can
also be used as an effective commitment mechanism when, under behavior-
based price discrimination, ‘‘the seller may be better off if it can commit to
ignore information about the buyer’s past decisions’’ (Fudenberg and
Villas-Boas, 2005, p. 2).
Acknowledgment
We thank the editor, Terrence Hendershott, for helpful comments.
References
Adams, W.J., J.L. Yellen (1976). Commodity bundling and the burden of monopoly. Quarterly Journal
of Economics 90(3), 475–498.
Anderson, S.P., L. Leruth (1993). Why firms may prefer not to price discriminate via mixed bundling.
International Journal of Industrial Organization 11, 49–61.
Ansari, A., S. Siddarth, C.B. Weinberg (1996). Pricing a bundle of products or services: the case of
nonprofits. Journal of Marketing Research 33(1), 86–93.
Armstrong, M. (1996). Multiproduct nonlinear pricing. Econometrica 64(1), 51–75.
Armstrong, M. (1999). Price discrimination by a many-product firm. Review of Economic Studies 66,
151–168.
Avery, C., T. Hendershott (2000). Bundling and optimal auctions of multiple products. Review of
Bakos, Y., E. Brynjolfsson (1999). Bundling information goods: pricing, profits and efficiency. Man-
agement Science 45(12), 1613–1630.
Bakos, Y., E. Brynjolfsson (2000a). Bundling and competition on the internet. Marketing Science 19(1),
63–82.
Bakos, Y., E. Brynjolfsson (2000b). Aggregation and disaggregation of information goods: implications
for bundling, site licensing and micropayment systems, in: B. Kahin, H.R. Varian (eds.), Internet
Publishing and Beyond. MIT Press, August 2000.
524 X. Geng et al.
Bork, R.H. (1978). The Antitrust Paradox. Basic Books, New York.
Bowman, W.S. (1957). Tying arrangements and the leverage problem. Yale Law Review 67, 19–36.
Carlton, D.W., M. Waldman (2002). The strategic use of tying to preserve and create market power in
evolving industries. RAND Journal of Economics 33, 194–220.
Chen, Y. (1997). Equilibrium product bundling. The Journal of Business 70(1), 85–103.
Choi, J.P., C. Stefanadis (2001). Tying, investment, and the dynamic leverage theory. RAND Journal of
Director, A., E. Levi (1956). Law and the future: trade regulation. Northwestern University Law Review
51, 281–296.
Eppen, G.D., W.A. Hanson, R.K. Martin (1991). Bundling—new products, new markets, low risk.
Sloan Management Review 32, 7–14.
Fang, H., Norman, P. (2005). To bundle or not to bundle. Rand Journal of Economics. Available at
faculty.arts.ubc.ca/pnorman/Research/Research.htm.
Fudenberg, D., Villas-Boas, J.M. (2005). Behavior-based price discrimination and customer recognition,
in: Hendershott (eds.), Handbook in Information Systems. Elsevier, Amsterdam, forthcoming.
Geng, X., M.B. Stinchcombe, A.B. Whinston (2005). Bundling information goods of decreasing value.
Gilbert, R.J., Riordan, M.H. (2003). Product improvement and technological tying in a winner-take-all
market. Working Paper, December 2003.
Guiltinan, J.P. (1987). The price bundling of services: a normative framework. Journal of Marketing
51(2), 74–85.
Hanson, W.A., R.K. Martin (1990). Optimal bundle pricing. Management Science 36(2), 155–174.
Heeb, R. (2003). Innovation and vertical integration in complementary markets. Journal of Economics
and Management Strategy 12(3), 387–417.
Hitt, L., P.-Y. Chen (2005). Bundling with customer self-selection: a simple approach to bundling low
marginal cost goods. Management Science 51(10), 1481–1493.
Hui, K.-L., I.P.L. Png (2005). The economics of privacy, in: Hendershott (eds.), Handbook in Infor-
mation Systems. Elsevier, Amsterdam, forthcoming.
Jedidi, K., S. Jagpal, P. Manchanda (2003). Measuring heterogeneous reservation prices for product
bundles. Marketing Science 22(1), 107–130.
Kopalle, P.K., A. Krishna, J.L. Assuncao (1999). The role of market expansion on equilibrium bundling
strategies. Managerial and Decision Economics 20(7), 365–377.
Lewbel, A. (1985). Bundling of substitutes or complements. International Journal of Industrial Organ-
ization 3, 101–107.
Matutes, C., P. Regibeau (1992). Compatibility and bundling of complementary goods in a duopoly.
The Journal of Industrial Economics 40(1), 37–54.
McAfee, R.P., J. McMillan (1988). Multidimensional incentive compatibility and mechanism design.
Journal of Economic Theory 46, 335–354.
McAfee, R.P., J. McMillan, M.D. Whinston (1989). Multiproduct monopoly, commodity bundling, and
correlation of values. Quarterly Journal of Economics 104(2), 371–383.
Mirman, L.J., D.S. Sibley (1980). Optimal nonlinear prices for multiproduct monopolies. Bell Journal of
Nalebuff, B. (2004). Bundling as an entry barrier. Quarterly Journal of Economics 119(1), 159–187.
Posner, R.A. (1976). Antitrust Law: An Economic Perspective. University of Chicago Press, Chicago.
Rubinstein, A., M. Yaari (1983). Insurance and moral hazard. Journal of Economic Theory 14,
441–452.
Salinger, M.A. (1995). A graphical analysis of bundling. Journal of Business 68(1), 85–98.
Schmalensee, R. (1982). Commodity bundling by a single-product monopolist. Journal of Law and
Schmalensee, R. (1984). Gaussian demand and commodity bundling. Journal of Business 57(1),
S211–S230.
Stigler, G.J. (1968). A note on block booking, in: G.J. Stigler (ed.), The Organization of Industries,
Irwin, pp. 165–170.
Taylor, C.R. (2002). Private demands and demands for privacy: dynamic pricing and the market for
customer information. Working Paper, Department of Economics, Duke University. Available at
http://www.econ.duke.edu/Papers/Other/Taylor/private.pdf.
Venkatesh, R., W. Kamakura (2003). Optimal bundling and pricing under a monopoly: contrasting
complements and substitutes from independently valued products. Journal of Business 76(2), 211–
231.
Venkatesh, R., V. Mahajan (1993). A probabilistic approach to pricing a bundle of products or services.
Journal of Marketing Research 30(4), 494–508.
Whinston, M.D. (1990). Tying, foreclosure, and exclusion. American Economic Review 80(4), 837–859.
Whinston, M.D. (2001). Exclusivety and tying in U.S. v. Microsoft: what we know, and don’t know.
Journal of Economic Perspectives 15(2), 63–80.
Wilson, L.O., A.M. Weiss, G. John (1990). Unbundling of industrial systems. Journal of Marketing
Research 27(2), 123–138.
Xia, M., G.J. Koehler, A.B. Whinston (2004). Pricing combinatorial auctions. European Journal of
Operational Research 154(1), 251–270.
Yadav, M.S. (1994). How buyers evaluate product bundles: a model of anchoring and adjustment. The
Journal of Consumer Research 21(2), 342–353.
Yarri, M. (1976). A law of large numbers in the theory of consumer’s choice under uncertainty. Journal
of Economic Theory 12, 202–217.
Chapter 11
Dynamic Pricing in the Airline Industry
R. Preston McAfee
California Institute of Technology
Vera L. te Velde
California Institute of Technology
Abstract
Dynamic price discrimination adjusts prices based on the option value of

future sales, which varies with time and units available. This paper surveys
the theoretical literature on dynamic price discrimination, and confronts the
theories with new data from airline-pricing behavior.
Computerized reservation systems were developed in the 1950s to keep

track of airline seat booking and fare information. Initially these were in-
ternal systems, but were soon made available to travel agents. Deregulation
of airline pricing in 1978 permitted much more extensive use of the systems
for economic activity, especially pricing. The initial development of dynam-
ically adjusted pricing is often credited to American Airlines’ Robert Crand-
all, as a response to the rise of discount airline People’s Express in the early
1980s. The complexity and opaqueness of airline pricing has grown over
time. As a result, the ‘‘yield management’’ system employed by airlines has
pricing have become one of the most arcane and complex information sys-
tems on the planet, and one with a very large economic component. Airline
pricing represents a great challenge for modern economic analysis because it
is so distant from the ‘‘law of one price’’ level of analysis. This paper surveys
the theoretical literature, which is mostly found in operations research jour-
nals, develops some new theory, assesses the holes in our knowledge, and
describes some results from a new database of airline prices.
Dynamic pricing, which is also known as yield management or revenue
management, is a set of pricing strategies aimed at increasing profits. The
527
528 R. P. McAfee and V. L. te Velde
techniques are most useful when two product characteristics coexist. First,
the product expires at a point in time, like hotel rooms, airline flights,
generated electricity, or time-dated (‘‘sell before’’) products. Second, ca-
pacity is fixed well in advance and can be augmented only at a relatively
high marginal cost. These characteristics create the potential for very large
swings in the opportunity cost of sale, because the opportunity cost of sale
is a potential foregone subsequent sale. The value of a unit in a shortage
situation is the highest value of an unserved customer. Forecasting this
value given current sales and available capacity represents dynamic pricing.
Yield management techniques are reportedly quite valuable. One estimate
suggests that American Airlines made an extra $500 million per year based on
its yield management techniques (Davis, 1994). This number may be inflated
for several reasons. First, it includes sales of yield management strategy to
others, as opposed to American’s own use of the techniques, although the
value of American’s internal use is put at just slightly less. Second, it incor-
porates ‘‘damaged good’’ considerations in the form of Saturday-night stay-
over restrictions, as well dynamic pricing. Such restrictions facilitate static
price discrimination, and are reasonably well-understood in other contexts
(Deneckere and McAfee, 1996). Nevertheless, there is little doubt that dy-
namic price discrimination is economically important. The pricing systems
used by most major airlines are remarkably opaque to the consumer, which is
not surprising given one estimate that American Airlines changes half a mil-
lion prices per day. The implied frequency of price changes seems especially
large given that American carries around 50,000 passengers per day.
There is surprisingly little research in economics journals concerning yield
management, given its prominence in pricing in significant industries and
the economic importance attached to it. This paper contributes to our un-
derstanding of yield management in five ways. First, it provides an extensive
survey of yield management research in operations research journals. Sec-
ond, we explore an existing model of Gallego and van Ryzin (1994) that has
a number of desirable properties, including closed form solutions and sharp
predictions, to address dynamic pricing considerations. Third, most of the
literature assumes demand takes a convenient but unlikely form. We con-
sider the implications of constant elasticity of demand and demonstrate
some new inequalities concerning this more standard case. We examine this
case in the context of an efficient allocation, rather than the profit-max-
imizing allocation, and show that many of the conclusions attributed to
profit-maximization are actually consequences of the dynamic efficiency.
Fourth, we take a new look at dynamic pricing from the perspective of
selling options. A problem that airlines face is of late arrivals, which may
have significantly higher value than early arrivals, suggesting the airline
ought to sell two kinds of tickets: a guaranteed use ticket and a ticket that
can be delayed at the airline’s request. Fifth, we’ve collected airline pricing
data and generated stylized facts about the determinants of pricing, facil-
itating the evaluation of models.
Ch. 11. Dynamic Pricing in the Airline Industry 529
1 Airline pricing
Airline pricing in the United States is opaque. It is not uncommon for

one-way fares to exceed round-trip prices. The difference in price between
refundable and non-refundable tickets is often a factor of four or five.
Prices change frequently, with low fares on a particular flight being avail-
able, then not, and then available again. Average prices for round-trips
between Phoenix and Los Angeles differ depending on whether they orig-
inate in Los Angeles or in Phoenix. This is particularly mysterious in that
the same airlines fly these round-trips with the same set of offerings.
We collected data on fares for 1,260 flights from American Airlines, Orbitz,
and Travelocity. Initially we collected data on each flight eight times daily,
but when American Airlines objected to the volume of searches, we stopped
using American’s site and scaled back the other two sites to once per day per
flight. As we completed some of the searches, we scaled the frequency up.
Nevertheless, this represents a more intensive look at dynamic pricing than is
available from any other source to our knowledge. We will take up our
findings after presenting the theory to put the findings in perspective.
We will use the following terminology. We will use ‘price dispersion’ to
refer to ‘static randomization’ by firms in pricing, in which each customer
pays an identical price, but that price is determined by a random process.
We will use ‘dynamic price discrimination’ to refer to charging different
customers distinct markups over marginal cost based on the time of pur-
chase; when such pricing is efficient (maximizes the expected present value
of the gains of trade), we’ll call it ‘dynamic pricing’ rather than ‘dynamic
price discrimination’. Restrictions like Saturday-night stayovers, that create
less valuable products, involve static price discrimination.
2 Existing literature
Readers seeking a general discussion of yield management are referred to

Williams (1999), Brooks and Button (1994), and Kimes (1989).1 Kimes
(1989) discusses situations appropriate for yield management solutions,
particular issues involved such as demand forecasting and overbooking,
solution techniques, and managerial implications of yield management sys-
tems. Brooks and Button (1994) discuss the rise of yield management dur-
ing and after deregulation of the 1970s and 1980s, using the shipping
industry as a detailed example, and Williams (1999) discusses yield man-
agement in terms of the interactions between the firm, resources, products,
customers, and competitors. Talluri and van Ryzin (2004) have created a
thorough textbook for the students of yield management.
1
A thorough literature review is contained in McGill and Van Ryzin (1999), along with a complete
glossary of terms applicable to yield management, particularly in the airline industry. This literature
review is largely distinct from theirs and may be considered in tandem.
Yield management applications in the made-to-order (MTO) manufac-

turing industry include Harris and Pinder (1995), Sridharan (1998), and
Barut and Sridharan (2004). Both MTO firms and service providers such as
airlines face the problem of effectively utilizing a fixed capacity under un-
certain or high demand in order to maximize revenue, and thus many yield
management results are applicable to the MTO manufacturing industry.
However, MTO manufacturing is different on the key points of finite time
horizon and unchanging capacity. The horizon is infinite, since the factory
never stops all operations at a specific time or sets a common deadline for
all activity, and the capacity is not fixed, in that as orders are completed,
capacity is replenished. Thus, the MTO problem is more of a ‘‘stock out’’
problem than a yield management problem. Harris and Pinder (1995) dis-
cuss the applicability of traditional yield management to MTO manufac-
turing and its managerial implications and develop a relevant theoretical
framework using price classes based on unit-capacity rates. Sridharan
(1998) describes the use of yield management in manufacturing situations
with higher demand than capacity, discussing three methods of increasing
efficiency and revenue: capacity rationing based on price classes, increased
coordination between marketing and manufacturing, and subcontracting.
Barut and Sridharan (2004) further explore capacity rationing by develop-
ing a dynamic capacity apportionment procedure based on discriminating
in favor of projects with a higher expected profit margin.
Rather than a continuous stream of one-time manufacturing requests,
Carr and Lovejoy (2000) consider a non-competitive firm that agrees to
commitments of a normally distributed random annual demand. This ‘‘in-
verse newsvendor problem’’ matches a known capacity with a desired ag-
gregate demand distribution. They also consider the effect of multiple price
classes. Fan and Wang (1998) study a similar annual capacity management
problem: developing an optimal harvesting policy of a renewable resource.
The resource population is modeled as a time-dependent logistic equation
with periodic coefficients, and the maximum annual-sustainable yield is
determined along with the corresponding harvesting policy.
The most common setting for yield management research is motivated by
the airlines and hotel industries. The remainder of the literature review focuses
on these applications, and the rest of the paper deals with the airline industry
exclusively.
Botimer (1996) and Belobaba and Wilson (1997) investigate effects of
yield management external to the firm using it. Botimer (1996) presents
arguments for efficiency of yield management pricing in the airline industry,
and Belobaba and Wilson (1997) investigate the impacts of yield manage-
ment introduction in competitive airline markets.
Most yield management research, however, deals with how to actually
maximize revenue. One approach is to assume that customers arrive to re-
quest a flight, state the price they will pay, and then the firm decides whether
or not to serve them. Van Slyke and Young (2000) study this situation in
terms of filling a knapsack of fixed capacity with objects of known weights

(or vector weights, to fill multidimensional knapsacks) and value, where each
type arrives as a time-dependent Poisson process. With the goal of maxi-
mizing value, each object is accepted or rejected at the time of arrival. The
case of equal weights is applied to airline seat inventory control, since each
customer uses one seat, and the multidimensional knapsack is applied to the
problem of allocating seats in a multiple origin and destination flight net-
work. Sawaki (2003) addresses a similar question: customers arrive through-
out the continuous time horizon and state their price and demand size, and
the firm decides whether to accept the request. A semi-Markov decision
process is used, and an optimal policy and its analytical properties are found
when demand arrives as a semi-Markov process.
A more realistic way to treat customer price preferences with respect to the
airline industry, however, is to assume that the customer’s willingness to pay
is unknown when they request a ticket. Gallego and van Ryzin (1994) use
intensity control theory (controlling the demand intensity with price changes)
to study dynamic pricing under imperfect competition (i.e., demand is price-
sensitive) and stochastic demand modeled as a Poisson process. A clo-
sed-form solution is found for exponential demand functions and an
upper-bound on revenue under general demand functions is found using a
deterministic heuristic. Also, monotonicity of the optimal price policy is
shown, as well as asymptotic optimality of a fixed-price heuristic with market
size. These results are then extended to allow compound Poisson demand,
discrete prices, time-dependent demand, overbooking, resupply and cancel-
lations, holding costs and discounting, and variable initial capacity.
Feng and Xiao (2000a) consider the problem from the perspective of
switching between a predetermined finite set of prices at calculated time
thresholds depending on remaining time and stock. Demand is a Poisson
process at each price level, and they find the optimal continuous time
switching policy of an arbitrary number of either price mark-ups or mark-
downs. Feng and Xiao (2000b) extend this to allow reversible price changes.
They show that any subset of prices in the optimal solution is part of the
maximum concave envelope contained in the full set of allowed prices. Feng
and Gallego (2000) also address the discrete price time-threshold problem,
but allow demand to be Markovian, i.e., fares and demand are not only
dependent on remaining time but also on prior sales. Chatwin (2000) ad-
ditionally allows demand (Poisson) to be time sensitive and provides the
option of re-stocking at a unit cost. Zhao and Zheng (2000) find structural
properties of the optimal price-switching policy (from a compact, but not
necessarily finite, price set) when demand is a non-homogenous Poisson
process and investigate solution methods in the case of discrete prices.
Badinelli (2000) considers product-differentiated market segments, each of
which is allowed one advertised price at a time, and formulates an efficiently
computed dynamic programming solution allowing general demand
functions. In his model, customers request a hotel room or flight with a
particular set of attributes and they are given a price quote based on re-
maining time and availability of all relevant commodities.
Rather than dynamically changing prices to maximize revenue, some au-
thors ration capacity with price classes to ensure that high-paying customers
are served, effectively implementing a mark-up policy based on remaining
capacity and, if seat allocation between classes is dynamically controlled,
remaining time. Ladany (1996), assuming deterministic non-linear demand
and a fixed cost for creating each price class, develops a dynamic-program-
ming solution for finding the optimal number of price classes, the optimal
number of capacity units (specifically, hotel rooms) allocated to each price
class, and the optimal price at each class. Bodily and Weatherford (1995)
allow uncertain demand, and study when to curtail low-price sales based on
the probability of spoilage and expected marginal revenue. They don’t allow
re-opening of price classes, but do not assume that all low-fare customers will
arrive before high-fare customers (thereby forcing all high-fare customers to
pay their full acceptable price). Li (2001) again considers deterministic de-
mand, and studies the use of sales restrictions (such as advance purchasing,
or minimum trip duration) for segmenting demand. He develops several key
properties of the optimal restriction policies, and applies those results to
airline pricing, where leisure travelers’ relative price elasticity compared to
business travelers permits the efficient use of restrictions.
Dana (1999a,b) studies aspects of market segmentation other than methods
of optimal allocation. Dana (1999a) studies the phenomenon of equilibrium
price dispersion, showing that the optimal pricing system under uncertain
demand is price dispersed in monopolistic, imperfectly and perfectly compet-
itive markets. He shows that the amount of price dispersion increases with the
level of competition, using this to explain the observation that routes served
by more airlines exhibit more price dispersion. Dana (1999b) demonstrates
how setting multiple price levels on flights at different times can shift demand
from the higher-demand departure time to the alternate flight, even when it is
unknown which time is the peak. He addresses the competitive and monopoly
cases and uses his model to show that high-fare customers may benefit from
price dispersion as well as low-fare customers.
Moving beyond single flight-leg analysis, several authors consider origin–
destination networks. Feng and Xiao (2001) consider a network of multiple
origins, one hub, and one destination. Prices for each flight leg are distinct,
and demand at each origin is a Poisson process. They use a nested fare
structure for inventory control and find the optimal time thresholds for clos-
ing origin–hub flights. They then extend her results to allow multiple fares on
each origin–destination flight and time-dependent demand. Kuyumcu and
Garcia-Diaz (2000) and Garcia-Diaz and Kuyumcu (1997) use a graph-theory
approach for allocating seats in a flight network. Garcia-Diaz and Kuyumcu
(1997) assumes a non-nested seat allocation system, normally distributed
random demand that is independent between price levels, and a fixed number
of fare classes. They develop an algorithm utilizing cutting-planes for
allocating seats throughout the origin–destination network and investigate

computational times. Kuyumcu and Garcia-Diaz (2000) use the same as-
sumptions except that demand on each day of the week may be different, and
airline capacities are considered. They develop three models for the network
pricing and seat allocation problem, the third being a polyhedral graph theory
approach, and a solution procedure that they test computationally. The doc-
toral thesis of Williamson (1992) compares the use of several network-level
seat inventory-control heuristics in simulation. She concludes that considering
network-wide effects may increase revenue by 2–4%, but only when the load
factor is very high. de Boer, Freling, and Nanda (2002) extend her research by
disputing the finding that deterministic approximation methods outperform
probabilistic heuristics, which they claim is due to a difference in fare-class
structure in modeling and simulation.
In these studies of dynamic price discrimination, there is a tension between
the practical and the insightful. Computation-based analyses in principle
could be used to solve real-world problems of optimal pricing, while simpler
theories elucidate potential principles for understanding pricing generally.
A common assumption in the literature posits a discrete price grid, or
even just two prices, which simplifies the problem to one of deciding when
to switch to another price (Feng and Gallego, 2000; Feng and Xiao, 2000a,
2000b; Sawaki, 2003). The assumption of a discrete price grid doesn’t seem
justified, either on theoretical grounds, since there is no economy of using
few prices, or on practical grounds, since airlines and hotels in fact employ a
large set of prices. Thus, the defense of the assumption relies on the sim-
plification of the mathematical problem. The main advantage of the as-
sumption is that proving existence of an optimal solution is trivial, but little
else is gained. An exception is Chatwin (1999), which exploits the finite grid
to develop a nice intuition for when prices rise and fall.
The use of a specific functional form for demand is common. The most
common assumption is exponential demand, q(p) ¼ aebp, for constants a
and b. This form of demand is useful to assume because q0 /q is a constant,
and thus marginal revenue is price minus a constant. As a result, a mon-
opolist would choose to charge a price which is marginal cost plus a con-
stant. This feature of exponential demand makes the solution to the
monopoly problem reduce to the problem of calculating marginal cost, a
simpler though nontrivial problem. Gallego and van Ryzin (1994) is the
best paper of the set of exponential demand papers, and we discuss it
extensively below and indeed ask some additional questions of this theory.
A variety of mechanisms for bringing customers to the seller are considered
in the literature. The most common assumption is a constant Poisson proc-
ess, but possibly with time-varying arrival rates (Zhao and Zheng, 2000). An
interesting variant on the arrival rate process is a Markov-switching model,
which involves a signal extraction problem: detecting from the behavior of
buyers whether demand conditions have changed (Feng and Gallego, 2002).
The connection between the classic peak-load pricing problem and dynamic
pricing is explored in Dana (1999b). This paper concludes that prices should
be increasing on average, because of a correlation between high demand and
high prices. As in Dana’s analysis, several authors posit multiple classes of
customers, who may arrive via distinct processes.
While the most common objective function is to maximize expected rev-
enue, Carr and Lovejoy (2000) consider the alternate assumption of pricing
to sell out. Pricing to sell out is a bad strategy for several reasons. First, a
profit-maximizing firm would not price to sell out. However, more alarm-
ingly, even the efficient solution that maximizes the gains from trade doesn’t
price to sell out.
Most of the literature focuses on the problem of a single flight, treating
competition and alternate flights as fixed, but several authors have made
headway into the problem of multiple flights and routes. Feng and Xiao
(2001) examine a simple Y-shaped pattern through a hub. Network issues
are also examined in de Boer et al. (2002).
Dana (1999a) is the only author to develop a full theory of competition
between dynamically pricing sellers. The theory, involving two firms pricing
over two periods, emphasizes that price dispersion may result from such
competition. This result is also available in static, one-period problems.
An important oversight of the literature is the absence of discounting.
Virtually, the entire literature presumes no discounting. In the hotel con-
text, zero discounting makes sense because even if one books a hotel in
advance, generally payment isn’t made until the time one stays in the hotel,
which implies the same discount factor whether one books the room early
or later. With airline tickets, however, generally payment is made at the
time of booking, not at the time of departure. This matters because the time
intervals are long enough for discounting to be significant, given the tickets
may be booked six months in advance.
3 Dynamic price discrimination with price commitment
The extent of price changes found in actual airline pricing is mysterious

because a monopolist with commitment ability, in a standard framework,
doesn’t want to engage in it at all! To develop this conclusion first proved
by Stokey (1979), we start with a simplified version of her analysis. This
analysis was dramatically extended by Board (2005). The seller sells either a
durable good or a one-time use good like an airplane trip. Time is com-
posed of discrete periods t ¼ 1, 2, y and buyers and the seller have a
common discount factor d. There are a continuum of potential buyers who
have values represented by a demand q, so that q(p) gives the measure of
buyers willing to pay p. We assume that a consumer’s demand persists until
the consumer purchases. The monopolist chooses a price sequence p1, p2, y
which can be taken to be non-increasing without loss of generality. A
consumer with a value v will prefer time t to time t+1 if

v pt 4dðv ptþ1 Þ. (1)
The equations producing indifference, vpt ¼ d(vpt+1), define a sequence
of critical values vt that make the buyer indifferent between purchasing at t
and purchasing at t+1. Note that the incentive constraint on buyers shows
that, if a buyer with value v chooses to buy before time t, then all buyers
with values exceeding v have also purchased by this time.
vt pt ¼ dðvt ptþ1 Þ. (2)
This set of equations can be solved for pt in terms of the critical values

pt ¼ ð1 dÞvt þ dptþ1 ¼ ð1 dÞvt þ d ð1 dÞvtþ1 þ dptþ2 ¼
(3)
X
1
¼ ð1 dÞ d j vtþj
j¼0
The monopolist sells q(vt)q(vt1) in period t, where v0 is defined so that

q(v0) ¼ 0. The monopolist’s profits are
!
X1 X
1 X
1
p¼ dt1 pt qt ¼ dt1 ð1 dÞ dj vtþj ðqðvt Þ qðvt1 ÞÞ
t¼1 t¼1 j¼0
" ! !#
X
1 X
1 X
1 X
1
t1 j t1 j
¼ ð1 dÞ qðvt Þd d vtþj qðvt1 Þd d vtþj
t¼1 j¼0 t¼1 j¼0
" ! !#
X
1 X
1 X
1 X
1
¼ ð1 dÞ qðvt Þdt1 dj vtþj qðvt1 Þdt1 dj vtþj
t¼1 j¼0 t¼2 j¼0
" ! !#
X
1 X
1 X
1 X
1
¼ ð1 dÞ qðvt Þdt1 dj vtþj qðvt Þdt dj vtþ1þj
t¼1 j¼0 t¼1 j¼0
" !#
X
1 X
1 X
1
¼ ð1 dÞ qðvt Þdt1 dj vtþj d dj1 vtþj
t¼1 j¼0 j¼1
X
1 X
1
¼ ð1 dÞ qðvt Þdt1 vt ¼ ð1 dÞ dt1 qðvt Þvt ð4Þ
t¼1 t¼1
Thus, the optimum level of vt is constant at the one-shot profit

maximizing level, which returns the profits associated with a static
monopoly. The ability to dynamically discriminate does not increase the
ability of the monopolist to extract rents from the buyers.
There is an important lesson to be drawn from Stokey’s theorem. If

dynamic price discrimination is playing a role, it is because of new
customers arriving, rather than an attempt to extract more profits from an
existing set of customers by threatening low-value customers with delayed
purchases. That is, dynamic price discrimination is driven by customer
dynamics rather than price discrimination over an existing set of customers.
4 Continuous time theory
Gallego and van Ryzin (1994) produced a closed form model of discrete
time optimal pricing. This model is very useful for its tractability, and we
reproduce some of their analysis here, as well as extend it by generating
predictions about the average path of prices.
Let l be the arrival probability of customers per unit of time, assumed
constant. A constant arrival can be assumed without loss of generality by
indexing time in terms of customer arrivals. Time will start at 0 and end at T.
If not sold, the product perishes at T, which might occur because a flight
takes off or the product is time-dated like a hotel room or time-share
condominium, where what is for sale is the use of the product on date T. For
technical reasons, no discounting is considered in this section. However, in
some applications, no discounting is the right assumption. For example,
hotels generally do not charge until the customer arrives, independently of
the time the room is booked, a situation which corresponds to no
discounting based on late booking. The marginal cost of the product is c,
a value which might include normal marginal costs (cleaning a hotel room, a
meal served on an airline), but could also include lost business—the
customer that takes a particular flight is less likely to take an alternative
flight by the same airline. Potential customers demand a single unit, and
their willingness to pay is given by a cumulative distribution function F.
The value of having n items for sale at time t is denoted by vn(t). Clearly
having nothing to sell conveys zero value. Moreover, if not sold by T, an
inventory of items also has zero-value, yielding
v0 ðtÞ ¼ vn ðTÞ ¼ 0. (5)
Consider a small increment of time, D, beyond a current time t. With
probability 1lD, no customer arrives, so that the current value becomes
vn(t+D). Alternatively, with probability lD, a customer arrives and the firm
either makes a sale or does not. For price p, the sale occurs with probability
1F(p). When a sale occurs, the value becomes pc+vn1(t+D), because
the inventory is decreased by one. Summarizing
vn ðtÞ ¼ max ð1 lDÞvn ðt þ DÞ
p
þ lDðð1 F ðpÞÞðp c þ vn1 ðt þ DÞÞ

þ F ðpÞvn ðt þ DÞ ð6Þ
or
vn ðtÞ vn ðt þ DÞ ¼ lD maxð1 F ðpÞÞðp c þ vn1 ðt þ DÞ
p
vn ðt þ DÞÞ. ð7Þ
Therefore, dividing by D and sending D to zero,
v0n ðtÞ ¼ l maxð1 FðpÞÞð p c þ vn1 ðtÞ vn ðtÞÞ. (8)
p
The expression for v0n ðtÞ is composed of two terms. First, there are profits
from a sale, pc. Second, there is the lost option of selling the unit in the
future, an option that has value vn ðtÞ vn1 ðtÞ: It is not possible to solve this
differential equation for an arbitrary demand function F. However, with a
convenient choice of F, it is possible to provide an explicit solution. Let
FðpÞ ¼ 1 eap . (9)
Note that (1eap)(pmc) is maximized at p ¼ 1/a+mc.2 Then
1
pn ðtÞ ¼ þ c þ vn ðtÞ vn1 ðtÞ, (10)
a
and
1
v0n ðtÞ ¼ leaðð1=aÞþcþvn ðtÞvn1 ðtÞÞ . (11)
a
The multiplicative constant b ¼ le1ac represents the arrival rate of buyers
willing to pay the static monopoly price 1/a+c. Thus, at time t, the
expected number of buyers willing to pay the monopoly price is b(Tt).
This observation helps explain why
Xn
ðbðT tÞÞj
Bn ðtÞ ¼ (12)
j¼0
j!
will appear in the solution. The first result characterizes the value function
and prices in a closed-form manner.
Lemma 1. (Gallego and van Ryzin, 1994): vn ðtÞ ¼ ð1=aÞ log ðBn ðtÞÞ and
pn ðtÞ ¼ ð1=aÞ log ðlBn ðtÞ=bBn1 ðtÞÞ:
At time zero, suppose there is an initial capacity k. Let qi(t) be the
probability that there are i units left for sale at time t.
Theorem 2. qn ðtÞ ¼ ððbtÞknP
Bn ðtÞÞ=ððk nÞ!Bk ð0ÞÞ: The expected number
of seats sold is Eðk nÞ ¼ kn¼0 ðk nÞqn ðtÞ ¼ btBBk1 ð0Þ
k ð0Þ
2
Let hðpÞ ¼ ð1 eap Þðp mcÞ: h0 (p) ¼ eap (1a(pc)) h0 (p) ¼ 0 implies h0 (p)r0. Thus, every
extreme point is a maximum, and so if there is an extreme point, it is the global maximum. Moreover,
p* is an extreme point.
Proofs are found in the appendix. Lemma 1 and Theorem 2 give a

complete, closed form description of profits, prices, and sales for the
dynamic monopolist practicing yield management. For example, the
probability that there is no available capacity at time T is
ðbTÞk =k!
q0 ðTÞ ¼ Pk j
. (13)
j¼0 ðbTÞ =j!
This formula ensures that, with sufficient time and a given capacity k, the
flight sells out, because
lim q0 ðTÞ ¼ 1. (14)
T!1
How does a thick market affect the outcome? To model this, consider
increasing both the capacity k and the arrival rate of customers, l,
proportionally. Let g ¼ bT/k, so that
!1 !1
Xk
k!ðbTÞjk X k
k!ðgkÞjk
q0 ðTÞ ¼ ¼
j¼0
j! j¼0
j!
( g1
g if g41
! ð15Þ
k!1 0 if g 1
What is interesting about this expression is that the probability of selling all
the capacity, q0(T), converges to the same level as would arise if the price
was just constant at the monopoly price 1/a+c. Since the price exceeds this
level always, because the option value is positive, the price must get very
close to the static monopoly price most of the time in order for the limiting
probabilities to coincide. Per unit profits of the dynamically discriminating
firm are
!
vk ð0Þ 1 1 Xk
ðgkÞj
¼ LogðBk ð0ÞÞ ¼ Log
k ak ak j¼0
j!
(
1 g if go1
! ð16Þ
k!1 a 1 þ LogðgÞ if g 1
As an alternative to yield management and dynamic price discrimination,

consider a firm that sets a price and leaves it fixed. Such a firm, a one-price
monopolist, will earn lower profits than the yield-management firm. Is the
profit reduction significant?
A monopolist who only offers one price p will have a flow-rate of sales of
m ¼ l(1F(p)). If Zi(t) is the number of future sales given capacity i at time t,
5
p1 (t )
p2(t )
p3(t ) Ep(t)
4
3
p1(t )
2 p10(t )
1
50 100 150 200 250 300 350
Fig. 1. Prices, and expected price, k ¼ 10, l ¼ 1/2, a ¼ 1, j ¼ 0, T ¼ 365.
then Z0(t) ¼ 0 and

Z T
Zj ðtÞ ¼ m emðstÞ 1 þ Zj1 ðsÞ ds. (17)
t
Pk1 ðmðTtÞÞj
Lemma 3. Zj ðtÞ ¼ k emðTtÞ i¼0 ðk jÞ j!
The profits associated with a single price can be numerically maximized.
Figure 1 provides an example of dynamic price discrimination, starting
with 10 units for sale. The parameters include zero marginal cost, a 365 day
period with an average of one customer every other day, and demand in the
form 1F(p) ¼ ep. The prices associated with various numbers of units to
be sold are illustrated, for example, p3 shows the optimal price associated
with three units to be sold. The average price, conditional on units
remaining for sale, is denoted by Ep; this represents the average of posted
prices. Note that units may not be available for sale, so that the expected
price is a bit of a fiction, although the comparison to the ‘‘one price
monopolist’’ is sensible since that price may also not be offered. The
optimal constant price is denoted in p1. Profits are 5.45% higher under
dynamic price discrimination than they are with a constant price. In this
figure, the monopoly price is the horizontal axis, at 1, and either scheme
improves substantially on the static monopoly price.
Figure 1 illustrates a common feature of simulations: the expected price
rises, then falls. The forces involved are as follows. First, prices must
eventually fall because there is a positive option value prior to time T, and
this option value is zero at time T. Thus, prior to T, prices are strictly larger
than the static monopoly price and converge to the static monopoly price at
0.8
0.6
0.4
0.2
50 100 150 200 250 300 350
Fig. 2. The probability of zero capacity.
T. While it is possible that prices fall for the entire time interval, they may
initially rise because early sales, by reducing available capacity, drive up the
prices.
Figure 2 illustrates the probability that all units sell under dynamic price
discrimination. This converges to slightly over 85.35% by the end of the
period.
Is dynamic price discrimination profitable when k and l are both very
large? Let y represent l/k. In this case, the solution for demand is
Zj ðtÞ
! minfyeap T; 1g (18)
k k!1
This is maximized at the static monopoly price, (1/a)+c, provided
g ¼ ye1acTo1. Otherwise, the optimal price satisfies yeapT ¼ 1. It fol-
lows that profits are
(
1 g if go1
a 1 þ LogðgÞ if g 1
and agree with discriminatory profits in the limit for large k. That is, per
unit gain in profits of dynamic price discrimination over an optimally cho-
sen constant price converges to zero, although the total gain will still be
positive. This happens because most sales take place at an approximately
constant price; dynamic price discrimination is advantageous only as the
probability of actually selling out changes, for a relatively small portion of
the very large number of sales. One can reasonably interpret these results to
say that dynamic price discrimination only matters on the last 20 or so
sales, so when a large number of units are sold, dynamic price discrimi-
nation doesn’t matter very much.
Dynamic price discrimination has a relatively modest effect when there
are 100 or more seats available. The kinds of profits predicted, for reason-
able parameter values, under dynamic price discrimination are not very
large, less than 1%, when compared to an appropriately set constant price.
An important aspect of this conclusion is that dynamic price discrimination
does not appear to account for the kinds of value claimed for American
Airlines’ use of yield management.
5 Efficiency in the Gallego and van Ryzin model
An efficient solution in this model has the property that the value func-
tion maximizes the gains from trade rather than the profit. The value
function, then, satisfies
vn ðtÞ ¼ max ð1 lDÞvn ðt þ lDÞ þ lD ðð1 F ðpÞÞ ðCðpÞ
p
þ vn1 ðt þ lDÞÞ þ FðpÞvn ðt þ lDÞÞ ð19Þ

where C(p) is the consumer’s value, plus seller profit, conditional on the
consumer’s value exceeding p. In this model,
Z 1
ð1 F ðpÞÞCðpÞ ¼ 1 F ðxÞdx þ ðp cÞð1 F ðpÞÞ
p
Z 1
¼ eax dx þ ðp cÞeap ¼ ðp þ ð1=aÞ cÞeap : ð20Þ
p
Thus, the efficient solution is the solution a monopoly whose costs are
reduced by 1/a, the static monopoly profit, would choose. The assumed
structure of demand insures that all the qualitative conclusions drawn apply
equally to efficiency as to monopoly. In particular, the shape of the time
path of prices and the conclusion that for large problems the gains of
dynamic price discrimination are small apply equally to efficient solutions
as they do to monopoly solutions.
In our view, the Gallego & van Ryzin model is not a very useful instru-
ment for examining efficiency, because of the similarity of the monopoly
price and efficient price. Reducing a monopolist’s marginal costs by the
monopolist’s markup does not produce the efficient price, but a higher price
under the standard regularity conditions like log concavity. However, it is
worth emphasizing that the efficient solution and the monopoly solution
lead to the same set of price paths, so that if costs are unobservable, the two
models are observationally equivalent.
6 Efficiently allocating limited capacity under uncertainty
The closed form of the Gallego & van Ryzin model facilitates compu-
tation, but obscures economic considerations because of the nature of the
demand. In this section, we consider efficiently solving the allocation prob-
lem, rather than maximizing profit. Efficiency is not unreasonable, since
airlines face competition and a perfectly competitive model would suggest

the efficient allocation of the available quantity. It seems likely that neither
the monopoly nor the fully efficient solution accurately represent the real
world, which probably falls somewhere in between the two extremes.
Borenstein and Nancy (1994) is perhaps the foremost example of an at-
tempt to assess airline price competition, but at the level of average fares.
Developing a theory that permits assessing the degree of competition at the
detailed level of the price path, in contrast, presents a formidable challenge.
Observations of quantities are going to be central to the assessment of the
degree of competition, because price observations alone are not sufficient to
identify the model.
This kind of efficient allocation problem was first studied in the context of
electricity generation. Costs of capacity induce a peak-load problem (Boit-
eaux, 1949; Williamson, 1966), so named because the marginal cost of
electricity generation is much higher when the plant capacity is reached
below that level, so that the peak load matters a lot to the economics of
electricity generation. The standard assumption in peak-load problems is
that demands in each period are independent, and that the full supply is
available in each period. The peak-load problem is really a capacity plan-
ning problem, where the needed capacity depends on prices. For a
two period model, let qi be the demand in period i. The firm’s profits are
given by
p ¼ p1 q1 þ p2 q2 b maxfq1 ; q2 g mcðq1 þ q2 Þ, (21)
where b is the marginal cost of capacity and mc is the marginal cost of
production.
In the airline and hotel context, the standard peak-load model is poorly
suited, because capacity sold in one period isn’t available in the subsequent
period. Thus, the standard model of the peak-load problem asks how large
a plane that flies round-trips between Dallas and Chicago should be, given
variation in total demand from day-to-day or from week-to-week. While
this problem represents an important logistical problem, it has little or
nothing to do with dynamic pricing. In contrast, airlines also face the
problem that seats on a particular flight can’t be occupied by two passen-
gers, and this problem of allocating limited capacity suggests a quite distinct
peak-load model. Moreover, the fact that future demand isn’t known at the
time of contracting in the first period is a crucial aspect of the peak-load
problem facing airlines and hotels. That is, airlines and hotels contract for
initial sales, not knowing the realization of overall demand.
We introduce a new model of random arrival that has some attractive
properties. The randomness comes in the form of random variables nt in
period t, with period t demand being, then, nt q(p) for price p. We also
assume that nt is observable at the beginning of period t, so that the firm can
either set price or quantity in that period; what is uncertain is the future
demand, not the current demand.3 We will focus on the constant elasticity
of demand case, q(p) ¼ ape, because this is a standard empirical model and
assists in tractability. We will refer to q(p) as per capita sales, and nt q(p) as
total sales, but this interpretation isn’t exact in the airline context. In the
airline context, if nt is the number of customers, q(p) would have to be the
probability of sale, in which case it should be no greater than one, which is
inconsistent with the constant elasticity assumption. At best, the model
represents an approximation for the airline context.
This section considers only the efficient allocation. The efficient alloca-
tion is interesting for several reasons. First, competition with other pro-
viders will generally push firms’ product offerings toward efficiency and
away from monopoly, so with vigorous competition, efficiency maximizat-
ion is probably a better model than monopoly profit maximization. Second,
the monopoly model has been better studied than the competitive, efficient
model. Third, some of the interesting behavior in the monopoly model does
not arise from price discrimination but because of the dictates of efficiency.
The source of the behavior is challenging to see without studying the effi-
cient allocation. That is, what appeared to be price discrimination is merely
variations in marginal cost.
The seller has a capacity K. We will focus on the two-period case
throughout this section, although this readily generalizes to more periods.
If per capita sales are s1 of the n1 first period demand, the number of seats
available for sale in the second period is Ks1n1. These are valued by
consumers at a price p2 satisfying
K s1 n1 ¼ n2 qðp2 Þ (22)
or

1 K s 1 n1
p2 ¼ q . (23)
n2
The cost of first period sales rises quite dramatically in the number of
sales. Suppose q has constant elasticity of demand, q(p) ¼ ape. We assume
e>1 so that the consumer surplus is finite. For any given n2,
1=
an2
p2 ¼ . (24)
K s1 n1
The price p1 that clears the market in period 1 satisfies s1 ¼ q(p1). The
customer’s value of quantity q is ða=ð 1ÞÞqð1Þ= : Thus the overall gains
3
Dana’s 1998 model is a special case, with q taking on two positive values, interpreted as a willingness
to pay for leisure and business travelers.
from trade are

a ð1Þ= n o
W¼ n1 s1 þ E n2 qðp2 Þð1Þ=
1 ( ð1Þ= )!
a ð1Þ= K s1 n 1
¼ n1 s 1 þ E n2
1 n2
a n o
ð1Þ= 1=
¼ n1 s1 þ E n2 ðK s1 n1 Þð1Þ= . ð25Þ
1
The gains from trade are maximized at
K
s1 ¼ n o . (26)
1=
n1 þ E n2
The per capita second-period sales are

8 n o 9
>
< K E n2
1= >
=
K s1 n1
Es2 ¼ E ¼E n o
n2 >
:n2 n1 þ E n1= >
;
2
8 n 1= o 9
< E n2 =
¼ s1 E ð27Þ
: n2 ;
Because pi ¼ ðsi =aÞ1= ; it is a rudimentary calculation to show that the

expected price in the first period and second period coincide. Equality of the
prices is a necessary condition for maximizing the gains from trade; oth-
erwise it would increase the gains from trade to allocate more of the seats to
the high-priced time.
Some seats are misallocated, because contracting with the early customers
occurs prior to knowing the realization of number of later customers.
Moreover, it is worth emphasizing that this misallocation is efficient—
maximizes the gains from trade—and not a consequence of monopoly
pricing. In our terminology, it is dynamic pricing but not dynamic price
discrimination. How large is the effect? The average reduction in the share
of period two customers satisfies
8 n 1= o 9
Es2 < E n2 =
¼E . (28)
s1 : n2 ;
Theorem 4. The second period share is larger on average than the first
period share, that is,
8 n 1= o 9
Es2 < E n2 =
¼E 1.
s1 : n2 ;
Theorem 4 shows that the share of customers served in the second period
is at least as great as are served in the first period. This is a consequence of
the possibility of small denominators leading to large shares; when few
customers arrive, the price is set low enough to insure the plane sells out. The
sales as a fraction of the average value of n2 are less than in period 1. That is,
n o
1=
Es2 n2 E n 2
¼ 1.
s1 En2 En2
Insight into the magnitude of this efficient
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi misallocation is given by the next
result. Let CV ¼ Varðn2 Þ=ðEn2 Þ2 be the coefficient of variation of n2.
1=
ðEfn2 gÞ
Theorem 5. 1 ð1 þ CV 2 Þ1 :
En2
The share of seats allocated to the first period may deviate from the
efficient equal share by as much as (1+CV2)1e; for the case of e ¼ 2 and
CV ¼ 1, this could be 50%. The bound in Theorem 5 is exact, in the sense
that for a binomial random variable with one value equal to zero, the right
inequality holds with equality, and at e ¼ 1, the left inequality holds with
equality.
The gains from trade generated are

0 n o 11=
1=
aK @ 1n þ E n2
W¼ A
1 K
aK ð1Þ= n
1=
o 1=
¼ n1 þ E n2 .
1
The firm’s revenue is
0 n o 11=
1=
s 1= n1 þ E n2
1
p1 K ¼ K ¼ a1= K @ A
a K
n o 1=
1=
¼ a1= K ð1Þ= n1 þ E n2
Thus, firm revenue is proportional to welfare under constant elasticity of

demand, and the analysis of revenue maximization (under efficiency, of
course) analyzes welfare maximization. To perform this analysis, we turn to
the log normal case.
The inefficiency in this model arises because contracting with early cus-
tomers occurs prior to the time that period two demand is realized. How
large is this inefficiency? To answer that question, we consider the case
where ni is log-normally distributed. In the log normal distribution case, the
welfare losses associated with the missing market can be very large.
7 The log normal case
If the distribution of n2 is log-normal, so that the log(n2) is normally

distributed with mean m and variance s2, many of the expressions have
closed forms, greatly facilitating computation. In particular,
2
n o 2 2
2
1=
En2 ¼ emþð1=2Þs and E n2 ¼ eðm=Þþðs =2 Þ ¼ emþðs =2Þ
(29)
The coefficient of variation for n is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
VarðnÞ En2 ðEnÞ2
CV ¼ ¼
ðEnÞ2 ðEnÞ2
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
En2 e2mþ2s2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
¼ 1 ¼ 1 ¼ es2 1. ð30Þ
ðEnÞ2 ðemþð1=2Þs2 Þ2
Thus,
n o
1=
E n2 2 ð1Þ=2
¼ es ð1Þ=2 ¼ 1 þ CV 2 ; and (31)
En2
8 n 1= o 9
< E n2 =
¼ es ð1þÞ=2 ¼ ð1 þ CV 2 Þð1þÞ=2
2
E (32)
: n2 ;
The expected gains from trade are
aK ð1=Þ n
1=
o 1=
W¼ n1 þ E n2
1
aK ð1=Þ 2
1=
¼ n1 þ emþðs =2Þ ð33Þ
1
ε
2 4 6 8 10
.95
σ2=3
0.9
σ2=5
σ2=4
.85
0.8
ΔW
Fig. 3. Efficiency loss relative to simultaneous contracting. Parameters: m1 ¼ 5, m2 ¼ 0.
How much does sequential contracting cost in terms of revenue or welfare?

We treat n1 at the expected value of a lognormal with the same variance but
different mean than n2. The proportion of the gains from trade preserved
through contracting over two periods is denoted %DW. It has the value
n o 1=
1=
n1 þ E n2
%DW ¼ n o . (34)
1=
E ðn1 þ n2 Þ
We show by example that %DW can be as low as 64%, that is, the inability
to contract in advance reduces welfare by a third relative to simultaneous
contracting.4
This example has a much larger future population of demand than first
period demand. However, the variance of demand is large too, and this
causes the losses associated with uncertainty to be very large. In particular,
1=
ðEfn2 gÞ is a negligible fraction of E{n2} in this example, a difference which
accounts for most of the efficiency loss. Even for more reasonable param-
eters, losses can be 15% or more. An example, with up to 20% losses, is
illustrated in Fig. 3. In these examples, the n1 takes on the values 0.6, 20 and
1808, while the expected value of n2 takes on the values 90, 2981 and
268337.
The main conclusion is that the important effects in dynamic pricing arise
not from an attempt to extract more money from the consumer, but from
4
Specifically, one set of parameters is e ¼ 3.474, s2 ¼ 11.651, m ¼ 45.153, and n1 ¼ 1:28 1028 : This
1=
set of parameters gives the following expected values for n2. Efn2 g ¼ 1:22 1049 ; and ðEfn2 gÞ ¼
28
1:24 10 :
incomplete markets, and in particular from impossibility of simultaneous

contracting with all potential buyers. Dynamic pricing is used primarily
to mitigate the costs of these missing markets. Moreover, welfare costs of
the missing markets are potentially quite large. This conclusion suggests
that options, which create markets for advance contracting, are an impor-
tant aspect of both maximizing revenue and of efficiently allocating
resources.
8 Options and interruptible sales
The welfare losses identified in the previous section arise because of in-
complete markets. In particular, it is not possible to contract with the
period 2 agents at the time that the period 1 market clears. This lack of
contracting leads to inefficiency because sometimes too many seats are sold
in the first period, when the demand in the second period is unusually high,
while sometimes too few are sold, because demand was unexpectedly low.
A solution to this problem generally is to sell contingent seats, that is, to sell
an option.
In this case, the main value of the option is to permit the seller to sell the
seat to a higher value buyer in the second period. This kind of option is little
used in hotels and airlines, but is quite common in pipeline transport where
it goes by the name of interruptible transmission rights, as opposed to firm
(guaranteed) rights. Priceline.com sold a kind of interruptible service, where
they sold the ticket well in advance but didn’t specify the time of the flight
until a day or so in advance.
To gain some insight into the sale of interruptible service, consider first
selling two qualities of service, j1>j2, where ji is the probability of serv-
ice, and these are sold at prices p1>p2. A consumer with type v values good
i at ji v pi : A type v consumer prefers type 1 if
j1 v p1 j2 v p2 , (35)
p1 p2
or v (36)
j1 j2
In addition, a consumer prefers good i to nothing at all if ji v pi 0: If

ðp1 =j1 Þ ðp2 =j2 Þ; then no consumer ever buys good 2, so we impose the
condition, without loss of generality, that ðp1 =j1 Þ ðp2 =j2 Þ: Let F be the
cumulative distribution of values. The demand for good 1 is 1 F ððp1 p2 Þ
=ðj1 j2 ÞÞ and the demand for good 2 is F ððp1 p2 Þ=ðj1 j2 ÞÞ
F ðp2 =j2 Þ:
Let ci be the marginal cost of service type i. The seller’s profits are

p p2
p ¼ Max p1 c1 1 F 1
p1 ;p2 j1 j2

p1 p2 p
þ p2 c2 F F 2 ð37Þ
j1 j2 j2
It is helpful to introduce the notation RðxÞ ¼ Maxp ðp xÞð1 F ðpÞÞ; which
is the profit maximization associated with one good. Let p ðxÞ ¼
arg Maxp ðp xÞð1 F ðpÞÞ be the solution to the one good maximization
problem. Marginal revenue is decreasing if p is an increasing function, so
that an increase in cost reduces the quantity and increases the price charged
by the monopolist. This is a standard assumption.
Theorem 6. Suppose marginal revenue is decreasing. A profit-maximizing
monopolist sells the low-quality good 2 if and only if ðc1 =j1 Þ ðc2 =j2 Þ; in
which case

c1 c2 c2
p ¼ j1 j2 R þ j2 R ,
j1 j2 j2

c1 c2 c2
p1 ¼ j1 j2 p þ j2 p
j1 j2 j2
and p2 ¼ j2 p ðc2 =j2 Þ
Otherwise; p ¼ j1 Rðc1 =j1 Þ and p1 ¼ j1 p ðc1 =j1 Þ:

The interruptible good problem breaks up into two separate maximizat-
ion problems, one for the low-quality good, and one for the difference of
the low-quality good and the high-quality good. This type of result is dis-
tinct from the usual result where a monopolist selling to two types of
consumers offers two qualities; here the monopolist is selling two goods to a
continuum of consumers. However, the present result is reminiscent of
Mussa and Rosen’s 1978 analysis.
Consider a seller that didn’t sell good 2, and then begins selling good 2.
Does that seller’s price rise, or fall? The price falls if

c1 c2 c2 c1
j1 j2 p þ j2 p oj1 p (38)
j1 j2 j2 j1
Thus, if p is concave, the price falls when the second good is introduced,
and the price rises when p is convex. In the case of uniform distributions of
values, or exponential distributions, or constant elasticity, p is linear in
cost, and thus introduction of the second good doesn’t change the price p1.
More generally, it may increase or decrease.
In the airline context, the cost of a sale is generally a shadow cost—how

much revenue is foregone because the seat is taken? Consequently, the
condition ðc1 =j1 Þ ðc2 =j2 Þ is automatically satisfied. The cost of a prob-
ability of service j1 entails a loss in flexibility relative to the service level j2.
That loss in flexibility means not rescheduling the service in circumstances
in which it would be desirable to do so. Thus, the cost c1 should be an
average of c2 over the flights when the interruptible service is offered as
scheduled, plus the cost in circumstances where it was profitable to inter-
rupt. Since the latter should exceed the former given optimal interruptions,
the cost of service for case 1 should exceed the cost in case 2, which means it
is always optimal to offer the interruptible service.
As a practical matter, the ‘‘right’’ interruptible service is probably a ticket
that lets the airline choose the time the passenger flies, but sets the flight in a
day. The need to coordinate air travel with hotels and other destination
activities restrict the ability of the most passengers to be flexible over mul-
tiple days. Nevertheless, flexibility with respect to the time of travel po-
tentially produces significant gains in the expected value of air travel, both
in the ability of the price system to identify and allocate seats to the high-
value passengers, and the ability to utilize more capacity.
The gains from trading options are larger when the firm posts a price
that is good for an appreciable amount of time. This phenomenon was
studied by Dana (1999a). Random demand will generally create a misal-
location, which we modeled by allocating first period seats without yet
knowing second period demand. However, under a posted price, seats will
be misallocated because the posted price will generally not be the correct
price to clear the market, either creating a surplus or a shortage. Consider
the model developed in the previous section, with a random demand in
the form nq(p) and constant elasticity q(p) ¼ ape. Let K be the capacity
and suppose that when capacity binds, the customers are a random selec-
tion of those willing to purchase. The gains from trade under a constant
price are
Z qðpÞ x1= dx
1
W ¼ Min K; nqðpÞ ¼ MinfKp ; ang
0 a qðpÞ 1 p
(39)
This equation arises because, when Konq(p), the capacity constraint binds,
and the gains from trade are the average of the value of seats over the set of
consumers willing to buy those seats. One interesting feature of this equa-
tion is that efficiency requires a positive probability of empty seats. If there
is a zero probability of empty seats, then the price is lower than the market-
clearing price. Low prices create random assignments when demand is
large, which entails a loss of efficiency; this loss is balanced against the extra
seats sold when demand is low. Indeed, the first order condition for
maximizing W can be expressed as
2
W¼ Kp1 Pr K nqð pÞ (40)
1
Thus, the probability that capacity binds, Krnq(p), is positive at the
price maximizing the gains from trade.
9 Actual airline pricing patterns
In order to assess the extent of dynamic price discrimination, we collected

data on four city pairs:
Los Angeles (LAX, BUR, SNA) to Las Vegas (LAS)
Bay Area (SFO, SJC, OAK) to Portland (PDX)
Dallas (DFW) to Oklahoma City (OKC)
Dallas (DFW) to Tulsa, OK (TUL)
We collected data on price offers from Orbitz, Travelocity and for part of
the period, from AA.com. Unfortunately, Southwest Airlines is not fea-
tured on any of these sources. Of these nine airport pairs, four (LAX and
BUR to LAS, SJC and OAK to PDX) are also served by Southwest. One
would expect Southwest to exert a moderating influence, given its relatively
low variance in prices, although Southwest has introduced a six fare struc-
ture, which may result in more dispersion rather than less.5 In any case,
with the data available, the effects of Southwest’s presence will have to
remain a topic for further research. Purchasing an American Airlines ticket
on AA.com is generally $5 cheaper, but showed no other difference. The
figures below only show fares quoted by Orbitz, which had the most ex-
tensive offerings.
We collected data on flights, all in the year 2004, departing on 8/26, 8/28,
8/29, 9/09, 9/11, 9/12, 9/23, 9/25, 9/26, 10/7, 10/9, 10/10, 10/21, 10/23, and
10/24. Six airlines represent over 99% of the data, with the following pro-
portions of the flights. Average prices and the frequency of airlines are
presented in Table 1.
Table 1
Average prices in the data
Airline AA Alaska Delta United AmWest USAir NW
Proportion of flights (%) 25.5 22.7 6.1 14.6 8.9 16.1 6.1
Average price $108.05 $126.59 $79.03 $44.06 $66.17 $239.01 $191.48
5
The six classes, with sample fares (BUR to LAS) are Refundable Anytime ($94), Restricted ($86),
Advance Purchase ($72), Fun Fares ($49), Promotional ($39), and Internet One-way ($39).
There are five major propositions that can be tested with the data.
1. Prices fall as takeoff approaches
2. Prices are rising initially
3. Competition reduces variance of prices
4. Prices change with each seat sold
5. Prices of substitutes are correlated
a. Substitute times/days
b. Substitute airports.
The first proposition is that prices are falling as takeoff approaches. This
is a remarkably robust prediction of theories that have identical customer
types arriving over time. The robustness follows from the following obser-
vation. The cost of selling a seat is composed of three components. First,
there is the cost of service. Second, there is the cost of not selling that seat
on a substitute flight. This second component includes any psychological
effects on consumer expectations—consumers may reason that current low
prices are indicators of likely low future prices, which tends to make de-
mand more elastic, to the airline’s detriment. Third, there is the foregone
option of selling the seat later on the same flight. The first two costs are
roughly constant over time, and certainly approximately constant as the
flight approaches. The third cost, however, goes to zero as the flight ap-
proaches. Thus, most theories will suggest that prices fall in the last mo-
ments before a flight, not necessarily to zero, but to some optimal price
reflecting the market power of the airline and the effect of low-price sales on
sales of future flights. But the price should still fall.
A more specialized prediction is that the average price rises initially. This
is a feature of the Gallego–van Ryzin model already discussed, and others.
It is more intuitive in models in which consumers can search than in the
more common model where consumers who don’t purchase disappear; if
prices are falling on average, consumers will tend to wait to purchase. Late
falling prices aren’t so obviously exploitable because of the significant like-
lihood of the plane filling up. Thus, consumers must weigh the expected
gains from delay against the costs of failing to acquire a seat at all; the latter
cost is negligible with a sufficiently long time horizon.
To assess these propositions, we ran the following regression to predict
the price. We have dummy variables for days with one, two, three, and four
weeks remaining, airline-specific dummies (American omitted), hour of
takeoff (overnight omitted) and city-pair identifiers (LAX to LAS omitted).
We omitted Burbank, Long Beach and SNA airports based on few flights
over the period. We only examined one-way economy fares, and have
138371 observations. The 38 variables account for 73.5% of the variation.
Prices rise $50 in the week before takeoff, which is on top of a rise of
$28.20 the previous week. The penultimate fortnight before takeoff ac-
counts for $16 in price increases. There is a slight, 1.9 cent per day increase
prior to that. Thus, the main time-path prediction of the standard theory
appears to be empirically false. These effects are quite significant, thanks to

the size of the database. (We have not attempted to control for correlation
in observations, so that the t-statistics, which treat observations as inde-
pendent, likely overstate the significance.) The important conclusion is that
changing demand is a salient feature of the data, and models that assume
that the shape of demand is constant over time are empirically invalid.
A few other remarks follow from the regressions. Although American
Airlines is a low-priced airline in the data, it is high-priced adjusting for
fixed effects, though not as expensive as US Air. Prices are highest mid-day
and early evening. LAX is a relatively inexpensive airport, while San Fran-
cisco is high-priced. This suggests that competition and thick markets do
not provide a good account of pricing, since San Francisco has a large
number of flights. Southwest’s presence at LAX, OAK, and SJC might help
explain the average pricing. Note, however, that airport dummies are also
accounting for the length of flights, so comparisons should only be made
among distinct airport pairs for a given city-pair (Table 2).
In another paper (Carbonneau, R.P.McAfee, Mialon, & Mialon, 2004), it
is shown that the third proposition is not supported in a large dataset of
airline prices. Indeed, more competition was weakly correlated with more
dispersion, not less. This finding is consistent with Borenstein and Rose’s
1994 findings.
Some theories posit a ‘‘two price’’ structure, others permit prices to vary
continuously. In the data, some airlines clearly lean toward a two-price
structure. For example, American Airlines’ prices on OAK to PDX show
evidence of two main prices, with a third just prior to takeoff. This is
illustrated in Fig. 4.6 Most of Northwest’s fares show just two prices.
However, other flights show evidence of almost continuous adjustment,
and indeed sometimes American appears almost to randomize its offers.
Fig. 5 shows an example of two American Airlines flights, AA 1038 on Sept
23 and Sept 25, with remarkable variation.
So the evidence in favor of this proposition—continuous adjustment of
prices—is mixed. Why do airlines use two prices? A standard economic
explanation is that there is a value in price commitment to assist in con-
sumer planning. This is the reason given for restaurants, movies, and the
like to maintain a constant price, or two prices, in the face of predictably
fluctuating demand. However, prices don’t seem nearly predictable enough
for predictability to be the reason for using only two prices. A more plau-
sible theory is that airlines use two prices because the theory is better de-
veloped for two-price systems. This is plausible for a consistent airline like
Northwest, but implausible for American Airlines, which has extraordinary
price adjustments on other routes.
6
The figures identify flights Airline_Flight Number_Departure Date. All data in the figures comes
from Orbitz’s website.
Table 2
Price regression
Estimate SE TStat PValue
Constant 54.1 0.9906 54.6 1.1 10642

Days 0.0192 0.0067 2.9 0.0042
WK1 50.0 0.6071 82.4 6.4 101441
WK2 28.2 0.6091 46.3 1.7 10464
WK3 10.6 0.6511 16.3 0.
WK4 5.8 0.6364 9.1 0.
Alaska 4.6 0.5006 9.2 4.3 1020
Delta 24.1 0.6112 39.4 6.4 10337
United 6.3 0.5894 10.6 2.4 1026
Amwest 6.5 0.6725 9.6 0.
USAir 86.8 0.5801 149.6 3.1 104508
NW 16.9 0.7070 23.9 8.6 10126
AM6 15.5 0.8189 18.9 2.0 1079
AM7 6.3 0.8380 7.5 5.5 1014
AM8 3.7 0.8348 4.4 0.
AM9 1.4 0.8806 1.6 0.1173
AM10 9.9 0.9115 10.8 0.
AM11 3.7 0.8221 4.5 6.1 106
Noon 5.9 0.8504 6.9 4.7 1012
PM1 3.6 0.8696 4.2 0.
PM2 1.3 0.9844 1.3 0.1932
PM3 2.6 0.7849 3.3 0.0009
PM4 1.2 0.9832 1.3 0.2044
PM5 7.4 0.8073 9.2 0.
PM6 3.4 0.8424 4.1 0.
PM7 3.5 0.8061 4.3 0.
PM8 5.8 0.9118 6.4 2.0 1010
PM9 2.0 1.0 2.0 0.0496
TUL2DFW 41.9 0.7870 53.2 4.7 10611
OKC2DFW 41.9 0.8412 49.8 1.2 10535
DFW2TUL 41.5 0.8068 51.4 7.4 10571
DFW2OAK 31.8 0.8381 37.9 4.5 10313
SFO2PDX 149.6 0.5144 290.8 3.9 1014331
SJC2PDX 44.6 0.8246 54.1 3.1 10632
OAK2PDX 49.4 0.7120 69.4 8.7 101031
PDX2SFO 152.1 0.5264 288.9 6.4 1014184
PDX2SJC 43.7 0.8253 53.0 1.3 10605
PDX2OAK 46.5 0.7352 63.2 1.8 10858
LAS2LAX 6.9 0.4877 14.1 2.9 1045
200
180
160
140
Average of price
120
100
80
Flight 60
AA _6825_9/23
AA_6912_9/23 40
AA_6966_9/23
AA_7480_9/23 20
0
85 80 75 70 65 60 55 50 27 22 16 11 6 0
Days Til
Fig. 4. American OAK to PDX, September 23.
350
Flight
AA_1038_9/23 300
AA_1038_9/25
250
Average of Price
200
150
100
50
0
85 80 75 70 65 60 55 42 30 25 20 15 10 5 0
Days Til
Fig. 5. American 1038 prices plotted against days prior to 9/25.

Table 3
Correlation coefficients
Flight pair r Flight pair r
(1) SFO2PDX, SJC2PDX 0.167 PDX2SFO, PDX2SJC 0.596

(2) SFO2PDX, OAK2PDX 0.190 PDX2SFO, PDX2OAK 0.568
(3) SJC2PDX, OAK2PDX 0.867 PDX2SJC, PDX2OAK 0.954
The final prediction of the theories is that the prices of substitutes should
be correlated. This, again, is a robust prediction and requires only con-
sumer substitution as a hypothesis. To assess this, we looked at the cor-
relation between prices of flights from the bay area to Portland, and
Portland to the bay area. Based on travel times, it seems that SFO and SJC
should be the closest substitutes, followed by SFO and OAK, followed by
SJC and OAK. In fact, however, the correlations are almost exactly the
reverse.
Table 3 shows the correlation coefficient of average prices, with the cor-
relation between mean price offers on each route as a function of the
number of days prior to departure. The prediction is that the first row of the
table should have the highest correlation, followed by the second, and fol-
lowed by the third. However, approximately the opposite arises. The pres-
ence of Southwest at San Jose and Oakland, but not at San Francisco, may
be the explanatory factor, especially in light of the fact that the price level at
SFO is much higher than at the other two airports.
It is more challenging to assess whether there is strong correlation be-
tween substitute flights on the same route. The following picture shows
some correlation over time of the substitute flights between Portland and
Oakland at 6 pm, 7 pm and 8 pm (71/2 hour). These flights have an
average correlation coefficient of about 0.8 overall (Fig. 6).
On the other hand, Fig. 7 presents a similar examination of economy
flights from LAX to LAS. No such correlation is readily apparent and these
flights are weakly correlated. Afternoon flights, presented in Fig. 8, are
similar.
Overall, the major predictions—including the most robust predictions—
of the theories appear to be violated in the data. Consequently, new the-
ories, and probably a new approach to the analysis, are needed.
10 Research projects and mysteries
There are many research projects that flow from the analysis of this
paper. First, much of the existing work is founded on a monopoly de-
scription. The analysis of this paper suggests that recasting the model as a
competitive model may increase the tractability of the model without sac-
rificing either empirical relevance or complexity of behavior.
180
160
140
120
Average of Price
100
80
60
Hour
18 40
19
20
20
0
115 110 105 100 95 90 85 80 75 70 65 60 55 50 45 40 35 30 25 20 15 10 5 0
Days Til
Fig. 6. PDX to OAK, return, at various times.
180
160
140
120
Average of Price
100
80
60
Hour 40
7
8 20
9
10
0
115 110 105 100 95 90 85 80 75 70 65 60 55 50 45 40 35 30 25 20 15 10 5 0
Days Til
Fig. 7. LAX to LAS economy class, 6–9 am.
Second, the sale of options present the possibility of enhancing the effi-
ciency of advance contracting, a possibility that has been little explored.
Continuous time techniques from the finance literature may play an im-
portant role in developing a theory of options for perishable goods. This
200
180
160
140
Average of Price
120
100
80
60
40
Hour
15
17 20
18
0
115 110 105 100 95 90 85 80 75 70 65 60 55 50 45 40 35 30 25 20 15 10 5 0
Days Til
Fig. 8. Afternoon LAX to LAS Flights by time of departure.
seems a fruitful approach, because options themselves are perishable, and

continuous time stochastic calculus techniques have played an important
role in simplifying similar problems.
The existing theory fared poorly when confronted with the data. In par-
ticular, the failure of prices to fall as takeoff approaches is devastating to
theories, leaving standing only those theories in which late arriving poten-
tial passengers have a relatively high willingness-to-pay. While this is a
reasonable hypothesis, it nonetheless needs further development, and it is
challenging to think of natural assumptions to impose order on the prob-
lem, once demand can be almost anything.
The data present a variety of mysteries. The gains to searching are oc-
casionally enormous. While prices rise as takeoff approaches, occasionally
they bounce around considerably. This finding is mirrored in Etzioni et al.
(2003) which empirically examined the gains to searching for flights. They
find an average savings of 27% by using a search algorithm, relative to just
booking the flight the first time it is checked.
What is perhaps even more mysterious, and illuminating, is an incident
that occurred in mid-July on Alaska 101, a flight from OAK to PDX. For
approximately a week, the price of AL101 departing 9/23 nearly doubled,
while the prices of the 9/25 and 9/26 departures were approximately un-
changed. The 9/26 departure had been more expensive. In fact, the price of
the 9/26 departure then fell briefly. This is illustrated in Fig. 9.
Given its departure time, American Airlines 6825 is a major substitute for
AL101. During this time, AA 6825’s flight experienced only a very modest
increase in price, which is illustrated in Fig. 10.
180
160
140
120
Average of Price
100
80
60
40
Flight
AL_101_9/23
AL_101_9/25 20
AL_101_9/26
0
87 82 77 72 67 62 57 42 28 22 17 12 6 1
Day
Fig. 9. Prices of AL101, departing 9/23, 9/25 and 9/26.
200
180
160
140
Average of Price
120
100
80
60
40
Flight
AA_6825_9/23 20
AL_101_9/23
0
86 81 76 71 66 61 56 51 27 22 16 11 6 0
Day
Fig. 10. Prices of AL101, AA 6825.
The price of AA6825 had been substantially higher than the price of
AL 101, but when AL 101 went up, the price of AL 101 was much higher.
The price of the AA 6825 went up slightly, then came back down to the
previously prevailing levels. What makes this so mysterious is that AA 6825
200
180
160
140
Average of Price
120
100
80
60
40
Flight
AA_6825_9/25
AL _101_9/25 20
0
88 83 78 73 68 63 58 42 28 21 16 11 5 0
Day
Fig. 11. AL101, AA 6825 departing 9/25.
and AL 101 represent the same airplane—AA 6825 is a code-share

flight operated by Alaska airlines. That is, the real mystery is why these
two identical fights are being marketed at such different prices, when
the fact that the flights are code-shared is clearly indicated on the
websites.
To make matters even more mysterious, the blip in prices of AL101
coincides with a slight increase in the price of AA 6825 for flights departing
on other days, but not the price of AL101 on other days. This is illustrated
in Fig. 11.
It seems likely that American reacted to the price of its rival. The facts
that (i) the identical flight was being offered for so much more when booked
as a code-share on American rather than on Alaska; (ii) Alaska’s price
increase wasn’t reflected in the prices of alternate days; and (iii) the price of
American’s identical flight only rose slightly and also rose on alternative
days, suggest that (1) consumers bring a substantial brand loyalty to the
flights; (2) airlines use a fare program that reacts to pricing by others across
a set of substitute flights; and (3) the process is poorly understood by
researchers.
The ease of collecting data for non-commercial use suggests that reverse-
engineering the pricing mechanisms of airlines is a feasible and potentially
very interesting research project.
11 Conclusion
This paper has considered so many different theories and data that it is
worth emphasizing a few highlights.
Dynamic price discrimination is primarily driven by customer dynamics
rather than price discrimination over an existing set of customers.
With a large number of units to sell, the per unit gain in profits of dynamic
price discrimination over a constant price is small, although the total
gain will still be positive. Most sales take place at an approximately constant
price; dynamic price discrimination is advantageous only as the prob-
ability of actually selling out changes, for a relatively small portion of a
large number of sales. One way to summarize this conclusion is that dy-
namic price discrimination only matters significantly on the last 20 or
so sales.
Monopoly and efficient dynamic pricing may be observationally equiv-
alent, and are in the important model of Gallego and van Ryzin (1994).
Directly solving for efficient solutions presents an alternative approach with
potential empirical merit and relative tractability.
The most important effects in dynamic price discrimination arise not
from an attempt to extract more money from the consumer, but from
addressing incomplete markets, and in particular from the value and costs
of advance contracting. Options, which create markets for advance con-
tracting, are an important aspect of both maximizing revenue and of effi-
ciently allocating resources.
The interruptible good problem breaks up into two separate maximizat-
ion problems, one for the low-quality good, and one for the difference of
the low-quality good and the high-quality good. In the airline context, an
interruptible good is one that provides the airline greater flexibility with
respect to the time of flight. The cost of delivering a seat reserving greater
airline flexibility is automatically lower, and thus is part of both profit
maximization and efficient provision of services.
Efficiency requires a positive probability of empty seats. Pricing to sell
out is inefficient.
Systematic changes in demand are a salient feature of the data, and
models that assume that early and late arrivals are identical are empirically
invalid.
Appendix
Proof of Theorem 2. The evolution of the probability that there are ex-
actly i unsold seats is governed by the probability of selling a seat when
there are i+1 unsold, and the probability of selling a seat when there are i
unsold, so that

q0i ðtÞ ¼ l 1 F ðpiþ1 ðtÞÞ qiþ1 ðtÞ l 1 F ðpi ðtÞÞ qi ðtÞ
¼ leðpiþ1 ÞðtÞ qiþ1 ðtÞ lepi ðtÞ qi ðtÞ
B0i ðtÞ B0 ðtÞ
¼ qi ðtÞ iþ1 qiþ1 ðtÞ
Bi ðtÞ Biþ1 ðtÞ
Given a capacity k at time 0, qk(0) ¼ 1, qk+1(t) ¼ 0, and
Bk ðtÞ
qk ðtÞ ¼ .
Bk ð0Þ
This is used as the base of an induction to establish the theorem. Suppose
the theorem is true at nrk. Then
B0n1 ðtÞ B0 ðtÞ

q0n1 ðtÞ ¼ qn1 ðtÞ n qn ðtÞ
Bn1 ðtÞ Bn ðtÞ
B0n1 ðtÞ B0 ðtÞðbtÞkn
¼ qn1 ðtÞ n ð41Þ
Bn1 ðtÞ Bk ð0Þðk nÞ!
This linear differential equation, with the boundary condition qn1(0) ¼ 0,
gives
Z t
1 B0n ðsÞðbsÞkn
qn1 ðtÞ ¼ Bn1 ðtÞ ds (42)
0 Bn1 ðsÞ Bk ð0Þðk nÞ!
Z t
bðbsÞkn ðbtÞknþ1
qn1 ðtÞ ¼ Bn1 ðtÞ ds ¼ (43)
0 Bk ð0Þðk nÞ! Bk ð0Þðk n þ 1Þ!
The expected number of unsold items, kn, satisfies
X
k X
k1
ðbtÞkn Bn ðtÞ
Eðk nÞ ¼ ðk nÞqn ðtÞ ¼ ðk nÞ
n¼0 n¼0
ðk nÞ!Bk ð0Þ
X
k1
ðbtÞ kn1
Bn ðtÞ
¼ bt
n¼0
ðk n 1Þ!B k ð0Þ
Bk1 ð0Þ X
k1
ðbtÞkn1 Bn ðtÞ
¼ bt
Bk ð0Þ n¼0 ðk n 1Þ!Bk1 ð0Þ
Bk1 ð0Þ X
k1
Bk1 ð0Þ
¼ bt qk1
n ¼ bt : Q:E:D: ð44Þ
Bk ð0Þ n¼0 Bk ð0Þ
Lemma A1.
8 x
>
< x1 if x41
X
k
k!ðxkÞ jk
!
j! k!1>
:
j¼0 1 if xo1
Proof of Lemma A1. Note

X
k Z Z
k!ðxkÞjk 1 1
¼ exk kx tk exkt dt ¼ kx ðtexðt1Þ Þk dt
j¼0
j! 1 1
Let cðtÞ ¼ texðt1Þ : If xo1, c(t) Z 1 for t A [1,1/x], so

Z 1
kx ðtexðt1Þ Þk dt ! 1.
1 k!1
Now, suppose x>1. For tZ1, c is decreasing. Locally, around 1, cðtÞ

1 þ ð1 xÞðt 1Þ: Thus,
Z 1 Z 1þ
kx ðtexðt1Þ Þk dt kx ð1 þ ð1 xÞðt 1ÞÞk dt
1 1
x k x
¼ 1 ð1 þ ð1 xÞÞk ! . ð45Þ
x 1k þ 1 k!1 x1
Pk
The proof for x ¼ 1 can be handled by observing that j¼0 ðk!ðxkÞjk Þ=j! is
non-increasing in x.
Q.E.D.
Lemma A2.
8
! > x if xo1
1 X
k
ðxkÞj <
log !
k j! k!1>
:
j¼0 1 þ LogðxÞ if x1
Proof of Lemma A2.

! xk Z 1
1 Xk
ðxkÞj 1 e k t
Log ¼ Log t e dt
k j¼0
j! k k! xk
xk Z 1
1 e k kt
¼ Log ðktÞ e kdt
k k! x
Z !
1 exk kk 1 t k
¼ log ðte Þ kdt
k k! x
For x>1, and large k, the term tet inside the integral is decreasing in t and
thus approximately equal to its first order taylor expansion ex ðx þ ð1 xÞ
ðt xÞÞ ¼ ex ðð1 xÞt þ x2 Þ: Therefore, for x>1,
Z !
1 exk kk 1 t k
¼ Log te kdt
k k! x
Z 2 !
1 exk kk x =ð1xÞ x
2 k
Log e ð1 xÞt þ x kdt
k k! x
Z 2 !
1 kk x =ð1xÞ k
¼ Log ð1 xÞt þ x2 kdt
k k! x
k !
1 kk ð1 xÞx þ x2
¼ Log k
k k! ð1 xÞðk þ 1Þ
!
1 kk xk
¼ Log k
k k! ðx 1Þðk þ 1Þ

1 ek xk k
Log pffiffiffiffiffiffiffiffi ! 1 þ LogðxÞ. ð46Þ
k 2pkðx 1Þðk þ 1Þ k!1
pffiffiffiffiffiffiffiffi
The last approximation invokes Stirling’s formula, k! 2pkkk ek :
For xo1, the proof is similar, applying Stirling’s formula and obtaining
Z ! k Z 1
1 exk kk 1 t k 1 e t k
Log ðte Þ kdt ¼ x þ Log pffiffiffiffiffiffiffiffi ðte Þ kdt
k k! x k 2pk x
Z 1
1 1 1t k
x þ Log pffiffiffiffiffiffiffiffi te kdt
k 2pk x
Z 1þð1=2Þpffiffi2 !
1 k
2 k
¼ x þ Log pffiffiffiffiffiffiffiffi pffiffi 1 2ð1 tÞ dt
k 2pk 1ð1=2Þ 2
pffiffiffi !
1 kGðk þ 1Þ
¼ x þ Log !x Q:E:D ð47Þ
k 2Gðk þ 1 þ ð1=2ÞÞ
Justification for some assertions in the text

If the price is constant at ð1=aÞ þ c; the probability of a sale given that a
customer has arrived is e1ac : Thus, let pi(t) be the probability of making at
least i sales in the period [t, T]. p0(t) ¼ 1, and pi(T) ¼ 0 for i>0.
Z T

pi ðtÞ ¼ lelðstÞ e1ac pi1 ðsÞ þ 1 e1ac pi ðsÞ ds (48)
t

p0i ðtÞ ¼ lpi ðtÞ l e1ac pi1 ðtÞ þ 1 e1ac pi ðtÞ

¼ b pi ðtÞ pi1 ðtÞ ¼ b 1 pi1 ðtÞ ð1 pi ðtÞÞ
This solves for
X
i1
ðbðT tÞÞj
pi ðtÞ ¼ 1 ebðtTÞ (49)
j¼0
j!
Thus, the probability that the flight sells out is

8
< 1
> if g41
X
k1
ðbTÞ j X
k1
ðgk Þ j
pk ð0Þ ¼ 1 ebT ¼ 1 egk ! 1=2 if g¼1
j¼0
j! j¼0
j! >
: 0 if go1
(50)
Proof of Lemma 3. First note that if capacity k is zero, then sales are
zero, verifying the base of an induction. Now suppose Lemma 3 is true
for k1.
Z T
Zk ðtÞ ¼ m emðstÞ 1 þ Zk1 ðsÞ ds
t
Z T
¼ m emðstÞ
t
!
X
k2
ðmðT sÞÞj
mðTsÞ
1þk1e ðk 1 jÞ ds
j¼0
j!

¼ k 1 emðTtÞ
Z T !
X
k2
ðmðT sÞ Þj
m emðTtÞ ðk 1 jÞ ds
t j¼0
j!
X
k2
ðmðT sÞÞjþ1
¼ k 1 emðTtÞ emðTtÞ ðk 1 jÞ
j¼0
j þ 1!
X
k1
ðmðT sÞÞj
¼ k 1 emðTtÞ emðTtÞ ðk jÞ
j¼1
j!
X
k1
ðmðT sÞÞj
¼ k emðTtÞ ðk jÞ Q:E:D
j¼0
j!
Proof of Theorem 4. Let x be a non-negative random variable, and
q ¼ ð þ 1Þ=; p ¼ þ 1: p and q are conjugate exponents. Let f ¼ x1=p

and g ¼ x1=ð qÞ : Then from the Holder inequality,
p 1=p q 1=q
E f E g E fg
or
1 1=ðþ1Þ n 1= o=ðþ1Þ n o

E x E x E xð1=ðþ1ÞÞþð1=ðþ1ÞÞ ¼ 1
or
( )
E x1=
E 1 Q:E:D
x
Proof of Theorem 5. Let x be a non-negative random variable, and q ¼

ð2 1Þ=ð 1Þ; p ¼ 2ð 1Þ=: p and q are conjugate exponents. Let
f ¼ x1/ep and g ¼ x2/q. Then from the Holder inequality,
p 1=p q 1=q
E f E g E fg
or
n o=ð21Þ ð1Þ=ð21Þ n o
E x1= E x2 E x1=ð21Þþ2ð1Þ=ð21Þ ¼ Efxg
or
n o 1
E x1= E x2 ðE fxgÞ21
1= 2 1
E x E x 1
2
¼ 1 þ CV 2 Q:E:D:
E fxg ðE fxgÞ
Proof of Theorem 6.

p p2
p ¼ Max p1 c1 1 F 1
p1 ;p2 j1 j2

p1 p2 p
þ p2 c2 F F 2
j1 j2 j2

p1 p2
¼ Max p1 c1 ðp2 c2 Þ 1 F
p1 ;p2 j1 j2

p2
þ ðp2 c2 Þ 1 F
j2

p1 p2 c1 c2 p1 p2
¼ Max j1 j2 1F
p1 ;p2 j1 j2 j1 j2 j1 j2

p c2 p
þ j2 2 1F 2
j2 j2 j2

c1 c2 c2
¼ j1 j2 R þ j2 R
j1 j2 j2
We assume the standard condition that marginal revenue is decreasing, in

which case there are sales of good 2 if and only if
p1 p
2
j1 j2
if and only if
p1 p2 p
2
j1 j2 j2
if and only if
c1 c2 c2

j1 j2 j2
if and only if
c1 c2
.
j1 j2
References
Badinelli, R.D. (2000). An optimal, dynamic policy for hotel yield management. European Journal of
Operational Research 121(3), 476–503.
Barut, M., V. Sridharan (2004). Design and evaluation of a dynamic capacity apportionment procedure.
European Journal of Operational Research 155(1), 112–133.
Belobaba, P.P., J.L. Wilson (1997). Impacts of yield management in competitive airline markets. Journal
of Air Transport Management 3(1), 3–9.
Board, S. (2005). Durable-goods monopoly with varying cohorts, presented at 2005 Meeting of the
Society for Economic Dynamics, Budapest.
Bodily, S.E., L.R. Weatherford (1995). Perishable-asset revenue management: generic and multiple-
price yield management with diversion. International Journal of Management Science 23(2),
173–185.
Boiteaux, M. (1949). La tarification des demandes en point: application de la theorie de la vent au cout
marginal. Revue Generale de l’Electricite 58(August), 321–340. (English translation available in
Journal of Business 33(April), 1960.)
Borenstein, S., N. Rose (1994). Competition and price dispersion in the U.S. airline industry. Journal of
Political Economy 103(August), 653–683.
Botimer, T.C. (1996). Efficiency considerations in airline pricing and yield management. Transportation
Research Part A 30(4, July), 307–317.
Brooks, M., K.J. Button (1994). Yield management: a phenomenon of the 1980s and 1990s?.
International Journal of Transport Economics 21(2, June), 177–196.
Carbonneau, S., R.P.McAfee, H. Mialon, S. Mialon (2004). Price discrimination and market power.
unpublished.
Carr, S., W. Lovejoy (2000). The inverse newsvendor problem: choosing an optimal demand portfolio
for capacitated resources. Management Science 46(7, July), 912–927.
Chatwin, R.E. (2000). Optimal dynamic pricing of perishable products with stochastic demand and a
finite set of prices. European Journal of Operational Research 125(1), 149–174.
Dana, J.D. Jr. (1999). Equilibrium price dispersion under demand uncertainty: the roles of costly
capacity and market structure. RAND Journal of Economics 30(4, Winter), 632–660.
Dana, J.D. Jr. (1999). Using yield management to shift demand when the peak time is unknown. RAND
Journal of Economics 30(3, Autumn), 456–474.
Davis, P. (1994). Airline ties profitability yield to management. SIAM News 27(5) (May/June).
de Boer, S.V., R. Freling, N. Piersma (2002). Mathematical programming for network revenue man-
agement revisited. European Journal of Operational Research 137(1), 72–92.
de Harris, B.F.H., J.P. Pinder (1995). A revenue management approach to demand management and
order booking in assemble-to-order manufacturing. Journal of Operations Management 13(4),
299–309.
Deneckere, R., R.P. McAfee (1996). Damaged goods. Journal of Economics and Management Strategy
5(2, Summer), 149–174.
Etzioni, O., C.A. Knoblock, R. Tuchinda, A. Yates (2003). To buy or not to buy: mining airline fare
data to minimize ticket purchase price, in: Proceeding of the 9th ACM SIGKDD International Con-
ference on Knowledge Discovery and Data Mining, Washington, DC.
Fan, M., K. Wang (1998). Optimal harvesting policy for single population with periodic coefficients.
Mathematical Biosciences 152(2), 165–177.
Feng, Y., G. Gallego (2000). Perishable asset revenue management with markovian time dependent
demand intensities. Management Science 46(7, July), 941–956.
Feng, Y., B. Xiao (2000). Optimal policies of yield management with multiple predetermined prices.
Operations Research 48(2, March–April), 332–343.
Feng, Y., B. Xiao (2000). A continuous-time yield management model with multiple prices and re-
versible price changes. Management Science 46(5 May), 644–657.
Feng, Y., B. Xiao (2001). A dynamic airline seat inventory control model and its optimal policy.
Operations Research 49(6 November–December), 938–949.
Gallego, G., G. van Ryzin (1994). Optimal dynamic pricing of inventories with stochastic demand over
finite horizons. Management Science 40(8 August), 999–1020.
Garcia-Diaz, A., A. Kuyumcu (1997). A cutting-plane procedure for maximizing revenues in yield
management. Computers and Industrial Engineering 33(1–2), 51–54.
Kimes, S.E. (1989). Yield management: a tool for capacity-constrained service firms. Journal of Op-
erations Management 8(4 October), 348–363.
Kuyumcu, A., A. Garcia-Diaz (2000). A polyhedral graph theory approach to revenue management in
the airline industry. Computers & Industrial Engineering 38(3), 375–396.
Ladany, S.P. (1996). Optimal market segmentation of hotel rooms-the non-linear case. International
Journal of Management Science 24(1), 29–36.
Li, M.Z.F. (2001). Pricing non-storable perishable goods by using a purchase restriction with an ap-
plication to airline fare pricing. European Journal of Operational Research 134(3), 631–647.
McGill, J.I., G.J. Van Ryzin (1999). Revenue management: research overview and prospects. Trans-
portation Science 33(2, May), 233–256.
Mussa, M., S. Rosen (1978). Monopoly and product quality. Journal of Economic Theory 18, 301–317.
Sawaki, K. (2003). Optimal policies in continuous time inventory control models with limited supply.
Computers & Mathematics with Applications 46(7), 1139–1145.
Sridharan, ‘Sri’.V. (1998). Managing capacity in tightly constrained systems. International Journal of
Production Economics 56–57, 601–610.
Stokey, N.L. (1979). Intertemporal price discrimination. The Quarterly Journal of Economics 93(3,
August), 355–371.
Talluri, K.T., G.J. van Ryzin (2004). The Theory and Practice of Revenue Management. Kluwer Ac-
ademic Publishers, Dordrecht, Netherlands.
Van Slyke, R., Y. Young (2000). Finite horizon stochastic knapsacks with applications to yield man-
agement. Operations Research 48(1 January–February), 155–172.
Williamson, E. L. (1992). Airline network seat control. Ph.D thesis, MIT Flight Transportation Lab-
oratory, Cambridge, Mass.
Williams, L. (1999). Revenue management: microeconomics and business modeling. Business Economics
34(April), 39–45.
Williamson, O. (1966). Peak load pricing and optimal capacity under indivisibility constraints. American
Economic Review 56(September), 810–827.
Zhao, W., Yu-.S. Zheng (2000). Optimal dynamic pricing for perishable assets with nonhomogeneous
demand. Management Science 46(3 March), 375–388.
Chapter 12
Online Auctions
Axel Ockenfels
Department of Economics, University of Cologne, Albertus Magnus Platz, Cologne, Germany
David H. Reiley
Department of Economics, University of Arizona, Tucson, AZ, USA
Abdolkarim Sadrieh
Faculty of Economics and Management, University of Magdeburg, Magdeburg, Germany
Abstract
The economic literature on online auctions is rapidly growing because of

the enormous amount of freely available field data. Moreover, numerous
innovations in auction-design features on platforms such as eBay have created
excellent research opportunities. In this article, we survey the theoretical,
empirical, and experimental research on bidder strategies (including the timing
of bids and winner’s-curse effects) and seller strategies (including reserve-price
policies and the use of buy-now options) in online auctions, as well as some of
the literature dealing with online-auction design (including stopping rules and
multi-object pricing rules).
1 Why do information systems make auctions (even) more popular?
Long before any electronic information systems were in place, people

used auctions to trade all kinds of goods and services. In his comprehensive
overview of the history of auctions, Cassady (1967) reports auctions of
items of almost every size, from jewels and spices to ships and provinces.
Auctions have also sold a wide range of services, including anything from a
date with a bachelor (for charity fundraising) to the lifetime work of a slave.
571
572 A. Ockenfels et al.
Despite being widespread, auctions have not been the most common way of
trading because the costs of conducting and participating in an auction
have been quite high: the buyers would gather in the same place, at the same
time, and the seller would pay for the use of an auctioneer’s services. For
many purposes, fixed prices are quite convenient: buyers know they can buy
at the posted price, and they don’t have to decide how much to bid. The
auctions depends on a trade off between the advantage of price discovery
(i.e., discovering the market-clearing price level) and the disadvantage of
having high transaction costs. Because of this, auctions are most valuable
when the party running the auction is highly uncertain of the item’s market
value and, thus, receives a considerable advantage from the price discovery
afforded by the auction.
With the emergence and spread of electronic information systems, both
the advantages and disadvantages have changed radically. On the one hand,
the transaction costs associated with conducting and participating in an
auction have decreased so substantially that auctions seem worthwhile even
when the expected advantage of detecting the true market value of the item
is relatively low. On the other, the expected advantage of price discovery
has increased sharply because many more potential auction participants can
virtually meet in an online auction house than at any physical location. The
effect is magnified by the fact that software agents in virtual auction houses
enable participants to interact ‘‘on time’’ without having to be present in
person. The time and space emancipation provided by electronic informa-
tion systems has increased the expected number of potential participants at
an auction, making it more likely for the auctioneer to meet participants
with especially high valuations.
Since transaction costs have decreased and potential gains of trade have
increased, it comes as no surprise that the volume of online auction trades
has exploded ever since the service was first offered. Meanwhile, there is no
doubt that online auctions constitute a major market with growing signifi-
cance for the global economy.
Not only size matters. In addition to the significance of trade volume,
online auctions have also taken a central role in market research and in-
novation. No other online economic institution has induced as much in-
novation and created as many research opportunities as auctions have. New
features concerning the design of online auctions are proposed and dis-
cussed almost on a daily basis. At the same time, the enormous amount of
market data generated in online auctions and recorded electronically has
enabled researchers to empirically address many issues that previously were
not researchable. Finally, the open access to online auctions has also
opened a door for field experiments.
In this chapter, we provide an overview of some of the theoretical, em-
pirical, and experimental research on online auctions. The literature in this
field is expanding so quickly that comprehensive coverage is impossible.
Nevertheless, we have tried to cover all major aspects of the research in the
Ch. 12. Online Auctions 573
field with the one notable exception of reputation in online auctions. An

entire other chapter of this book deals extensively with the issue of online
reputation.
The rest of this chapter is organized as follows. In Section 2, we present
some of the foundations of auction theory for the simplest case of single-
object auctions. The theoretical results are compared to experimental find-
ings and related to empirical observations in online auctions. In Section 3,
we discuss theoretical and empirical aspects of using public and secret re-
serve prices in online auctions. Furthermore, we discuss the research on shill
bids representing an alternative to secret reserves. In Section 4, we present
the research on late and incremental bidding in online auctions. This re-
search deals with the fact that in some online auctions many bids are sub-
mitted within the last few moments. An interesting aspect of this strand
of research is that it is entirely motivated by phenomena that were first
discovered in online auctions. In Section 5, we present the research on the
buy-now option, which is widespread in online auctions, but rarely ob-
served in classical auctions. The buy-now option creates a hybrid market
situation that allows bidders to choose between normal bidding or accept-
ing a posted sales price. The buy-now option is, in effect, an outside option
for the buyers relative to the bidding process. In Section 6, we examine
other types of buyer outside options, such as multiple parallel auctions and
market posted-prices. In Section 7, we present the research on multi-object
auctions, which are both theoretically and empirically more difficult to deal
with than the single-object auctions. In Section 8, we conclude the chapter
with some general remarks on the design of online auctions.
2 Single-object auctions: theory and experiments
Auction theory has been remarkably influential on the design of elec-

tronic market mechanisms. It has also motivated much of the empirical
research on auction behavior that we are surveying in this chapter. This
section, together with our discussion of multi-object auctions in Section 7,
reviews some of the central theoretical and experimental results relevant to
online auctions.1
2.1 Standard auction mechanisms and models
An auction elicits information, in the form of bids, from potential buyers

regarding their willingness to pay. The outcome—who wins and pays how
1
For more comprehensive and mathematical treatments of auction theory see, e.g., Klemperer (1999),
Krishna (2002), Milgrom (2004), and Menezes and Monteiro (2005).
much—is then determined based on this information. In a single-object

auction, one indivisible object is for sale. There are four single-object auc-
tion types, which are widely used and analyzed both in theory and practice:
the ascending-price auction (sometimes called English auction), the de-
scending-price auction (sometimes called Dutch auction),2 the first-price
sealed-bid auction, and the second-price sealed-bid auction (sometimes
called Vickrey auction).
The ascending-price auction is probably the best-known auction proce-
dure: the price is raised until only one bidder remains. This remaining
bidder wins the object at the price at which the strongest competitor
dropped out. There are many ways to run ascending-price auctions: having
the seller announce prices, the bidders announce prices, or the price con-
tinuously rising on a ‘‘price clock’’. In the latter version, which is the one we
will refer to when we speak of ascending-price auctions, bidders can quit the
auction at any price and observe other bidders quitting. Because the price
clock determines the price path, there is no possibility for bidders to speed
up or slow down the auction process, or to employ ‘‘jump bids’’ as a
signaling device.
The descending-price auction works in the opposite way: the auction
starts at a high price, which a price clock then lowers. The first bidder to call
out his acceptance of the displayed price immediately stops the clock. This
bidder wins the object and pays the price at which the clock stopped. Note
that while in the ascending-price auction the winner pays a price determined
by his strongest competitor (the price on the clock when the second-to-last
bidder exits), the winner in the descending-price auction determines the final
price (the price on the clock which he was the first to accept).
In the first-price sealed-bid auction, bidders independently submit a
single bid, without seeing the others’ bids. There is no open, dynamic
bidding. The bidder who submits the highest bid wins and pays a price
equal to his bid. In the second-price sealed-bid auction, again the bidder who
submits the highest bid wins, but here he pays a price equal to the second-
highest bid.
In addition to the four types of auctions, there are two standard models
of how bidders value an item: the private-value and the common-value
model. In the private-value model, each bidder knows the value (his max-
imum willingness to pay) that he assigns to the object, but different bidders
may have different values. For auctions to be a meaningful selling mech-
anism, the seller does not know the values of the potential buyers.3
Furthermore, there is typically asymmetric information among bidders:
2
Outside economics, the term ‘‘Dutch auction’’ is often used for different formats like for eBay’s
ascending-bid, multi-unit auction or, in investment banking, for uniform-price sealed-bid auctions such
as in the context of the Google IPO.
3
If the seller knew bidders’ values, he would simply make a take-it-or-leave-it offer to the bidder with
the highest value.
one’s value is private information to oneself.4 Provided that there are no

entry fees or other costs involved in bidding, the auction winner’s net gain is
his value of the object minus the final auction price. The losers’ net gain is
zero since they neither get nor pay anything.
In the common-value model, the value of the object is the same to all
bidders, but bidders have different information about what is the actual
value. For example, the ‘‘true’’ value of an antique statue may be very
similar to all bidders, but bidders may get different ‘‘signals’’ about whether
the statue is genuine or a fake. In such situations, bidders typically want to
change their estimates of the value when they learn about the competitors’
signals. In the private value model, on the other hand, bidders’ values are
unaffected by learning the competitors’ information. There are also more
general modeling approaches, encompassing both special cases of private-
value and common-value. In these models, each bidder gets a private signal,
and the value to the bidder is a function of all signals.
2.2 Bidding behavior and auction outcomes in theory
We begin with the private-value model. In the ascending-price auction, it

is a ‘‘dominant strategy’’ to stay in the auction until the price reaches the
bidder’s value. That is, no other strategy may yield a strictly higher expected
payoff regardless of the competitors’ strategies. It follows that the bidder
with the highest value wins (the auction outcome is efficient). The price paid
by the winner (the auction revenue) equals the value of the strongest com-
petitor, which is the second-highest value of all bidders.
In the second-price sealed-bid auction, a losing bid will determine the
price. Thus, a bid only affects whether the bidder wins or not, but not his
payoff. A bid equal to the bidder’s value makes sure that the bidder wins if
and only if the price is below his value, regardless of the strategies employed
by the competitors. Thus, bidding one’s value is a dominant strategy (as
first observed by Vickrey, 1961). It follows that, as in the ascending-price
auction, the bidder with the highest value wins at a price equal to the value
of the strongest competitor.
The first-price sealed-bid and the descending-price auction formats are
strategically equivalent. In the descending-price auction, each bidder de-
cides on the price at which he plans to stop the auction. This plan will not
be conditioned on other bidders’ bids, because the first bid immediately
ends the auction. Similarly, in the first-price sealed-bid auction a bidder
must submit a bid without knowing the competitors’ bids. Furthermore, in
both auction formats, the winner pays a price equal to his bid, and thus the
outcomes of both auctions are equivalent.
4
Thus, the appropriate equilibrium concept is Bayesian Nash-equilibrium, in which each bidder
maximizes his expected payoff given the competitors’ strategies (which are functions of the respective
bidder’s information) and his beliefs about the other bidders’ information.
However, bidding in the first-price sealed-bid and the descending-price

auctions is more difficult than bidding in the other auctions. The only way
to make a strictly positive gain is to ‘‘shade the bid,’’ that is to bid less than
one’s value. For example, in the simple case when the bidders know the
values of their competitors, the bidder with the highest value will bid just
above the second-highest value. This ensures that he will win at the lowest
possible price.5 If values are private information, however, each bidder
faces a risk-return trade-off. The equilibrium strategy then depends on what
bidders believe about other bidders’ values as well as their own risk atti-
tudes. The more competition he expects from other bidders or the more
risk-averse is a bidder, the higher the optimal bid.
To be more specific, let us assume that bidders are risk-neutral and that
values are independent draws from the same distribution, which is the so-
called symmetric independent private-values model. Then, in equilibrium,
each bidder will bid his expectation of the value of his strongest competitor
conditional on having the highest value (only in this case is the bid payoff-
relevant). As a result, the bidder with the highest value wins (the auction
outcome is efficient). The auction revenue is, on average, equal to the sec-
ond highest value. That is, under our assumptions, the descending-price and
first-price sealed-bid auctions yield the same expected revenue as the as-
cending-price and second-price sealed-bid auctions. Bidders adjust their
behavior to changes in the auction rules in a way such that winners do not
pay more than what they need to in order to win: the value of the strongest
competitor. This is the famous Revenue Equivalence Theorem by Vickrey
(1961) that has later been generalized by Myerson (1981) and Riley and
Samuelson (1981), among others.6
The revenue equivalence theorem not only holds for the private-value
model, but also for common-value models, if the signals are independent. In
auctions with common-value components, however, bidding is more com-
plicated because bidders must take into account that they run the risk of
suffering from the winner’s curse. Each bidder must recognize that (in
symmetric equilibria) she wins only if she has the highest estimate of the
value. In this sense, winning the auction is bad news—it implies that all
other bidders have information that led them to bid more cautiously, so
that the winner would probably have revised her value downwards if she
had access to competitors’ information. The winner’s curse refers to the fact
that winners may not anticipate this bad news that comes with victory.
They run the danger of systematically paying more, on average, than the
5
If he bids below the second-highest value, the best response of the strongest competitor would be to
outbid him.
6
More formally, one version of the revenue equivalence theorem states that if each bidder has a
privately known signal (in the private-value model: his private value), independently drawn from a
common strictly increasing distribution, then any efficient auction mechanism in which any bidder with
the lowest-feasible signal expects zero surplus, yields the same expected revenue.
actual value. Of course, this cannot happen in equilibrium with rational

bidding, where bidders would adjust their bids downwards.
Outside the controlled laboratory and field experiments of the sort we
survey in the next subsection, the theoretical results described above typ-
ically do not directly apply to online auctions, because the circumstances
often differ from those assumed in the theory. Bidders may be neither
rational nor risk-neutral. They may collude, endogenously enter the auc-
tion, or they may be asymmetric with respect to value distributions. Sellers
may employ reservation prices or impose participation costs. The same or
competing auctioneers may simultaneously or sequentially sell substitutes
or complements of the object. There might also be technical risks involved
in electronic bidding, and anonymous and geographically dispersed inter-
action in online auctions may create moral hazard and adverse selection
issues. We will address many of these complicating factors later in this
chapter, but first, the next subsection deals with empirical tests of the simple
theory.
2.3 Bidding behavior in controlled laboratory and field experiments
Auction experiments in the laboratory and in the field serve as a test of

auction theory (Kagel, 1995); as an empirical foundation of new ap-
proaches in behavioral game theory and other disciplines concerned with
economic decision-making (Camerer, 2003); as a test-bed for alternative
auction rules (Roth, 2002); and as a teaching tool (Asker et al., 2004).
Auction experiments have been conducted in highly controlled laboratory
settings to reproduce as accurately as possible the environmental assump-
tions of a given theoretical model. Particularly in the online-auction era,
field experiments can also take place in natural environments, increasing
external validity and decreasing the amount of control by the experimenter.
In this section, we concentrate on experimental tests of simple auction
theory, focusing on those laboratory and field experiments most relevant to
online auctions.
While individual values for the auctioned object are typically unobserv-
able in the field, they can be controlled in experimental settings with the
help of the so-called induced-value method (Smith, 1976). The trick is to sell
money. In a laboratory experiment, bidders compete to win a fictitious
good. The bidder who wins the good may subsequently redeem it with the
experimenter for a specified amount of cash. This redemption value is typ-
ically different for each bidder. So, for example, a bidder with a value of $30
who wins the auction at a price of $21 will earn a cash payment of $9 from
the experimenter. By inducing values and giving information to the bidders
about the probability distribution of other subjects’ values, the experi-
menter may impose the assumptions of a given theoretical model.
Laboratory experiments with induced private values demonstrate
that bidders tend to bid up to their values in ascending-price auctions, in
agreement with theoretical predictions. However, in first-price sealed-bid

auctions, bids are typically higher than predicted given the assumptions of
the revenue equivalence theorem. This overbidding is a robust effect ob-
served in numerous experiments. Bidder risk aversion was the earliest pro-
posed theoretical explanation for this behavior, but this theory also
generated quite a bit of skepticism; see the comprehensive discussion in the
survey by Kagel (1995). More recent studies propose explanations based on
emotions and bounded rationality. For example, the theoretical papers by
Engelbrecht-Wiggans (1989) and Morgan et al. (2003) and the experimental
papers by Ockenfels and Selten (2005) and Engelbrecht-Wiggans and
Katok (2005) argue that overbidding is consistent with concerns for relative
standing, spite, or regret.7 Kirchkamp and ReiX (2004), on the other hand,
provide experimental evidence suggesting that overbidding is an artifact of
laboratory testing, which often allowed over—but not underbidding on the
whole value range.
Laboratory experiments also found that, contradicting the revenue
equivalence theorem, open auctions generally raise less revenue but are
more efficient than the equivalent sealed-bid auctions (Kagel, 1995; see also
Engelmann and Grimm, 2004 for an analogous result in multi-unit auc-
tions). In particular, descending-price auctions typically raise less revenue
than first-price sealed-bid auctions, and bidders in second-price auctions
often overbid with respect to their dominant strategy and rarely underbid
(Kagel and Levin, 1993, among others). However, recent experimental
studies in environments that are closer to ‘‘naturally occurring’’ online
auction environments sometimes seem to qualify these findings.
In the first online-auction experiment, Lucking-Reiley (1999) auctioned
off collectible trading cards over the Internet. By going into the field, he
sacrificed some control—e.g., he did not induce values, he allowed for en-
dogenous entry, etc.—in order to assess the predictive power of the theory
in a ‘‘real-world’’ auction. He found that descending-price auctions earn
approximately 30% more revenue than first-price sealed-bid auctions,
which is inconsistent with both the revenue equivalence theorem and pre-
vious laboratory results. He could not find a significant difference between
the ascending-price and the second-price sealed-bid formats.
Other experiments exploit the fact that many online auction platforms,
such as eBay, operate much like second-price auctions. eBay asks the bid-
ders to submit maximum bids (called ‘‘proxy bids’’) and explains that ‘‘eBay
will bid incrementally on your behalf up to your maximum bid, which is kept
secret from other eBay users.’’ That is, once a bidder submits his (proxy)
bid, eBay displays the currently winning bid as the minimum increment
7
Recent social comparison models such as Fehr and Schmidt (1999) and Bolton and Ockenfels (2000)
proved to be quite successful in capturing ‘‘anomalous’’ behavioral patterns in a wide range of economic
situations.
above the previous high proxy bid.8 At the end of the auction, the bidder
who submitted the highest bid wins the auctioned item and pays a price
equal to the second-highest bid plus the minimum increment.9 In the anal-
ysis that follows, we shall assume the minimum increment amount to be
negligibly small.
To understand the connection between the single-unit eBay auction and
the second-price sealed-bid auction, assume for the moment that nobody
learns about the proxy bids of other bidders until the auction is over. Then,
in fact, eBay becomes a second-price sealed-bid auction, in which each
bidder has a dominant strategy to bid his value. eBay explains the eco-
nomics of second-price auctions to their bidders along these lines, and
extends the conclusion to its own auctions, in which bids are processed as
they come in: ‘‘eBay always recommends bidding the absolute maximum that
one is willing to pay for an item.’’ Ockenfels and Roth (forthcoming) support
this view based on a game-theoretic analysis of a continuous-time model of
eBay’s single-object auction. They show that, while there is no dominant
strategy in eBay’s open second-price format, strategies that involve bidding
above one’s value are dominated, and that bidders ‘‘sooner or later’’ will
always bid their value.10
Garratt et al. (2004) investigated bidding behavior of eBay buyers and
eBay sellers, experienced with eBay’s second-price rule, using induced val-
ues in a second-price sealed-bid auction experiment. While even highly
experienced eBay bidders tend to not bid exactly equal to their values, there
was no greater tendency to overbid than to underbid as previously observed
in laboratory experiments. Furthermore, Garratt et al. found that, on av-
erage, subjects with selling experience on eBay bid significantly less than
subjects with only buying experience on eBay. One important implication of
this study is that background and experience with online auctions seems to
affect bidding behavior.
Ariely et al. (2005) investigated eBay bidding behavior in the laboratory
by programming an experimental auction that emulated eBay’s central
auction rules as described above. In one study, they compared the per-
formance of eBay’s open second-price auction format with an analogous
second-price sealed-bid format. For inexperienced bidders, they reported
that the median sealed bids were substantially lower than in the open auc-
tion. Consistent with previous laboratory studies, the open format results in
8
While proxy bidding became widespread with the advent of eBay, Lucking-Reiley (2000b) documents
similar rules being used by auctioneers who wished to allow absentee bids, dating back at least to the
1800s. Ward and Clark (2002) compare the success of bidders using proxy bids to those, who bid
spontaneously. They find that more experienced bidders learn to use the proxy bids, placing them at an
ever later point in time.
9
If the second-highest bid plus one increment exceeds the highest bid, then the highest bidder pays his
own bid.
10
To be more specific, this holds for all Bayesian equilibria in undominated strategies of their private-
value auction model.
more efficient outcomes, but, inconsistent with previous results, the open
format also yields higher revenues. One reason for the better performance
of the open format is that learning in the sealed-bid auctions only takes
place across auctions, while learning in the dynamic auctions also takes
place within auctions. For example, a bidder who imagines that he can win
with a low bid does not learn that he is mistaken in a sealed-bid auction
until after the auction is over. However, in open auctions, he can revise his
bid as soon as someone outbids him. For experienced bidders, Ariely et al.
reported median final bids in both the sealed-bid and the open auctions that
converge to the equilibrium prediction: 100% of values.
Ockenfels (2005) studied bidder behavior in a one-shot field experiment,
using eBay as the experimental platform. He invited eBay bidders through
posters and e-mail to register for online, eBay experiments; the only re-
quirement was a valid eBay account. Bidders were privately informed about
their induced values. All communication and data collection were auto-
mated and via the Internet. Due to the second-price rule employed by eBay,
final bids should theoretically be equal to induced values. Ockenfels found
that, on average, the losers’ final bid (eBay’s interface does not reveal the
winners’ maximum bids) were indeed close to the induced values.
Regarding common-value auctions, laboratory experiments consistently
show that inexperienced bidders are subject to the winner’s curse: on av-
erage, winners typically overpay (Kagel and Levin, 2001a and 2002).
Learning to avoid the winner’s curse appears difficult. Ball et al. (1991)
explored the winner’s curse in a bilateral bargaining game with asymmetric
information, and found virtually no adjustment to the winner’s curse over
20 trials (see also Grosskopf et al., 2003). Kagel and Levin (1986) observed
that experienced bidders in first-price common-value auctions can over-
come the winner’s curse only with relatively few bidders, but succumb again
with larger number of bidders. Observe that a larger number of bidders
heightens the winner’s curse problem—winning is worse news when there
are more competitors who think the object is not worth the final price. So,
in theory, more bidders require further shading of bids, while, in fact, the
subjects in the experiment by Kagel and Levin (1986) tended to bid more
aggressively in larger bidding groups. Kagel et al. (1995) replicated this
finding in laboratory, second-price common-value auctions. Bidders again
failed to respond in the right direction to more competitors.
Given the laboratory evidence, it is reasonable to expect that the winner’s
curse also play a role in online auctions because many online auctions have
common-value components. For one, online auctions often make it too
costly for buyers to inspect in person the object being sold, so that an
assessment of the ‘‘true’’ condition of the object can be difficult and may
vary across bidders, depending on the sources and quality of individual
information. Moreover, there is often a risk of being exploited by a fraud-
ulent seller on C2C auction platforms. Because bidders may differ with
respect to their assessment of a seller’s trustworthiness, winning the auction
may imply bad news regarding the seller’s intentions. Finally, online auc-
tions often attract large numbers of potential bidders, which further am-
plifies the winner’s curse; winning an object on eBay means, in the extreme,
that thousands of other users do not think that the object is worth its final
price.
A couple of papers looked for winner’s curse effects on eBay. While direct
evidence from field experiments supports the laboratory findings that win-
ner’s curse effects are a real and robust phenomenon,11 indirect statistical
analyses of eBay data also support the view that bidders, to some extent,
respond in a strategic way to take the risk of overpayment into account.
Jin and Kato (2004, 2005) conducted a field experiment searching for the
winner’s curse on eBay. They bid on eBay auctions for upgraded, baseball
cards, and then let a professional grading service evaluate the cards. They
found that claims of high-value result in more fraud (i.e., default or sending
counterfeits) and no better card quality. However, eBay buyers were willing
to pay 23–54% more for cards that claimed quality of mint or better.
Overall, Jin and Kato concluded that some buyers have fallen prey to the
winner’s curse by having underestimated the probability of fraud (see also
Bajari and Hortac- su, 2004 for a discussion of these findings). A more in-
direct test by Bajari and Hortac- su (2003b) relied on analyzing strategic
bidding within a common-value auction model of eBay. The model is a
second-price sealed-bid auction with an uncertain number of bidders. The
sealed-bid aspect of the model implies that no bidder can learn about the
value from the bids of others.12 Using field data of collectible coin auctions
and applying various statistical instruments to account for the endogeneity
of the number of bidders, the authors found that bids decline with the
number of competing bidders. This is in line with theory but, inconsistent
with laboratory research. Regarding experience effects, Bajari and Hortac-
su found that experienced bidders are slightly more cautious. On the other
hand, Ow and Wood (2004) reported, also based on a field study of rare-
coin eBay auctions, that more experience leads to more aggressive bidding.
However, this is not necessarily inconsistent with the winner’s curse story
put forward by Bajari and Hortac- su, since Ow and Wood argue that the
winner’s curse effect is partly overlaid by an ‘‘initial lack of institution-
based trust’’ that decreases with experience. Other evidence in favor of
rational reactions to winner’s curse effects comes from survey information
on completed eBay auctions by Yin (2005). She showed that the winning
bid in a sample of eBay auctions is negatively correlated with the variance
11
A related but distinct phenomenon is price dispersion. Bidders often pay a higher price for an object
than prices for an identical object offered at the same time by a different seller; see, e.g., Ariely and
Simonson (2003) and the survey by Morgan et al. (2005) in this Handbook.
12
The authors justify this modeling approach by noting that many bids are submitted in the last
minute of the auctions.
of the self-reported, ex ante estimates of the objects’ values.13 That is, the
more certain a bidder is of the item’s value, the more aggressively they bid.
The data surveyed here suggest that auction theory and laboratory results
are sometimes, but not always, a good predictor of online auction behavior.
For instance, auction theory has difficulties capturing overbidding in first-
price auctions and overpaying in common-value environments. Field
behavior seems to differ from laboratory behavior when it comes to over-
bidding in second-price auctions and to experience and number effects in
common-value environments. Some causes for these discrepancies suggest
themselves: lack of bidder experience (Garrat et al., 1994); small decision
costs and stakes (List and Lucking-Reiley, 2002); uncontrolled institutional
differences, self-selection and subject-pool effects; and presentation
effects.14 While, to date, there is only little research on the external valid-
ity of theoretical and experimental auction research, it appears likely to us
that online auctions will play an important role in investigating how labo-
ratory and field behavior are linked to each other (Bolton and Ockenfels,
2006).
3 Reserve prices, minimum bids, and shill bids
Superficially, it may seem that all activity in an auction comes only from
bidders submitting offers, while the seller running the auction simply waits
and hopes that the auction will be profitable. However, in addition to the
initial choice of auction mechanism, the seller can also choose a reserve
price that prevents low-revenue sales and stimulates competition. Strictly
speaking, a reserve price defines the minimum amount that a seller will
accept for the auction to end with a sale.15
The most straightforward behavior for the seller is to set a reserve price
equal to her true willingness to accept and announce it at the outset of the
auction. In this case, the reserve price would serve to enhance efficiency,
preventing the item from being sold if none of the buyers values it as highly
as the seller. However, the seller may also make strategic use of the reserve
price by setting it higher than her willingness to accept. The theoretical
results discussed in the next subsection will show that optimal reserve prices
are often—but not always—set strictly higher than the seller’s willingness to
accept.
13
High variances occur in auctions with poorly designed web pages or where the object had am-
biguous characteristics.
14
For the latter, observe, for instance, that eBay’s ‘‘proxy bidding’’ explanation of second-price
auctions (y think of the bidding system standing in for you as a bidder at a live auction) appears much
more understandable than a typical explanation used in a laboratory experiment (You pay the amount of
the second-highest bid—i.e., the highest bid placed by any other participant in the auction).
15
For convenience, here and elsewhere, we restrict our attention to the case where the auction is held
by the seller, but the same discussion applies equally well (with obvious adjustments) to procurement
auctions held by a buyer.
In most online auctions, the seller may make a strategic choice not only of
the amount of the reserve price, but also whether the reserve price should be
secret or public and, if made public, at what point in the auction. Finally,
although it violates the formal rules of the auction at eBay and most other
online auction sites,16 the seller may effectively camouflage and dynamically
adjust the reserve price during the auction by using shill bids or bids cov-
ertly placed by the seller which are by the seller’s confederates to artificially
inflate the final sale price. Clearly, any of these strategic options (or com-
binations thereof) may be used by the seller to increase the expected revenue
from the auction.
The most popular type of reserve price used in online auctions is the
minimum bid (sometimes also called the opening bid or starting bid in dy-
namic auctions). A minimum bid defines the lowest bid that a bidder may
submit at the outset of an auction. Because it is publicly announced before
the auction begins and cannot be adjusted later, a minimum bid represents a
static public reserve price.
When the seller sets the minimum bid below her valuation, she often
combines this strategy either with a secret reserve price or with shill bidding.
Neither of these reserve-price variants is made public; in fact, shill bidding is
a type of fraud. However, both have a similar effect as a public reserve
price: a trade only occurs if the final highest bid is above the secret reserve
price or the shill bid. The three instruments differ in their informational and
dynamic aspects. Secret reserve prices are fixed with the auctioneer before
the auction starts. On eBay and most online auction sites, bidders are
informed that an auction has secret reserve, and whether or not it has yet
been met by the bidding. (In an interesting contrast, traditional, live auction
houses like Christie’s and Sotheby’s do not inform bidders whether the
secret reserve price has yet been exceeded.)
By contrast with secret reserve prices, shill bids—like all bids in dynamic
auctions—can be submitted and updated at any time during the course of
the auction. Bidders are not informed of the presence of shill bidding, but
obviously, they might guess that shill bidding is taking place. These differ-
ences in the informational and dynamic features are not only theoretically
relevant, but in many countries also have legal consequences. It is impor-
tant to note, however, that online shill bidding can be well organized (e.g.,
with a relatively large set of paid colluders, a rotating scheme with peer
sellers, or through the use of a false online identity) and hard to detect. The
16
Most online auctions explicitly forbid shill bidding on their sites. eBay, for example, writes: ‘‘Shill
bidding is the deliberate use of secondary registrations, aliases, family members, friends, or associates to
artificially drive up the bid price of an item. (This is also known as ‘‘bid padding.’’) Shill bidding is
strictly forbidden on eBay. eBay members found shill bidding may receive warnings or indefinite sus-
pensions of their accounts.’’ And: ‘‘Shill Bidding undermines trust in the marketplace since prices are
being artificially manipulated. Furthermore, it is illegal in many locations throughout the world. To
ensure that buyers’ and sellers’ trust in eBay’s marketplace is appropriately placed, Shill Bidding is not
permitted.’’
possibility to use any number of anonymous online identities has substan-

tially simplified undetectable shill bidding.
3.1 Theoretical considerations
A public reserve price can increase the revenue of a seller, both in the
independent-value setting and in the affiliated or common-value settings. In
the first case, the seller should choose a reserve price that maximizes his
expected income by extracting the expected surplus from the highest bidder.
This optimal reserve price will typically be well-above the seller’s valuation.
Since the seller does not know the realized buyer valuations, the optimal
reserve price will sometimes turn out to be higher than the highest bidder’s
valuation, in which case no trade will occur. Hence, analogous to the case of
monopoly pricing, the optimal reserve price raises expected revenues for the
seller but leads to inefficiency through reducing the quantity of trade. A
number of authors have argued that this inefficiency cannot persist, because
it entails incentives for re-negotiations and re-auctioning, unless the item is
extremely perishable. The modeling of post-auction re-sale leads to different
conclusions about equilibrium auction outcomes and optimal reserve prices
(Haile, 2000; Zheng, 2002).
In the case of affiliated and common values, where bidders do not know
their own values with certainty, the seller’s revenue will usually increase
with a public reserve price. Since the announcement of the reserve price may
improve the bidders’ information about the true value of the auctioned
item, the bidders can reduce their cautious bid shading and submit higher
bids (Milgrom and Weber, 1982). At the same time, however, a public
reserve price may reduce the amount of information available to the active
bidders in an ascending bid auction. Since, in this case the bidders with
signals below the reserve price cannot participate, their signal information
is not revealed in their bids.
The theoretical results mentioned above are based on the assumption of a
fixed and known number of bidders who incur no cost of entry. When the
number of bidders is endogenous (i.e., bidders can choose whether or not to
participate) and bidders have some cost of entry, it may be advantageous
for the seller to set the reserve price no higher than her valuation in order to
encourage efficient levels of entry (Samuelson, 1985; Engelbrecht-Wiggans,
1987; McAfee and McMillan, 1987; Levin and Smith, 1996).
The theoretical effects of secret reserve prices are also rather mixed. The
obvious market-design question is whether the use of a secret reserve price
is more beneficial than a public reserve price (minimum bid). Tatsuya
Nagareda (2003) models a second-price sealed-bid auction where the seller
can either set a public or a secret reserve price. He finds that no symmetric
equilibrium exists in which secret reserve prices increase the expected rev-
enue of the seller. Other researchers, such as Elyakime et al. (1994) analyze
an independent-value first-price auction and conclude that a seller is strictly

worse off using a secret reserve price versus a minimum bid.
Not all theoretical models predict a disadvantage to secret reserve pricing.
Li and Tan (2000) focus on risk-averse bidders rather than risk-neutral
bidders. The authors find that with risk-averse bidders, a secret reserve may
increase the seller’s revenue in an independent, private-value, first-price
auction. On the other hand, in second-price and English auctions, risk
preference does not play a role and the seller should be indifferent between
a private or public reserve price. The work of Vincent (1995) provides an
example where setting a secret reserve price in an English or second-price
auction can increase a seller’s revenue in an affiliated-values setting. He
argues that since a nonzero minimum bid can cause some bidders to avoid
the auction entirely, the attending bidders will have less information than in
an auction with a secret reserve price, but no minimum bid. As usual, less
information on other bidders’ signals in an affiliated-values auction leads to
more cautious equilibrium bidding and hence lower prices.
With shill bidding, the informational situation is rather different than in
the case of secret reserve prices. For one thing, the bidders in such auctions
receive no explicit notice that seller is effectively imposing a reserve. Of
course, in an institution in which shill bidding is possible, buyers may expect
it to happen. In fact, Izmalkov (2004) shows that in an English auction with
asymmetric independent private values, there exists an equilibrium with shill
bidding that has an equivalent outcome to that of Myerson’s (1981) optimal
mechanism. The intuition for this result is best described by Graham et al.
(1990), who show that setting a reserve price dynamically, that is after having
observed some bidding, can increase the seller’s revenue. The effect is due to
the fact that the longer the seller can observe bidding in the auction, the more
precise becomes the seller’s information buyers’ values. In the absence of
penalties for shilling, the seller should weakly prefer an adjustable reserve
price (e.g., a shill bid) to an ex ante fixed reserve price.
Similar, more detailed results for the independent-private-value ascending
auction are shown by Sinha and Greenleaf (2000) and Wang et al. (2004).
For example, Sinha and Greenleaf’s discrete-bidding model shows that the
advantage to the seller of shilling depends on a number of parameters,
including the number of active bidders and their ‘‘aggressiveness’’ as well as
the sequence of buyer and shill bids. They find that, in some cases, sellers
should optimally refrain from shill bidding and credibly commit to their
abstinence. This commitment ensures that bidders shade their bids less than
when they fear shill bidding occurs.
Chakraborty and Kosmopoulou (2004) derive a similar result for the case
of dynamic shill bidding in a common-value setting with an ascending
auction. Their model uses a pool of bidders who are only able to see a
binary signal for the value of the good (high or low). They show that as a
seller increases her rate of shill bids, while holding bidder behavior con-
stant, she increases his selling price, since common-value bidders infer
higher item value from greater participation. However, since bidders guess
that the seller will employ shill bidding, they decrease the amount of their
bids, which lowers the final sale price. Furthermore, when the seller happens
to win her own auction with a shill bid, she must pay a fee to the auctioneer
without actually making a sale. If these two negative effects outweigh the
potentially higher selling price, the seller would prefer to commit to a policy
of no shill bidding. However, the seller has a credibility problem with
committing to not shill because in any auction where bidders do not believe
shill bidding is occurring, a seller has a clear incentive to shill in order to
increase the final price. Given the seller’s lack of credibility, bidders should
always believe that shill bidding will occur and lower their bids accordingly.
Finally, some features unique to online auctions make shilling behavior
more attractive to the seller. Engelberg and Williams (2005), for example,
argue that dynamic shill bidding is strongly supported by eBay’s system of
‘‘proxy bidding’’ to approximate a second-price auction. The bidder with
the highest bid in an eBay auction is called the ‘‘high bidder’’ and holds the
‘‘current bid’’ that usually is equal to the next highest proxy bid plus a fixed
bidding increment. One important exception is that if the proxy bid of the
high bidder is not large enough to provide a full minimum increment over
the second-highest bid, then the current bid is set to exactly the value of the
high bidder’s proxy bid. In that event, the second-highest bidder can infer
that the high bidder’s proxy bid amount has just been reached.
Engelberg and Williams point out that this feature of eBay’s rules facil-
itates a ‘‘discover-and-stop’’ shilling strategy.17 They observe that most bid-
ders enter bids of whole- or half-dollar amounts, so a particularly effective
shilling strategy would place bids with unusual decimal parts, for example,
making all shill bids end in 37 cents. Continuously increasing the shill bid by
the minimum bid increment up to the point in which the current bid is no
longer increased by the full amount of the minimum bid increment, allows
sellers to squeeze the full value of the item from the highest bidder, while
minimizing the chances of winning the auction and receiving no revenue.
For example, suppose that the high bidder has submitted a proxy bid
of $7.50 and that the minimum bid increment is 50 cents. The seller first
submits a shill bid of $5.37, which results in eBay showing the high bid
as $5.87. He continues with a series of shill bids increasing in one-dollar
steps: $6.37, $7.37. At this point, the system reports a current bid of $7.50
instead of $7.87, revealing the high bidder’s maximum bid to be exactly
$7.50. At this point, the shill bidder stops, extracting the bidder’s full will-
ingness to pay, without risking overbidding the true bidder and failing to
sell the item. Engelberg and Williams (2005) conjecture that a bidder’s
best response to this discover-and-stop shill strategy is one of sniping, or
17
Shah et al. (2002) describe a similar pattern of ‘‘uncovering’’ behavior, without the special eBay
details.
withholding their true bid until the last seconds before the closing of the
auction (see Section 4).18
Overall, it seems clear that the theoretical predictions concerning the
different variants of reserve prices depend on many details. Clearly, all of
the reserve-price variants can, under some circumstances, increase the
seller’s revenue. However, the rational reaction of the buyers often involves
increased bid shading. As the sellers become smarter in their effort to un-
cover bidder’s valuation information, bidders should become ever more
cautious about revealing information in their bids. Hence, empirical and
experimental work is needed to assess the actual economic impact of reserve
prices and to compare the effects of different reserve-price strategies.
3.2 Empirical and experimental observations
Many of the theoretical results concerning public and secret reserve prices
depend on the details of the models used. The effects of reserve prices are
especially sensitive to theoretical assumptions about the information and
valuations of bidders. While the empirical and experimental research that
we introduce in this subsection provides some help in sorting out and
matching the appropriate models and auction situations, we will also see
many questions left open for further research.
3.2.1 Entry and revenue effects of public and secret reserve prices
Although details may differ, there are a few predictions that are shared by
all theoretical models. These basic predictions are among the earliest stud-
ied in the field. The first basic hypothesis is that reserve prices (whether
public or secret) should lead to a decrease in the number of bids and the
number of bidders in an auction. The second hypothesis is that the number
of auctions ending without a trade should increase when reserve prices are
used. What overall effect these two observations have on average prices
(i.e., revenues) depends on the details of the theory used.
An early test of these hypotheses was performed by Reiley (2006). In his
field experiment, collectible trading cards from the game ‘‘Magic: The
Gathering’’ were sold in first-price sealed-bid auctions on Internet news-
groups. The size of the minimum bid (public reserve price) was varied
systematically as a fraction of each card’s book value, or reference price.
18
In an environment without proxy bidding, Kirkegaard (2005) discovers a related theoretical result:
bidders may wish to submit jump bids as a defense against shilling. By mixing jump bids and normal
bids, bidders obscure the information about their willingness-to-pay, making dynamic shilling less
effective. Note that proxy bidding precludes jump bidding, because a single bidder cannot cause the
current bid to jump above the second-highest bid; someone wishing to execute a real jump bid on eBay
would have to use two different bidder accounts. Note also that Kirkegaard uses the term phantom
bidding, which is technically distinguished from shill bidding because the former refers to the auctioneer
making up a nonexistent (i.e., ‘‘phantom’’) bid, while the latter refers to the seller or a confederate
actually placing a bid with the auctioneer.
The main results of the experiment are consistent with the basic hypotheses
above: holding all other variables constant, the use of a public reserve price
(1) reduces the number of bidders, (2) increases the frequency with which
goods go unsold, and (3) increases the revenues received on the goods
conditional on their having been sold. Furthermore, bidders clearly exhibit
strategic behavior in their reactions to the public reserve prices. High-value
bidders, for example, raise their bids above the reserve in anticipation that
rival bidders will do the same.
Ariely and Simonson (2003) study eBay auction prices for tickets to the
2000 Rose Bowl (a popular, American, collegiate football game). They
found that the minimum bid and the total number of bids have a positive
correlation to the price. Unfortunately, the authors do not report the in-
teraction between the minimum bid and the number of bids. According to
the first basic hypothesis and given the evidence from the other empirical
studies, we should expect that the number of bids will depend on the value
of the minimum bid and cannot simply be viewed as an exogenous param-
eter. In fact, in a follow-up field experiment on VHS, DVD, CD, and book
sales on eBay reported in the same paper, the authors observe that bidder
activity and the number of bids submitted were greater for a low minimum
bid than for a high one. This clearly indicates that activity measures such as
the number of bids should be treated as endogenous parameters in empir-
ical work on auctions.
With their field experiment, Ariely and Simonson (2003) show that there
is another exogenous parameter that may affect the market activity level
and may interact with the effect of the minimum bid on prices. This ex-
ogenous parameter is a measure of the amount of supply by other sellers.
When many sellers are offering identical (or very similar) items at the same
time,19 then auctions with both high and low minimum bids end at roughly
the same price. Thus, a high degree of supply by other sellers reduces the
effect of the public reserve price. By contrast, when there are few other
sellers offering the same item, a high minimum bid yields empirically higher
auction prices. Note that this effect is not in line with standard auction
theory. In a standard auction model, we would expect high seller-side
competition to lead to a lower number of bidders per auction. This, in turn,
should make the use of a public reserve price more valuable to the seller
because it helps to drive up the price especially in ‘‘thin’’ markets. Hence,
standard auction theory would predict the greatest difference between auc-
tions with low and high minimum bids when seller-side competition is high.
The authors explain their results by claiming that the low seller-side com-
petition reduces the probability that bidders compare competing auctions
and, thus, enables the minimum bid in each auction to be a much more
effective psychological ‘‘anchor’’ for the bidding behavior.
19
Ariely and Simonson term this ‘‘high comparability.’’
The anchoring hypothesis finds some support in the field and laboratory
auction experiments reported by Häubl and Popkowski Leszczyc (2003). In
their field experiments, they auction identical postage stamps, while sys-
tematically varying the minimum bid, and the shipping and handling costs.
They find that the total selling price substantially increases with the min-
imum bid and the fixed cost of shipping. The effect seems especially strong
when the true value of the item is harder to estimate. However, an external
reference point does not alleviate the effect. Hossain and Morgan (2006)
conduct eBay field experiments for Xbox games and music CDs, system-
atically varying the shipping costs. They find that for Xbox games, setting a
low minimum bid and a high shipping cost yields more revenue than doing
the reverse. Buyers do not seem to take the extra shipping cost as much into
account as the stated minimum bid in the auction. They do not discover the
same effect for music CDs; when the shipping costs were a substantial
fraction of the item’s selling price, the bidders took shipping costs into
account just as much as minimum bids. These results are consistent with
psychological anchoring effects being present but limited in scope.
Anchoring can obviously be effective only with public reserves, not secret
ones. Hence, if the anchoring hypothesis is true, we should expect to ob-
serve a greater correlation between a public reserve price and the auction
outcome, than between a secret reserve price and the auction outcome.
Unfortunately, we know of no study so far, that systematically varies the
amount of the secret reserve price. There are, however, a number of studies
comparing public and secret reserve prices.
In an early empirical study, Bajari and Hortac- su (2003a) examined the
effects of minimum bids and of secret reserve prices in all 407 auctions for
mint and proof sets of US coins that occurred on eBay during a week in late
September of 1998. Only 14% of the observed auctions used a secret re-
serve, with the average book value of these items being more than 20 times
higher than the average of items without a secret reserve. While 84% of the
items without a secret reserve sold, only 49.1% of the items with a secret
reserve sold. Surprisingly, the average number of bidders for auctions with
a secret reserve was substantially higher (5.0) than in the other auctions
(2.8). This correlation disappears when the confounding effect of item value
is taken into consideration: high-value items generally induce more bidding
activity, and are also more likely to have secret reserves. However, these
high-activity auctions are also associated with lower bids relative to book
values. Overall, the results suggest that a secret reserve has less of an entry-
deterring effect than a public reserve, but a secret reserve does have a
positive effect on revenue. Hence, the authors suggest that a combination of
a low minimum bid and a secret reserve is likely to be the optimal con-
figuration from a seller’s point of view, especially in auctions of high-value
items.
Dewally and Ederington (2004) also analyzed eBay auctions to measure
the impact of secret reserve prices on bidding strategies and final auction
prices. Their data on 5,275 auctions of classical Silver Age comic books
was gathered in 2001 and 2002. Unlike Bajari and Hortac- su (2003a), they
find clear statistical evidence that the use of a secret reserve reduces the
number of active bidders in an auction and, thus, has a negative impact on
the seller’s revenue. This result is strongly supported and extended by
Katkar and Reiley (2005), who report on a field experiment in which they
auctioned on eBay 50 matched pairs of Pokémon trading cards. One card
of each pair was auctioned with a minimum bid, while the other was auc-
tioned with an equivalent secret reserve price. On average, the secret-reserve
auctions return 10% less revenue and are more than 30% less likely to end
in a sale.
3.2.2 Auction fever

Auction fever is one of the most frequently discussed issues concerning
online auctions.20 In general, auction fever is thought to be an excited and
competitive state-of-mind, in which the thrill of competing against other
bidders increases a bidders’ willingness to pay in an auction, beyond what
the bidder would be willing to pay in a posted-price setting. Since auction
fever supposedly derives from the thrill of competition, one might reason-
ably expect the effect to increase with the number of active bidders. This
theory may explain why some auctioneers prefer a low minimum bid, per-
haps lower even than the auctioneer’s true willingness to accept. The low
minimum bid would attract as many bidders as possible, in an attempt to
promote auction fever. (In case auction fever is insufficient, shill bidding
could prevent the item being sold below the seller’s reservation price.)
It is important to note that auction fever (sometimes also referred to as
competitive arousal, bidding frenzy, or bidding war) generates a diametrically
opposite prediction to both the standard auction-theoretic argument for
reserve prices and the anchoring hypothesis described above. While those
imply that installing a high public reserve price will increase seller’s rev-
enues, auction fever predicts that a low public reserve price will create a
more competitive atmosphere, which in turn leads to bidders’ arousal,
higher bids, and higher auction revenues.
Häubl and Popkowski Leszczyc (2004) run a series of experiments in
which they vary the frequency of bid arrivals and the perceived total
number of active bidders in order to find evidence for auction fever.21 They
find both parameters to have positive, significant effects on revenues,
20
In many of its ads, eBay actually advertises with an image of aroused bidders enjoying the thrill of
bidding and the joy of winning.
21
To be able to manipulate these parameters in a controlled manner, Häubl and Popkowski Leszczyc
(2004) let each subject play against bidding robots that are programmed to create each treatment
environments. In order not to disturb the emergence of auction fever, the authors mislead their subjects
to believe that they are bidding against other human players. Strictly speaking, it is not clear to what
extent subjects from a subject pool that has been exposed to deception in experiments, will actually
believe any part of the information that they are given.
indicating that auction fever may be effectively pushing up prices. No such

effect is observed when bid increments and time pressure are varied.
Ku et al. (2005) explore field and survey data of online and offline auc-
tions to look for evidence for competitive arousal. The survey results seem
to provide evidence for auction fever. In addition to the evidence for auc-
tion fever, the authors also find evidence of overbidding due to an attach-
ment effect, which is when long bidding durations and other sunk costs
intensify the desire to win the auction, leading to increased revenues for the
seller. Both effects are also observed in a controlled laboratory experiment,
in which the sunk-cost parameter and the number of bidding rivals were
varied.
Heyman et al. (2004) also examine these two phenomena of competition
and attachment. They use the term ‘‘opponent effect’’ to describe the
arousal due to competing with others and the term ‘‘quasi-endowment’’ for
increased valuation due to having been attached to the item as the high
bidder over a long period. In two experiments, one comparing different
hypothetical scenarios, and one examining real transactions in the labora-
tory, they vary the number of rival bids and the duration of quasi-endow-
ment (i.e., time spent as the high bidder). Both an increase in the number of
rival bids and an increase of the duration of quasi-endowment have a pos-
itive effect on the final price. The authors conclude that sellers may be able
to increase revenues by increasing the total auction duration and by low-
ering the minimum bid in order to induce more ‘‘feverish’’ bidding.
The evidence to date suggest that auction fever may be a real phenom-
enon. This implies that sellers might be able to increase revenues by setting
very low minimum bids in order to increase the number of active bidders;
however, that prediction has not yet been tested directly.22 These studies
also report somewhat higher bids from those bidders who were the high
bidder for a longer duration (and hence would have more opportunity to
become ‘‘attached’’ to the item).
3.2.3 Shill bids

Identifying shill bids in field data is an extremely difficult task, even
though most online auction sites provide detailed bid-history informa-
tion for every auction. The most crucial issue is that online identities are
easily created and cannot be tracked back to the physical identities
without inside information, nor is it easy to prove collusion among le-
gitimate seller identities. This means that, in theory, a seller can create
22
Reiley (2006) presents some evidence along these lines in field experiments involving first-price
sealed-bid auctions: realized prices were slightly higher with no minimum bid than with a minimum bid
at 20% of book value. However, ‘‘auction fever’’ typically refers to ascending-bid auctions where
bidders can observe the number of competitors, not to sealed-bid auctions where the number of com-
petitors is uncertain.
any number of online buyer identities and have them bid on the items
auctioned. Obviously, creating online identities is not free of cost: at
the very least, the seller will incur opportunity and effort costs. In
practice, this means that sellers who shill will have some incentive to try
to minimize the number of fake identities they use to cover up the shill
bidding. But even with a small number of fake online identities, iden-
tifying the connection between a seller and the seller’s confederates
remains a difficult empirical task. Proving such a connection is even
more difficult.
Some authors have presented clever approaches to this difficult pro-
blem. Kauffman and Wood (2003) gathered data from eBay auctions of
rare coins. Their central instrument for the detection of shill bidding
consists of the search for ‘‘questionable bids,’’ meaning bids that appear
to be strictly dominated from the bidder’s point of view, but could be
rationalized as a seller’s shill bid. Kauffman and Wood (2003) consider
the following criteria to detect questionable bids: (1) there are two
auctions of identical items at about the same time, where auction A ends
before auction B; (2) the questionable bid is placed in auction B, even
though an equal or lower bid in A would have been the highest bid in A;
(3) the bidder who submitted the questionable bid in B, did not bid in
A. Clearly, a buyer who places such a questionable bid could have done
better by submitting a lower or equal bid to auction A, which terminates
earlier than B.
Since questionable bids might appear for reasons other than shill bidding,
Kauffman and Wood introduce additional requirements for identifying shill
behavior. The criteria for detecting shill bids consist of (1) shill bids are
questionable bids; (2) shill bids are submitted by buyers who concentrate
their bidding on the auctions of very few unique sellers; (3) shill bids are
submitted earlier than average; (4) shill bids increase the current bid by
more than the average bid; and (5) shill bids are less likely to win the
auction than the average bid. Of the more than 10,000 auctions examined
by Kauffman and Wood (2003), 622 auctions (i.e., 6%) met their criteria for
shill bidding. The authors also find that the probability of observing shill
bidding in an auction increases when the minimum bid is relatively low,
when the book value of the item is higher, when the auction duration is
relatively long, and when the seller’s other auctions appear to feature shill
bidding. In a second paper, Kauffman and Wood (2004) show that when
they identify an auction as having a shill bid, that auction tends to yield a
higher price.
In another empirical study of shill bidding, Engelberg and Williams
(2005) concentrate on finding evidence for the discover-and-stop technique
described above. They examine almost 40,000 event-ticket auctions on
eBay in September 2004, collecting additional information about the other
auctions participated in by the buyers and sellers in that sample. In
their 40,000 auctions, they find that 3% of all bids are discover-and-stop
bids,23 and estimate that half of these, or 1.5%, are intentional shill bids.
They also find that when bidders bid frequently on the same seller’s auc-
tions, the observed bids are more likely to be discover-and-stop shill bids.24
Hoppe and Sadrieh (2006) take a completely different approach than
those of the empirical studies. They conduct a field experiment in which
they auction blockbuster-movie DVDs and collectible-coin bundles. They
auction three copies of each item simultaneously in three different treat-
ments: (1) an auction with the minimum bid set to the lowest possible value,
(2) an auction with the minimum bid set to about 60% of the book value,
and (3) an auction with the minimum bid set to the lowest possible value,
but with a later shill bid at about 60% of the book value. Both the literature
on auction fever and the notion of avoiding the auctioneer’s reserve-price
fee suggest that the last treatment with the shill bid should result in the
highest seller profit. In fact, the experiment shows that sellers in online
auctions may have good reasons to use this latter setup. Although the
observed prices are indistinguishable between treatments, seller profits are
significantly higher in the two treatments with low minimum bids because
those strategies avoid paying additional fees for setting a high public reserve
price. Interestingly, the effects are very similar no matter whether the auc-
tioned item is in a ‘‘thick’’ market (blockbuster DVD) or in a ‘‘thin’’ market
(unique collector’s bundle).
Hoppe and Sadrieh (2006) observe no additional bid shading in the auc-
tions with a shill bid; bidders in this environment would have no idea which
auction (if any) involved shill bidding. By contrast, Kosmopoulou and De
Silva (2005) verify the theoretical prediction of Chakraborty and Ko-
smopoulou (2004) that providing the subjects with knowledge about con-
federate bidding would induce them to increase their bid shading. In their
laboratory experiment, subjects first participated in a number of ascending-
price auctions without bid shilling, before the sellers were allowed to par-
ticipate in bidding. Once the ability of the seller to participate was an-
nounced, the bidders’ bid levels dropped, and average seller profit dropped
from 97.5% to 88.9% of the item value. These findings highlight the seller’s
dilemma: sellers make more profit when there is no possibility of shilling,
yet they always have an incentive to shill when bidders believe that shilling
is not taking place. Once a seller’s ability to shill is recognized by the
bidders, bid levels drop, regardless of whether shill bids can actually be
observed.
23
Engelberg and Williams define a discover-and-stop bid to be a bidder who bids twice, incrementally,
within 10 min, and stops bidding as the second-highest bidder with evidence that the high bidder’s bid is
less than one increment higher.
24
Shah et al. (2002) use data-mining techniques, on 12,000 eBay auctions of video-game consoles, to
look for suspicious relationships between buyers and sellers, finding evidence suggestive that sellers often
use multiple pseudonyms or confederates to execute their shilling. They point out the possibility of a
shilling strategy of ‘‘unmasking,’’ similar to Engelberg and Williams’ discover-and-stop strategy, and
propose (but do not execute) an analysis similar to that later performed by Engelberg and Williams.
To summarize the results on shilling, we note that experimental research

has shown that shill bids are effective tools for sellers in ascending-bid auc-
tions, and that shill bids allow sellers to increase profits merely by escaping
reserve-price fees. However, theory and experiments also suggest that sellers
might be better off if the auctioneer could find some way to guarantee the
absence of shill bidding. Empirical analysis suggests that shilling probably
takes place in more than 1% but less than 10% of eBay auctions. eBay could
probably reduce the rate of shilling not only by tracking close relationships
between certain buyer and seller usernames, but also by changing its bidding
system to thwart the discover-and-stop shilling strategy. For example, one
possible remedy would be to eliminate the kinds of odd-number bids most
effective in the discover-and-stop strategy, requiring all bids to be an even
multiple of the current bid increment. Enforcement of rules against shilling
can be quite tricky, especially as sellers become more sophisticated: a new
service called XXLSell.com now provides a service (at least to German-
speaking online-auction sellers) that automates the process of shilling, ap-
parently using thousands of other XXLSell members’ usernames to execute
the shills in a way that is more difficult to detect.25
4 Late and incremental bidding
Many researchers found that bids on eBay, where auctions run typically
for a week, often arrive very near to the closing time—a practice called
‘‘sniping.’’ For instance, in the sample of computer and antiques auctions
with at least two bidders, Roth and Ockenfels (2002) found that about 50%
of all auctions still have bids in the last 5 minutes; 37% in the last 1 minute;
and still 12% in the last 10 seconds. Bajari and Hortac- su (2003a) found that
32% of the bids in their sample are submitted after 97% of the auction has
passed. Anwar et al. (2006) noted that more than 40% of the bids in their
eBay sample are submitted during the final 10% of the remaining auction
time. Simonsohn (2005) reported that in his sample almost 20% of all
winning bids are placed with just 1 min left in the auction, and Hayne et al.
(2003b) reported that bidding in the last minute occurs on average in 25%
of their sample of 16,000 auctions. Regarding the whole distribution of the
timing of bids, Roth and Ockenfels (2000) and Namazi (2005) observed that
bid-submission times on eBay follow a power-law distribution with most
bids concentrated at the closing time. Shmueli et al. (2005) added that the
start of an auction also sees an unusual amount of bidding activity.26
25
Thanks to Henning Krieg for pointing out this interesting new development to us.
26
There appear to be differences with respect to sniping frequencies across countries. Wu (2004) noted
that there is much less sniping on eBay’s Chinese platform Eachnet. However, one might speculate that
Eachnet was not as mature as other platforms at the time of the study; the feedback score of most
bidders was zero, and there were almost no competing auctions. Hayne et al. (2003a) reported that in
their sample bidding occurs in the last minute of an auction with, for instance, 12% probability in
United Kingdom and 36.5% probability in Sweden.
At first glance, last-minute bidding of this sort cannot easily be reconciled

with economic theory. As explained in Section 2, eBay makes available a
software bidding agent, called ‘‘proxy bidding,’’ to make bidding simple
for bidders without having to be constantly vigilant or online at the close
of the auction. As a consequence, not the last bid (as in ascending-price
auctions) but the highest bid wins, regardless of submission time. Further-
more, there is a risk involved in late bidding in online auctions. Because the
time it takes to place a bid may vary considerably due to erratic Internet
traffic or connection times, last-minute bids have a positive probability of
coming in too late (after the close of the auction).27 eBay explains the simple
economics of second-price auctions and the risks involved in late bidding
and comes to the conclusion: eBay always recommends bidding the abso-
lute maximum that one is willing to pay for an item early in the auction.
(y) If someone does outbid you toward the last minutes of an auction, it
may feel unfair, but if you had bid your maximum amount up-front and let
the Proxy Bidding system work for you, the outcome would not be based on
time.
However, Ockenfels and Roth (forthcoming) demonstrated within an
auction theoretic model that sniping on eBay could be a best response to
a variety of strategies. In particular, inexperienced, ‘‘naı̈ve’’ bidders might
mistake eBay’s proxy system for an ascending-price auction, and thus
continually raise their bids to maintain their status as the high bidder. In
an eBay style auction that closes at a predetermined deadline (‘‘hard
close’’), bidding very late might be a best response to ‘‘incremental bid-
ding’’ (or multiple bidding) of this sort. That is, bidding very near the end
of the auction would not give the incremental bidder sufficient time to
respond, so a sniper competing with an incremental bidder might win the
auction at the incremental bidder’s initial, low bid. In contrast, bidding
one’s value early in the auction, when an incremental bidder is present,
would win the auction only if one’s value were higher than the incre-
mental bidder’s, and in that case one would have to pay the incremental-
bidder’s value.
Late bidding may also be a best response to other incremental bidding
strategies for private-value model environments. One of these strategies is
shill bidding by confederates of the seller in order to push up the price
beyond the second-highest maximum bid. Barbaro and Bracht (2004),
among others, argue that bidding late may protect a bidder from certain
shill bidding strategies. Engelberg and Williams (2005) demonstrate how
27
In a survey of 73 bidders who successfully bid at least once in the last minute of an eBay auction, 63
replied that it happened at least once to them that they started to make a bid, but the auction was closed
before the bid was received (Roth and Ockenfels, 2002). Human and artificial bidders do not differ in
this respect. The last-minute bidding service esnipe.com, which offers to automatically place a prede-
termined bid a few seconds before the end of an eBay auction, acknowledged that it cannot make sure
that all bids are actually received on time by eBay.
shill bidders may use incremental bids and eBay’s proxy-bid system to make
bidders pay their full valuations.
An additional reason for rational late bidding is given by Rasmusen
(2003), where multiple bidding is caused by uncertainty over one’s own
private valuation (see also Hossain, 2006). He argues within a game-
theoretic model that bidders are ignorant of their private values. Thus,
rational bidders may refrain from incurring the cost of thinking hard
about their values until the current price is high enough that such thinking
becomes necessary. Note that this, in turn, creates incentives for bidding
late, because it prevents those incremental bidders from having time to
acquire more precise information on their valuation of the object being
auctioned.
Another rational reason for incremental bidding is that bidders may be
reluctant to report their values, fearing that the information they reveal
will later be used against them (see Rothkopf et al., 1990). While the
highest maximum bid is kept secret on eBay, it sometimes happens that
the winner defaults and that then the seller contacts the bidder who sub-
mitted the second-highest bid. If this bidder revealed his value during the
auction, the seller can make a take-it-or-leave-it offer squeezing the whole
surplus from trade. By bidding incrementally, private information can be
protected—but only at the risk that a sniper will win at a price below
one’s value. Other papers refer to emotional factors as explanations for
incremental bidding, such as ‘‘auction fever’’ (Heyman et al., 2004),
escalation of commitment and competitive arousal (Ku et al., 2005). An-
other explanation along these lines is the ‘‘pseudo-endowment effect’’
(Wolf et al., 2005), which posits that temporarily being the high bidder
during an auction increases the bidder’s value. Note that not only is
late bidding a good strategy to avoid incremental bidding wars with
other /ce:italic> emotional bidders, but that late bidding may also serve as a
self-commitment strategy to avoid one’s own bids being affected by auction
fever and endowment effects.
The evidence in the laboratory and the field indicates that incremental
bidding is common, and that sniping is likely to arise in part as a response to
incremental bidding. Wilcox (2000) indicates that the average bidder submits
1.5–2 bids. Ockenfels and Roth (forthcoming) report that 38% of the bid-
ders submit a bid at least twice. Among these bidders, the large majority
submits a new bid after being outbid. In particular, 53% of the last bids of
incremental bidders are placed after the previous bid was automatically
outbid by eBay’s proxy bidding agent (i.e., by another bidder’s proxy that
was submitted earlier in the auction), 34% are placed after the previous bid
was outbid by a newly submitted proxy bid of another (human or artificial)
bidder, and only 13% are placed by the current high bidder (so that the
current price is not changed). Bids per bidder increase with the number of
other bidders who bid multiple times in an auction, which suggests that
incremental bidding may induce bidding wars with like-minded incremental
bidders.28 In a regression study using eBay field data, Wintr (2004) found
that the presence of incremental bidders leads to substantially later bids,
supporting the view that sniping is reinforced by incremental bidding.
Ariely et al. (2005) investigated the timing of bids in a pure private-value
laboratory setting. They observed that early bids are mostly made in in-
cremental bidding wars, when the low bidder raises his bid in an apparent
attempt to gain the high bidder status, while late bids are made almost
equally often by the current high bidder and the current low bidder. That is,
late bids appear to be planned by bidders regardless of their status at the
end the auction.
Incremental bidding is not the only possible cause for late bidding. Roth
and Ockenfels (2002) and Ockenfels and Roth (forthcoming) demonstrate
that there can be equilibria where all bidders submit only one bid late in the
auction, even in purely private-value auctions and even though this risks
failing to bid at all. This kind of equilibrium can be interpreted as collusion
against the seller because it has the effect of probabilistically suppressing
some bids, and hence giving higher profits to the successful bidders. Several
researchers tested the implications of the model, but the model could gen-
erally not be supported. Using eBay field data, Bajari and Hortac- su (2003a)
could not statistically confirm whether early bids lead to higher final prices.
Hasker et al. (2004) as well as Wintr (2004) could not find evidence that the
distribution of final prices is different for winning snipes and winning early
bids on eBay. In a controlled field experiment, Gray and Reiley (2004)
found 2.54% lower prices when the experimenter submitted the bid just 10 s
before the end of the auction compared to when the bid was submitted
several days before the end, though the difference was not statistically sig-
nificant. Finally, in their laboratory study, Ariely et al. (2005) found that
when the risk of sniping is removed, the amount of late bidding goes up.
This evidence also contradicts the ‘‘implicit collusion’’ explanation, for late
bidding should decrease when there is no chance of suppressing bids
through sniping. However, most of the studies reported substantial
amounts of multiple bidding. This again suggests that parts of the snip-
ing behavior can be interpreted as a response to incremental bidders.
Another explanation for incremental bidding without positing inexperi-
ence on the part of the bidders is to note that, if an auction is common-
value rather than private-value, bidders receive information from others’
bids that causes them to revise their willingness to pay. In general, late bids
motivated by information about common values arise either so that bidders
28
They also note that naive English-auction bidders may also have an incentive to come back to the
auction close to the deadline in order to check whether they are outbid. However, the data indicate that
among those bidders who submit a bid in the last 10 min of an eBay auction, one-bid bidders submit
their bid significantly later than incremental bidders. The data also reveal that bidders with a larger
feedback score tend to submit less bids per auction, suggesting that incremental bidding is reduced with
experience. However, in a study by Hayne et al. (2003b) the bidders who submitted multiple bids had a
higher average feedback score than the average for all bidders.
can incorporate into their bids the information they have gathered from the
earlier bids of others, or so bidders can avoid giving information to others
through their own early bids. Bajari and Hortac- su (2003a) formalize this
idea in a symmetric common value model; Ockenfels and Roth (forthcom-
ing) give an example of equilibrium sniping in a simple common-value
model with asymmetrically informed bidders.
Roth and Ockenfels (2002) provide survey evidence, and Ockenfels and
Roth (forthcoming) provide field evidence, which supports the common
value explanation. They show that there is less last-minute bidding on eBay
computer auctions than on eBay antiques auctions, which supposedly pos-
sess more common value elements. However, the fact that Ariely et al.
(2005) observed substantial sniping in the laboratory for a pure private-
value context strongly suggests the common-value explanation that bids are
interpreted as value signals does not entirely explain the motivations for
sniping behavior.
Another direction for explaining late and multiple bidding is based on the
multiplicity of listings of identical objects, which may create incentives to
wait until the end of an auction in order to see how prices develop across
auctions. Peters and Severinov (forthcoming) propose a model with simul-
taneously competing auctions and argue that late bidding is consistent with
this model. Stryszowska (2005a, see also 2005b) models online auctions as
dynamic, private-value, multi-unit auctions. By submitting multiple bids,
bidders coordinate between auctions and thus avoid bidding wars. In one
class of Bayesian equilibria, multiple bidding also results in late bidding,
even when late bids are accepted with probability smaller than one. Wang
(2003) shows theoretically that in a twice-repeated eBay auction model,
last-minute bidding is in equilibrium and offers some field evidence for this.
Anwar et al. (2006) provide evidence suggesting that eBay bidders tend to
bid across competing auctions and bid on the auction with the lowest
standing bid. This seems to support the idea that the incentives to bid late
are amplified when there are multiple listings of the same item.
Some observers of eBay believe that the amount of sniping will decrease
over time because it is mainly due to inexperience and unfamiliarity with
eBay’s proxy bidding system. This section showed, however, that there are a
variety of rational, strategic reasons for sniping. It is a best response to
naı̈ve and other incremental bidding strategies, and can even arise at equi-
librium in both private- and common-value auctions. In fact, Wilcox
(2000), Roth and Ockenfels (2002), Wintr (2004), and Ariely et al. (2005)
observed, both in laboratory and field studies, that more experienced bid-
ders snipe more often than less experienced bidders.29 Thus, as long as the
29
Simonsohn (2005) investigated the consequences of such lateness on the strategic behavior of sellers.
The idea is that because many bidders snipe, an auction’s ending-time is likely to influence the number of
bidders it receives. In fact, he found that a disproportionate fraction of sellers set the ending-time of
their auctions to hours of peak-demand.
auction rules remain unchanged, it seems likely that late bidding will remain
a persistent phenomenon on eBay.
5 The buy-now option
A feature of online auctions increasing its share of market transactions is

the buy-now option. This option, also known as buyout option allows any
bidder to end the auction early at a buy-now price previously specified by the
seller. When a bidder exercises a seller’s buy-now option, he pays the spec-
ified price (some authors prefer to call this price the buy-out price or the buy
price), in order to receive the item immediately and shut out other bidders.
On eBay, the buy-now price is called the Buy it Now price, whereas on Yahoo
it is called the Buy price. The buy-now option used by eBay and Yahoo differ
by more than just the name. The Buy-it-Now option on eBay is a temporary
buy-now option, available only so long as no bid has yet been placed on the
item.30 The Buy-it-Now option disappears after the first bid. In contrast, the
Buy-price option on Yahoo is a permanent buy-now option and available for
the entire duration of the auction. Other online auctioneers use these options
as well: a survey by Matthews (2005) shows that temporary buy-now options
are available on eBay, LabX, and Mackley & Company, whereas Yahoo,
uBid, Bid or Buy, MSN, and Amazon offer permanent buy-now options.31
The practice appears to be much more common in the online world than in
traditional English auctions. We are not aware of any documented use of
buy-now prices in traditional auctions, perhaps because live auctions usually
take place on much smaller timescales (seconds versus days).
The buy-now option has become increasingly popular with sellers. Ac-
cording to Mathews and Katzman (2006), of the items for sale on eBay,
30% were listed with a buy-now option in the first quarter of 2001 and 35%
in the second quarter of 2001. By December 2001, 45% of items for sale on
eBay were listed with a buy-now option. Hof (2001) generally agrees with
these findings, reporting that about 40% of the items for sale on eBay had
the buy-now option toward the end of 2001. The economic relevance of the
buy-now option can easily be seen in eBay’s quarterly reports. The reports
for 2005 show that eBay’s fixed-price sales, consisting largely of Buy-it-Now
purchases, accounted for $3.2 billion (30% of the gross merchandise sales)
in quarter 1, $3.2 billion (29%) in quarter 2, $3.4 billion (32%) in quarter 3,
and $4.0 billion (34%) in quarter 4.
The existence and increasing popularity of the buy-now option are puz-
zling from the point of view of auction theory. An auction’s primary benefit
30
If the seller has set a secret reserve price, the Buy it Now option remains active until the reserve price
is overbid.
31
This practice developed even before commercial auction websites. Lucking-Reiley (2000a) docu-
ments the use of ‘‘buyout prices’’ in English auctions run by individuals on online newsgroups before the
advent of eBay.
is that it relieves the seller of the job of determining an item’s price, instead
allowing bidders to determine the price by competing with each other.
Introducing a buy price could potentially decrease a seller’s revenue because
when exercised it rules out the possibility of higher prices reached by com-
petitive bidding. If exercised by a bidder with less than the highest value, the
buy price can similarly reduce efficiency. Given the potential inefficiency
and loss of revenue, why is the buy-now option so popular with sellers?
5.1 Explaining the buy-now option with risk-aversion
One of the first proposed explanations for the observed popularity of the
buy-now option is the risk aversion of bidders or sellers. Budish and Take-
yama (2001) show that adding a permanent buy-now option to an ascend-
ing auction can increase the seller’s revenue, in a model with two risk-averse
bidders with only two possible valuations. Reynolds and Wooders (2003)
extend the result to a continuous uniform distribution of valuations, dem-
onstrating that it holds for both types of buy-now options. They show that
optimally chosen buy-now prices are never exercised in equilibrium when
the bidders are risk-neutral. In contrast, when the bidders are risk-averse,
the optimal buy-now prices are exercised with a positive probability, pro-
viding insurance value to risk-averse bidders and increasing the risk-neutral
seller’s revenue.
Though a buy price may in principle allow a lower-value bidder to shut
out the highest-value bidder, Hidvégi et al. (2006) show that in ascending
auctions with a permanent buy-now option and uniformly risk-averse bid-
ders, such displacement of the high valuation bidder will not occur in
equilibrium. Intuitively, with a permanent buy-now price set optimally by
the seller, no bidder immediately jumps to the buy-now price. Instead,
bidders with valuations that are high enough to accept the buy-now price
will first engage in straightforward bidding until the current bid level has
reached a certain threshold. If bidding reaches some bidder’s threshold, that
bidder ends the auction by accepting the buy-now price. Assuming all bid-
ders have exactly the same degree of risk aversion, their thresholds decrease
monotonically in their valuations (i.e., the bidder with the highest valuation
will be the first to jump to the buy-now price). This monotonicity of
threshold values ensures an efficient outcome of the auction, and the op-
timally chosen buy price yields at least as much expected revenue as an
auction without a buy price.32
Hidvégi et al. (2004) note that their efficiency and neutrality results break
down if the permanent buy-now option is replaced by a temporary buy-now
option. Temporary buyout options do not allow for the type of threshold
32
Interestingly, the risk-averse bidders in this model do not receive higher expected utility from the
presence of a buy-now price. Though they receive some benefit from reduced uncertainty, the seller
manages to extract the added bidder surplus with an optimally chosen buy-now price.
strategies discussed above. Both efficiency and seller revenue are lower than
in an auction with a permanent buy price. Permanent buy prices also pro-
duce higher revenues than temporary buy prices in the risk-averse-bidder
model of Reynolds and Wooders (2003) and the impatient-bidder model of
Gupta and Gallien (2006).
Using a slightly different approach and focusing on a temporary buy-
now option, Mathews and Katzman (2006) show that buy-now prices may
increase expected utility for risk-averse sellers facing risk-neutral bidders.
The intuition here is simply that a risk-averse seller may be willing to give
up part of the expected auction revenue to reduce the volatility of auction
revenue. In an extreme case, an infinitely risk-averse seller can choose a
buy-now price low enough for even the lowest-valued buyer type to
accept. This guarantees immediate trade at a fixed price. A seller with less
extreme risk aversion will choose a higher buy-now price, earning higher
expected revenues with nonzero variance. The buy price produces higher
expected utility for the seller even though it may result in an inefficient
allocation.
5.2 Explaining the buy-now option with impatience and other transaction
costs
An alternative explanation for the prevalence of the buy-now option is

the impatience of the trading agents. An impatient bidder may be willing to
pay a premium to receive the item quickly. Similarly, an impatient seller
may be willing to accept a lower price to end the auction early. Indeed, eBay
cited impatience as a reason for introducing the Buy-it-Now price (Den-
nehy, 2000). Mathews (2003) studies both bidder and seller impatience in an
independent-private-valuations auction with a temporary buy-now option.
He shows that impatience on either side of the market creates incentives for
the seller to set a buy-now price that would be exercised with positive
probability. The optimal buy-now price increases directly with bidders’
impatience, inversely with seller’s impatience, and inversely with the
number of bidders.
Gupta and Gallien (2006) also examine the case of ‘‘time sensitive’’ (i.e.,
impatient) bidders in an independent-private-value auction with bidders
arriving at the auction via a Poisson process. In addition to comparing
temporary and permanent buy-now options, they investigate the theoretical
possibility of a ‘‘dynamic’’ buy-now price whose level can be changed by the
seller during the course of the auction. The authors first solve for the Nash
equilibrium in the bidders’ strategies for each of the auction formats, then
use numerical simulations to derive results on the optimal seller choice. The
results show that a dynamic buy-now price barely improves seller utility
compared to a static one, which perhaps explains the lack of dynamic buy
prices observed in the field. The simulations also show that a permanent
buy-now option can enhance revenues far more than temporary buy-now
option can.33
In the models discussed so far, impatience makes auction participation
costly for a bidder relative to a fixed-price purchase. Wang et al. (2004)
suggest that a consumer may have other substantial transaction costs as-
sociated with participation in an auction. Among others, they describe the
cognitive effort that is necessary to observe, plan, and execute the bidding in
an auction.34 In the presence of these bidder transaction costs, buy prices
produce benefits similar to those derived in the above models of bidder
impatience.
5.3 Explaining the buy-now option with a sequence of transaction

opportunities
An alternative explanation for the observed frequent use of buy-now

prices in online auctions rests on the idea that sellers may use them to
optimize revenues intertemporally when identical units of the item come up
for sale at different points in time. Kirkegaard and Overgaard (2003), for
example, examine a sequence of single-unit auctions, run either by the same
seller or by multiple sellers. In the latter case, the first seller can increase her
revenues and decrease the revenues of the subsequent sellers by choosing
the optimal temporary buy-now price. In the former case, when the seller is
a monopolist, the optimal strategy prescribes not to use the buy-now option
for the first auction, but to announce its use for later auctions. Thus, the
buy-now price in this model is a valuable instrument to the seller even
though all agents are risk-neutral and no one is impatient.
Etzion et al. (2003) describe a similar result for the case of a multi-unit
seller who faces a stream of randomly arriving bidders and can sell any
number of items either in unrestricted auctions or with a permanent buy-
now price. The seller uses both mechanisms to effectively price discriminate
between buyers with low versus high willingness-to-pay. The former would
earn negative surplus at the buy-now price, so they merely participate in the
auction. The latter, who could earn positive surplus at the buy-now price,
generally do not do so immediately: they bid in the auction at first, and only
later consider accepting the buy-now price. In a closely related model, Ca-
ldentey and Vulcano (forthcoming) derive a similar equilibrium with a more
complex model of bidders’ utility.
Many online auctions represent clearance sales of overstock items. That
is, both on eBay and on merchandise sites such as uBid.com, auctions take
place for items that failed to sell at a posted price in ordinary sales channels.
33
Interestingly, the model also predicts increased late bidding when a permanent buy price exists. This
result is related to the fact that the auction price approaches the permanent buy price toward the end of
the auction.
34
See Engelbrecht-Wiggans (1987) for additional discussion of auction entry costs.
Obviously, if buyers rationally expect the possibility of purchasing an item

in a later overstock auction, that might affect their initial purchase decision.
Note that this possibility can be modeled as a multiple-item auction with a
temporary buy-now price. If all items sell immediately at the buy-now price,
no auction takes place, but any ‘‘leftover’’ items are sold at auction.
Bose and Daripa (2006) analyze this scenario. They model buyers’ val-
uations rather unconventionally: a buyer either has a pre-specified high
valuation or a valuation randomly drawn from a continuous distribution
strictly lower than the high valuation. The seller’s problem is to price dis-
criminate by making the buy-now option attractive enough to the high-
value buyers so that they do not wish to wait for the auction. Bose and
Daripa show that the seller cannot perfectly price discriminate; the optimal
mechanism involves a temporary buy-now price set low enough to be at-
tractive to some of the low-valuation buyers as well.
5.4 Empirical and experimental evidence
The growing body of alternative theoretical explanations calls for em-

pirical and experimental testing. Some preliminary research is available, but
much work remains to be done, both on drawing out the testable impli-
cations that might distinguish the competing theories, and on collecting
empirical data to test them.
The earliest empirical studies of buy-now prices have mainly generated
descriptive data about the practice. Mathews (2003), for example, reports
the prevalence and execution of the buy-now option in two specific product
categories on eBay (software for the Sony PlayStation). He finds the buy-
now option available for more than half of the items (59%). The option was
exercised in about 27% of the cases in which it was available. In about 62%
of these cases, the buy-now price was below the auction price.
Reynolds and Wooders (2003) also provide frequencies with which the
buy-now option has been used on eBay and Yahoo. They sample a total of
31,142 eBay auctions and 1,282 Yahoo auctions in the categories of au-
tomobiles, clothes, DVD-players, VCRs, digital movie cameras, and TV-
sets. Forty percent of eBay auctions used the temporary Buy-it-Now op-
tion, while 65% of Yahoo, auctions used the permanent Buy-Price option.
Since the permanent buy-now option of Yahoo is chosen more frequently
than the temporary buy-now option of eBay, this gives some empirical
support to those theories predicting that the permanent buy-now option
may be more beneficial for sellers than the temporary option.
Standifird et al. (2004) examine a sample of 138 auctions of American
silver dollars on eBay. They find that the 41 auctions using a buy-now price
(temporary since it is on eBay) result in significantly higher selling prices
(on average $10.27) than the auctions without a buy-now price (on average
$9.26). From these findings, the authors conclude that the buy-now
option increases seller revenue, consistent with the idea that there are some
risk-averse or impatient bidders willing to pay a premium to guarantee a

win or end the auction early.
Anderson et al. (2004) collected data on about 1,000 Palm Pilot Vx auc-
tions on eBay. The goal of the study is to identify and relate typical seller
profiles to typical seller strategies. One of the most interesting findings is
that the high-volume sellers typically use a buy-now price in combination
with a very low minimum bid.
Hendricks et al. (2005) analyze data from almost 3,000 Texas Instruments
TI-83 Graphing Calculator auctions on eBay, where 831 (roughly 30%)
offered a buy-now price. The auctions with buy-now prices appeared to
produce significantly higher revenue than the auctions without buy-now
prices. However, by contrast with the behavior of the high-volume sellers
studied by Anderson et al., a great majority of the auctions with a buy-now
price also featured a high minimum bid, 90% or more of the buy-now price.
Because the buy-now auctions also tended to have higher minimum bids
than the nonbuy-now auctions, it is hard to tell whether the cause was the
buy-now price or the higher minimum bid (see Section 3).
A number of studies have used experimental methods to study the
buy-now option. Standifird et al. (2004) report a field experiment in which
they auctioned 84 American Eagle silver dollars on eBay to study the
impact of varying buy-now prices. Surprisingly, they find that eBay buyers
hardly made use of the buy-now option, even when the buy-now price was
substantially below the prevailing market price. The authors suggest that
buyers may be reluctant to use the buy-now option, in order not to forfeit
the entertainment benefit associated with the participation in an eBay
auction.
Shahriar and Wooders (2005) report laboratory experiments examining
whether sellers can profitably use the buy-now option in independent-
private-value auctions with risk-averse bidders. Common value auctions, in
which the buy-now option theoretically creates no advantage, are also
examined in a control treatment. The authors find that suitably chosen
buy-now prices raise sellers’ revenues in both treatments. They speculate
that the unexpected positive effect of the buy-now price on the sellers’
revenues in the common-value treatment may be due to winner’s-curse-type
overbidding.
Using a similar experimental design, Seifert (2006) reports a strong in-
teraction effect between the market size and the revenue impact of a tem-
porary buy-now option. The experiment shows that buy-now prices have
the expected positive impact on sellers’ revenues with five active bidders,
but the effect is lost when the number of active bidders falls to three.
All in all, the buy-now option remains rather poorly understood. We see
few robust findings, and little ability to discriminate between the different
proposed theoretical explanations. As a relatively new economic phenom-
enon, the buy-now option clearly remains an area in with many opportu-
nities for exciting empirical research.
6 Parallel markets and other outside options
Before the advent of electronic markets, most auctions took place as

fairly isolated events. The items sold in an auction were often not easily
available elsewhere, at least not at feasible time and travel costs.
Furthermore, bidders were constrained to being at one auction at a time
and usually could not bid on several auctions in parallel.35 In general, the
only alternative for a bidder, who did not receive the item in an auction,
was to wait for a subsequent auction. Hence, early auction literature has
been fairly extensive on sequential auctions,36 but has almost completely
ignored settings in which bidders can simultaneously participate in
multiple auctions or use other sales channels to purchase the item. In
such settings, the seller in an auction is no longer a monopolist, but faces
competition by the sellers in other auctions or markets. The bidders do
not only have an entry choice, as in isolated auctions with endogenous
entry decisions, but also have an outside option in the alternative sales
channel.
To see the empirical relevance of online auctions with parallel markets, it
suffices to start up any Internet browser and open one window on a typical
online auction (e.g., eBay or Yahoo!) and another window on a typical
shopbot (e.g., mysimon.com or shopping.com). For hundreds of products,
there are numerous auctions running in parallel and numerous online shops
offering fixed price sales at the same time. Auction sellers are often in a
highly competitive environment that is substantially different from the
monopoly position they are assumed to have in classical auction theory.
Buyers must choose the type of market (e.g., auction or posted offer), the
specific instance of that type (which auction or which posted offer), and—if
necessary—the bid to submit. These are all more complicated strategic de-
cision situations than those that have typically been studied in the auction
theory literature.
The fact that the strategic interaction quickly turns overly complicated,
when competition and outside options are added to the standard auction
models, has limited the amount of theoretical work on the topic. McAfee
(1993) shows that equilibria in games in which sellers compete by offering
different direct mechanisms, may not be feasible in general, due to the
possible nonconvexity and discontinuity of the sellers’ profit functions.
However, under some restrictive assumptions, McAfee (1993) shows the
existence of an equilibrium in which all sellers offer second-price auctions
with the reserve prices set to their marginal cost.
35
Telephone bidding clearly presented the first step toward bidder presence at multiple auctions.
However, since telephone bidding is quite costly and only a small portion of all bidders use it, the
auctioneers often plan their auctions in a way to reduce parallel auctions of the same category of goods.
36
See the overviews in Klemperer (1999) and Krishna (2002).
Peters and Severinov (1997) use a limit equilibrium concept to explicitly

characterize the symmetric equilibrium that arises when the seller’s
competition is restricted to choosing only among auctions (instead of
choosing any arbitrary direct mechanism). The author’s approach also
allows them to examine the efficiency aspects of the auction competition
equilibria. They show that the efficiency of the auction competition
market depends on the seller’s ability to advertise his reserve prices and on
the timing of the buyers’ knowledge of the own valuation. An efficient
market performance, in the analyzed setting, is achieved when sellers can
advertise their reserve prices and buyers learn about the realization of
their valuation only after having chosen a specific auction. If buyers are
informed of their valuations before they select an auction, the reserve
prices are driven down to the sellers’ marginal costs and there is ineffi-
ciently excessive entry. All these results, however, are derived under the
restrictive symmetry assumption that buyers randomize their purchases
over all sellers who offer the same conditions. This is a restrictive as-
sumption; it excludes all possible equilibria in which buyers systematically
sort among sellers.
Building on the two studies discussed above, Peters and Severinov (2006)
characterize a dynamic adjustment mechanism that provides a perfect Bay-
esian equilibrium for a market where sellers compete by offering different
reserve prices in their independent ascending second-price auctions. Buyers
bid in multiple rounds, costlessly and independently adjusting their bids
and moving from one auction to the other, whenever their previous bid was
not successful. The perfect Bayesian equilibrium derived for this decentral-
ized trading institution induces an efficient set of trades at a uniform trad-
ing price.
The strength of this equilibrium is that the decentralized equili-
brium bidding has extremely low informational requirements and neither
depends on a buyer’s beliefs about the other buyers’ valuations, nor on
the number of buyers and sellers. The equilibrium bidding rule only
requires that any buyer, who currently does not hold a high bid, should
bid in the auction with the lowest current high bid, raising the high bid
as slowly as possible and exiting the market if the calculated next bid
is greater than one’s valuation. This minimal increment bidding is a feasible
strategy for most of the existing online auctions because the only
two pieces of information that a bidder needs to follow this strategy
are generally available: (1) the information whether one’s bid is currently
the high bid of an auction and (2) knowledge of which high bids are
currently associated with each of the auctions. Note, however, that while
minimal increment bidding is feasible, it is hardly ever observed in online
auctions.
Given that bidders use minimal increment bidding, Peters and Severinov
(2006) show that in equilibrium all sellers will set their reserve prices equal
to their marginal costs if the number of traders in the market is sufficiently
large. Under these circumstances, the induced equilibrium of the parallel

auctions market is efficient and sequentially optimal at every stage.37
The empirical evidence concerning parallel online auctions is mixed. Tung
et al. (2003) tracked simultaneous online auctions of identical consumer
electronic items, controlling for seller reputation and quality. They report
large price disparities and, hence, substantial arbitrage opportunities across
the auctions. They observed only a few cross-bidders, i.e., bidders switching
from one auction to another. Interestingly, they note that none of the cross-
bidders they identified ever succeeded to buy an item. Furthermore, they
observed that bidders who were outbid in one auction did not switch to
another, even though their unsuccessful bid was substantially higher than
the winning bid of the latter auction. All this seems to indicate that bidding
in online parallel auctions little resembles the minimal-increment bidding
required in the equilibrium specified by Peters and Severinov (2006).
There is, however, also empirical evidence in support of the Peters and
Severinov (2006) equilibrium analysis. Anwar et al. (2006) collected data
from competing eBay auctions for CPUs. Controlling for all auction pa-
rameters, three samples were generated, which differed only in closeness of
the auctions ending times: same day, same hour, and same minute. As
suggested by minimal increment bidding, bidders tend to bid on the auction
with the lowest high bid. Furthermore, the smaller the difference between
the ending times, the more cross bidding is observed. Finally, the authors
show that bidders using cross-bidding strategies on average pay only 91%
of the price paid by those not using cross bidding. This evidence indicates
that the bidders understand the strategic situation of competing auctions
and react to it in an appropriate way. However, the fact that bids are far
from being increased at the smallest increment indicates that bidders do not
actually employ the predicted minimal increment bidding. Additionally, the
price differences among auctions of identical goods seem to suggest that the
observed behavior in the competing auction markets does not perfectly
match the equilibrium behavior predicted by Peters and Severinov (2006).
Stryszowska (2005a) takes a slightly different approach to competing
auctions. She analyzes two simultaneous, second-price, private-value auc-
tions for an identical item.38 The interesting twist in the derived equilibria is
that bidders may use early, low, nonconsequential bids to identify them-
selves and to coordinate across the auctions. This early coordination effort
can explain the multiple bidding that is frequently observed in online
37
Burguet and Sákovics (1999) point out that efficiency breaks down in a setting in which the number
of competing sellers is small enough to allow for strategic interaction. Examining the case of two auction
sellers, who strategically set their reserve prices in a mutual best response, they show that equilibrium
reserve prices are well above the marginal cost, inefficiently excluding some bidders.
38
Instead of examining two parallel auctions, some authors analyze the situation with a sequence of
two auctions that is known to the bidder. Zeithammer (2003) shows that in such a setting, bidders will
‘‘bargain-hunt’’ and reduce their bids, if a more preferred item is next up for sale. A similar approach
can be found in ReiX and Schoendube (2002) and Brosig and ReiX (forthcoming).
auctions. Furthermore, Stryszowska (2005a) shows that in some equilibria

of the game, all crucial bids arrive early, thus inducing an efficient allo-
cation with identical prices in both auctions. However, in another type of
equilibrium, in which bidders send last-minute bids, the auction outcome is
inefficient and prices may be dispersed.
A completely different path for modeling the outside option is taken by
ReiX (2004). Instead of examining parallel auctions that mutually affect one
another, ReiX (2004) aggregates whatever outside opportunity a bidder has
into a simple payoff value that the bidder receives if he is not successful in
the auction. This payoff, for example, can represent the utility gain of
buying the item at a posted-offer shop. Given the model with a bidder
outside option, ReiX (2004) shows that the optimal auction reserve price is
decreases as the outside option increases. In a related model, Kirchkamp et
al. (2004) implement the bidder outside option as fixed payments to all the
bidders who are not successful in the auction. In the experiments, the the-
oretical finding that increasing the outside options decreases aggressive
bidding is reproduced. But, the experiments also show that bidders in sec-
ond-price auctions manage to fully expropriate their outside option, while
those in the first-price auctions fail to do so. Hence, outside options seem to
amplify the well-documented effect that first-price auctions generate more
revenues than second-price auctions. In terms of efficiency, however, nei-
ther auction type is strongly affected, leaving the differences between the
two designs insubstantial.
7 Multi-item auctions
With few exceptions, previous sections dealt primarily with single-item

online auctions. However, almost all online auction platforms also offer
multi-item auction mechanisms. In recent years, multi-item auctions have
received increasing attention, mainly because of their accelerated use in B2B
(business to business) commerce and government allocation procedures. Yet,
both the theoretical and empirical literature is less developed and contains
only few general results. This is partly because when items are heterogeneous
or bidders demand multiple items, new difficulties such as market power and
strategic and computational complexities arise. Here we present the standard
auction mechanisms for selling multiple items, and we discuss some central
intuitions as well as empirical evidence on bidding behavior. For an in-depth
overview of multi-item auction theory, see Milgrom (2004).
7.1 Standard multi-unit auction mechanisms
For selling multiple units of one item (such as car tires, financial secu-
rities, energy products, environmental permits, etc.), there are, analogous to
the single-object case, four standard mechanisms: the descending-price
auction, the sealed-bid pay-as-bid auction, the ascending-price auction, and

the sealed-bid uniform-price auction, plus some variations and extensions.39
In the ascending-price multi-unit auction, the price gradually increases
while bidders indicate how many units they want at each price. The final
price is set and the auction closes when aggregate demand equals the
number of units supplied. All bidders pay the same final price, which is the
price at which the auction closed. In the corresponding sealed-bid version,
bidders independently submit a whole demand curve. That is, each bidder
indicates how much he is willing to pay for the first unit he acquires, the
second unit, etc. Then, the outcome of the auction is determined by finding
the first price at which aggregate demand equals supply. All bids above or
equal to this price win, and bidders must pay their bid price.40 In the sealed-
bid uniform-price auction, all units have the same price: the market-clearing
price.
Note that the ascending-price and the sealed-bid, uniform-price mech-
anisms enforce uniform prices for all units sold. Other auction formats
endogenously promote ‘‘similar prices for similar objects’’ by encouraging
arbitrage. The best-known auction mechanism in this category is the si-
multaneous ascending auction (SAA) developed by Milgrom, Wilson, and
McAfee, who proposed the design in the context of the US radio spectrum
auction. All items, which may or may not be identical, are simultaneously
offered in different auctions. Bidding on all objects takes place simultane-
ously in rounds subject to an activity rule (see Section 8.2). Bidders observe
prices throughout the auction, and this information allows them to arbit-
rage among substitute licenses, and to piece together complementary pack-
ages. The auction ends when a round passes with no new bids on any
licenses (see, e.g., Cramton, 2002, 2004 for more details). For multiple,
divisible items, Ausubel and Cramton (2004) advocate the simultaneous
clock auction, where a price clock for each divisible good indicates its ten-
tative price per unit quantity. Bidders express the quantities desired at the
current prices, and the price is then repeatedly increased by an increment
until demand is made equal to supply, at which point the tentative prices
and assignments become final. This auction also yields similar prices for
similar items by encouraging arbitrage. On the other hand, the next two
multi-unit auction formats are discriminatory; identical units are sold at
different prices.
In the decreasing-price multi-unit auction, the price gradually decreases
while bidders indicate the price at which they are willing to buy one or more
units. At each price, bidders are informed about the supply left at that
39
All auction mechanisms considered here are simultaneous auctions; for a brief discussion of se-
quential auctions, see Section 8.3.
40
When the number of units is an integer, the clearing price may be the lowest accepted bid or the
highest rejected bid. We also note that if the units being sold are not substitutes in the eyes of the
bidders, then market clearing prices can fail to exist; see Milgrom (2004).
point. The auction closes when no supply is left. Each winner pays the price
at which he indicated he was willing to buy. In the sealed-bid pay-as-bid
auction, bidders independently submit a demand curve. Every winning
bidder pays his bid for each unit, provided that the bid is above the clearing
price.
Another multi-unit auction with nonuniform prices was proposed in the
seminal paper by Vickrey (1961). Suppose there are k units for sale. As
before, the highest k bids are accepted, but the pricing rule of the Vickrey-
auction determines that for the kth unit awarded, bidders have to pay the
amount of the kth highest losing bid.41 This rule generalizes Vickrey’s sec-
ond-price auction rule for the single-item auction, where the winner pays
the largest losing bid, to the multi-unit case. In fact, analogous to the single-
object case, all bidders have a dominant strategy to bid true values for all
units.
Search engines such as Google typically use multi-item auctions, distantly
related to the Vickrey auction, to sell online ads. Search results are typically
shown along with sponsored links, which in turn are shown in decreasing
order of bids. If a user of the search engine then clicks on an ad in position k,
that advertiser is charged by the search engine an amount equal to the next
highest bid, i.e., the bid of an advertiser in position k+1. Because there are
multiple positions available, there are many winners, and each winner pays
the next highest bidder’s bid. Edelman et al. (2005) show that this ‘‘gen-
eralized second-price’’ auction generally does not have an equilibrium in
dominant strategies. But, it has an unique ex post equilibrium, resulting in
the same payoffs as the dominant strategy equilibrium of the Vickrey auc-
tion (see Varian, 2006 for another analysis of Google’s ‘‘position auction’’).
7.2 Bid shading and demand reduction in multi-unit auctions
When bidders do not demand more than one unit, the analysis of the
single-item case straightforwardly generalizes. For instance, bid shading
will occur in sealed-bid pay-as-bid auctions (reflecting the trade-off between
the probability of winning and the surplus from winning), and ‘‘truth-
revealing’’ in the ascending-price and the uniform-price auction. In fact,
just as in the single-item case, the standard auction rules are revenue-
equivalent under some appropriate assumptions. However, with multi-unit
demand, bidding incentives can be quite different, and revenue equivalence
fails to hold.
Maybe the most important intuition from the literature is that uniform-
price auctions do not share the desirable properties of the second-price
41
Suppose, e.g., that there are three bidders A, B, and C competing for three units of an object. Bidder
A bids 14, 10, and 2, bidder B bids 12, 9, and 0, and bidder C bids 8, 5, and 4, respectively. Then bidder
A is awarded two units and bidder B one unit. Bidder A pays 17 ( ¼ 9 + 8) for both units, and bidder B
pays 9 for his unit.
auction in the single-item case. The reason is that if a bidder can demand
more than one unit, there is a positive probability that his bid on a second
or later unit will be pivotal, thus determining the price for the first and
possibly other units. With discrete goods, the bidder will bid his true value
on the first unit, but strictly less on all subsequent units. As a consequence,
in equilibrium, bidders understate their values, or (equivalently) reduce
demand quantities which hampers revenue and efficiency.42 Furthermore,
uniform-price auctions typically facilitate (tacit or explicit) collusion. Sup-
pose the bidders agree on a collusive agreement and each bidder bids higher
prices for smaller quantities than his collusively agreed share. Then, if any
bidder attempts to obtain more, all bidders would have to pay high prices.
This stabilizes collusion. So, a key concern with uniform-price auctions is
the possibility of low price equilibria.
Several field studies provide direct evidence of strategic demand reduction
and collusive behavior in electronic auction markets, such as in the German
auction of GSM spectrum (Grimm et al., 2003), in the Austrian auction of
third generation mobile wireless licenses (Klemperer, 2004), in the FCC’s
Nationwide Narrowband Auction (Cramton, 1995), in the UK electricity
market (Wolfram, 1998), and in the California electricity market (Boren-
stein et al., 2002). This field evidence is strongly supported by laboratory
evidence (e.g., Kagel and Levin, 2001b; Engelmann and Grimm, 2004) and
controlled field experiments (List and Lucking-Reiley, 2000). It has also
been shown that, in line with theory, the amount of demand reduction
decreases with the number of bidders (Engelbrecht-Wiggans et al., 2006).
There is, however, little research on multi-unit bidding and demand reduc-
tion in online auctions.
The two most common online multi-unit formats are the ‘‘Yankee auc-
tion’’ (as used by Onsale.com and also called ‘‘Multiple Item Progressive
Electronic Auction’’; see Bapna et al., 2000) and eBay’s ‘‘Dutch auction’’ as
used in the US. Both auction mechanisms allow the seller to simultaneously
auction off two or more units of an item. Bidders must specify in their bid
the price per unit and the number of units desired. That is, unlike in the
standard formats described above, bidders are not allowed to express a
whole demand curve with prices as a function of quantities, but only one
price–quantity pair. Bids are then ranked by price, then by quantity, and
finally by the timing of the bid (earlier bids take precedence). There is no
proxy bidding. During the auction, bids can be improved according to an
improvement rule requiring that the pair value (price times quantity) must
increase with any new submitted price–quantity pair. The most important
difference between Yankee and Dutch auctions is that in Yankee auctions
42
This is similar to market power effects in monopsony. The ranking of the uniform-price auction and
the pay-as-bid auction, where bidder too shade their bids, is ambiguous in both efficiency and revenue
terms (Engelbrecht-Wiggans and Kahn, 1998; Ausubel and Cramton, 2002).
all winning bidders pay their own bids, while in Dutch auctions all winning
bidders pay the same, uniform price, which is the lowest successful bid.43
7.3 Complementarities and combinatorial auctions
Auctioning multiple items quickly becomes complicated when there are

complementarities between items. Complementarities exist when the value
of a bundle of items is larger than the sum of values of each object sep-
arately. This is the case in many applications including auctions for the
radio spectrum, electricity, airport-landing-slot, supply chains, and trans-
portation services. In such cases, a bidder may end up stuck with items that
are worth little because he failed to win complementary items (exposure
problem), or he may quit early, fearing that he may fail to win comple-
mentary items (holdup problem). As a result, inefficiencies are likely to arise
in multi-item auctions where bidders cannot ensure winning complementary
items. Theory suggests that in these situations, a combinatorial auction, in
which bidders can place bids for one or more packages of items, can in-
crease revenue and efficiency. The underlying reason is that these auctions
allow bidders to more fully express their preferences. Applications of com-
binatorial auctions include truckload transportation, bus routes, industrial
procurement, airport arrival and departure slots auctions, radio spectrum
auctions, and course registration at Chicago Business School (see Cramton
et al., 2006; Kwasnica et al., 2005 and the references cited therein).
The most famous combinatorial auction is the Vickrey–Clarke–Groves
(VCR) mechanism, sometimes called generalized Vickrey auction, which
works as follows. Bidders bid on all possible packages. The items are then
allocated to bidders such that efficiency (the sum of realized values) is
maximized according to the stated bids. Each winner pays the smallest
(fictitious) bid such that he would still have won his part of the allocation.
The resulting price for a bidder equals the external costs (social shadow
costs) of winning, in the sense that it is the (stated) value of the awarded
package for the other bidders. Observe that this holds analogously for the
single-object second-price auction introduced in Section 2 and Vickrey’s
multi-unit auction introduced in the last subsection. The VCG mechanism
generalizes these formats. In particular, bidding one’s values for each
package is a dominant strategy.44
However, the VCR mechanism suffers from a number of practical prob-
lems in the presence of complementarities that seem to seriously limit its
43
As long as demand is smaller than supply, the price equals the seller’s reservation price.
44
Bidders must be risk-neutral and are not allowed to face binding budget constraints. Ausubel (2006)
developed an open auction version of the VCG mechanism. Another well-known combinatorial auction
is the pay-as-bid package auction by Bernheim and Whinston (1986), which is relatively easy and
transparent, but strategically much more complex, as it is typically the case with pay-as-bid auction
formats.
usefulness for many applications (Ausubel and Milgrom, forthcoming).45

One is that the VCG mechanism does not maximize revenues. In fact,
revenues can be very low when items are not substitutes though competition
is substantial. This alone disqualifies the auction for many purposes.46
Furthermore, the VCG mechanism makes it easy for losing bidders to col-
lude, and individual bidders can sometimes profit from bidding under
pseudonyms—something that appears to be particularly problematic for
online auctions, where identities can be more easily manipulated. Another
problem is computational complexity. The number of potential bids per
bidder is exponentially growing with the number of items auctioned. There
are 2N1 packages of N items. Bidding on all packages can be too de-
manding for human bidders, even though the VCG mechanism removes all
strategic complexity by implementing dominant strategies. Computational
complexity is also an issue for the auctioneer. Finding the efficiency (or
revenue) maximizing allocation of objects in a general combinatorial auc-
tion is difficult (more precisely: NP-hard; see de Vries and Vohra, 2003),
though researchers succeeded in demonstrating that the ‘‘winner determi-
nation’’ problem can often be reasonably addressed (e.g., Rothkopf et al.,
1998; Sandholm (forthcoming)). Finally, it has been shown that no general
equivalent of the VCG mechanism exists in common-value environments,
and second-best mechanisms have not yet been identified (Jehiel and
Moldovanu, 2001).
That said, there has been much progress in practical multi-object auction
design in recent years. Much of the literature focuses on open, progressive
auctions, which can reduce strategic and computational complexity. Some
researchers argue, based on applied research in electricity and other infra-
structure industry markets, that when complementarities are weak and do
not strongly differ across bidders, auction formats like the simultaneous
ascending auction may work satisfactorily, even though they do not allow
combinatorial bids (e.g., Ausubel and Cramton, 2004). A well-known pro-
gressive auction format that includes combinatorial bidding features is
Banks et al.’s (1989) continuous-package bid auction that tries to reduce
both value and strategic computation (see Kwasnica et al., 2005 for a recent
advancement of this format). Another interesting format to deal with
complementarities is the ascending-proxy auction (Ausubel et al., 2006).
This is a hybrid auction that begins with a simple and transparent clock
phase, not unlike the simultaneous-clock auction, and that ends with a final
proxy auction round based on package bids. Similar to eBay’s proxy bid-
ding system, bidders in the proxy phase submit values to an artificial proxy
agent who then bids on their behalf to maximize profits. It can be shown
45
Maskin (2004) entertains a more positive picture of the potential practical importance of the VCG
mechanism.
46
General results about revenue maximizing auction mechanisms in the private-value multi-object
environment do not exist.
that including such a proxy phase may handle many of the complications
that we discussed above, including the exposure problem.
Some experimental studies, starting with Banks et al. (1989; see also Led-
yard et al., 1997; Plott, 1997), investigate bidding when complementarities
are present. However, to our knowledge, the only experimental paper that
relates its design directly to online auctions is Katok and Roth (2004). They
compared the performance of an auction designed to resemble eBay’s multi-
unit ‘‘Dutch’’ auction to the descending-price auction. The laboratory set-
ting used a set of value environments that include more or less strong
complementarities among homogenous objects. Overall, eBay’s ascending
Dutch auction performed relatively poorly because of the exposure problem.
Recall that while the eBay mechanism guarantees a uniform price for all
units, it does not guarantee a winning bidder the entire quantity on which he
bids. On the other hand, the descending Dutch auction avoids the exposure
problem because a bidder who stops the clock obtains the full quantity he
desires at the price he stopped the auction. In this sense, the descending
Dutch auction can be interpreted as a simple version of a combinatorial
auction in case of homogeneous goods. Katok and Roth (2004) conclude
that eBay’s Dutch auction is susceptible to the exposure problem in envi-
ronments with synergies, but they also mention that synergies may not be
very relevant for B2C and C2C auctions such as eBay. We add that eBay
gives winners the right to refuse to purchase ‘‘partial quantities’’—a rule that
has not been accounted for in the experiment. That is, if a bidder only wins
some of the desired object, he does not have to buy any of them. This rule is
meant to protect eBay users from the exposure problem (but might create
other strategic complications as we will briefly note in the next section).
8 Design of online auctions
Auction design matters. In the previous sections, we have shown that the
choice of the auction format, the reservation price, the buy-it-now price and
other auction parameters may systematically and significantly affect rev-
enue, efficiency, and bidder participation. In this section, we discuss some
further auction mechanism choices relevant to online auctions, which have
been studied in the literature.47
8.1 The advantages of long, open auctions
Unlike offline auctions that typically last only a few minutes, Internet
auctions such as those on eBay, Yahoo and Amazon last many days.48
47
For the question how to promote trust and trustworthiness in online auctions by clever design
choices, see e.g., Dellarocas (2003 and in this handbook), as well as Brosig et al. (2003), Bolton et al.
(2004a,b, forthcoming), Güth et al. (2005), among many others.
48
Google’s and Yahoo’s auctions of online ads are even always accepting bids.
Since bidders may enter an auction from anywhere, and at anytime, a

longer auction time allows more bidders to spot an item and bid on it.
Lucking-Reiley et al. (1999) and Hasker et al. (2004) observed that longer
auction durations on eBay tend to attract more bidders and lead to higher
prices. Lucking-Reiley et al. (1999) reported that 7-day auction prices are
approximately 24% higher than shorter auctions, and 10-day auctions are
42% higher, on average. Hasker et al. (2004) observed that the change in
the final sales price achieved by extending the auction from three to ten days
is about 10.9%.49
Long durations also create challenges because bidders cannot be expected
to continually monitor the auctions. Many auction houses, including eBay,
respond to this by providing bidders with artificial proxy agents. These
agents bid on the bidders’ behalf, automatically responding as other bids
come in, and thus free bidders from the necessity of following the auctions
and the price discovery process themselves.
A related important design question is whether the auction should be
conducted by sealed bid. That is, should bidders submit their (proxy) bids
over an extended period of time, but without the opportunity to react to the
bidding activity of other human or proxy bidders, or should bidding be
open, so that bidders can see how bidding activity evolves during the course
of auction? This way, bidders would retain the right to change (proxy) bids
in response to the bid history.50 Many online auction houses, such as eBay,
chose an open format. From a theoretical point of view, open ascending-
price auctions tend to reduce the force of the ‘‘winner’s curse’’ in environ-
ments with a common-value element because the competitors’ bidding ac-
tivities may convey relevant information that the bidders use in revising
their estimates of value. Thus, uncertainty is reduced and so is the winner’s
curse, and bidders can bid more aggressively. This, in turn, can result in
higher revenues in open auctions (see Milgrom and Weber, 1982 for the
theory and Klemperer, 1999 for a more precise intuition behind the effect).
Recently, Compte and Jehiel (2004) showed that open auctions are also the
preferred choice in private-value environments, if the bidders do not know
ones’ value a priori. So the rationale for using open formats appear quite
robust across auction models.
On the other hand, note that such results, derived in simple auction
models, cannot directly be applied to online auctions. One characteristic of
eBay is that bidders can enter and leave the auction at any point they wish.
So bidding activity, or nonactivity, has less information value than in the
ascending-price auction described above, in which entry and exit decisions
49
Hasker et al. (2004) also reported that experienced sellers respond to these incentives in that they sell
more valuable objects in longer auctions. Simonsohn (2005) found, on the other hand, that too many
sellers set their auctions to end during peak-demand hours such that the probability of sale during such
hours is actually lower.
50
We assume here that the open auction is ascending.
are perfectly observable. Yet, the fact that bidders condition their behavior
on others’ activities (see, e.g., Sections 3 and 4) suggests that open online
auctions reveal some helpful information.
Another argument for open auction formats comes from laboratory ex-
periments. It has been shown that the feedback delivered in open second-
price auctions such as eBay substantially accelerates the speed of learning
compared to second-price sealed-bid auctions (Ariely et al., 2005). This
improves the price discovery process and increases competition among
bidders so that efficiency and revenues can be enhanced, even in purely
private-value environments. In line with this finding, Ivanova-Stenzel and
Salmon (2004) report that, when having the choice between sealed-bid and
open, ascending-bid auctions, laboratory subjects in a private-value envi-
ronment have a strong preference for the open format. Finally, Cramton
(1998) notes that in practical applications, the dynamic price-discovery
process of an open auction most often does a better job than sealed bidding.
This is, of course, in particular true in multi-object auctions, where the
dynamics facilitate arbitrage and packaging.
However, there are also disadvantages that come with open bidding.
Open auctions are more susceptible to various forms of collusion and fraud.
Bapna (2003) argues that open auctions facilitate collusive bidding against a
repeat seller (and has other more technological disadvantages). He, there-
fore, recommends that eBay run sealed-bid auctions. The literature on
spectrum auctions, however, demonstrated that certain auction design fea-
tures can address and mitigate many of these problems. For instance, a
history of bids that conceals bidder identities can, to some extent, suppress
bidder collusion against sellers and rival bidders.51 Furthermore, the an-
onymity and the number of potential bidders, as well as free entry in online
auctions, seem to make coordination, signaling, and communication among
bidders more difficult than in many offline auction environments.
Other concerns are probably more serious: open auctions can lead to
lower revenues when bidders are risk-averse (as we mentioned in Section 2),
and when ex ante asymmetries among bidders are strong or competition is
weak (e.g., Cramton, 1998). This might be part of the reason why eBay
recently introduced a sealed-bid format as an option for sellers; in the best
offer format, bidders can make sealed-bids, and sellers can accept any bid at
any time they wish.
8.2 Controlling the pace of bidding
As we have seen in Section 4, bidders in eBay auctions tend to bid late.

This may distort the virtues of long, open auctions described above. One
way to avoid late bidding and to control the pace of auctions is to create
51
See Klemperer (2004) for a review of design recommendations for the European spectrum auctions
to avoid collusive behavior.
pressure on bidders to bid actively from the start. Milgrom and Wilson
designed an activity rule that was applied to the US spectrum auctions
(McAfee and McMillan, 1996). The activity rule requires a bidder to be
‘‘active’’ (that is to be the current high bidder or to submit new bids) on a
predetermined number of spectrum licenses. If a bidder falls short of the
required activity level, the number of licenses it is eligible to buy shrinks.
Thus, bidders are prevented from holding back. However, activity rules of
this sort are incompatible with the flexibility needed on global auction
platforms.
Roth and Ockenfels (2002) observed that the rule by which online auc-
tions end may have a substantial effect on the timing of bids and price
discovery. On eBay, auctions end at a predetermined time: a ‘‘hard close.’’
In contrast, Amazon emulates the ‘‘Going, Going, Gone’’ feature of tra-
ditional auction houses. That is, Amazon automatically extends an auction
if a bid comes in late, so that all bidders always have the opportunity to
respond to the opponents’ bids.52
Ockenfels and Roth (forthcoming) show that, although the risks of last-
minute bidding remain, the strategic advantages of last-minute bidding are
eliminated or severely attenuated in Amazon-style auctions. That is, a bid-
der who waits to bid until the last seconds of the auction still runs the risk
that his bid will not successfully be transmitted in time. However, if his bid
is successfully transmitted, the auction will be extended for 10 min, so that,
no matter how late the bid was placed, other bidders will have time to
respond. Thus on Amazon, an attentive incremental bidder, for example,
can respond whenever a bid is placed.53 The differences in the strategic
environment are reflected in the data of Roth and Ockenfels (2002): there is
significantly more late bidding on eBay than on Amazon. For instance,
40% of eBay-computer auctions and 59% of eBay-antiques auctions in the
sample have last bids in the closing 5 min, compared to about 3% of both
Amazon computer and Amazon antiques auctions that have last bids in the
final 5 min before the initially scheduled deadline or later. Further analysis
reveals that while the impact of the bidders’ feedback numbers on late
bidding is significantly positive in eBay, it is negative in Amazon, suggesting
that more experienced bidders on eBay bid later than less experienced bid-
ders, but experience in Amazon has the opposite effect.
52
In Amazon’s own words: ‘‘We know that bidding can get hot and heavy near the end of many
auctions. Our Going, Going, Gone feature ensures that you always have an opportunity to challenge
last-second bids. Here’s how it works: Whenever a bid is cast in the last 10 min of an auction, the auction
is automatically extended for an additional 10 min from the time of the latest bid. This ensures that an
auction can’t close until 10 min have passed with no further bids.’’ On Yahoo’s auction platform, the
seller decides whether he wishes a hard or a soft close. Otherwise, all three platforms employ similar
auction rules.
53
However, there are other, non-strategic reasons for late bidding, including procrastination, use of
search engines that make it easy to find auctions about to end, endowment effects, or management of
bidding in multiple auctions in which similar objects may be offered. These motives for late bidding
should be relatively unaffected by the difference in closing rules between eBay and Amazon.
Experiments by Ariely et al. (2005) replicate these findings in a controlled

laboratory private-value setting in which the only difference between auc-
tions is the ending rule. The experiment thus controls for differences other
than the closing rule that might affect behavior on Amazon and eBay, such
as the number of auctions being conducted at a time and the number of
potential bidders. The experiment also demonstrates that, ceteris paribus,
‘‘early’’ prices on Amazon are an increasingly good predictor for final
prices, whereas price discovery on eBay became increasingly frenzied. Sim-
ulation experiments by Duffy and Ünver (2005) with artificial adaptive
agents who can update their strategies via a genetic algorithm, replicate
these findings and thus provide another robustness check.
Controlled field experiments, on the other hand, seem to have more diffi-
culties finding evidence for the impact of the ending rule. Brown and Morgan
(2005) and Houser and Wooders (2005) took advantage of the fact that
Yahoo sellers are allowed to choose whether to end the auction with a hard
or a soft close. In both studies, identical items were sold using both ending
rules. None of these studies found a significant effect of the ending rule on the
amount of late bidding.54 However, Houser and Wooders (2005) observed—
as Ariely et al. (2005) and Duffy and Ütku (2005)—that, ceteris paribus,
hard-close auctions tend to raise less revenue than soft-close auctions.55
Online market design also includes the design of artificial agents, such as
eBay’s ‘‘proxy bidder.’’ Because late bidding involves a good deal of plan-
ning and effort, artificial agents can also help executing late-bidding strat-
egies. In fact, there is a market for artificial sniping agents that will allow a
bidder not only to submit a proxy bid, but also to do so at the last moment.
Sniping agents take two forms: downloadable programs that run on the
bidder’s own computer, and web-based services like esnipe.com to which a
bidder can subscribe. Both offer bidders the ability to choose their max-
imum bid early in the auction, record when the auction is scheduled to end,
and decide how many minutes or seconds before the end of the auction the
sniping agent should submit the bid.
Recall that whether the timing of bids matters depends on the rules of the
game. Artificial last-minute bidding agents (like esnipe.com) might support
human bidders in eBay auctions, but they would hardly help on Amazon,
where the closing rule removes or greatly attenuates the incentives to snipe.
By the same token, human bidders on Amazon have more reason to make
use of the proxy bidding agent provided by the auction houses than bidders
on eBay, where the fixed deadline may create incentives to submit the bid
late, depending on other (human or artificial) bidders’ behavior. Thus, how
54
In a recent laboratory experiment, in which three sudden termination variants of hard-close auction
(a.k.a. candle auction) were examined, Füllbrunn and Sadrieh (2006) find that the extent of late-bidding
crucially depends on the first stage in which the probability of sudden termination is greater than zero.
55
In a theoretic model of sequential auctions, Stryszowska (2005b) identified a situation in which soft-
close auctions should be expected to yield smaller revenues.
well different kinds of artificial agents perform depends on how the auction
rules are designed.
Note also, that as sniping by human and artificial agents become more
widespread on eBay, eBay will be gradually transformed into a sealed-bid
second-price auction. If a large part of the late-bidding activity takes place
on third-party sites like esnipe.com, eBay faces a number of design and rule
choices; one is to ban sniping services. In fact, eBay.de (Germany) banned
third-party sniping services in its general terms and conditions (which is, of
course, difficult to enforce), because, according to them, bidders who use
sniping services have an ‘‘unfair advantage’’ over people who bid manually.
A second choice would be just the opposite: recapturing the sniping market
by offering a sniping option on eBay itself. Under this option, last-minute
bids submitted in advance directly to eBay could all be counted at the same
time, immediately after the auction close. This would give bidders certainty,
both that their bids would be successfully transmitted, and that there would
be no time for other bidders to react. Of course, if all bidders used this
option, the auction becomes a sealed-bid auction (Ockenfels and Roth,
2002). As we have argued above, eBay might prefer not to encourage this
development toward sealed-bids, given the advantages of open auctions, yet
even now, eBay is considering a sniping service that would enable last-
minute bidding via phone (http://www.unwiredbuyer.com/). While bidding
by phone will still involve risks that the bid fails to successfully register, it
will likely further increase the number of snipes. Finally, eBay could con-
sider changing the ending rule of the auction to a soft close. This, however,
may also cause adverse effects such as lowering the entertainment value of
eBay.56
There is at least one other design choice influencing auction design: the
speed of the price clock in decreasing-price auctions. As we mentioned in
Section 2, Lucking-Reiley (1999) found that in a controlled field study of
online auctions, the single-item descending-price format yields higher rev-
enues than corresponding sealed-bid auctions—just the opposite of what
has been found in some laboratory experiments (Cox et al., 1982, 1983).
Lucking-Reiley observed that his descending-price auctions took much
longer than the experiments and speculated that the higher revenues are
56
Ockenfels (2003) noted that online negotiation sites that promise dispute resolution (such as e-
commerce disputes and traditional litigation) via electronic and standardized communication have to
deal with related design problems. One of the more prominent online negotiation sites, clicknsettle.com,
experimented in 1999 with round-by-round demands and offers. But this format did not prove to be
effective, because a deadline effect similar to what has been observed on eBay and to what has been
observed in experimental bargaining games (Roth et al., 1988) hindered efficient negotiations: After
reviewing the early results with our clients, we discovered that in most negotiations, the first two rounds
were being ‘wasted’ and the disputing parties really only had one opportunity to settle the case, the final
round. (http://www.clicknsettle.com/onlinenegmodel.cfm 2003). eBay also works together with a dispute
resolution provider. A recent study by Brett et al. (2005) investigated the time it takes to resolve a
dispute in an online setting. By analyzing 582 eBay-generated disputes they find that the opening moves
can be critical in accelerating or delaying resolution to disputants.
because bidders may be impatient to complete their purchase, or having

more time allows bidders to more accurately determine their value of the
item. These ideas are supported in a laboratory experiment by Katok and
Kwasnica (2004). In descending-price auctions with a fast clock, revenue
turned out to be significantly lower, and with a slow clock significantly
higher, than in the corresponding sealed-bid version. As the authors show,
this bidding pattern is in line with a simple model of impatient bidders.
Carare and Rothkopf (2005) come to similar conclusions in both a decision-
theoretic and game-theoretic model that incorporates a ‘‘cost of returning
to the auction site.’’ Bidders prefer to bid sooner, yielding higher prices,
when the cost is higher. These results suggest that, without seller compe-
tition and without impatience on the side of the seller, sellers would prefer
to implement a slow clock rather than a fast clock, or a sealed-bid mech-
anism. We are not aware of any online auction that allows a choice like this,
but eBay’s buy-it-now option may be a substitute choice, because it gives
impatient bidders the opportunity to immediately end the auction at a
higher price.
8.3 Design aspects in multi-unit auctions
There are potentially many ways to sell multiple items in online auc-
tions. One way is to sell them simultaneously or sequentially in a series
of single-object auctions. Another way is to sell them through a single
auction, tailored to selling multiple units. The latter approach has several
advantages. For one, selling multiple units through one auction can re-
duce transaction costs for both buyers and sellers. Second, simultaneous
and sequential auctions impose strategic complexities and coordination
problems on bidders because bidders must guess prices of the other ob-
jects in order to realize arbitrage and to efficiently package objects. Wrong
guessing may hamper revenue and efficiency (Cramton, 2004; Milgrom,
2004).
Laboratory and field studies of sequential auctions strongly support this
view. For instance, sequential auctions typically fail to generate similar
prices for similar items. Rather, prices display a downward drift. This phe-
nomenon is called declining price anomaly (Ashenfelter, 1989). One possible
explanation is related to the winner’s curse: those bidders who win the early
units are those who overestimated the prices realized in later auctions.
Related problems arise in simultaneous eBay auctions. Since eBay bidders
have only weak incentives to contribute to the price discovery process early
in the auction—especially in the presence of simultaneous, multiple listings
of identical items—the decision where and what to bid is complex and may
lead to random or erroneous entry decisions close to the ending time.
The result is coordination failure. Stryszowska (2005a,b,c) investigated
coordination (failure) in simultaneous, overlapping and sequential Internet

auctions.
In principle, multi-unit auction formats, such as eBay’s ‘‘Dutch’’ auction,
diminish coordination problems by reducing the number of auctions. They
also enforce uniform prices for identical objects, reducing the risk associated
with price dispersion. However, details in the design matter. Ockenfels
(2005) noted that even in the simplest case of unit-demand, the price rule of
eBay’s ‘‘Dutch’’ auction makes bidding much more complicated than in the
single-object auction. Assuming that no bidder demands more than one unit,
the ‘‘natural’’ extension of eBay’s second-price single-object auction mech-
anism is the Vickrey auction described earlier, in which the final price is
equal to the highest losing bid (plus a small increment). Facing Vickrey’s
price rule, bidders should ‘‘sooner or later’’ just bid their values, independent
of the other bidders’ behavior. In eBay’s ‘‘Dutch’’ auction, however, the final
price is equal to the lowest winning bid, so that one of the winners will
eventually determine the price. This creates incentives for bid shading.
Specifically, winners can minimize the price paid by not bidding more than a
small increment above the highest losing bid.57 But, because the highest
losing bid is usually not known before the auction is over, the outcome of the
auction again depends on the accurateness of the bidders’ estimations. In
fact, Ockenfels (2005) found more bid shading in eBay’s ‘‘Dutch’’ auction
than in eBay’s single-object auction in a controlled field experiment on eBay.
These kinds of arguments convinced eBay Germany to change the multi-
unit format in the summer of 2005. The multi-unit auction format is now
called ‘‘multiauktion’’ and to a large extent analogous to the single-object
auction. Most importantly, there is now proxy bidding (proxy bids are
concealed to other bidders), and the price equals the highest losing bid—
analogous to the single-object format. However, there are other issues with
eBay’s multi-unit auction having to do with the fact that neither the old, nor
the new format avoid demand reduction and exposure problems for multi-
unit demand. We believe that designing a robust multi-unit auction that
takes complex preferences and incentives of the bidders into account is still
an important challenge for online auctions. The changes in Germany are
moving in the right direction.
Acknowledgements
We thank David Caballero and Jason McCoy for their research assist-
ance. Axel Ockenfels gratefully acknowledges support from the German
Science Foundation (DFG). David Reiley gratefully acknowledges support
from NSF grant SES-0094800.
57
In the single-object auction, finding this price is the job of the proxy bidder.
References
Anderson, S.T., D. Friedman, G.H. Milam, N. Singh (2004). Buy it now: a hybrid internet market
institution. UC Santa Cruz Economics Working Paper No. 565. Available at http://ssrn.com/ab-
stract=533182.
Anwar, S., R. McMillan, M. Zheng (2006). Bidding behavior in competing auctions: evidence from
eBay. European Economic Review 50(2), 307–322.
Ariely, D., A. Ockenfels, A.E. Roth (2005). An experimental analysis of ending rules in internet auc-
tions. The RAND Journal of Economics 36, 790–809.
Ariely, D., I. Simonson (2003). Buying, bidding, playing or competing?: value assessment and decision
dynamics in online auctions. Journal of Consumer Psychology 13(1–2), 113–123.
Ashenfelter, O. (1989). How auctions work for wine and art. Journal of Economic Perspectives 3, 23–36.
Asker, J., B. Grosskopf, C.N. McKinney, M. Niederle, A.E. Roth, G. Weizsäcker (2004). Teaching auction
strategy using experiments administered via the internet. Journal of Economic Education 35(4), 330–342.
Ausubel, L.M. (2006). An efficient ascending-bid auction for multiple objects. American Economic
Review 96(3), 602–629.
Ausubel, L.M., P. Cramton (2002). Demand reduction and inefficiency in multi-unit auctions. Working
Paper. University of Maryland.
Ausubel, L.M., P. Cramton (2004). Auctioning many divisible goods. Journal of the European Economic
Association 2, 480–493.
Ausubel, L.M., Cramton, P., P. Milgrom (2006). The clock-proxy auction: a practical combinatorial
auction design, Chapter 5. In: Cramton, P., Steinberg, R., Shoham, Y. (eds.), Combinatorial Auc-
tions, 115–138.
Ausubel, L.M., and P. Milgrom (forthcoming). The lovely but lonely vickrey auction, in: P. Cramton, R.
Steinberg, Y. Shoham (eds.), Combinatorial Auctions.
Bajari, P., A. Hortac- su (2003a). Winner’s curse, reserve prices and endogenous entry: empirical insights
from eBay auctions. Rand Journal of Economics 2, 329–355.
Bajari, P., A. Hortac- su (2003b). Economic insights from internet auctions: a survey. NBER Working
Paper No. W10076. Available at http://ssrn.com/abstract=467544.
Bajari, P., A. Hortac- su (2004). Economic insights from internet auctions. Journal of Economic Literature
XLII, 457–486.
Ball, S., M.H. Bazerman, J.S. Carroll (1991). An evaluation of learning in the bilateral winner’s curse.
Organizational Behavior and Human Decision Processes 48, 1–22.
Banks, J., J. Ledyard, D. Porter (1989). Allocating uncertain and unresponsive resources: an exper-
imental approach. RAND Journal of Economics 20(1), 125.
Barbaro, S., B. Bracht (2004). Shilling, squeezing, sniping: explaining late bidding in online second-price
auctions. Working Paper. University of Mainz, Germany.
Bapna, R. (2003). When snipers become predators: can mechanism design save online auctions?. Com-
munication of the ACM 46(12), 152–158.
Bapna, R., P. Goes, A. Gupta (2000). A theoretical and empirical investigation of multi-item on-line
auctions. Information Technology and Management 1(1), 1–23.
Bernheim, D., M. Whinston (1986). Menu auctions, resource allocation and economic influence. Quar-
terly Journal of Economics 101, 1–31.
Bolton, G.E., E. Katok, A. Ockenfels (2004a). How effective are online reputation mechanisms? An
experimental investigation. Management Science 50(11), 1587–1602.
Bolton, G.E., E. Katok, A. Ockenfels (2004b). Trust among internet traders. A behavioral economics
approach. Analyse und Kritik 26, 185–202.
Bolton, G.E., E. Katok, A. Ockenfels (forthcoming). Bridging the trust gap in electronic markets. A
strategic framework for empirical study, in: E Akcali, J. Geunes, P.M. Pardalos, H.E. Romeijn, Z.J.
Shen (eds.), Applications of Supply Chain Management and E-Commerce Research in Industry,
Kluwer Academic Publishers.
Bolton, G., A. Ockenfels (2000). ERC—a theory of equity, reciprocity and competition. American
Economic Review 90(1), 166–193.
Bolton, G., A. Ockenfels (2006). Bridging laboratory and field behavior: bargaining and auctions on Ebay.
University of Cologne, Mimeo.
Borenstein, S., J. Bushnell, F. Wolak (2002). Measuring market inefficiencies in California’s wholesale
electricity industry. American Economic Review 92(5), 1376–1405.
Bose S., A. Daripa (2006). Buy-now option in auctions. Working Paper. University of Illinois, Urbana-
Champaign.
Brett, J.M., M. Olekalins, N. Goates, R.A. Friedman, C.C. Lisco (2005). Tortoise or hare? A study of
the speed to resolution in on-line disputes. Working Paper. Kellogg School of Management, North-
western University.
Brosig, J., A. Ockenfels, J. Weimann (2003). The effect of communication media on cooperation.
German Economic Review 4(2), 217–241.
Brosig, J., J.P. Reiss (forthcoming). Entry decisions and bidding behavior in sequential first-price pro-
curement auctions: an experimental study. Games and Economic Behavior.
Brown, J., J. Morgan (2005). How much is a dollar worth? Tipping versus equilibrium coexistence on
competing online auction sites. Working Paper. University of California, Berkeley.
Budish, E., L. Takeyama (2001). Buy prices in online auctions: irrationality on the internet?. Economics
Letters 73, 325–333.
Burguet, R., J. Sákovics (1999). Imperfect competition in auction design. International Economic Review
40(1), 231–247.
Caldenty, R., G. Vulcano (forthcoming). Online auction and revenue management. Management Science.
Camerer, C.F. (2003). Behavioral Game Theory: Experiments in Strategic Interaction. Princeton Uni-
versity Press, Princeton NJ, USA.
Carare, O., M. Rothkopf (2005). Slow Dutch auctions. Management Science 51(3), 365–373.
Cassady, R. (1967). Auctions and Auctioneering. California University Press, California.
Chakraborty, I., G. Kosmopoulou (2004). Auctions with shill bidding. Economic Theory 24(2), 271–287
Springer.
Cox, J., B. Roberson, V.L. Smith (1982). Theory and behavior of single object auctions. Research in
Experimental Economics 1, 61–99.
Cox, J., V.L. Smith, J.M. Walker (1983). A test that discriminates between two models of the Dutch-first
auction non-isomorphism. Journal of Economic Behavior and Organization 4, 205–219.
Cramton, P. (1995). Money out of thin air: the nationwide narrowband PCS auction. Journal of Eco-
nomics and Management Strategy 4, 267–343.
Cramton, P. (1998). Ascending auctions. European Economic Review 42(3–5), 745–756.
Cramton, P. (2002). Spectrum auctions, in: M. Cave, S. Majumdar, I. Vogelsang (eds.), Handbook of
Telecommunications Economics, Elsevier Science B.V, Amsterdam, pp. 605–639.
Cramton, P. (2004). Simultaneous ascending auctions. Working Paper. University of Maryland.
Cramton, P., Y. Shoham, R. Steinberg (2006). Introduction to combinatorial auctions, in: P. Cramton,
Y. Shoham, R. Steinberg (eds.), Combinatorial Auctions, MIT Press, pp. 1–14.
Compte, O., P. Jehiel (2004). Auctions and information acquisition: sealed-bid or dynamic formats?
Working Paper. C.N.R.S. Paris.
Dellarocas, C. (2003). The digitization of word-of-mouth: promise and challenges of online reputation
mechanisms. Management Science 49(10), 1407–1424.
de Vries, S., R. Vohra (2003). Combinatorial auctions: a survey. INFORMS Journal on Computing 5(3),
284–309.
Dennehy, M. (2000). eBay adds ‘buy-it-now’ feature. AuctionWatch.com, October 26, 2000.
Dewally, M., L.H. Ederington (2004). What attracts bidders to online auctions and what is their
incremental price impact? Available at http://ssrn.com/abstract=589861.
Duffy, J., U. Ünver (2005). Internet auctions with artificial adaptive agents: a study on market design.
Working Paper. University of Pittsburgh.
Edelman, B., M. Ostrovsky, M. Schwarz (2005). Internet advertising and the generalized second-price
auction: selling billions of dollars worth of keywords. Working Paper. Harvard University.
Elyakime, B., J.J. Laffont, P. Loisel, Q. Vuong (1994). First-price sealed-bid auctions with secret res-
ervation prices. Annales D’Economie et de Statistique 34, 115–141.
Engelberg, J., J. Williams (2005). A license to shill: incremental bidding and eBay’s proxy system.
Working Paper. Kellogg School of Management.
Engelbrecht-Wiggans, R. (1987). On optimal reservation prices in auctions. Management Science 33(6),
763–770.
Engelbrecht-Wiggans, R. (1989). The effect of regret on optimal bidding in auctions. Management
Science 35(6), 685–692.
Engelbrecht-Wiggans, R., C.M. Kahn (1998). Multi-unit auctions with uniform prices. Economic Theory
12, 227–258.
Engelbrecht-Wiggans, R., E. Katok (2005). Regret in auctions: theory and evidence. Working Paper.
Available at http://lema.smeal.psu.edu/katok/regretTheory_post.pdf.
Engelbrecht-Wiggans, R., J.A. List, D.H. Reiley (2006). Demand reduction in multi-unit auctions with
varying numbers of bidders: theory and evidence from a sports car field experiments. International
Engelmann, D., V. Grimm (2004). Bidding behavior in multi-unit auctions—an experimental investi-
gation and some theoretical insights. Available at http://home.cerge-ei.cz/engelmann/WorkingPa-
pers/DerExGEB.pdf.
Etzion, H., E. Pinker, A. Seidmann (2003). Analyzing the simultaneous use of auctions and posted prices
for online selling. Working Paper CIS-03-01. University of Rochester.
Fehr, E., K. Schmidt (1999). A theory of fairness, competition, and cooperation. Quarterly Journal of
Economics 114, 817–868.
Füllbrunn, S., A. Sadrieh (2006). Sudden termination auctions. Working Paper. University of Ma-
gdeburg.
Garratt, R., M. Walker, J. Wooders (2004). Behavior in second-price auctions by highly experienced
eBay buyers and sellers. Working Paper 04-04. Department of Economics, UCSB.
Graham, D.A., R.C. Marshall, J.-F. Richard (1990). Phantom bidding against heterogeneous bidders.
Economics Letters 32, 13–17.
Gray, S., D. Reiley (2004). Measuring the benefits to sniping on eBay: evidence from a field experiment.
Working Paper. University of Arizona.
Grimm, V., F. Riedel, E. Wolfstetter (2003). Low price equilibrium in multi-unit auctions: the GSM
spectrum auction in Germany. International Journal of Industrial Organization 21, 1557–1569.
Grosskopf, B., Y. Bereby-Meyer, M.H. Bazerman (2003). On the robustness of the winner’s curse
phenomenon. Harvard Business School. Working Paper No. 03119.
Güth, W., F. Mengel, A. Ockenfels (2005). The dynamics of trust and trustworthiness on eBay. An
evolutionary analysis of buyer insurance and seller reputation. Working Paper. University of Cologne.
Gupta S., J. Gallien (2006). Temporary and permanent buyout prices in online auctions. MIT Sloan
Working Paper 4608-06.
Haile, P.A. (2000). Partial pooling at the reserve price in auctions with resale opportunities. Games and
Economic Behavior 22, 231–248.
Hasker, K., R. Gonzales, R.C. Sickles (2004). An analysis of strategic behavior and consumer surplus in
eBay auctions. Working Paper. Rice University. Available at http://www.ruf.rice.edu/rsickles/pa-
per/auction.pdf.
Hayne, S.C., C.A.P. Smith, L.R. Vijayasarathy (2003a). Sniping in eBay: a cultural analysis. Working
Paper. University of Colorado.
Hayne, S.C., C.A.P. Smith, L.R. Vijayasarathy (2003b). Who wins on eBay: an analysis of bidders and
their bid behaviours. Electronic Markets 13(4), S.282–S.293.
Häubl, G., P.T.L. Popkowski Leszczyc (2003). Minimum prices and product valuations in auctions.
Marketing Science Institute Reports 03-117(3), 115–141.
Häubl, G., P.T.L. Popkowski Leszczyc (2004). Bidding frenzy: intensity of competitive interaction
among bidders and product valuation in auctions. Advances in Consumer Research 31, 90–93.
Hendricks, K., I. Onur, T. Wiseman (2005). Preemption and delay in eBay auctions. Working Paper.
University of Texas at Austin.
Heyman, J.E., Y. Orhun, D. Ariely (2004). Auction fever: the effect of opponents and quasi-endowment
on product valuations. Journal of Interactive Marketing 18(4), 7–21.
Hidvégi, Z., W. Wang, A.B. Whinston (2006). Buy-price English auction. Journal of Economic Theory
129(1), 31–56.
Hof, R. (2001). The people’s company. Businessweek Online.
Hoppe, T., A. Sadrieh (2006). An experimental assessment of confederate reserve price bids in online
auctions. Working Paper. University of Magdeburg.
Hossain, T. (2006). Learning by bidding. Working Paper. Hong Kong University of Science and Tech-
nology.
Hossain, T., J. Morgan (2006). Plus shipping and handling: revenue (non)equivalence in field exper-
iments on eBay. Advances in Economic Analysis & Policy 6(2), 1–27.
Houser, D., J. Wooders (2005). Hard and soft closes: a field experiment on auction closing rules, in: R.
Zwick, A. Rapoport (eds.), Experimental Business Research, Vol. II. Kluwer Academic Publishers,
Boston, MA, USA, pp. 279–287.
Ivanova-Stenzel, R., T. Salmon (2004). Bidder preferences among auction institutions. Economic Inquiry
42, 223–236.
Izmalkov, S. (2004). Shill bidding and optimal auctions. Working Paper. MIT.
Jehiel, P., B. Moldovanu (2001). Efficient design with interdependent values. Econometrica 69, 1237–1259.
Jin, G.Z., A. Kato (2004). Consumer frauds and the uninformed: evidence from an online field ex-
periment. Working Paper. University of Maryland. Available at http://www.glue.umd.edu/ginger/
research/eBay-Jin-Kato-0305.pdf.
Jin, G.Z., A. Kato (2005). Quality and reputation: evidence from an online field experiment. Working
Paper. University of Maryland.
Kagel, J.H., D. Levin (1993). Independent private value auctions: bidder behavior, in first-, second- and
third price auctions with varying numbers of bidders. Economic Journal 103(July), 868–879.
Kagel, J., D. Levin (1986). The Winner’s curse and public information in common value auctions.
American Economic Review 76(5), 894–920.
Kagel, J.H., D. Levin (2002). Common Value Auctions and the Winner’s Curse. Princeton University
Press, .
Kagel, J.H. (1995). Auctions: a survey of experimental research, in: J.H. Kagel, A.E. Roth (eds.), The
Handbook of Experimental Economics, Princeton University Press, Princeton, NJ, pp. 501–585.
Kagel, J.H., D. Levin (1995). Comparative static effects of number of bidders and public information on
behavior in second-price common value auctions. International Journal of Game Theory 24, 293–319.
Kagel, J.H., D. Levin (2001a). The winner’s curse and public information in common value auctions.
Kagel, J.H., D. Levin (2001b). Behavior in multi-unit demand auctions: experiments with uniform price
and dynamic auctions. Econometrica 69(2), 413–454.
Kagel, J., D. Levin, R. Harstad (1995). Comparative static effects of number of bidders and public infor-
mation in second-price common value auctions. International Journal of Game Theory 24, 293–319.
Katkar, R., D.H. Reiley (2005). Public versus secret reserve prices in eBay auctions: results from a
Pokémon field experiment. Working Paper. University of Arizona.
Katok, E., A.M. Kwasnica (2004). Time is money: the effect of clock speed on seller’s revenue in Dutch
auctions. Working Paper. Pennsylvania State University.
Katok, E., A.E. Roth (2004). Auctions of homogenous goods with increasing returns: experimental
comparison of alternative ‘‘Dutch’’ auctions. Management Science 50(8), 1044–1063.
Kauffman, R.J., C.A. Wood (2003). Running up the bid: detecting, predicting, and preventing reserve
price shilling in online auctions. Working Paper. MIS Research Center. Carlson School of Man-
agement. University of Minnesota. Available at http://misrc.umn.edu/workingpapers/fullPapers/
2003/0304_022003.pdf.
Kirchkamp, O., P. ReiX (2004). The overbidding-myth and the underbidding-bias. SFB 504 Working
Paper No. 04-32. University of Mannheim. Available at http://www.kirchkamp.de/pdf/04-32.pdf.
Kirkegaard, R. (2005). The phantom made me jump! Preemptive jump bidding in (takeover) auctions.
Working Paper, Brock University, October.
Kirkegaard, R., P.B. Overgaard (2003). Buy-out prices in online auctions: multi-unit demand. Univer-
sity of Aarhus. Mimeo.
Kirchkamp, O., E. Poen, J.P. ReiX (2004). Outside options: another reason to choose the first-price
auction. Working Paper. University of Magdeburg.
Klemperer, P. (1999). Auction theory: a guide to the literature. Journal of Economic Surveys 13, 227–286.
Klemperer, P. (2004). Auctions: Theory and Practice. The Toulouse Lectures in Economics. Princeton
University Press, Princeton, NJ, USA.
Kosmopoulou, G., D.G. De Silva (2005). The effect of shill bidding upon prices: experimental evidence.
Available at SSRN: http://ssrn.com/abstract=643143.
Krishna, V. (2002). Auction Theory. Academic Press, San Diego, CA, USA.
Ku, G., D. Malhotra, J.K. Murnighan (2005). Towards a competitive arousal model of decision-
making: a study of auction fever in live and internet auctions. Organizational Behavior and Human
Decision Processes 96, 89–103.
Kwasnica, A., J. Ledyard, D. Porter, C. DeMartini (2005). A new and improved design for multi-object
iterative auctions. Management Science 51(3), 419–434.
Ledyard, J.O., D. Porter, A. Rangel (1997). Experiments testing multiobject allocation mechanisms.
Journal of Economics and Management Strategy 6, 639–675.
Levin, D., J.L. Smith (1996). Optimal reservation prices in auctions. Economic Journal 106, 1271–1282.
Li, H., G. Tan (2000). Hidden reserve prices with risk averse bidders. Working Paper. University of
British Columbia.
List, J.A., D. Lucking-Reiley (2000). Demand reduction in multi-unit auctions: evidence from a
Sportscard field experiment. American Economic Review 90(4), 961–972.
List, J.A., D. Lucking-Reiley (2002). Bidding behavior and decision costs in field experiments. Economic
Inquiry 40(44), 611–619.
Lucking-Reiley, D. (1999). Using field experiments to test equivalence between auction formats: magic
on the internet. American Economic Review 89(5), 1063–1080.
Lucking-Reiley, D. (2000a). Auctions on the internet: what’s being auctioned, and how?. Journal of
Lucking-Reiley, D. (2000b). Vickrey auctions in practice: from nineteenth-century philately to twenty-
first-century e-commerce. Journal of Economic Perspectives 14(3), 183–192.
Lucking-Reiley, D., D. Bryan, N. Prasad, D. Reeves (1999). Pennies from eBay: the determinants of
price in online auctions. Working Paper. University of Arizona.
Maskin, E. (2004). The unity of auction theory. Journal of Economic Literature 42(4), 1102–1115.
Mathews, T. (2003). A risk averse seller in a continuous time auction with a buyout option. Brazilian
Electronic Journal of Economics 5(2), 1–26.
Mathews, T. (2005). Bidder welfare in an auction with a buyout option. Available at http://
buslab5.csun.edu/tmathews/pdf/BidderWelfare.pdf.
Mathews, T., B. Katzman (2006). The role of varying risk attitudes in an auction with a buyout option.
Economic Theory 27, 597–613.
McAfee, R.P., J. McMillan (1987). Auctions with entry. Economics Letters 23(4), 343–347.
McAfee, R.P. (1993). Mechanism design by competing sellers. Econometrica 61, 1281–1312.
McAfee, R.P., J. McMillan (1996). Analyzing the airwaves auction. Journal of Economic Perspectives
10(1), 159–175.
Menezes, F.M., P.K. Monteiro (2005). An Introduction to Auction Theory. Oxford University Press,
Oxford, NY, USA.
Milgrom, P.R. (2004). Putting Auction Theory to Work. Cambridge University Press, Cambridge, UK.
Milgrom, P.R., R.R. Weber (1982). A theory of auctions and competitive bidding. Econometrica 50(5),
1089–1122.
Morgan, J., K. Steiglitz, G. ReiX (2003). The spite motive and equilibrium behavior in auctions. Con-
tributions to Economic Analysis & Policy 2(1), Article 5. Available at http://www.bepress.com/bejeap/
contributions/vol2/iss1/art5).
Myerson, R.B. (1981). Optimal auction design. Mathematics of Operations Research 6, 58–73.
Nagareda, T. (2003). Announced reserve prices, secret reserve prices, and winner’s curse. Available at
http://www.e.u-tokyo.ac.jp/cirje/research/workshops/micro/micropaper02/gs/nagareda.pdf.
Namazi, A. (2005). Emergent behavior and criticality in online auctions. Ph.D. dissertation, University of
Cologne.
Ockenfels, A. (2003). New institutional structures on the internet: the economic design of online auc-
tions, in: M.J. Holler, H. Kliemt, D. Schmidtchen, M. Streit (eds.), Jahrbuch für Neue Politische
Ökonomie, Vol. 20. Mohr Siebeck, Tübingen, pp. 57–78.
Ockenfels, A. (2005). Bidding in private value eBay auctions: a controlled field experiment. Mimeo.
Economics Department, University of Cologne.
Ockenfels, A., A. E. Roth (2002). The timing of bids in internet auctions: market design, bidder be-
havior, and artificial agents. Artificial Intelligence Magazine (Fall), 79–87.
Ockenfels, A., A. E. Roth. (forthcoming). Late and multiple bidding in second price internet auctions:
theory and evidence concerning different rules for ending an auction. Games and Economic Behavior,
55, 297–320.
Ockenfels, A., R. Selten (2005). Impulse balance equilibrium and feedback in first price auctions. Games
and Economic Behavior 51, 155–179.
Ow, T.T., C.A. Wood (2004). Trust versus the winner’s curse—the effects of bidder experience on prices
in online auctions. Working Paper. University of Notre Dame. Available at http://www.nd.edu/
cwood1/research/BidderX.pdf.
Peters, M., S. Severinov (1997). Competition among sellers who offer auctions instead of prices. Journal
of Economic Theory 75(1), 141–179.
Peters, M., S. Severinov (2006). Internet auctions with many traders. Journal of Economic Theory 130(1),
220–245.
Plott, C.R. (1997). Laboratory experimental testbeds: applications to the PCS auction. Journal of Eco-
nomics and Management Strategy 6, 605–638.
Rasmusen, E. (2003). Strategic implications of uncertainty over one’s own private value in auctions.
Working Paper. Indiana University. Available at http://php.indiana.edu/erasmuse/papers/auc-
tion.pdf.
Reiley, D.H. (2006). Field experiments on the effects of reserve prices in auctions: more magic on the
internet. RAND Journal of Economics 37(1), 195–211.
ReiX, J.P. (2004). Optimal auctions with public outside options. Working Paper. University of Ma-
gdeburg.
ReiX, J.P., J.R. Schoendube (2002). Entry and bidding behavior in sequential procurement auctions.
Working Paper. University of Magdeburg.
Reynolds, S.S., J. Wooders (2003). Auctions with a buy price. Working Paper. University of Arizona.
Available at http://www.u.arizona.edu/sreynold/buynow.pdf.
Riley, J.G., W.F. Samuelson (1981). Optimal auctions. American Economic Review 71, 381–392.
Roth, A.E. (2002). The economist as engineer: game theory, experimental economics and computation
as tools of design economics. Econometrica 70(4), 1341–1378.
Roth, A.E., J.K. Murnighan, F. Schoumaker (1988). The deadline effect in bargaining: some exper-
imental evidence. American Economic Review 78(4), 806–823.
Roth, A.E., Ockenfels, A. (2000). Last minute bidding and the rules for ending second-price auctions:
theory and evidence from a natural experiment on the internet. NBER Working Paper 7729.
Roth, A.E., A. Ockenfels (2002). Last-minute bidding and the rules for ending second-price auctions:
evidence from eBay and Amazon auctions on the internet. American Economic Review 92(4), 1093–
1103.
Rothkopf, M.H., T.J. Teisberg, E.P. Kahn (1990). Why are Vickrey auctions rare?. Journal of Political
Economy 98(1), 94–109.
Rothkopf, M.H., A. Pekec, R.M. Harstad (1998). Computationally manageable combinational auc-
tions. Management Science 44, 1131–1147.
Samuelson, W.F. (1985). Competitive bidding with entry costs. Economics Letters 17, 53–57.
Sandholm, T. (2006). Optimal winner determination algorithms, in: P. Cramton, R. Steinberg, Y.
Shoham (eds.), Combinatorial Auctions. MIT Press, Cambridge, MA, USA, 337–368.
Seifert, S. (2006). Posted Price Offers in Internet Auction Markets. Lecture Notes in Economics and
Mathematical Systems. Springer, Berlin.
Shah, H.S., N.R. Joshi, P.R. Wurman (2002). Mining for bidding strategies on eBay. Working Paper.
Department of Computer Science, North Carolina University.
Shahriar, Q., J. Wooders (2005). An experimental study of auctions with a buy price under private and
common values. Working Paper. Department of Economics, University of Arizona, Tucson.
Shmueli, G., R.P. Russo, W. Jank (2005). Modeling bid arrivals in online auctions. Working Paper.
Robert H. Smith School of Business, University of Maryland.
Simonsohn, U (2005). Excess entry into high-demand markets: evidence from the timing of online
auctions. Working Paper. The Wharton School, University of Pennsylvania.
Sinha, A.R., E.A. Greenleaf (2000). The impact of discrete bidding and bidder aggressiveness on sellers’
strategies in open English auctions: reserves and covert shilling. Marketing Science 19(3), 244–265.
Smith, V.L. (1976). Experimental economics: induced value theory. American Economic Review, Papers
and Proceedings. 274–279.
Standifird, S., M. Roelofs, Y. Durham (2004). eBay’s ‘‘buy-it-now’’ function and the impact on bidder
behavior. International Journal of Electronic Commerce 9(2), 167–176.
Stryszowska, M. (2005a). Late and multiple bidding in simultaneous and overlapping second price
internet auctions. CentER discussion paper.
Stryszowska, M. (2005b). On the ending rule in sequential internet auctions. CentER discussion paper.
Stryszowska, M. (2005c). Coordination failure in internet auctions with costly bidding. CentER dis-
cussion paper.
Tung, Y.A., R.D. Gopal, A.B. Whinston (2003). Multiple online auctions. IEEE Computer 36(2), 88–90.
Varian, H. R. (2006). Position auctions. Working Paper. Berkelee.
Vickrey, W. (1961). Counterspeculation, auctions and competitive sealed tenders. Journal of Finance 16,
8–37.
Vincent, D.R. (1995). Bidding off the wall: why reserve prices may be kept secret. Journal of Economic
Theory 65, 575–584.
Wang, T.J. (2003). Is last minute bidding bad? Working Paper. UCLA.
Wang, W., Z, Hidvégi, A. B. Whinston (2004). Shill-proof fee (SPF) schedule: the sunscreen against
seller self-collusion in online English auctions. Working Paper. Goizueta Buisness School, Emory
University. Available at http://www.wenli.net/Application/Shill.pdf.
Wang, X., A.L. Montgomery, K. Srinivasan (2004). When auction meets fixed price: a theoretical and
empirical examination of buy-it-now auctions. Working Paper. Carnegie Mellon University. Avail-
able at http://www.andrew.cmu.edu/user/alm3/papers/buyitnow.pdf.
Ward, S.G., J.M. Clark (2002). Bidding behavior in on-line auctions: an examination of the eBay
Pokemon card market. International Journal of Electronic Commerce 6(4), 139–155.
Wilcox, R.T. (2000). Experts and amateurs: the role of experience in internet auctions. Marketing
Letters 11(4), 363–374.
Wintr, L. (2004). Some evidence on late bidding in eBay auctions. Working Paper. Clark University,
Worcester, MA.
Wolf, J.R., H.R. Arkes, W.A. Muhanna (2005). Is overbidding in online auctions the result of a pseudo-
endowment effect? Working Paper. Ohio State University.
Wolfram, C.D. (1998). Strategic bidding in a multiunit auction: an empirical analysis of bids to supply
electricity in England and Wales. RAND Journal of Economics 29, 703–725.
Wu, J. (2004). Late bidding in fixed-deadline internet auctions. Working Paper, Renmin University of
China.
Yin, P.-L. (2005). Information dispersion and auction prices. Working Paper. Stanford University.
Zeithammer, R. (2003). Forward-looking bidders in online auctions: a theory of bargain-hunting on
eBay. Working Paper. University of Chicago.
Zheng, C. (2002). Optimal auction with resale. Econometrica 70(6), 2197–2224.
Chapter 13
Reputation Mechanisms
Chrysanthos Dellarocas
R. H. Smith School of Business, University of Maryland, College Park, MD 20742, USA
Abstract
Reputation mechanisms harness the bidirectional communication capabilities

of the Internet in order to engineer large-scale word-of-mouth networks. Best
known so far as a technology for building trust and fostering cooperation in
online marketplaces, such as eBay, these mechanisms are poised to have a
much wider impact on organizations. This paper surveys our progress in
understanding the new possibilities and challenges that these mechanisms
represent. It discusses some important dimensions in which Internet-based
reputation mechanisms differ from traditional word-of-mouth networks and
surveys the most important issues related to their design, evaluation, and use.
It provides an overview of relevant work in game theory and economics on
the topic of reputation. It discusses how this body of work is being extended
and combined with insights from computer science, marketing, and psychol-
ogy in order to take into consideration the special properties of online en-
vironments. Finally, it identifies opportunities that this new area presents for
information systems research.
1 Introduction
The Internet offers buyers and sellers unprecedented opportunities to

access resources that were difficult, or impossible, to locate in previous
times. Whether these resources are obscure books, highly specialized serv-
ices, or faraway trading partners, the global connectivity of the Web has the
potential to bring them within everyone’s reach, significantly enriching our
economic and cultural lives.
629
630 C. Dellarocas
To fully reap the benefits of global connectivity and arms-length trans-

actions with faraway partners our societies thus need to develop new trust
mechanisms capable of ensuring cooperation and efficiency in a universe of
strangers. Several of the mechanisms through which cooperation is induced
in offline settings, such as the legal system and stable partnerships, do not
work as well on the global, decentralized Internet (Kollock, 1999).
Reputation networks constitute an ancient solution to the problem of
trust-building. The historical appeal of these networks has been their power
to induce cooperation without the need for costly enforcement institutions.
Before the establishment of formal law and centralized systems of contract
enforcement backed by the sovereign power of a state, most ancient and
medieval communities relied on reputation as the primary enabler of eco-
nomic and social activity (Benson, 1989; Milgrom et al., 1990; Greif, 1993).
Many aspects of social and economic life still do so today (Klein, 1997).
It is a little ironic that these most ancient of mechanisms are emerging as
one of the most promising solutions to the problem of building trust on the
Internet. Online reputation mechanisms, also known as reputation systems
(Resnick et al., 2000; Dellarocas, 2003), are using the Internet’s bidirectional
communication capabilities in order to artificially engineer large-scale word-
of mouth networks where individuals share opinions and experiences on a
wide range of topics, including companies, products, services, and even world
events.
For example, eBay’s feedback mechanism is the primary means through
which eBay elicits honest behavior and, thus, facilitates transactions among
strangers over the Internet (Resnick and Zeckhauser, 2002). Several other
communities also rely on reputation mechanisms to promote trust and
cooperation. Examples include eLance (online community of freelance
professionals), Slashdot (online discussion forum where reputation scores
help prioritize and filter postings), and Epinions (online consumer report
forum where user feedback helps evaluate the quality of product reviews).
Table 1 lists several noteworthy examples of such mechanisms in use
today.
Online reputation mechanisms have a lot in common with their offline
counterparts. Their design and implementation, thus, has a lot to gain from
a substantial body of prior work on reputation formation in economics
and psychology. On the other hand, online mechanisms possess a number
of unique properties, whose implications are not yet fully understood.
Specifically:
Global reach enables new applications. Scale is essential to the effectiveness
of reputation networks. In an online marketplace, for example, sellers care
about buyer feedback primarily to the extent that they believe that it might
affect their future profits; this can only happen if feedback is provided by a
sufficient number of current customers and communicated to a significant
portion of future prospects. Theory predicts that a minimum degree of
participation in reputation communities is required before reputation
Table 1
Examples of commercial reputation mechanisms (in use as of June 2005)
Web site Category Summary of reputation Format of solicited feedback Format of published
mechanism feedback
Citysearch Entertainment Users rate restaurants, bars, Users rate multiple aspects of Weighted averages of ratings
guide clubs, hotels, and shops reviewed items from 1 to 10 per aspect reflecting both
and answer a number of user and editorial ratings;
yes/no questions; readers user reviews can be sorted
rate reviews as ‘‘useful,’’ according to ‘‘usefulness’’
Ch. 13. Reputation Mechanisms

‘‘not useful,’’ etc.
eBay Online auction Buyers and sellers rate one Positive, negative, or neutral Sums of positive, negative,
house another following rating plus short comment; and neutral ratings received
transactions ratee may post a response during past 6 months
eLance Professional Contractors rate their Numerical rating from 1 to 5 Average of ratings received
services satisfaction with plus comment; ratee may during past 6 months
marketplace subcontractors post a response
Epinions Online opinions Users write reviews about Users rate multiple aspects of Averages of item ratings; %
forum products/services; other reviewed items from 1 to 5; of readers who found a
members rate the readers rate reviews as review ‘‘useful’’
usefulness of reviews ‘‘useful,’’ ‘‘not useful,’’ etc.
Google Search engine Search results are ordered A Web page is rated based on No explicit feedback scores
based on how many sites how many links point to it, are published; ordering acts
contain links that point to how many links point to as an implicit indicator of
them (Brin and Page, 1998) the pointing page, etc. reputation
Slashdot Online discussion Postings are prioritized or Readers rate posted No explicit feedback scores
board filtered according to the comments are published; ordering acts
ratings they receive from as an implicit indicator of
readers reputation
631
632 C. Dellarocas
effects can induce any cooperation. Once this threshold is reached, however,
the power of reputation immediately springs to life and high levels of
cooperation emerge in a discontinuous fashion (Bakos and Dellarocas, 2002).
Therefore, the vastly increased scale of Internet-based reputation mecha-
nisms is likely to render them powerful institutions in environments where
traditional word-of-mouth networks were heretofore considered ineffective
devices.1 The social, economic, and perhaps even political consequences of
such a trend deserve careful study.
Information technology enables systematic design. In offline settings,
word-of-mouth emerges naturally and evolves in ways that are difficult
to control or model. The Internet allows this powerful social force to be
precisely measured and controlled through proper engineering of the infor-
mation systems that mediate online reputation communities. Such auto-
mated feedback mediators specify who can participate, what type of
information is solicited from participants, how it is aggregated, and what
type of information is made available to them about other community
members. Through the proper design of these mediators, mechanism
designers can exercise precise control over a number of parameters that are
very difficult or impossible to influence in brick-and-mortar settings. For
example, feedback mediators can replace detailed feedback histories with a
wide variety of summary statistics; they can apply filtering algorithms to
eliminate outlier or suspect ratings; they can weight ratings according to
some measure of the rater’s trustworthiness, etc. Such degree of control
can impact the resulting social outcomes in non-trivial ways. Understand-
ing the full space of design possibilities and the impacts of specific design
choices on the resulting social outcomes is an important research challenge
introduced by these new systems.
Online interaction introduces new challenges. The disembodied nature of
online environments introduces several challenges related to the interpre-
tation and use of online feedback. Some of these challenges have their roots
in the subjective nature of consumer feedback. Offline settings usually pro-
vide a wealth of contextual cues that assist in the proper interpretation of
opinions and gossip (such as familiarity with the person who acts as the
source of that information, the ability to draw inferences from the source’s
facial expression or mode of dress, etc.). Most of these cues are absent from
online settings. Readers of online feedback are thus faced with the task of
1
Three recent incidents illustrate the growing power of online opinion forums to exert influence on
corporations and other powerful institutions of our society. In December 2002, criticism of controversial
remarks made by US Senator Trent Lott by authors of Web Logs (blogs) eventually led to his res-
ignation from his post as majority leader. In 2003, Intuit Corporation was forced to remove unpopular
copy-protection spyware from its Turbo Tax software following a wave of very negative reviews posted
by customers in online product forums. In September 2004, scrutiny by blog authors revealed inac-
curacies in a story aired by long-serving and respected CBS anchor Dan Rather. The ensuing events
culminated into Dan Rather’s retirement from the channel.
Ch. 13. Reputation Mechanisms 633
evaluating the opinions of complete strangers. Other challenges to feedback

interpretation have their root in the ease with which online identities can be
changed. This opens the door to various forms of strategic manipulation.
For example, community members can build a good reputation, milk it by
cheating other members, and then disappear and reappear under a new
online identity and a clean record (Friedman and Resnick, 2001). They can
use fake online identities to post dishonest feedback and thus try to inflate
their reputation or tarnish that of their competitors (Dellarocas, 2006b).
Finally, the mediated nature of online reputation mechanisms raises ques-
tions related to the trustworthiness of their operators. An important pre-
requisite for the widespread acceptance of online reputation mechanisms is,
therefore, a better understanding of how such systems can be compromised
as well as the development of adequate defenses.
This chapter surveys our progress so far in understanding the new pos-
sibilities and challenges that these mechanisms represent. Section 2 intro-
duces a framework for understanding the role of reputation mechanisms in
various settings. Section 3 provides an overview of relevant past work in
game theory and economics. Section 4 then discusses how this stylized body
of work is being extended in order to take into consideration the special
properties of online environments. Sections 5 surveys empirical and exper-
imental work on reputation mechanisms. Finally, Section 6 summarizes the
main points of the paper and discusses the opportunities that this new area
presents for information systems research.
2 Signaling and sanctioning role of reputation mechanisms
The primary objective of reputation mechanisms is to enable efficient

transactions in communities where cooperation is compromised by post-
contractual opportunism (moral hazard) or information asymmetries
(adverse selection). It is instructive to distinguish the role of reputation
mechanisms with respect to moral hazard from their role with respect to
adverse selection.
Moral hazard can be present any time two parties come into agreement
with one another. Each party in a contract may have the opportunity to
gain from acting contrary to the principles laid out by the agreement. For
example, on eBay, the buyer typically sends money to the seller before
receiving the goods. The seller then is tempted to keep the money and not
ship the goods, or to ship goods that are inferior to those advertised.
Reputation mechanisms can deter moral hazard by acting as sanctioning
devices. If the community follows a norm that punishes traders with his-
tories of bad behavior (by refusing to buy from them, or by reducing the
price they are willing to pay for their products) and if the present value
of punishment exceeds the gains from cheating, then the threat of public
634 C. Dellarocas
revelation of a trader’s cheating behavior in the current round provides

rational traders with sufficient incentives to cooperate.
Adverse selection is present in situations where sellers have information
(about some aspect of their innate ability, product quality, etc.) that buyers
do not (or vice versa). Such situations often arise in markets for experience
goods. Consider, for example, an online hotel-booking site where hotels of
different qualities advertise rooms. Consumers cannot be certain about the
true quality offered by each hotel until they have actually stayed there. On
the other hand, hotels do not have an incentive to advertise any of their
weak points. Knowing this, consumers will assume that all hotels are of
average quality and will not be willing to pay more than the average price.
Akerlof (1970) shows that such a situation will eventually drive all, except
the lowest quality sellers, out of the market.
Reputation mechanisms alleviate adverse selection issues by acting as
signaling devices. For example, by soliciting and publishing experiences of
consumers who have stayed in advertised hotels, they help the community
learn the true quality of each hotel. This, in turn, allows a better matching
of buyers and sellers and a more efficient market.
The most important distinction between (pure) moral hazard and (pure)
adverse selection settings is that, in the former, all sellers are capable of the
same type of behavior (e.g., cooperate, cheat), whereas in the latter case
seller behavior is completely constrained by their innate ‘‘type.’’ The role of
reputation mechanisms in pure moral hazard settings is to constrain
behavior, whereas the role of such mechanisms in pure adverse selection
settings is to induce learning.
In some real-life settings, moral hazard and adverse selection considera-
tions are simultaneously present: sellers differ in their intrinsic ability levels
but, in addition, have a choice of behavior (which is partially, but not
completely, conditioned by their type). For example, certain attributes of
the customer experience (location, size of rooms, etc.) can be considered
as part of a hotel’s immutable ‘‘type,’’ whereas other attributes (cleanliness
of facilities, professionalism, and politeness of staff, etc.) are the result of
the hotel’s level of ‘‘effort’’ and can be varied strategically on a daily basis.
In such settings, reputation mechanisms play both a sanctioning and a
signaling role, revealing the hotel’s true immutable attributes while pro-
viding incentives to the hotel to exert reasonable effort.
In other settings, one of the two roles is dominant. For example, Amazon
Reviews primarily serves a signaling role: it spreads information about the
(initially privately known, but essentially ‘‘immutable’’) qualities of the
products (books, CDs, DVDs, etc.) being reviewed. eBay, on the other
hand, is an example of a mechanism that primarily acts as a sanctioning
device. Under the assumption that all eBay sellers are equally capable of
acting in honest and dishonest ways, eBay’s problem is to deter moral
hazard. Accordingly, eBay users do not rate sellers on the absolute quality
of their products but rather on how well they were able to deliver what was
promised on the item description. The role of eBay’s reputation mechanism

is to promote honest trade rather than to distinguish sellers who sell high-
quality products from those that sell low-quality products.
The distinction between sanctioning and signaling is central in reputation
mechanisms. Throughout this article we shall see that several principles of
reputation mechanism design depend on this distinction. Designers should,
therefore, be conscious of their mechanism’s primary objective and care-
fully make design choices that maximize the resulting market efficiency,
given that objective.
3 Reputation in game theory and economics
Reputation formation has been extensively studied by economists using

the tools of game theory. This body of work is, perhaps, the most promising
foundation for developing an analytical discipline of online reputation
mechanisms. This section surveys past work on this topic, emphasizing the
results that are most relevant to the design of online feedback mechanisms.
Section 4 then discusses how this stylized body of work is being extended to
address the unique properties of online environments.
3.1 Basic concepts
According to Wilson (1985), reputation is a concept that arises in

repeated game settings when there is uncertainty about some property
(the ‘‘type’’) of one or more players in the mind of other players. If ‘‘unin-
formed’’ players have access to the history of past stage-game outcomes,
reputation effects then often allow informed players to improve their long-
term payoffs by gradually convincing uninformed players that they belong
to the type that best suits their interests. They do this by repeatedly choos-
ing actions that make them appear to uninformed players as if they were of
the intended type, thus ‘‘acquiring a reputation’’ for being of that type.
The existence of some initial doubt in the mind of uninformed players
regarding the type of informed players is crucial in order for reputation
effects to occur. To see this, consider a repeated game between a long-run
player and a sequence of short-run (one-shot) opponents. In every stage
game, the long-run player can choose one of several actions but cannot
credibly commit to any of those actions in advance. If there is no uncer-
tainty about the long-run player’s type,2 rational short-run players will then
always play their stage-game Nash equilibrium response. Such behavior
typically results in inefficient outcomes.
2
In other words, if short-run players are convinced that the long-run player is a rational utility-
maximizing player whose stage-game payoffs are known with certainty.
636 C. Dellarocas
Consider, for example, the following stylized version of a repeated ‘‘on-

line auction’’ game. A long-lived seller faces an infinite sequence of sets of
identical one-time buyers in a marketplace where there are only two kinds
of products:
low-quality products that cost 0 to the seller and are worth 1 to the
buyers, and
high-quality products that cost 1 to the seller and are worth 3 to the
buyers.
Each period the seller moves first, announcing the quality of the product
he promises to buyers. Since high-quality products are more profitable, the
seller will always promise high quality. Buyers then compete with one
another in a Vickrey auction and therefore bid amounts equal to their
expected valuation of the transaction outcome. The winning bidder sends
payment to the seller. The seller then has the choice of either ‘‘cooperating’’
(delivering a high-quality good) or ‘‘cheating’’ (delivering a low-quality
good). It is easy to see that this game has a unique subgame perfect equi-
librium. In equilibrium the seller always cheats (delivers low quality), buy-
ers each bid 1, each buyer’s expected payoff is zero and the seller’s expected
payoff is 1.
The ability to build a reputation allows the long-run player to improve
his payoffs in such settings. Intuitively, a long-run player who has a track
record of playing a given action (e.g., cooperate) often enough in the past
acquires a reputation for doing so and is ‘‘trusted’’ by subsequent short-
run players to do so in the future. However, why would a profit-maxi-
mizing long-term player be willing to behave in such a way and why would
rational short-term players use past history as an indication of future
behavior?
To explain such phenomena, Kreps and Wilson (1982), Kreps et al.
(1982), and Milgrom and Roberts (1982) introduced the notion of ‘‘com-
mitment’’ types. Commitment types are long-run players who are locked
into playing the same action.3 An important subclass of commitment types
are Stackelberg types: long-run players who are locked into playing the so-
called Stackelberg action. The Stackelberg action is the action to which the
long-run player would credibly commit if he could. In the above ‘‘online
auction’’ example the Stackelberg action would be to cooperate; cooper-
ation is the action that maximizes the seller’s lifetime payoffs if the seller
could credibly commit to an action for the entire duration of the game.4
Therefore, the Stackelberg type in this example corresponds to an ‘‘honest’’
3
Commitment types are sometimes also referred to as ‘‘irrational’’ types because they follow fixed,
‘‘hard-wired’’ strategies as opposed to ‘‘rational’’ profit-maximizing strategies. An alternative way to
justify such players is to consider them as players with non-standard payoff structures such that the
‘‘commitment’’ action is their dominant strategy given their payoffs.
4
If the seller could commit to cooperation (production of high quality), buyers would then each bid 2
and the seller’s expected per period payoff would be 2.
seller who never cheats. In contrast, an ‘‘ordinary’’ or ‘‘strategic’’ type

corresponds to an opportunistic seller who cheats whenever it is advanta-
geous for him to do so.
Reputation models assume that short-run players know that commitment
types exist, but are ignorant of the type of the player they face. An
additional assumption is that short-run players have access to the entire
history of past stage-game outcomes.5 A player’s reputation at any given
time then consists of the conditional posterior probabilities over that
player’s type, given a short-run player’s prior over types and the repeated
application of Bayes’ rule on the history of past stage-game outcomes.
In such a setting, when selecting his next move, the informed player must
take into account not only his short-term payoff, but also the long-term
consequences of his action based on what that action reveals about his type
to other players. As long as the promised future gains due to the increased
(or sustained) reputation that comes from playing the Stackelberg action
offset whatever short-term incentives he might have to play otherwise, the
equilibrium strategy for an ‘‘ordinary’’ informed player will be to try to
‘‘acquire a reputation’’ by masquerading as a Stackelberg type (i.e.,
repeatedly play the Stackelberg action with high probability).
In the ‘‘online auction’’ example, if the promised future gains of repu-
tation effects are high enough,6 ordinary sellers are induced to overcome
their short-term temptation to cheat and to try to acquire a reputation for
honesty by repeatedly delivering high quality. Expecting this, buyers then
place high bids, thus increasing the seller’s long-term payoffs.
In general, reputation effects benefit the most patient player in the game:
the player who has the longest time horizon (discounts future payoffs less)
is usually the one who is able to reap the benefits of reputation. Fudenberg
and Levine (1992) show that this result holds even when players can observe
only noisy signals of each other’s actions, so that the game has imperfect
public monitoring. They prove that, if short-run players assign positive
prior probability to the long-run player being a Stackelberg type, and if that
player is sufficiently patient, then an ordinary long-run player achieves an
average discounted payoff close to his commitment payoff (i.e., his payoff if
he could credibly commit to the Stackelberg action). In order to obtain this
payoff, the ordinary player spends long periods of time choosing the
Stackelberg action with high probability.7
5
The traditional justification for this assumption is that past outcomes are either publicly observable
or explicitly communicated among short-run players. The emergence of online reputation mechanisms
provides, of course, yet another justification (but see discussion of complications arising from the private
observability of outcomes in such systems in Section 4.1).
6
In this of game, this requires that the remaining horizon of the seller is long enough and that the
profit margin of a single transaction is high enough relative to the discount factor.
7
This result also requires that the stage game is either a simultaneous move game, or in a sequential-
move game, that short-run players always observe whether or not the Stackelberg strategy has been
played.
638 C. Dellarocas
3.2 Reputation dynamics
In most settings where reputation phenomena arise, equilibrium strate-

gies evolve over time as information about the types of the various players
accumulates. In general, the derivation of closed-form solutions in repeated
games with reputation effects is complicated. Nevertheless, a small number
of specific cases have been studied. The general lesson is that reputation-
based performance incentives are highly dynamic: agents tend to behave
differently in different phases of the game.
Initial phase. In most cases, reputation effects begin to work immediately
and in fact are strongest during the initial phase, when players must work hard
to establish a reputation. Holmstrom (1999) discusses an interesting model of
reputational considerations in the context of an agent’s ‘‘career’’ concerns.
Suppose that wages are a function of an employee’s innate ability for a task.
Employers cannot directly observe an employee’s ability. However, they can
keep track of the average value of past task outputs. Outputs depend both on
ability and labor. The employee’s objective is to maximize lifetime wages while
minimizing the labor put in. At equilibrium, this provides incentives to the
employee to work hard right from the beginning of a career in order to build a
reputation for competence. In fact, these incentives are strongest at the very
beginning of a career when observations are most informative.
During the initial phase of a repeated game, it is common that some
players realize lower or even negative profits, while the community ‘‘learns’’
their type. In those cases, players will only attempt to build a reputation if
the losses from masquerading as a Stackelberg type in the current round are
offset by the present value of the gains from their improved reputation in
the later part of the game. In trading environments, this condition usually
translates to the need for sufficiently high profit margins for ‘‘good-quality’’
products so that the promise of future gains from sustaining a reputation is
persuasive enough to offset the short-term temptation to cheat. This was
first pointed out by Klein and Leffler (1981) and explored more formally by
Shapiro (1983).
Another case where reputation effects may fail to work is when short-run
players are ‘‘too cautious’’ vis-à-vis the long-run player and therefore
update their beliefs too slowly in order for the long-run player to find it
profitable to try to build a reputation. Such cases may occur when, in
addition to Stackelberg (‘‘good’’) types, the set of commitment types also
includes ‘‘bad’’ or ‘‘inept’’ types: players who always play the action that
the short-run players like least. In the ‘‘online auction’’ example, a ‘‘bad’’
type corresponds to a player who always cheats (because, for example, he
lacks the capabilities that would enable him to deliver high quality.) If
short-run players have a substantial prior belief that the long-run player
may be a ‘‘bad’’ type, then the structure of the game may not allow them to
update their beliefs fast enough to make it worthwhile for the long-run
player to try to acquire a reputation.
Diamond’s (1989) analysis of reputation formation in debt markets

presents an example of such a setting. In Diamond’s model, there are three
types of borrowers: safe borrowers, who always select safe projects (i.e.,
projects with zero probability of default); risky borrowers, who always
select risky projects (i.e., projects with higher returns if successful but with
nonzero probability of default); and strategic borrowers, who will select the
type of project that maximizes their long-term expected payoff. The ob-
jective of lenders is to maximize their long-term return by offering com-
petitive interest rates, while at the same time being able to distinguish
profitable from unprofitable borrowers. Lenders do not observe a bor-
rower’s choice of projects, but they do have access to her history of defaults.
In Diamond’s model, if lenders believe that the initial fraction of risky
borrowers is significant, then, despite the reputation mechanism, at the
beginning of the game, interest rates will be so high that strategic players
have an incentive to select risky projects. Some of them will default and will
exit the game. Others will prove lucky and will begin to be considered as safe
players. It is only after lucky strategic players have already acquired some
initial reputation (and therefore begin to receive lower interest rates) that it
becomes optimal for them to begin ‘‘masquerading’’ as safe players by con-
sciously choosing safe projects in order to sustain their good reputation.
Steady state (or lack thereof). In their simplest form, reputation games
are characterized by an equilibrium in which the long-run player repeat-
edly plays the Stackelberg action with high probability and the player’s
reputation converges to the Stackelberg type.
The existence of such steady states crucially depends on the ability to
perfectly monitor the outcomes of individual stage games. For example,
consider the ‘‘online auction’’ game that serves as an example throughout
this section with the added assumption that buyers perfectly and truthfully
observe and report the seller’s action. In such a setting, the presence of even
a single negative rating on a seller’s feedback history reveals the fact that
the seller is not honest. From then on, buyers will always choose the low bid
in perpetuity. Since such an outcome is not advantageous for the seller,
reputation considerations will induce the seller to cooperate forever.
The situation changes radically if monitoring of outcomes is imperfect. In
the online auction example, imperfect monitoring means that even when the
seller produces high quality, there is a possibility that a buyer will post a
negative rating, and, conversely, even when the seller produces low quality,
the buyer may post a positive rating. A striking result is that in such
‘‘noisy’’ environments reputations cannot be sustained indefinitely: if a
strategic player stays in the game long enough, short-run players will even-
tually learn his true type and the game will inevitably revert to one of the
static Nash equilibria (Cripps et al., 2004).
To see the intuition behind this result, note that reputations under perfect
monitoring are typically supported by a trigger strategy. Deviations from
the equilibrium strategy reveal the type of the deviator and are punished by
640 C. Dellarocas
a switch to an undesirable equilibrium of the resulting complete-informa-

tion continuation game. In contrast, when monitoring is imperfect, indi-
vidual deviations neither completely reveal the deviator’s type nor trigger
punishments. A single deviation has only a small effect on the beliefs of the
short-term players. As a result, a player of normal type trying to maintain a
reputation as a Stackelberg-type incurs only a small cost (in terms of altered
beliefs) from indulging in occasional deviations from Stackelberg play. In
fact, it is clear that always playing the Stackelberg action cannot be an
equilibrium strategy, because if the short-term players expect long-term
players of normal type to behave that way, then they can actually deviate at
no cost, since any bad outcome will be interpreted by the short-run players
as a result of imperfect monitoring. But the long-run effect of many such
small deviations from the commitment strategy is to drive the equilibrium
to full revelation.
These dynamics have important repercussions for reputation systems in
settings with both moral hazard and adverse selection (for example, eBay,
under the assumption that there exist rational and honest seller types).
According to the Cripps, Mailath and Samuelson result, if eBay makes the
entire feedback history of a seller available to buyers and if an eBay seller
stays on the system long enough, once he establishes an initial reputation
for honesty he will be tempted to cheat buyers every now and then. In the
long term, this behavior will lead to an eventual collapse of his reputation
and therefore of cooperative behavior. I revisit the implications of this
result for reputation mechanism design in Section 4.2.
Endgame considerations. Since reputation relies on a trade-off between
current ‘‘restraint’’ and the promise of future gains, in finitely repeated
games, incentives to maintain a reputation diminish and eventually disap-
pear as the end of the game comes close.
One solution to this problem is to assign some postmortem value to rep-
utation, so that players find it optimal to maintain it throughout the game.
For example, reputations can be viewed as assets that can be bought and sold
in a market for reputations. Tadelis (1999) shows that a market for repu-
tations is indeed sustainable. Furthermore, the existence of such a market
provides ‘‘old’’ agents and ‘‘young’’ agents with equal incentives to exert
effort (Tadelis, 2002). However, the long-run effects of introducing such a
market can be quite complicated since good reputations are then likely to be
purchased by ‘‘inept’’ agents for the purpose of depleting them (Mailath and
Samuelson, 2001; Tadelis, 2002). Further research is needed in order to fully
understand the long-term consequences of introducing markets for reputa-
tion as well as for transferring these promising concepts to the online domain.
3.3 When is reputation bad?
In traditional reputation theory, publication of a long-term player’s past

history of outcomes is good for the long-term player. One, therefore, is
tempted to assume that implementation of reputation mechanisms is always

a ‘‘good thing.’’ Ely et al. (2006), henceforth referred to as EFL, challenge
this assumption and show that there exist settings where the presence of
public histories of past outcomes is unambiguously bad.
EFL generalize an example provided by Ely and Valimaki (2003), hence-
forth referred to as EV. EV describe a setting where a mechanic of unknown
‘‘character’’ (rational, dishonest) is facing a sequence of customers who
bring their cars to him for repair. Each car might need a tune-up or an
engine change; only the mechanic is able to determine the correct type of
repair. Rational mechanics are assumed to have a payoff structure that
induces them to perform the correct type of repair in a one-stage game.
Dishonest mechanics, on the other hand, always perform engine changes.
Assume, now, that customers have access to the history of past repairs
performed by the mechanic and will only contract with him if they are
sufficiently confident that he is going to perform the correct type of repair.
Given the above assumptions, if there is a positive prior probability that the
mechanic is dishonest, after histories with many engine changes, the short-
run players will become sufficiently convinced they are facing such a bad
type and exit. In order to avoid these histories, a rational mechanic who has
had the ‘‘bad luck’’ of having many customers that need an engine change
may then begin to recommend tune-ups to customers who need engine
changes; foreseeing this, the short-run players will chose not to enter.
Observe that, whereas, in the absence of reputation, rational types play
friendly actions, the presence of a public history of past outcomes induces
them to behave in ways that short-run players like least. Foreseeing this,
short-run players choose not to enter; the presence of reputational infor-
mation then causes the market to break down.
EFL show that settings where reputation is bad are characterized by the
following properties: (i) a long-run player of privately known type is facing
an infinite sequence of short-run opponents, (ii) commitment types include
‘‘unfriendly’’ types who play ‘‘unfriendly’’ actions, that is, actions that
short-run players dislike, (iii) short-run players will only enter the game if
they are sufficiently confident that the long-run player is going to play one
of the friendly actions, (iv) there exist ‘‘bad signals’’ that are most likely to
occur when unfriendly actions are played but also occur with positive
probability when friendly actions are played (in EV the bad signal is ‘‘engine
change’’). Finally, (v) there are some actions that are not friendly, but reduce
the probability of bad signals (such as ‘‘always perform tune-up’’ in EV).
The main result of EFL is that, in a bad reputation game with a suffi-
ciently patient long-run player and likely enough unfriendly types, in any
Nash equilibrium, the long-run player gets approximately the payoff that
results from non-participation (exit) of short-run players.
EFL show that bad reputation games arise in various ‘‘expert advice’’
settings. This includes consulting a doctor or stockbroker, or in the mac-
roeconomics context, the decision whether or not to turn to the IMF for
642 C. Dellarocas
assistance. In EV, the short-run players observe only the advice, but not the
consequences of the advice. EFL consider what happens when the short-run
players observe the consequences as well. They show that the bad repu-
tation effect persists so long as this additional information is not perfectly
accurate.
The EFL/EV result delivers a pessimistic message with respect to the
application of reputation mechanisms in expert advice settings, an area
where the author believes that such mechanisms have a lot to offer. On the
other hand, the result crucially depends on the assumption that short-term
players will exit as soon as the long-term player’s reputation falls below a
threshold. If the long-term player can persuade them to participate (by
charging lower prices or, perhaps, paying them a participation reward) the
result breaks down: if a rational player who has had a stream of ‘‘bad luck’’
has a way to restore his reputation, he will continue to engage in friendly
actions. Nevertheless, reputation mechanism designers must be aware of the
EFL result and its implications. It is important to understand the range
of practical settings in which these results might apply, and therefore to
identify classes of settings for which the development of online reputation
mechanisms may not be a good idea.8
3.4 Other extensions to the basic theory
The basic theory assumes that uninformed players are short-term. Facing
longer-lived opponents may be worse for the informed player and generally
results in less sharp predictions about reputation effects (Cripps and
Thomas, 1995; Cripps et al., 1996). Quite interestingly, however, in repeated
games where a patient player faces one or more long-lived but less patient
opponents, if the more patient player does not observe the less patient
players’ intended actions but only sees an imperfect signal of them, rep-
utation effects once again become strong and result in lower bounds that
are even higher than in the case where all opponents are myopic (Celentani
et al., 1996). This last case is equivalent to a situation where a long-run
player faces a sequence of long-run but ‘‘infrequent’’ players. This is, per-
haps, an even more realistic model of relationships in online communities
and therefore an area that deserves further study.
8
A related, but not identical, problem arises when the establishment of a reputation mechanism
induces long-run players to change their behavior in ways that improve their payoff but reduce social
welfare. Dranove et al. (2003) examine the consequences of public disclosure of patient health outcomes
at the level of the individual physician and/or hospital. The intention behind this measure was to address
informational asymmetries in markets for health care. However, it also gave doctors and hospitals
incentives to decline to treat more difficult, severely ill patients. Using national data on Medicare
patients at risk for cardiac surgery, Dranove et al. find that cardiac surgery report cards in New York
and Pennsylvania led both to selection behavior by providers and to improved matching of patients with
hospitals. On net, this led to higher levels of resource use and to worse health outcomes, particularly for
sicker patients.
Another assumption underlying most game theoretic models of reputa-

tion is that all players have identical prior beliefs and that behavior is
consistent with the concept of Bayesian Nash equilibrium. These assump-
tions are probably too stringent and unrealistic in environments as diverse
as large-scale online communities. Fortunately, reputation phenomena arise
under significantly weaker assumptions on the knowledge and behavior of
players. Watson (1993, 1996) and Battigalli and Watson (1997) demon-
strated that reputation effects do not require equilibrium. They are implied
by a weak notion of rationalizability along with two main conditions on the
beliefs of players: first, there must be a strictly positive and uniform lower
bound on the subjective probability that players assign to the Stackelberg
type. Second, the conditional beliefs of short-run players must not be ‘‘too
dispersed.’’
4 New opportunities and challenges of online mechanisms
I began this chapter by discussing a number of differences between online

reputation mechanisms and traditional word-of-mouth networks. This sec-
tion surveys our progress in understanding the opportunities and challenges
that these special properties imply.
4.1 Eliciting sufficient and honest feedback
Most game theoretic models of reputation formation assume that stage-

game outcomes (or imperfect signals thereof) are publicly observed. Most
online reputation mechanisms, in contrast, rely on private monitoring of
stage-game outcomes and voluntary self-reporting.9
This introduces two important new considerations: (a) ensuring that
sufficient feedback is, indeed, provided, and (b) inducing truthful reporting.
Economic theory predicts that voluntary feedback will be underprovided.
There are two main reasons for this. First, feedback constitutes a public
good: once available, everyone can costlessly benefit from it. Voluntary
provision of feedback leads to suboptimal supply, since no individual takes
account of the benefits that her provision gives to others. Second, provision
of feedback presupposes that the rater will assume the risks of transacting
with the ratee. Such risks are highest for new products. Prospective con-
sumers may, thus, be tempted to wait until more information is available.
However, unless somebody decides to take the risk of becoming an early
evaluator, no feedback will ever be provided.
Avery et al. (1999) analyze mechanisms whereby early evaluators are paid
to provide information and later evaluators pay in order to balance the
9
For a comprehensive introduction to games with imperfect private monitoring see Kandori (2002)
and other papers contained in the same issue.
644 C. Dellarocas
budget. They conclude that, of the three desirable properties for such a
mechanism (voluntary participation, no price discrimination, and budget
balance), any two can be achieved, but not all three.10
Since in most reputation mechanisms, monitoring of transaction out-
comes is private, an additional consideration is whether feedback is honest.
No generally applicable solution to this important problem currently exists.
Nevertheless, several researchers have proposed mechanisms that induce
truth-telling in restricted settings.
Jurca and Faltings (2004) propose a mechanism that limits false reporting
in settings with pure moral hazard. They consider a bilateral exchange
setting where long-run buyers and long-run sellers transact repeatedly. Each
period a buyer is asked to rate a transaction only if the corresponding seller
has claimed to have successfully provided the service. If the two reports
disagree, at least one of the traders must be lying; the center then fines both
transacting parties different (fixed) amounts.11 Jurca and Faltings show
that, if a buyer always reports feedback on a particular seller truthfully, the
seller also finds it optimal to truthfully report transaction outcomes, with
the exception of a finite number of transactions.
Papaioannou and Stamoulis (2005) propose a similar mechanism that is
suitable for peer-to-peer environments. In such environments side payments
are usually not possible and peers are able to exchange roles. After each
transaction both peers submit a rating, with each peer not knowing the
rating submitted by the other. A credibility metric is maintained for each
peer regarding his overall truthfulness record in rating transactions. In case
of agreement between two peers, the credibility metric of each peer is
improved. In case of disagreement, the credibility metric of each peer is
deteriorated and both peers are punished. Punishment amounts to not
allowing a peer to transact with others for a period that is exponential to the
peer’s credibility metric. This is enforced by publicly announcing a peer’s
punishment and by punishing other peers when they transact with him.
Performance of this mechanism is analyzed by means of experiments in
dynamically evolving peer-to-peer systems with renewed populations.
Miller et al. (2005) (MRZ) propose mechanisms for eliciting honest feed-
back in environments with pure adverse selection. Their mechanisms are
based on the technique of proper scoring rules (Cooke, 1991). A scoring
rule is a method that induces rational agents to truthfully reveal their beliefs
10
Empirical evidence has, so far, not confirmed the (rather pessimistic) predictions of theory. Con-
sumer participation in online feedback mechanisms is surprisingly high, even though, in most cases, such
mechanisms offer no concrete participation incentives. Such behavior is consistent with a large body of
empirical evidence (Dichter, 1966; Sundaram et al., 1998; Hennig-Thurau et al., 2004) that has identified
a variety of extra-economic motivations to explain why consumers engage in (offline and online) word-
of mouth (desire to achieve social status, utility from engaging in social interaction, altruism, concern for
others, easing anger, dissonance reduction, vengeance, etc.). I return to this point in Section 5.
11
Fines can be implemented by levying refundable listing fees from the two traders at the beginning of
each transaction and confiscating these fees in the case of conflicting reports.
about the distribution of a random variable by rewarding them on the basis

of how a future realization of the random variable relates to the distribution
they announced. A proper scoring rule has the property that the agent
maximizes his expected score when he truthfully announces his beliefs.
Assuming that (i) users rate a set of products, whose types (e.g., qualities)
remain fixed over time, (ii) each type maps to a distribution of outcomes
(e.g., satisfaction levels perceived by consumers) that can be statistically
distinguished from that of every other type, and (iii) all raters and the center
have a common set of prior beliefs over types, or, alternatively, each rater’s
prior beliefs have been communicated to the center, MRZ propose a side-
payment mechanism that asks each rater to announce the outcome a1 she
observed, and rewards her by an amount proportional to log p(a29a1) where
a2 is the outcome reported by a future rater. MRZ show the existence of a
Nash equilibrium where such a side-payment rule induces truth-telling.
They also show that their mechanism can be extended to not only induce
truth-telling, but also ratings of a given precision.
The above mechanisms represent promising first steps toward ensuring
the credibility of online feedback. On the other hand, they have several
limitations. First, such mechanisms only work as long as raters are assumed
to act independently; all break down if raters collude. Second, in addition
to the desirable truth-telling equilibria., all three mechanisms induce
additional equilibria where agents do not report the truth. Equilibrium
selection, thus, becomes an important consideration in practical implemen-
tations. The development of robust mechanisms for eliciting truthful feed-
back thus remains one of the most important open areas of research.
4.2 Exploring the design space of feedback mediators
Information technology has added a large degree of flexibility and pre-

cision to the design of reputation mechanisms. Online mechanism designers
can control a number of important parameters that are difficult, or impos-
sible, to influence in offline settings. Examples of such parameters include
the format of solicited feedback (eBay allows traders to rate a transaction as
‘‘positive,’’ ‘‘negative,’’ or ‘‘neutral,’’ Amazon Auctions supports integer
ratings from 1 to 5, other systems support even higher levels of detail), the
amount and type of information included in a trader’s reputation profile
(most systems publish the sum or arithmetic mean of all posted ratings,
some systems highlight recent ratings, other systems provide access to a
trader’s entire ratings history) as well as the frequency with which repu-
tation profiles are updated with new information (most current systems
make new ratings publicly available as soon as they are posted). These
parameters impact the consequences of a trader’s current behavior on the
community’s perception of him in the future, and thus, his incentives to
cooperate. This section summarizes our current understanding of how such
design choices affect trader behavior and market efficiency.
646 C. Dellarocas
Granularity of feedback. eBay’s mechanism solicits ternary feedback

(positive, neutral, negative), while Amazon Auctions asks traders to rate
transactions on a finer-grained scale of 1–5. What is the impact of the
granularity of feedback on the seller’s incentives to cooperate? In the special
case of settings with two seller actions (cooperate, cheat) and pure moral
hazard, Dellarocas (2005) shows that the equilibrium that maximizes
cooperation occurs when buyers divide feedback of arbitrary granularity
into two disjoint subsets (the ‘‘good ratings’’ subset and the ‘‘bad ratings’’
subset) and behave as if feedback was binary, i.e., they reward the seller by
the same amount if he receives any rating that belongs to the ‘‘good’’ set
and they punish him by the same amount if he receives any rating that
belongs to the ‘‘bad’’ set. Maximum efficiency is then inversely proportional
to the minimum likelihood ratio of punishment if the seller cooperates vs. if
he cheats over all possible ways of dividing ratings into ‘‘good’’ and ‘‘bad’’
subsets.12
Length of published feedback history. How much history should a mech-
anism publish about a trader’s past behavior? The answer here crucially
depends on the type of setting (moral hazard, adverse selection, or com-
bined). In pure adverse selection settings, the goal of the mechanism is to
promote social learning. More information is, thus, always better. In con-
trast, in pure moral hazard settings Dellarocas (2005) shows that the max-
imum efficiency that can be induced by a reputation mechanism is
independent of the length of published history: a mechanism that publishes
a trader’s entire history performs as well as a mechanism that only publishes
the trader’s single most recent rating. The intuition behind this result is that,
in pure moral hazard settings, the goal of the reputation mechanism is to
threaten players with future punishment if the public outcome of the current
transaction indicates that they cheated. The length of time during which a
rating persists in the seller’s reputation profile only affects the duration of
future punishment but not the total amount of punishment. This makes
efficiency independent of the length of published history.
Finally, in settings where both moral hazard and adverse selection are
present, if the objective of the mechanism is to induce long-term seller
cooperation, Cripps, Mailath and Samuelson’s result (see Section 3.2)
suggests that reputation mechanisms should not publish a seller’s entire
history. (Otherwise, once sellers have established a good reputation, they
will be tempted to occasionally cheat; in the long run, this will reveal their
opportunistic nature and will drive the system to the undesirable one-shot
Nash equilibrium.) On the other hand, since their role is to promote some
degree of learning about the seller’s type, reputation mechanisms should
12
In more general settings where both adverse selection and moral hazard are present, the answer is
generally more complex. Soliciting (and publishing) finer-granularity feedback might or decrease a
seller’s incentives to cooperate. See Dewatripont et al. (1999) for a precise statement of the relevant
conditions.
publish some among of past history. Fan et al. (2005) explore the idea of
discounting past ratings in such settings. Using simulation, they show that
such policies can sustain a seller’s incentives to cooperate. General design
guidelines for deciding what is the optimal length of history (or, equiva-
lently, what is the optimal discount factor of past ratings) constitute an
interesting open area of research.
Frequency of feedback profile updating. Most systems in use today update
a trader’s reputation profile with new evaluations as soon as these are
posted by users. Dellarocas (2006a) shows that, in pure moral hazard set-
tings, this is not necessarily the optimal architecture. Specifically, he shows
that, if ratings are noisy and the per-period profit margin of cooperating
sellers sufficiently high, a mechanism that does not publish every single
rating it receives but rather, only updates a trader’s public reputation profile
every k transactions with a summary statistic of a trader’s most recent
k ratings, can induce higher average levels of cooperation and market effi-
ciency than a mechanism that publishes all ratings as soon as they are
posted. The intuition behind the result is that delayed updating reduces
the impact of spurious negative ratings (because these are amortized over
k transactions). On the other hand, it also increases the trader’s temptation
to cheat (because he can cheat for k period before news of his behavior
become public). The optimal updating period k is derived from a trade-off
between these two opposite forces.
4.3 Coping with cheap online identities
In online communities it is usually easy for members to disappear and

re-register under a completely different online identity with zero or very low
cost. Friedman and Resnick (2001) refer to this property as ‘‘cheap pseu-
donyms.’’ This property hinders the effectiveness of reputation mecha-
nisms. Community members can build a reputation, milk it by cheating
other members and then vanish and re-enter the community with a new
identity and a clean record.
Friedman and Resnick discuss two classes of approaches to this issue:
either make it more difficult to change online identities, or structure the
community in such a way that exit and re-entry with a new identity becomes
unprofitable. The first approach makes use of cryptographic authentication
technologies and is outside the scope of this paper. The second approach is
based on imposing an upfront cost to each new entrant, such that the
benefits of ‘‘milking’’ one’s reputation are exceeded by the cost of subse-
quent re-entry. This cost can be an explicit entrance fee or an implicit cost
of having to go through a reputation building (or dues paying) stage with
low or negative profits. Friedman and Resnick show that, although dues
paying approaches incur efficiency losses, such losses constitute an inev-
itable consequence of easy name changes.
648 C. Dellarocas
Dellarocas (2005) shows how such a ‘‘dues paying’’ approach can be

implemented in pure moral hazard trading environments. He proves that, in
the presence of easy name changes, the design that results in optimal social
efficiency is one where newcomers pay an entrance fee and the mechanism
only publishes a trader’s single most recent rating. Dellarocas further dem-
onstrates that, although this design incurs efficiency losses relative to the
case where identity change is not an issue, in settings with two possible
transaction outcomes and opportunistic sellers that can freely change iden-
tities, its efficiency is the highest attainable by any mechanism.
4.4 Understanding the consequences of strategic manipulation
As online reputation mechanisms begin to exercise greater influence

on consumer behavior the incentive for strategically manipulating them
becomes correspondingly stronger. The low cost of submitting online feed-
back coupled with the relative anonymity of the rater makes such manip-
ulation a real problem that needs to be studied and addressed.
Dellarocas (2000, 2004) has pointed out a number of manipulation sce-
narios and has proposed a number of immunization mechanisms, based on
ideas from clustering and robust statistics, that reduce the effect of such
attacks if the fraction of unfair raters is reasonably small (up to 20–30% of
the total population). A number of commercial reputation mechanisms (for
example, Amazon Reviews and Epinions) are attempting to address such
problems through the concept of ‘‘rate the rater’’: members can rate how
useful other members’ feedback has been to them. Whereas this technique is
somewhat effective for separating high-quality from low-quality postings,
it is not effective for reducing strategic manipulation. Determined manip-
ulators can manipulate the ‘‘rate the rater’’ ratings as much as they can
manipulate the ratings themselves.
The ability to ‘‘cheaply’’ create multiple online identities further compli-
cates attempts to combat manipulation attacks. For example, an early
technique through which eBay traders attempted to manipulate their rep-
utation profile was for colluding traders to form a ring, engaging in re-
peated ‘‘fake’’ transactions with each other for the purpose of artificially
boosting each other’s feedback score. To combat such attacks, eBay is now
only counting unique ratings received from frequent partners toward a
trader’s reputation profile. Such defenses can be defeated, however, if trad-
ers can easily create multiple identities. Douceur (2002) has called this the
sybil attack. Cheng and Friedman (2005) provide formal definitions of sybil-
proofness, i.e., resistance to sybil attacks, and prove the impossibility of
developing sybilproof reputation mechanisms is a broad class of settings.
Since it appears that complete elimination of reputation mechanism
manipulation will remain a formidable task for some time, it is useful to
explore the consequences of such activity in strategic settings. In settings
with pure adverse selection, Mayzlin (2003) has analyzed the impact of
strategic manipulation in the context of Usenet groups where consumers

discuss products and services. Mayzlin’s setting involves two rival firms that
wish to influence consumer beliefs regarding the quality of their respective
products by posting costly ‘‘fake’’ promotional messages in Usenet groups.
Mayzlin’s basic result is that, if the ratio of profits to manipulation cost is
high enough, there exists an equilibrium in which both firms manipulate but
the low-quality firm manipulates more. Promotional chat thus decreases,
but does not completely destroy, the informativeness of online forums.
Dellarocas (2006b) extended Mayzlin’s results in more general settings. His
most interesting new result is that there exist equilibria where strategic
manipulation can increase forum informativeness. Such equilibria arise in
settings where firm revenues are sufficiently convex functions of consumer
perceptions of their quality. In such settings, the presence of honest con-
sumer opinions can induce firms to reveal their own, more precise, knowl-
edge of product qualities by manipulating the forums at relative intensities
that are proportional to their actual qualities. The impact of manipulation
then is to further separate the published ratings of the high-quality firm
from those of the low-quality firm, making it easier for consumers to infer
each firm’s true quality.
4.5 Distributed reputation mechanisms
Traditional reputation theory assumes the existence of a public history of

past outcomes. This implies a centralized architecture where outcomes are
either automatically recorded or explicitly self-reported. It also implies the
presence of a trusted mediator who controls feedback aggregation and dis-
tribution. Though the design possibilities of even that simple architecture
are not yet fully understood, centralized reputation mechanisms do not
nearly exhaust the new possibilities offered by information technology.
The growing importance of peer-to-peer (P2P) networks (Oram, 2001) is
introducing new challenges to reputation mechanism design. In P2P net-
works every entity can act both as a provider and consumer of resources.
Entities are assumed to be self-interested and cannot be trusted to engage in
cooperative behavior unless concrete incentives are in place. For example,
in file-sharing P2P networks, self-interested entities have short-term incen-
tives to free ride (consume as much content as they can without contrib-
uting any content themselves) or to contribute low-quality content.
Furthermore, there is, usually, no central, universally trusted entity that
can act as a repository of reputational information.
To cope with these challenges, several researchers have proposed decen-
tralized reputation mechanisms. Two lines of investigation stand out as
particularly promising:
Reputation formation based on analysis of ‘‘implicit feedback.’’ Traditional
reputation mechanisms rely on explicit solicitation of feedback from trans-
action participants. If reliable explicit feedback is not available, information
650 C. Dellarocas
about an agent’s type can often be inferred by analyzing publicly available

attributes of the network in which the agent is embedded.
Perhaps the most successful application of this approach to date is
exemplified by the Google search engine. Google’s PageRank algorithm
assigns a measure of reputation to each web page that matches the key-
words of a search request. It then uses that measure to rank order search
hits. Google’s page reputation measure is based on the number of links that
point to a page, the number of links that point to the pointing page, and so
on (Brin and Page, 1998). The underlying assumption is that if enough
people consider a page to be important enough in order to place links to
that page from their pages, and if the pointing pages are ‘‘reputable’’
themselves, then the information contained on the target page is likely to be
valuable. Google’s success in returning relevant results is testimony to the
promise of that approach.
Pujol et al. (2002) apply network flow techniques in order to propose a
generalization of the above algorithm that ‘‘extracts’’ the reputation of
nodes in a general class of social networks. Sabater and Sierra (2002)
describe how direct experience, explicit and implicit feedback can be
combined into a single reputation mechanism.
Basing reputation formation on implicit information is a promising so-
lution to problems of eliciting sufficient and truthful feedback. Careful
modeling of the benefits and limitations of this approach is needed in order
to determine in what settings it might be a viable substitute or complement
of voluntary feedback provision.
Distributed reputation mechanisms. The majority of decentralized repu-
tation mechanisms proposed so far are based on variations on the theme of
referral networks (Aberer and Despotovic, 2001; Yu and Singh, 2002;
Kamvar et al., 2003; Xiong and Liu, 2004). Since no centralized reputation
repository exists, each agent in a P2P network solicits referrals about a
target agent from a set of neighbors who might, in turn, ask their neighbors,
and so on. Referrals are weighted by the relative amounts of trust that
the soliciting agent places on his neighbors’ advice and combined with
personal experiences that the soliciting agent might have had with the target
agent in the past. Based on his subsequent experience with the target
agent, the soliciting agent dynamically adapts both his beliefs regarding the
target agent as well as the amounts by which she trusts her neighbors’
recommendations.13
Most work in this area is currently based on heuristics and evaluated
using simulation. The development of rigorous results on the efficiency
bounds and design principles of distributed reputation mechanisms consti-
tutes an important open area of research.
13
See Despotovic and Aberer (2004) for a more comprehensive overview of trust-building mechanisms
in P2P networks.
5 Empirical and experimental studies
A growing literature of empirical and experimental studies constitutes

an essential complement to game-theoretic and simulation-based analyses
of reputation mechanisms. This section surveys some of the main themes of
such work.
5.1 Empirical studies and field experiments
Most empirical studies of reputation mechanisms have focused on eBay’s

feedback mechanism. The majority of early studies has looked at the buy-
ers’ response to seller reputation profiles. In particular, a large number of
studies estimate cross-sectional regressions of sale prices and probabilities
of sale on seller feedback characteristics. Table 2 summarizes the main
results of these studies.
The following points summarize the principal conclusions derived from a
collective reading of these works:
Feedback profiles seem to affect both prices and the probability of sale.
However, the precise effects are ambiguous; different studies focus on
different components of eBay’s complex feedback profile and often
reach different conclusions.
The impact of feedback profiles on prices and probability of sale is
relatively higher for riskier transactions and more expensive products.
Among all different pieces of feedback information that eBay publishes
for a member, the components that seem to be most influential in
affecting buyer behavior are the overall number of positive and neg-
ative ratings, followed by the number of recently (last seven days, last
month) posted negative comments.
Resnick et al. (2006) point out the potential for significant omitted
variable bias in these cross-sectional regressions. They argue that the price
difference commanded by sellers of higher reputation might be due to sev-
eral other factors that tend to exhibit positive correlation with a seller’s
eBay score (for example, the clarity and aesthetics of the item listing, the
professionalism of the seller’s e-mail communications, etc.). To better assess
the premium attached to reputation they conduct a controlled field exper-
iment in which a seasoned seller sells identical postcards using his real name
and an assumed name. They find an 8% premium to having 2,000 positive
feedbacks and 1 negative over a feedback profile with 10 positive comments
and no negatives.
In another field experiment, Jin and Kato (2004) assess whether the
reputation mechanism is able to combat fraud by purchasing ungraded
baseball cards with seller-reported grades, and having them evaluated by
the official grading agency. They report that, while having a better seller
reputation is a positive indicator of honesty, reputation premia or discounts
652 C. Dellarocas
Table 2
Summary of early empirical studies on eBay
Shorthand Citation Items sold Remarks
BP Ba and Pavlou Music, software, Positive feedback increased

(2002) electronics estimated price, but
negative feedback did not
have an effect
BH Bajari and Coins Both positive and negative
Hortacsu feedback affect probability
(2003) of modeled buyer entry into
the auction, but only
positive feedback had a
significant effect on final
price
DH Dewan and Hsu Stamps Higher net score increased
(2004) price
E Eaton (2005) Electric guitars Negative feedback reduces
probability of sale, but not
price of sold items
HW Houser and Pentium chips Positive feedback increases
Wooders price; negative feedback
(2006) reduces it
KM Kalyanam and Palm pilot PDAs Positive feedback increases
McIntyre price; negative feedback
(2001) reduces price
KW Kauffman and Coins No significant effects, but
Wood (2000) negative feedback seems to
increase price (!) in
univariate analysis
LIL Lee et al. (2000) Computer Negative feedback reduces
monitors and price, but only for used
printers items
L Livingston (2002) Golf clubs Positive feedback increases
both likelihood of sale and
price; effect tapers off once
a record is established
LBPD Lucking-Reiley Coins No effect from positive
et al. (2006) feedback; negative
feedback reduces price
MA Melnik and Alm Gold coins Positive feedback increases
(2002) price; negative feedback
decreases price
MS McDonald and Dolls Higher net score (positives–
Slawson (2002) negatives) increases price
RZ Resnick and MP3 players, Both forms of feedback affect
Zeckhauser Beanie babies probability of sale but not
(2002) price contingent of sale
Shorthand Citation Items sold Remarks
RZSL Resnick et al. Vintage Controlled field experiment;

(2006) postcards established seller
commands higher prices
than newcomers; among
newcomers, small amounts
of negative feedback have
little effect
Source: Adapted from Resnick et al. (2006)
in the market do not fully compensate for expected losses due to seller
dishonesty.
Using panel data, Cabral and Hortacsu (2006) analyze the impact of
reputation on eBay sellers. They find that, when a seller first receives neg-
ative feedback, his weekly sales rate drops from a positive 7% to a negative
7%; subsequent negative feedback ratings arrive 25% more rapidly than the
first one and do not have as much impact as the first one. They also find
that a seller is more likely to exit the lower his reputation is; and that, just
before exiting, sellers receive more negative feedback than their lifetime
average.
Cabral and Hortacsu then consider a series of theoretical models (pure
moral hazard, pure adverse selection, combined moral hazard, and adverse
selection) and measure them against these empirical results. They are not
able to draw definite conclusions as to which theoretical model best explains
the data; they do, however, conclude that eBay’s reputation system gives
way to noticeable strategic responses from both buyers and sellers.
Another interesting area of empirical work relates to the motivations for
participation in online feedback mechanisms. Economic theory predicts
that, being public goods, evaluations will be underprovided unless evalu-
ators are provided with concrete incentives (Avery et al., 1999). Neverthe-
less, in systems such as eBay, more than 50% of transactions receive
feedback even though no direct incentives to rate are in place. Dellarocas
and Wood (2006c) use survival analysis techniques to study the motivations
and dynamics of voluntary feedback submission on eBay. They find that
reciprocity is an important driver of participation on eBay’s mechanism:
a trader’s propensity to rate a transaction increases after her partner posts a
rating. They also find that silence is often an indicator of an unsatisfactory
transaction and propose an empirical method for inferring the true number
of unsatisfactory transactions on eBay from the observed patterns of feed-
back submission. Using their method, they estimate that, whereas the per-
centage of negative feedback on eBay is less than 1%, the fraction of
transactions where, at least one of the two parties remained unsatisfied is
substantially higher.
654 C. Dellarocas
Product review forums (such as Amazon Reviews) have, so far, received

less attention. Nevertheless, their increasing popularity is attracting more
empirical research. Hennig-Thurau et al. (2004) conduct an online survey to
identify what motivates consumers to rate products in such forums. Their
study identifies four significant motives. In order of decreasing significance
the four motives are social benefits (genuine fun that results from the social
experience of participating in an online dialog), economic incentives (re-
wards offered by some sites for contributing content), concern for other
consumers (urge to help others by recommending good products and
warning against bad products), and extraversion/self-enhancement (positive
feelings that result from sharing one’s successes with others; enhancement
of one’s self-image by projecting oneself as intelligent shopper).
Another question of theoretical and practical consequence is the extent to
which such forums influence consumer behavior. Obtaining reliable answers
to this question is challenging. eBay has the advantage of being a ‘‘closed
universe’’: all transactions and feedback are mediated through the site and
are publicly visible. On the other hand, in the general case, consumers
can obtain product ratings from one forum and purchase the product
from another site, or from an offline store. It is, thus, difficult to establish
causality between ratings and sales.
The approach of Chevalier and Mayzlin (2003) constitutes a step in the
right direction. Chevalier and Mayzlin examine the effect of consumer re-
views on relative sales of the same set of books on Amazon.com and
BarnesandNoble.com, two large online booksellers.14 They find that an
improvement in book’s reviews on one site leads to an increase in relative
sales at that site. They also find that reviews are overwhelmingly positive
and, thus, that the impact of (less common) negative reviews is greater than
the impact of positive reviews.
5.2 Controlled experiments
Controlled experiments are a useful complement to empirical studies,

especially in relation to exploring individual-level perceptions of and reac-
tions to reputation mechanisms.
Keser (2003) reports a number of experiments based on the ‘‘trust game.’’
In this game, one player (the ‘‘buyer’’) can choose to send money to a
second (the ‘‘seller’’). This amount is then substantially increased and the
seller can choose to share some of the gain with the buyer. By removing
many of the complexities involved in market transactions, this game pro-
vides a simple context to study the effect of different information policies
about revealing past behaviors. Keser finds that the presence of a reputa-
tion mechanism significantly increases both the buyers’ level of investment
14
Book sales are inferred from the publicly available ‘‘sales rank’’ metric using an approach intro-
duced by Chevalier and Goolsbee (2003).
(trust) as well as the fraction that sellers share with their buyers (trustwor-
thiness). Furthermore, she finds that both trust and trustworthiness are
higher when the reputation mechanism publishes the entire history of each
player’s past behavior than when it publishes each player’s behavior in the
most recent transaction only.
Questions involving reputation’s effect on market efficiency require more
complex experimental scenarios. Bolton et al. (2004) (BKO) study trading
in a fixed-price market where buyers can choose whether to send the pur-
chase price to the seller and sellers have the option of not fulfilling their
contracts. They compare transaction completion rates in a setting with
random matching of players and public histories of trader fulfillment (rep-
utation), to a setting with random matching without reputation as well as to
a market where the same people interact with each other repeatedly (part-
ners market). They find that, while the presence of the reputation mech-
anism induces a substantial improvement in trading efficiency, it falls short
of the efficiency achieved in the partners market.
Chen et al. (2004) (CHW) conduct experiments similar to those per-
formed by BKO but provide a broader set of endogenous choices for the
players. First, players can explicitly decide who they wish to do business
with rather than being paired with a single other player by the experimenter.
Second, both buyers and sellers make fulfillment choices and so face a
moral hazard for which reputations are relevant. Third, in addition to
settings with automatically generated reputation, CHW examine games
where players self-report transaction outcomes, allowing them to misreport
their experiences as possible punishment for a poor report on their own
reputation. Fourth, prices and trading volumes are determined endog-
enously. The findings of CHW are consistent with the previous experiments:
the presence of reputational information led to a substantial increase of
transaction fulfillment. Interestingly, CHW found fulfillment rates to be
almost as high when traders self-reported transaction outcomes as when
reliable reputational information was automatically generated, indicating
that, in practice, private observability of outcomes might not be as big a
problem as theory suggests.
6 Conclusions: opportunities for IS research
Online reputation mechanisms harness the remarkable ability of the Web

to not only disseminate, but also collect and aggregate information from
large communities at very low cost, in order to artificially construct large-
scale word-of-mouth networks. Best known so far as a technology for
building trust and fostering cooperation in online marketplaces, these
mechanisms are poised to have a much wider impact on organizations.
The design of online reputation mechanisms can greatly benefit from the
insights produced by more than 20 years of economics and game theory
656 C. Dellarocas
research on the topic of reputation. These results need to be extended to

take into account the unique new properties of online environments, such as
their unprecedented scalability, the ability to precisely design the type of
feedback information that is solicited and distributed, and the volatility of
online identities. The following list contains what the author considers to be
the most important open areas of research in reputation mechanism design:
Scope and explore the design space and limitations of online reputation
mechanisms. Understand what set of design parameters work best in
what settings. Develop models and prototype implementations of such
systems.
Develop effective solutions to the problems of sufficient participation,
easy identity changes and strategic manipulation of online feedback.
Conduct theory-driven experimental and empirical research that sheds
more light on buyer and seller behavior vis-à-vis such mechanisms.
Compare the relative efficiency of reputation mechanisms to the effi-
ciency of more established mechanisms for dealing with moral hazard
and adverse selection (such as state-backed contractual guarantees and
advertising); develop theory-driven guidelines for deciding which set of
mechanisms to use when.
Understand how decision-makers must adapt their strategies to react
to the presence of such mechanisms in areas such as marketing, prod-
uct development, and customer service.
The power of online reputation mechanisms has its roots in the strategic
side effects brought about by the increased interdependencies they create
among firms, their customers, their partners, and their competitors. As the
mathematical study of interaction of self-interested agents, game theory is
the natural foundation for the study of online reputation mechanisms.
Other established paradigms of information systems research can play an
important role in translating the conceptual insights of game theoretic
models into concrete guidelines for building (and reacting to) large-scale
reputation mechanisms that can influence the dynamics of entire industries
or societies. For example, computational methods can help analyze games
that may be too complex to solve analytically. Laboratory experiments can
inform about how people will behave when confronted with these mech-
anisms, both when they are inexperienced and as they gain experience.
Finally, game-theoretic models can often be approximated by generally
more tractable, decision-theoretic approaches.15
15
See Zacharia et al. (2001) for an example of using simulation modeling to study the effects of
reputation mechanisms on markets with dynamic pricing. See Shapiro (1982) and Dellarocas (2001) for
examples of how the adoption of ‘‘reasonable’’ (commonsense or empirically driven) assumptions about
the behavior of other players permits the use of decision-theoretic approaches to model reputation
phenomena.
There is much work to be done. But it is important that research be

conducted now, in the formative phases of this technology and the social
practices surrounding it. There are likely to be path-dependent effects in the
deployment and use of online reputation mechanisms, so it is important
that researchers develop insights into the functioning and impacts of these
systems while they can still have a large impact on practice.
References
Aberer, K., Z. Despotovic (2001). Managing trust in a Peer-2-Peer information system. Proceedings of the
2001 ACM International Conference on Information and Knowledge Management (CIKM), pp. 310–317.
Akerlof, G. (1970). The market for lemons: quality uncertainty and the market mechanism. Quarterly
Avery, C., P. Resnick, R. Zeckhauser (1999). The market for evaluations. American Economics Review
89(3), 564–584.
Ba, S., P. Pavlou (2002). Evidence of the effect of trust building technology in electronic markets: price
premiums and buyer behavior. MIS Quarterly 26(3), 243–268.
Bajari, P., A. Hortacsu (2003). Winner’s curse, reserve prices and endogenous entry: empirical insights
from eBay auctions. Rand Journal of Economics 34(2), 329–355.
Bakos, Y., C. Dellarocas (2002). Cooperation without enforcement? A comparative analysis of litigation
and online reputation as quality assurance mechanisms, in: L. Applegate, R. Galliers, J.I. DeGross
(eds.), Proceedings of the 23rd International Conference on Information Systems (ICIS 2002), As-
sociation for Information Systems, Barcelona, Spain, pp. 127–142.
Battigalli, P., J. Watson (1997). On reputation refinements with heterogeneous beliefs. Econometrica 65,
369–374.
Benson, B. (1989). The spontaneous evolution of commercial law. Southern Economic Journal 55(Jan-
uary) 644–661, in: D.B. Klein (ed.), Reputation: Studies in the Voluntary Elicitation of Good Conduct.
University of Michigan Press, Ann Arbor, pp. 165–189 (Reprinted in 1997).
Bolton, G.E., E. Katok, A. Ockenfels (2004). How effective are online reputation mechanisms? An
experimental study. Management Science 50(11), 1587–1602.
Brin, S., L. Page (1998). The anatomy of a large-scale hypertextual web search engine. Computer
Networks and ISDN Systems 30(1–7), 107–117.
Cabral, L., A. Hortacsu (2006). The dynamics of seller reputation: theory and evidence from eBay (April
2004). CEPR Discussion Paper No. 4345. Available at http://ssrn.com/abstract=541161.
Celentani, M., D. Fudenberg, D. Levine, W. Pesendorfer (1996). Maintaining a reputation against a
long lived opponent. Econometrica 64, 691–704.
Chen, K–Y, T. Hogg, N. Wozny (2004). Experimental study of market reputation mechanisms. Pro-
ceeding of the ACM Conference on Electronic Commerce 2004, pp. 234–235.
Cheng, A., E. Friedman (2005). Sybilproof reputation mechanisms. Proceedings of the 3rd Workshop on
the Economics of Peer-2-Peer Systems (P2PECON), Philadephia, PA.
Chevalier, J.A., A. Goolsbee (2003). Price competition online: Amazon versus Barnes and Noble.
Quantitative Marketing and Economics 1(2), 203–222.
Chevalier, J.A., D. Mayzlin (2003). The effect of word of mouth on sales: online book reviews (August 6,
2003). Yale SOM Working Paper No’s. ES-28 & MK-15. Available at http://ssrn.com/ab-
stract=432481.
Cooke, R.M. (1991). Experts in Uncertainty: Opinion and Subjective Probability in Science. Oxford
University Press, New York, NY.
Cripps, M., G. Mailath, L. Samuelson (2004). Imperfect monitoring and impermanent reputations.
Econometrica 72, 407–432.
Cripps, M., K. Schmidt, J. Thomas (1996). Reputation in perturbed repeated games. Journal of Eco-
nomic Theory 69, 387–410.
658 C. Dellarocas
Cripps, M., J. Thomas (1995). Reputation and commitment in two-person repeated games without
discounting. Econometrica 63, 1401–1419.
Dellarocas, C. (2000). Immunizing online reputation reporting systems against unfair ratings and dis-
criminatory behavior. Proceedings of the 2nd ACM Conference on Electronic Commerce. Association
for Computing Machinery, Minneapolis, MN, pp. 150–157.
Dellarocas, C. (2001). Analyzing the economic efficiency of eBay-like online reputation reporting
mechanisms. Proceedings of the 3rd ACM Conference on Electronic Commerce. Association for
Computing Machinery, Tampa, FL, pp. 171–179.
Dellarocas, C. (2003). The digitization of word-of-mouth: promise and challenges of online feedback
mechanisms. Management Science 49(10), 1407–1424.
Dellarocas, C. (2004). Building trust on-line: the design of robust reputation mechanisms for online
trading communities, in: G. Doukidis, N. Mylonopoulos, N. Pouloudi (eds.), Social and Economic
Transformation in the Digital Era, Idea Group Publishing, Hershey, PA.
Dellarocas, C. (2005). Reputation mechanism design in online trading environments with pure moral
hazard. Information Systems Research 16(2), 209–230.
Dellarocas, C. (2006a). How often should reputation mechanisms update a trader’s reputation profile?
Information Systems Research, forthcoming.
Dellarocas, C. (2006b). Strategic manipulation of internet opinion forums: implications for consumers
and firms, Management Science, forthcoming.
Dellarocas, C., Wood, C.A. (2006c). The sound of silence in online feedback: estimating trading risks in
the presence of reporting bias. Working Paper (June 2006). Available at SSRN: http://ssrn.com/
abstract=923823
Despotovic, Z., K. Aberer (2004). Possibilities for managing trust in P2P networks. Swiss Federal
Institute of Technology (EPFL) Technical Report IC/2004/84, Lausanne, Switzerland.
Dewatripont, M., I. Jewitt, J. Tirole (1999). The economics of career concerns, part I: comparing
information structures. Review of Economic Studies 66(1), 183–198.
Dichter, E. (1966). How word-of-mouth advertising works. Harvard Business Review 44(November–
December), 147–166.
Dewan, S., V. Hsu (2004). Adverse selection in reputations-based electronic markets: evidence from
online stamp auctions. Journal of Industrial Economics 52(4), 497–516.
Diamond, D. (1989). Reputation acquisition in debt markets. Journal of Political Economy 97(4), 828–
862.
Douceur, J. (2002). The sybil attack. Proceedings of the 1st International Workshop on Peer-to-Peer
Systems (IPTPS ‘02), Cambridge, MA.
Dranove, D., D. Kessler, M. McClellan, M. Satterthwaite (2003). Is more information better? The
effects of ‘‘Report Cards’’ on health care providers. Journal of Political Economy 111(3), 555–588.
Eaton, D.H. (2005). Valuing information: evidence from guitar auctions on eBay. Journal of Applied
Economics and Policy 24(1), 1–19.
Ely, J.C., D. Fudenberg, D.K. Levine (2006). When is reputation bad? Games and Economic Behaviour,
forthcoming.
Ely, J.C., J. Valimaki (2003). Bad Reputation. Quarterly Journal of Economics 118(3), 785–814.
Fan, M., Y. Tan, A.B. Whinston (2005). Evaluation and design of online cooperative feedback mech-
anisms for reputation management. IEEE Transactions on Knowledge and Data Engineering 17(3),
244–254.
Friedman, E., P. Resnick (2001). The social cost of cheap pseudonyms. Journal of Economics and
Management Strategy 10(1), 173–199.
Fudenberg, D., D. Levine (1992). Maintaining a reputation when strategies are imperfectly observed.
Review of Economic Studies 59(3), 561–579.
Greif, A. (1993). Contract enforceability and economic institutions in early trade: the Maghribi Traders’
Coalition. American Economic Review 83(June), 525–548.
Hennig-Thurau, T., K.P. Gwinner, G. Walsh, D.D. Gremler (2004). Electronic word-of-mouth via
consumer-opinion platforms: what motivates consumers to articulate themselves on the internet?.
Journal of Interactive Marketing 18(1), 38–52.
Holmstrom, B. (1999). Managerial incentive problems: a dynamic perspective. Review of Economic

Studies 66(1), 169–182.
Houser, D., J. Wooders (2006). Reputation in auctions: theory and evidence from eBay. Journal of
Economics and Management Strategy 15(2), 353–370.
Jin, G.Z., A. Kato (2004). Blind trust online: experimental evidence from baseball cards. University of
Maryland Working Paper.
Jurca, R., B. Faltings (2004). CONFESS: eliciting honest feedback without independent verification
authorities. Sixth International Workshop on Agent Mediated Electronic Commerce (AMEC VI
2004), New York, USA, July 19, 2004.
Kalyanam, K., S. McIntyre (2001). Returns to reputation in online auction markets. Retail Workbench
Working Paper W-RW01-02, Santa Clara University, Santa Clara, CA.
Kamvar, S. D., M.T. Schlosser, H. Garcia-Molina (2003). The Eigen Trust algorithm for reputation
management in P2P networks, in: Proceedings of the 12th International World Wide Web Conference,
pp. 640–651.
Kandori, M. (2002). Introduction to repeated games with private monitoring. Journal of Economic
Theory 102(1), 1–15.
Kauffman, R. J., C. Wood (2000). Running up the bid: modeling seller opportunism in internet auc-
tions. Proceedings of the Americas Conference on Information Systems (AMCIS 2000), Association
for Information Systems, Long Beach, CA, pp. 929–935.
Keser, C. (2003). Experimental games for the design of reputation management systems. IBM Systems
Journal 42(3), 498–506.
Klein, B., K. Leffler (1981). The role of market forces in assuring contractual performance. Journal of
Klein, D. (editor), 1997, Reputation: Studies in the Voluntary Elicitation of Good Conduct. University
of Michigan PressAnn Arbor, MI, .
Kollock, P. (1999). The production of trust in online markets, in: E.J. Lawler, M. Macy, S. Thyne, H.A.
Walker (eds.), Advances in Group Processes, Vol. 16. JAI Press, Greenwich, CT.
Kreps, D., P. Milgrom, J. Roberts, R. Wilson (1982). Rational cooperation in the finitely repeated
prisoners’ dilemma. Journal of Economic Theory 27(2), 245–252.
Kreps, D., R. Wilson (1982). Reputation and imperfect information. Journal of Economic Theory 27(2),
253–279.
Lee, Z., I. Im, S.J. Lee (2000). The effect of negative buyer feedback on prices in internet auction
markets, in: W.J. Orlikowski, S. Ang, P. Weill, H.C. Krcmar, J.I. DeGross (eds.), Proceedings of the
21st International Conference on Information Systems (ICIS 2000), Association for Information
Systems., Brisbane, Australia, pp. 286–287.
Livingston, J. (2002). How valuable is a good reputation? A sample selection model of internet auctions.
Working Paper, University of Maryland, College Park, MD.
Lucking-Reiley, D., D. Bryan, Prasad, N. and Reeves, D. (2006). Pennies from eBay: the determinants
of price in online auctions. Journal of Industrial Economics, forthcoming.
Mailath, G.J., L. Samuelson (2001). Who wants a good reputation?. Review of Economic Studies 68(2),
415–441.
Mayzlin, D. (2003). Promotional chat on the internet, Working Paper #MK-14, Yale School of Man-
agement, New Haven, CT.
McDonald, C.G., V.C. Slawson Jr. (2002). Reputation in an internet auction market. Economic Inquiry
40(3), 633–650.
Melnik, M.I., J. Alm (2002). Does a seller’s reputation matter? Evidence from eBay auctions. Journal of
Industrial Economics 50(3), 337–349.
Milgrom, P.R., D. North, B.R. Weingast (1990). The role of institutions in the revival of trade: the law
merchant, private judges, and the champagne fairs. Economics and Politics 2(1), 1–23, in: D.B. Klein
(ed.), Reputation: Studies in the Voluntary Elicitation of Good Conduct. University of Michigan Press,
Ann Arbor, pp. 243–266 (Reprinted in 1997).
Milgrom, P., J. Roberts (1982). Predation, reputation and entry deterrence. Journal of Economic Theory
27(2), 280–312.
660 C. Dellarocas
Miller, N., P. Resnick, R. Zeckhauser (2005). Eliciting honest feedback: the peer-prediction method.
Oram, A. (editor), 2001, Peer-to-Peer: Harnessing the Power of Disruptive Technologies. O’Reilly
MediaSebastopol, CA, .
Papaioannou, T., G.D. Stamoulis (2005). An incentives’ mechanism promoting truthful feedback in
peer-to-peer systems, in Proceedings of IEEE/ACM CCGRID 2005 (Workshop on Global P2P
Computing), May 2005.
Pujol, J.M., R. Sanguesa, J. Delgado (2002). Extracting reputation in multi agent systems by means of
social network topology. Proceedings of the First International Joint Conference on Autonomous
Agents and Multiagent Systems, Association for Computing Machinery, Bologna, Italy, pp. 467–474.
Resnick, P., R. Zeckhauser, E. Friedman, K. Kuwabara (2000). Reputation systems. Communications of
the ACM 43(12), 45–48.
Resnick, P., R. Zeckhauser (2002). Trust among strangers in internet transactions: empirical analysis of
eBay’s reputation system, in: M.R. Baye (ed.), The Economics of the Internet and E-Commerce.
Advances in Applied Microeconomics, Vol. 11. JAI Press, Greenwich, CT.
Resnick, P., R. Zeckhauser, J. Swanson, K. Lockwood (2006). The value of reputation on eBay: a
controlled experiment. Experimental Economics 9(2), 79–101.
Sabater, J., C. Sierra (2002). Reputation and social network analysis in multi-agent systems. Proceeding
of the First International Joint Conference on Autonomous Agents and Multiagent Systems, Asso-
ciation for Computing Machinery, Bologna, Italy, pp. 475–482.
Shapiro, C. (1982). Consumer information, product quality, and seller reputation. Bell Journal of Eco-
nomics 13(1), 20–35.
Shapiro, C. (1983). Premiums for high quality products as returns to reputations. The Quarterly Journal
of Economics 98(4), 659–680.
Sundaram, D.S., K. Mitra, C. Webster (1998). Word-of-mouth communications: a motivational anal-
ysis. Advances in Consumer Research 25, 527–531.
Tadelis, S. (1999). What’s in a name? Reputation as a tradeable asset. The American Economic Review
89(3), 548–563.
Tadelis, S. (2002). The market for reputations as an incentive mechanism. Journal of Political Economy
92(2), 854–882.
Watson, J. (1993). A reputation refinement without equilibrium. Econometrica 61, 199–205.
Watson, J. (1996). Reputation in repeated games with no discounting. Games and Economic Behavior 15,
82–109.
Wilson, R. (1985). Reputations in games and markets, in: A. Roth (ed.), Game-Theoretic Models of
Bargaining, Cambridge University Press, Cambridge, UK, pp. 27–62.
Xiong, L., L. Liu (2004). Peer trust: supporting reputation-based trust in peer-to-peer communities.
IEEE Transactions on Knowledge and Data Engineering 16(7), 843–857.
Yu, B., M. Singh (2002). An evidential model of distributed reputation management. Proceedings of the
First International Joint Conference on Autonomous Agents and Multiagent Systems, Association for
Computing Machinery, Bologna, Italy, pp. 294–301.
Zacharia, G., T. Evgeniou, A. Moukas, P. Boufounos, P. Maes (2001). Economics of dynamic pricing in
a reputation brokered agent mediated marketplace. Electronic Commerce Research 1(1), 85–100.
Subject Index
Admission control, 64–66, 73–76, 83, 101, Broadcast mechanism, 207–210, 212–213, 217,
108–109, 115, 127 221, 223, 230, 233, 235, 238
Adverse selection, 480, 577, 633–634, 640, 644, Bundling, 129, 453, 466, 499–523
646, 648, 653, 656 Buy-now, 571, 573, 599–604
Advertising fees, 355–356
Affiliated-value, 584–585 Cameras, 150, 327, 452, 603
Aggregation, 5, 60, 72, 119–120, 123, 246–251, Capacity, 54, 57, 61, 63–65, 67–69, 72, 74, 76–78,
253, 256, 649 81, 84–97, 103–106, 108–110, 114, 117, 119,
Alaska Airlines, 560 124–128, 145, 154, 193, 197, 332, 349, 431,
American Airlines, 447, 527–529, 541, 551, 553, 558 528, 530–532, 537–538, 540–543, 550–551,
Anticommons, 296 562, 565
Ascending-price, 574–578, 593, 595, 609–610, 615 Change agents, 8, 27
Assignment, 101, 127, 138–139, 148, 152, 155–173, Chicago school, 473–476, 483, 485
176, 179, 186–192, 196, 279, 485, 487 Clearinghouse, 323, 331–332, 348–359, 363,
Assimilation, 19–20 367–369, 439
ATM, 36, 60–61, 78, 127 Clock auction, 609, 613
Auction, 69, 71–72, 96, 118–123, 139–140, Coase theorem, 486
143–144, 146, 151, 171, 182–184, 186, 196, Code-share flight, 560
222, 323, 343–344, 358, 428, 477, 505, Co-invention, 4, 12, 23, 25–28, 42
571–621, 631, 636–639, 652 Commitment type, 390, 636–638, 641
Auction design, 123, 571, 613–614, 616, 619 Common value, 411, 584, 597–598, 604
Auction fever, 590–591, 593, 596 Communication, 1–3, 5, 7–9, 11, 13–15, 17, 19–25,
Autonomy, 473, 476, 483 27–29, 31, 33, 35–37, 39, 41–42, 54–55, 57, 59,
63–64, 67, 69, 94–95, 100, 103, 112–113, 118,
Bandwagon effects, 30, 34–35 122–123, 137–139, 141–143, 146–149,
Bandwidth sharing, 107–111, 114, 116 151–154, 156, 165–166, 169–170, 173–175,
Behavior-based price discrimination, 377, 181, 188, 196, 220, 223, 226, 228–229, 237,
379–385, 387, 389–391, 393–395, 397–399, 242, 272, 280, 432, 466, 483–484, 488, 580,
401, 403, 405, 407, 409, 411, 413, 415, 417, 616, 619, 629–630
419, 421, 423, 425, 427, 429, 431, 488, 522–523 Communication burden, 226, 228
Berkeley software distribution licence, 286 Communication complexity, 220, 280
Bertrand model, 362 Communication networks, 55, 63–64
Best effort service, 61–63, 66–67, 70, 72, 75, 77, 80, Compact discs, 325, 327, 329–330
83–85, 87–88, 99, 101–102, 104–108, 112, 117, Compatibility, 18, 102, 105, 150, 396, 418,
129 438–439, 445, 447, 449, 453–454, 466,
Blocking probability, 74, 77, 87 505, 521
Books, 236–237, 325, 327, 329–330, 370, 450, 534, Complementarities, 12–14, 120, 612–614
590, 629, 634, 654 Complementary investments, 26, 447
Bounded rationality, 96, 330, 358–360, 578 Computation, 92, 122, 137–138, 141–143,
Brand, 412, 439, 445–446, 450–451, 453–455, 560 147–151, 153, 156, 166, 168–169, 181, 188,
661
662 Subject Index
196, 258–259, 261, 264, 280, 385, 533, 541, Diamond paradox, 338, 340, 342
546, 613 Differentiated products, 137–140, 143, 153,
Congestion, 54, 58–68, 73, 78, 81, 83–97, 99, 160–164, 167, 169, 172–173, 179–180,
105–114, 117–118, 121, 124–125 183–185, 188, 191–193, 412, 463, 466, 502
Consumer attention, 483 Differentiated services, 54, 62–63, 65–66, 70, 72,
Consumer behavior, 432, 449, 490, 648, 654 74, 78–79, 84, 88, 103, 115, 117, 127
Consumer electronics, 327–330, 368, 370 Digital divide, 40–41, 354, 360
Convergence hypothesis, 369–370 Digital product, 447
Cookie, 472 Direct network externalities, 8, 30–33
Coordination costs, 10, 13–16, 21, 447, 451 Direct revelation, 202, 209
Core, 42–43, 56, 71, 73, 125, 127, 147, 149, Discounting, 531, 534, 536, 647
166–170, 184, 186, 489 Discrete choice model, 490
Corporate information, 492 Do not call, 484, 486, 489, 491
Cost heterogeneities, 337, 340–341, 357–358 Do not contact, 486, 489
Cost saving, 24, 174, 500 Double auction, 139, 181–184, 186
Crandall, Robert, 527 Drug discovery
Credit markets, 379, 406, 422, 425–429, 432 and open source, 315–316
Cross bid, 580, 607, 613 Duopoly competition, 500, 515–522
Cross-country diffusion, 39 Durable goods, 146, 393–395, 399
Cumulative innovation, 288 Dutch auction, 118, 574, 611–612, 614
Customer profiling, 454 DVD, 422, 588, 593, 603
Customer recognition, 377, 379, 381–385, 387, Dynamic pricing, 53, 56, 70, 72, 75, 80–82, 88,
389, 391, 393–397, 399, 401, 403, 405, 407, 100, 105–106, 109, 115–116, 128, 359, 472,
409, 411, 413, 415, 417, 419, 421–425, 427, 527–529, 531, 533, 535, 537, 539, 541–545,
429, 431, 454 547–549, 551, 553, 555, 557, 559, 561, 563,
Customer relationship management, 431, 439 565, 656
Customer retention, 437–443, 452, 455, 457–458,
461, 464–466 e-commerce, 16–17, 28, 40–41, 53, 56, 377, 472,
Customization, 430–431, 476, 490 490, 522
Customized pricing, 379, 422, 429, 431, 433 EDI, 2, 13, 32–35, 151
Education
Damaged good, 528 and open source, 295
Data Effective bandwidth, 74, 76–82
geographical, 286 Efficiency, 9, 37, 56, 87, 100, 107, 119–120,
Data communications, 53–55, 57, 59, 61, 63, 65, 122–123, 137–139, 141–142, 148, 152–154,
67–69, 71, 73, 75, 77, 79, 81–85, 87, 89, 91, 93, 156, 158–161, 164, 170, 174, 189, 192–193,
95–97, 99, 101, 103, 105, 107, 109, 111, 113, 195–197, 228, 241, 398, 400, 404, 407,
115, 117–121, 123, 125, 127–128 421–423, 473–476, 481, 485, 500, 528, 530,
Data loss, 54, 58, 61, 63–64, 67, 85, 87–88, 106 541, 543, 546–547, 550, 557, 561, 582,
Data transport, 53, 55–57, 59–60, 67–68, 600–601, 606–608, 611–614, 616, 620, 630,
128–129 635, 645–648, 650, 655–656
Decentralization, 10, 12, 208, 266, 268–270, Elasticity, 74, 77, 94, 341, 398, 423, 528, 532, 543,
272–274 546, 549–550
Decentralization penalty, 266, 268–270, 272–273 Electronic monitoring, 12
Delay, 58–65, 67, 69–70, 72–73, 75–76, 79, Embedded LINUX, 288, 292
84–88, 90–92, 94–101, 103–106, 109, 115–117, English auction, 574, 585, 599
120–122, 240–243, 245–246, 249, 258–264, Epidemic diffusion models, 5
279–280, 552 European Union Directive 95/46/EC on data
Delta Airlines, 552, 554 protection, 474
Demand reduction, 610–611, 621 Exchange economy, 206, 214, 235
Demand, constant elasticity, 528, 543, 546 Exposure problem, 612, 614, 621
Demand, exponential, 531, 533 Externality, cross-market, 483, 486
Descending price, 574–576, 578, 608, 614, 619–620 Externality, direct, 474
Subject Index 663
Externality, indirect or consequential, 474 465, 478, 485, 487, 490–491, 584, 596, 598,
Externality, peer-to-peer, 485, 487 601, 610, 615, 618, 620–621, 634, 637–638,
Externality, within-market, 476 640, 642, 644–647, 649, 653–654
Incomplete information, 53, 56, 69–72, 78, 95,
Fair Credit Reporting Act of 1970, 472, 488 109, 111, 117, 128, 193, 394
Fair Information Practices, 472 Incremental bid, 573, 594–598, 617
False reporting, 644 Independent private value, 119, 585
Feedback mechanism, 61, 66, 83, 107, 630, 635, Indirect network externalities, 8, 30–33
644, 651, 653 Information acquisition, 323, 422
Feedback, 22, 61–62, 65–66, 83–84, 87–88, Information good, 511, 514, 522–523
106–108, 111–112, 594, 597, 616–617, Information technology, 2, 9, 12, 20, 22, 27, 147,
630–633, 635, 639–640, 643–654, 656 201–202, 408, 437, 439, 441, 443, 445, 447,
Field experiments, 490, 572, 577, 581, 589, 591, 449, 451, 453, 455, 457, 459, 461, 463, 465,
611, 618, 651 632, 645, 649
Finite mechanism, 230–231, 233 Information transmission, 8, 12, 367, 490
Firm size and IT adoption, 9, 17 Informational costs, 201–205, 207, 264–266, 275,
Flow control, 61, 65–67, 83, 99–100, 107–109, 278
111–114, 116–117 Information-intensive markets, 438, 440, 445–447,
Fooling set, 217, 220 451, 464–466
Forking Informed consumers, 352–354
and open source, 294, 298 Innovation
Free recall, 338 cumulative, 288
Free-rider problem, 356–357 user, 290
Free riding Internet, 2–4, 13, 16–17, 20, 22–28, 33–35,
and open source, 310 37–43, 53–55, 59–61, 63–64, 66, 68–69,
Frictions, 10, 16, 338, 363, 451 83–84, 90, 101, 104, 107, 111, 113–114, 118,
124, 126–127, 129, 142–143, 146, 150–151,
154, 194, 197, 201, 324, 330–331, 354–355,
Gatekeeper, 332, 355–357
General purpose license, 286 359–360, 365–366, 370, 382, 431, 446,
General purpose technology (GPT), 9 448–449, 451–452, 464–466, 472, 488–490,
Geographical data 505, 551, 578, 580, 587, 595, 605, 614, 621,
629–630, 632
and open source, 286, 315
Gift culture, 297 IT and organizational change, 9, 12, 14, 26, 39
Goal function, 221, 242–244, 246–249, 251,
253–256, 258, 261, 264 Late bid, 595–599, 602, 616–618
Law of demand, 334
Graph, 138, 151–156, 158–162, 164–165, 170–171,
173–175, 195–196, 223, 229, 243–244, Law of one price, 170, 324, 330, 370, 527
252–254, 258–259, 261, 532–533 Learning, 5, 18–19, 29, 31, 34, 36, 70, 202, 204,
241, 273, 279, 408, 410, 419, 425, 445–447,
Graph theory, 138, 151–153, 533
449, 453, 456, 464–465, 575, 580, 616,
634, 646
Homogeneous product, 323–324, 332, 355, 362,
Learning costs, 19, 279, 446–447, 453, 464
364, 370 License
Honest feedback, 633, 643–644 general purpose, 286
viral, 286
Impatient bidder, 601, 604, 620 open source, 302
Implementation, 37, 55–56, 63, 71, 79, 96, Link, 3, 9, 17, 61–62, 64–67, 69, 72–79, 84,
108, 112–113, 118–119, 122, 124–126, 129, 88–89, 93–94, 101–103, 106, 111, 114, 118,
264–266, 455, 630, 641 124, 142–143, 152–155, 157, 165, 194–196,
Implicit feedback, 649–650 275–277, 399
Incentives, 3, 12, 14–17, 24, 32–33, 54, 68, 79, LINUX
82, 89, 91, 104, 184, 201, 204, 207–208, 229, embedded, 288, 292
264–265, 268, 352, 354, 356, 384, 394, 419, Long-term contracts, 378–380, 390–392, 395–397,
423, 429, 431, 448–449, 451–452, 455–456, 399, 406–408, 432, 445
664 Subject Index
Loyalty, 331, 439, 443, 445, 450–451, 454–456, Networks, 5–6, 8, 20, 32, 55, 57, 62–65, 67, 69, 78,
464–466, 560 82–83, 93, 96, 104, 106, 109, 116, 118, 124,
Loyalty program, 445, 455–456, 465–466 126, 129, 137–139, 141–147, 149–157, 159,
161–163, 165–167, 169, 171, 173, 175, 177,
Marginal cost, 64, 77, 93, 332, 337–338, 340, 179, 181, 183, 185, 187–189, 191, 193–197,
342–347, 349–350, 353–354, 357–358, 203, 238, 242–243, 275–277, 279, 438,
360–362, 366, 380, 399, 411–412, 416, 430, 448, 454, 532, 629–630, 632, 643,
438, 477, 492, 500, 502–503, 505, 507–508, 649–650, 655
510, 512–514, 516–517, 522, 528–529, 533, Node, 80, 92, 97–100, 103, 111, 116–117,
536, 539, 541–543, 549, 605–607 119, 155, 157, 195, 229, 243–245, 254–255,
Market making, 137–138, 141–143, 149, 258–261, 275
181–183 Norms
Market microstructure, 140, 152 and open source, 298, 304
Marriage theorem, 155–156, 159–161, 164, 171, Northwest Airlines, 553
173
Matchmaking, 137–138, 141–144, 147–149, 181, OECD guidelines on privacy protection and
183 transborder data flow, 471
Mechanism, 73, 76, 78–79, 81, 94, 101, 104, Online identity, 583, 633, 647
108, 112, 115, 119, 121–123, 125, 140, Online marketplace, 146, 629–630, 655
142, 151, 156, 166, 170, 182, 184–188, 201, Open auction, 578–580, 612, 614–616, 619
203, 205, 207–238, 240–246, 249, 251–254, Open source, 285
258–267, 279–280, 344, 378, 487, 493, 511, Options, 26, 28–29, 88, 102, 129, 444, 505, 528,
517, 522–523, 574, 576, 582, 585, 603, 548, 550, 557–558, 561, 571, 573, 583,
606, 609, 612–614, 620–621, 630–632, 599–601, 605, 608
634–635, 639–640, 642, 644–651, Organization for Economic Co-operation and
653–656 Development (OECD), 471
Message space, 120, 208–209, 211, 213, 215–221,
226, 228, 230–232, 234–235, 265 Parallel auction, 573, 605, 607–608
Minimum bid, 582–593, 604 Patent thickets
Modularity, 294, 310 and open source, 296
Moral hazard, 577, 633–634, 640, 644, 646–648, Peak-load problem, 542
653, 655–656 People’s express, 527
Multidivisional firm, 202, 206 Perfect Bayesian equilibrium, 389, 606
Multi-item auction, 608, 610, 612 Personal information, 471–492
Multi-unit auction, 574, 578, 598, 608–610, 612, Personalization, 431, 454, 461, 466
620–621 Pharmaceuticals
and open source, 286
Net present value (NPV), 29 Platform, 53, 150, 580, 594, 617
Network, 2–4, 7–8, 20, 29–37, 42–43, 53–70, Poaching, 378, 400, 402–403, 405–408, 420, 432
72–75, 77–81, 83–86, 89–94, 96–102, 104–119, Price comparison site, 324, 331, 348, 359,
121, 124–128, 137–139, 141–143, 150, 368, 370
152–157, 161, 164–165, 171–177, 180–181, Price cycles, 387, 390, 394–395
191–197, 203, 222–229, 233, 238, 242, Price discovery, 572, 615–618, 620
275–279, 413, 419, 432, 438–440, 447–450, Price discrimination, 100, 105, 377–385, 387,
454–455, 463–464, 531–534, 650 389–391, 393–395, 397–399, 401, 403–405,
Network effects, 2–4, 30–37, 42–43, 305, 438, 440, 407–409, 411–413, 415, 417, 419, 421–423,
447–448, 454–455, 464 425, 427, 429–431, 454, 456, 472, 477–478,
Network externalities, 4, 7–8, 20, 30–37, 43, 59, 480, 488, 491, 499–500, 502, 505–506, 510,
94, 295, 413, 419, 432, 439, 448 513–514, 522–523, 527–529, 533–534, 536,
Network flow, 650 538–541, 543–544, 551, 561, 644
Network mechanism, 222–229, 233 Price dispersion, 145–146, 323–333, 335–339,
Network theory, 137–138, 152, 154, 156–157, 341–343, 345–349, 351–370, 453, 529, 532,
197 534, 581, 621
Subject Index 665
Price mechanism, 69, 72, 78, 92, 95–96, 118, 121, Reverse first-price auction, 344
123, 125, 214–215, 217, 221, 224–226, 228, Routing, 64–66, 74, 80, 84, 93, 101, 103, 106, 112,
230, 235–236, 609 114
Price persistence, 369
Priorities, 79, 86, 101, 104–105, 121, 127 Scheduling, 65–67, 74, 79, 84, 86, 91, 94, 97, 101,
Privacy, 208–209, 213, 218, 223, 234–235,
103–104, 108, 115–116
237–238, 266, 280, 379, 422–426, 432, Sealed bid, 579, 615–616
471–485, 487–493, 523 Search, 14, 33, 137–139, 141–145, 148, 151,
Privacy Act of 1974, 471, 473 188–189, 191–194, 196, 236–237, 239, 254,
Private information, 54, 68–69, 102, 104, 358, 380,
266–268, 270–274, 323–325, 327, 329–349,
390, 392–394, 423, 426, 429, 475, 482–483, 351–355, 357–359, 361, 363, 365–367, 369,
575–576, 596 437–438, 445–446, 450–453, 456, 459–461,
Probit (rank) diffusion models, 6
465–466, 500, 520, 552, 558, 592, 610, 617,
Product design, 413, 422, 452–453, 456
631, 650
Product line, 379, 413, 444–445, 449, 453–454, Search costs, 14, 33, 148, 188, 272–274, 323–324,
464–465 330, 333, 335, 340, 342, 345, 354, 359, 361,
Productivity, 8–9, 13, 25, 37, 39, 58, 148, 163, 165,
365–367, 438, 445–446, 451–452, 456,
179, 191, 271, 474, 480–481, 504 459–461, 465
Proper scoring rule, 644–645 Search technology, 266–268, 270, 272–273, 451
Property rights, 17, 32, 471, 474–475, 478, Seclusion, 473–474, 476, 483
485–487
Second-best, 165, 476, 492, 613
Proprietary software Secrecy, 473, 476, 482–483
and open source, 312 Self-regulation, 485, 491
Proportional fairness, 107, 109–110, 113–116 Service design, 55–56, 59, 62, 68–72, 79, 82, 88,
Protocol, 61, 83–84, 107, 124, 127, 142, 150,
102, 105–106, 118, 128
259–261 Shill bid, 573, 582–587, 590–596
Proxy bid, 578–579, 582, 586–587, 595–596, 598, Shirking, 268–270, 276–277
611, 613, 615, 618, 621
Shopbot, 324, 450, 605
Public good, 229, 264–265, 481, 492, 643, 653 Shopper, 349, 654
Purchase history, 379, 404, 477–478 Signaling
and open source, 293
Quality of service, 53, 57, 107–108, 127, 203 Site licensing, 504
Queueing, 63, 65–67, 85–86, 98, 103, 108, Smoothness requirements, 218–219, 222, 237
112–113, 115, 117, 121, 129 Sniping, 586, 594–598, 618–619
Software, 18–19, 28, 31, 58, 65, 147, 150, 325,
Random network, 139, 149, 188, 193–196 329–330, 419, 439, 442, 446–447, 449, 451,
Ratchet effect, 381, 417–418, 432 453–454, 484, 493, 504, 510–511, 515–516,
Real options, 26, 28–29 519–520, 522, 572, 595, 603, 632, 652
Referral network, 650 proprietary, and open source, 312
Regulation, 65–67, 84, 101, 127, 194, 456, 471, Solicitations, 474, 476, 483–484, 491
473–475, 477–480, 484–485, 487–489, SourceForge, 290
491–492 Spam, 472–473, 489, 491, 493
Repeated game, 405, 635, 638, 640, 642 Speak-once-only mechanism, 203, 240–243,
Reputation, 380, 390–391, 394, 493, 573, 607, 245–246, 253–254, 256, 258–260, 267, 279
629–651, 653–657 Spillovers
Reputation mechanism, 629–635, 637, 639–651, and open source, 292
653–657 Squandering, 268–272, 276–277
Reservation price, 122, 337–343, 345, 487, 577, Stable assignment, 139, 168–174, 179
590, 612, 614 Stable network, 276–277
Reserve price, 121, 573, 582–585, 587–590, 593, Standards
599, 605–608 and open source, 295
Revenue Equivalence Theorem, 344, 576, 578 Subscription fees, 124, 175–177, 179–181,
Revenue management, 527 354–355
666 Subject Index
Supply-chain, 28, 35 U.S. Federal Trade Commission (FTC), 472

Switching costs, 18, 145, 378–380, 397, 399, Uniform price, 166, 429, 431, 609–610, 612, 614,
408–413, 432–433, 437–466 621
Sybil attack, 648 Uninformed consumers, 353–354
Uniqueness property, 217–222
Tariffs, 68–73, 75, 79–80, 82, 95, 106, 124–125, United Airlines, 551, 554
129 Unix, 286
TCP, 61, 83–84, 107, 114, 150, 448, 451 Upgrades and buybacks, 414, 417
Team, 13, 202, 238–239, 270, 272 Urban agglomeration and IT, 21–22
Technology acceptance Model (TAM), 20 Urban/rural differences and IT, 21–25
Telemarketing, 473, 484, 489, 491 User innovation, 290
Theory of teams, 238, 270, 272, 278
Transaction costs economics, 17 Value, private, 119, 293, 479, 574–577, 579, 585,
Transaction costs, 14, 16–17, 137–138, 141, 595–598, 601, 604, 607, 613, 615–616, 618
146–148, 151–154, 157–158, 168, 170, 179, Value, social, 293, 311, 479–480
182, 188, 193, 355, 438, 445–446, 460, 572, Verification scenario, 209, 212–213
601–602, 620 Vertical integration, 10, 13, 15–17
Transactions, 14–16, 56, 58, 138, 140–141,
143, 146–148, 151–154, 156–157, 159, 161,
Wikipedia, 315
172, 174, 177–178, 181, 188, 196–197,
334–335, 342, 351–352, 354, 360, 362, 366, Willingness to accept, 487, 582, 590
437, 446, 454–455, 472, 591, 599, 630–631, Willingness to pay, 58, 112, 138, 153, 174, 183,
633, 644, 646–648, 651, 653–654 349, 379, 426, 473, 477, 487, 531, 536, 543,
573–574, 586, 590, 597
Tree, 229, 242–243, 245–246, 254–255, 257–258,
261, 267, 275–276 Winner’s curse, 576, 580–581, 615, 620
Two-sided market, 137–149, 151, 153, 155, 157, Word-of-mouth, 629, 632, 643, 655
159, 161, 163, 165–169, 171, 173, 175, 177,
179, 181, 183–185, 187–189, 191, 193, 195 Yankee auction, 611
Tying, 501, 505, 516, 518 Yield management, 527–530, 538, 541

Economics and Information Systems, Volume 1

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Economics and Information Systems, Volume 1

Hochgeladen von

Copyright:

Verfügbare Formate

HANDBOOKS IN INFORMATION SYSTEMS

Advisory Editors Editor

Gupta, Alok Volume 1

First edition 2006

Copyright r 2006 Elsevier B.V. All rights reserved

No part of this publication may be reproduced, stored in a retrieval system

Library of Congress Cataloging-in-Publication Data

ISBN-13: 978-0-444-51771-5 (this volume)

For information on all Elsevier publications

Printed and bound in The Netherlands

2.3. Service production 63

4.1. Assignments with homogenous products 157

3.2. Models in which the organization’s mechanism is partly designed by

Long-term contracts 391

8. Design of online auctions 614

Subject Index 661

Technological advances in processing and communicating information

Information has risen to prominence in economics, for example, the 2001

factors, they examine ﬁrms’ geographic location—both rural versus urban

In Open Source Software: The New Intellectual Property Paradigm, Step-

Diffusion of Information and Communication

We survey the literature on the adoption and diffusion of information and

The invention of new technology is only the ﬁrst step in economic

and the beneﬁt of ICT investment to the organization. By investigating how

literature on the diffusion of other technologies (Hall and Khan, 2003). By

How adopter characteristics inﬂuence the net beneﬁts of

Source of variation in Adoption costs Adoption beneﬁts

2 ICT diffusion and its impact

2.1 Diffusion modeling

Before examining the economics of ICT diffusion, it is necessary to give a

review on diffusion modeling in economics is Stoneman (2002). In this

Fig. 1. The typical pattern of diffusion.

state. That is, we rarely observe organizations ‘‘unadopting’’ a new tech-

has attempted to understand how direct and indirect network externalities

2.2 The impact of ICT diffusion

There is considerable evidence that ICT investment has a positive impact

3 ICT adoption and organizational characteristics

The decision to adopt a technology can be inﬂuenced by factors relating

3.1 Adoption, internal firm organization, and organizational change

Bresnahan and Trajtenberg (1995, p. 84) deﬁne general purpose technol-

Open question Papers Results

Harris, 1998; Forman et al., 2005).8 A GPT is an enabling technology

literature, there remains relatively little empirical adoption research in this

Empirical work on ICT adoption and organizational change remains a

3.2 Adoption and firm boundaries

costs than on internal coordination costs. This leads to the conclusion

Building on the work of Hubbard (2000), Baker and Hubbard (2003,

Gertner and Stillman (2001) examine how vertical integration inﬂuences

3.3 Adoption and size

Prior research in the diffusion of innovations literature has consistently

Measures of size are typically included as a control in recent research on

Some research has also emphasized how prior IT investments inﬂuence

3.5 Adoption, assimilation, intra-firm diffusion, and usage

demographic characteristics of early adopters are very different from those

4 Geographic differences in adoption

Open question Papers Results

4.1 Adoption of ICT across urban and rural areas

Adoption papers in this literature generally proceed by examining differ-

4.2 Evidence of how ICT use influences location patterns

4.3 Future research

The relationship between ICT adoption and ﬁrm location remains a

5 Trade-offs between organization and environment

Many adoption decisions depend on the interaction of factors that are

Open Question Papers Results