Asymmetric Social Interactions in Physician Prescription Behavior: The Role of Opinion Leaders

Harikesh Nair Assistant Professor of Marketing, Graduate School of Business Stanford University, 518 Memorial Way, Stanford, CA 94305 Phone: 650-723-9675; Fax: 650-725-7979 Puneet Manchanda Associate Professor of Marketing, Ross School of Business University of Michigan, 701 Tappan St., Ann Arbor, MI 48109 Phone: 734-5591; Fax: 734-936-8716 Tulikaa Bhatia Assistant Professor of Marketing, Rutgers Business School 94 Rockefeller Road, Piscataway, NJ 08854, Phone: 732-445-5274 Past versions: June 2006, May 2008 This version: December 3, 2008 Abstract We quantify the impact of social interactions and peer effects in the context of prescription choices by physicians. Using detailed individual-level prescription data, along with selfreported social network information, we document that physician prescription behavior is significantly influenced by the behavior of research-active specialists, or opinion leaders in the physicians reference group. We leverage a natural experiment in the category, whereby new guidelines released about the therapeutic nature of the focal drug generated conditions where physicians were more likely to be influenced by the behavior of specialist physicians in their network. We find important, statistically significant peer effects that are robust across model specifications. We use the estimates to measure the incremental value to firms of directing targeted sales-force activity to these opinion leaders, and present estimates of the social multiplier of detailing in this category. Key-words Key-words: Social Interactions, Peer effects, Social Multiplier, Contagion, Physician

Prescription Behavior, Pharmaceutical Industry.

* The authors are listed in reverse alphabetical order. The authors would like to thank Ron Burt, Pradeep Chintagunta, Tim Conley, Wes Hartmann, Peter Reiss, Christophe Van den Bulte, Raphael Thomadsen and seminar participants at Berkeley (Haas), Chicago GSB (O&M), Christian Albrechts University, Erasmus, Michigan (Information Systems; College of Pharmacy) and participants at Frank M. Bass (Dallas, 2007), INFORMS Practice of Marketing Science (Wharton, 2007), IOFest (Stanford, 2006), Marketing Science (Pittsburgh, 2006) and SICS (Berkeley, 2006) conferences, and at the 7th Choice Symposium (Wharton, 2007) for feedback; and an anonymous pharmaceutical company for providing the data. Manchanda would like to thank the Kilts Center for Marketing and the True North Communications faculty research fund at the University of Chicago for research assistance. The authors can be contacted via e-mail at harikesh.nair@stanford.edu (Nair), pmanchan@bus.umich.edu (Manchanda) and tulikaa@rci.rutgers.edu (Bhatia). The usual disclaimer applies.

1.

Introduction

Marketers, sociologists and economists have traditionally been interested in the role of interpersonal communication (i.e., communication outside the firms control) on consumer choice and consumption behavior. These interactions have been variously labeled as peereffects, contagion and word-of-mouth effects. In this paper, we test and provide empirical support for asymmetric peer-effects. These effects arise when some consumers exert a stronger influence on the attitudes and behavior of other consumers than vice versa. Such consumers have typically been labeled opinion leaders in the literature (Rogers 2003, Chapter 8). There is little research in marketing that has tested for the existence of these asymmetric peer-effects. The context of our analysis is prescription drug choice by physicians. An asymmetric social interaction or peer effect arises in this setting because non-specialist physicians may rely on prominent physicians, the opinion leaders, to help reduce the uncertainty around their prescription choices. The role of opinion leaders becomes most salient when changes occur in the therapeutic environment, as these typically lead to increased uncertainty about drug efficacy among the non-specialist physicians. The pharmaceutical industry believes in the existence of such opinion leaders, and has invested in targeting marketing activities at opinion leaders (CIE 2004). However, to date, there is little empirical evidence that opinion leaders matter i.e., significantly influence the opinions and behavior of other physicians. Coleman, Katz and Menzel 1966 (the classic study in this field) found no asymmetries in peer effects between nominators and their opinion leaders adoption pattern for a new drug. Recent work using the same data as that study found no peer effects at all (Van den Bulte and Lillien 2001). Finally, using simulations based on computational models of network tipping, Watts and Dodds (2007) also find little or no role for opinion leaders. Our main empirical question therefore is to test for, and to measure the extent of asymmetric peer effects in this category. We then use our analysis to explore the implications of these peer effects for targeted allocation of marketing effort in the form of personal sales-calls or detailing to these opinion leaders. More generally, the issues we address in econometrically identifying and measuring peer effects are relevant across a broad range of social networking situations in which firms are interested in understanding the return on investment of marketing activity to opinion leaders (e.g. Godes and Mayzlin 2004). Asymmetric social interactions have important implications for the allocation of marketing effort by firms. If present, they increase the return-on-investment to marketing

activity targeted at agents having stronger influence. In the pharmaceutical context, if actions of opinion leaders have a true causal effect on the prescription behavior of other physicians, then marketing effort directed at the opinion leader will generate a multiplier effect. The multiplier arises because an incremental sales-call to an opinion leader increases the opinion leaders prescriptions, and on the margin, induces the physicians he influences to prescribe more. The extent to which net prescriptions are higher due to these cross-physician spillovers is the social multiplier (c.f. Becker and Murphy 2000). Given that the

pharmaceutical industry in the US currently sets physician-level detailing based on past prescription volume (c.f. Manchanda, Rossi and Chintagunta 2004), the presence of significant social multipliers would imply that the return-on-investment of detailing to opinion leaders may be much higher than is suggested by just the opinion leaders prescription volume. We use our estimates to measure the social multiplier of detailing in our data. To test for these effects, we leverage a novel dataset that is based on a combination of primary (survey) and secondary (behavioral) data. Broadly speaking, there are five major challenges that arise in measuring these effects. First, some effort needs to be made to identify the opinion leaders that constitute the reference group for a given physician. Second, once these opinion leaders have been identified, some change in the environment needs to take place in a manner that this change affects the attitudes and/or behavior of the opinion leader. Third, these changes then needs to be transmitted to the agents whose opinions and/or behavior is affected by the opinion leaders behavior (or opinions). Fourth, there should be a resultant change in the behavior of these consumers. Finally, we need to be able to distinguish between correlation and causation in the observed behavior of physicians and their opinion leaders. As we discuss below, correlation in behavior can arise from three possible sources endogenous group formation, correlated unobservables and/or simultaneity and we need to be able to control for these explanations. As the past literature on social interactions has pointed out (c.f. Manski 1993, Moffitt 2001), solving this identification problem is a formidable challenge. Our dataset, which comes from the pharmaceutical industry, enable us to formulate empirical strategies that address most, if not all, of these issues. Our data contain survey information on the social networks of physicians, as well as individual-level panel data on the prescription behavior of these physicians and the doctors they nominate. We believe our identification strategy is novel to the literature, and is relevant across a broad range of situations involving the analysis of data arising from social interactions.

Our analysis of the data reveals significant evidence for peer effects. These effects persist after allowing for endogenous group formation, targeted marketing activity, correlated unobservables and simultaneity, and are also robust to functional form. We find that opinion leaders behavior significantly affects physician behavior after an exogenous change in the market that resulted in a change in the therapeutic environment. Interestingly, in the changed environment, we find that peer effects have more marginal impact on prescriptions than targeted detailing. Our empirical results also find that peer effects in this category are asymmetric in the sense that opinion leaders prescriptions are not statistically significantly affected by the prescription pattern of the physicians they influence. We then use our results to explore the implications of targeting detailing at the opinion leaders. We quantify the direct effect of detailing on opinion leader prescriptions, as well as the indirect effect on prescriptions by the corresponding nominating physician generated via peer influence. We find that for the average opinion leader, who influences 1.56 physicians, social interactions alone provide an additional 5% increase in prescription revenue. This implies a social multiplier of detailing in this category of about 1.05. For the top opinion leader, who influences 17 physicians, we find a social multiplier of 1.35. The large differences underscore the importance of both the correct identification of opinion leaders, as well the identification of top influencers among these opinion leaders in order to make optimal resource allocation decisions. In summary, our key contributions are as follows. First, we document the existence of asymmetric peer effects amongst a specific social network. Specifically, we provide evidence for these effects in the domain of physician prescription decisions. Further, we document the finding that peer influence can significantly impact on behavior even in stable, mature categories. These are novel findings that add to the literature on peer effects in the presence of marketing, especially in the pharmaceutical industry. Second, we discuss and clarify how the identification issues that arise in measuring and testing for causal peer effects may be overcome for data-rich settings such as ours. Third, we measure empirically the extent to which peer effects matter in driving prescriptions of both physicians and opinion leaders, and show that these are robust to functional form and alternative specifications of peer effects. Further, we use our estimates to derive implications for marketing resource allocation for firms in the industry, and present estimates of the social multiplier effects of detailing. Finally, the increasing salience of social networks in the economy has generated renewed interest among theorists in many fields to incorporate aspects of social networks into their

frameworks. Robust evidence of social effects increases the practical, real-world relevance of these models. The rest of the paper is organized as follows. The next section discusses the industry background and reviews the relevant literature. Sections 3 and 4 present the model and describe the data. Section 5 presents results from estimation. The last section concludes.

2.
2.1

Industry Background and Related Literature

Industry Background

As mentioned above, the pharmaceutical industry strongly believes that opinion leaders (also referred to as thought leaders) play an important role in the adoption and usage of new products by practicing physicians. These opinion leaders are typically believed to be physicians who have an academic title with the department of a medical school and have contributed peer-reviewed publications (Tan 2003). Both these characteristics are believed to lend credibility and authority to their opinions and beliefs about various products. As would be expected, the industry believes that the role of the opinion leaders is the strongest when a new product is launched (or is about to be launched). For example, the industry spends an estimated 24% of their new product commercialization budget on opinion leader activities (CIE 2004). The same study also showed that the 15 largest pharmaceutical manufacturers spend 32% of their total marketing expenditures on opinion leaders. Opinion leader activity is also stepped up when environmental changes occur. In the pharmaceutical industry, these are typically the launch of a new competitive drug, the withdrawal of a drug, issue of new guidelines by the Department of Health and Human Service and/or the National Institutes of Health (NIH) or the emergence of new scientific evidence on the efficacy of a drug or class of drugs. Physicians also socialize at meetings and symposia, and exchange knowledge through scientific and medical journal articles. These interpersonal

communications between physicians can provide information to a physician about the efficacy of new drugs in trials and in practice, new trends in the treatment of particular diseases of interest, availability of generic substitutes, etc. As noted before, these information flows can potentially affect the prescribing behavior of the influenced physicians. Marketing to opinion leaders is typically managed by direct contact with these physicians through detailing. In some pharmaceutical companies, special teams consisting of higher caliber detailers carry out most of this detailing activity. Members of such teams are typically designated Medical Scientific Liaisons (MSLs). A typical team in a large

pharmaceutical company consists about of about 45 MSLs. Industry estimates suggest that about half of the large pharmaceutical companies have MSL teams (CIE 2004). In conclusion, the existence of opinion leaders is taken for granted in the pharmaceutical industry. While firms try and manage their relationships with these opinion leaders via marketing, anecdotal evidence suggests that the identification of opinion leaders and the extent to which they impact other physicians are issues that the industry grapples with. Specifically, first, firms may not usually have a clear idea about who these opinion leaders are.1 Second, there is little systematic understanding of the mechanisms through which opinion leaders and nominating physicians interact. Finally, there is little quantification of the return on investment from targeting these opinion leaders.

2.2.

Related Literature

Our work is related to the sociology, economics and marketing literature on social networks (e.g. Burt 1987; Coleman, Katz and Menzel 1966, Van den Bulte and Lilien 2001, Duflo and Saez 2001). The main focus of this related literature has been to test for social interactions and peer effects using micro-level data.2 There is some work that has postulated the existence of asymmetric peer effects. For example, Reingen and Kernan (1986) focus on identifying the links of a given social network via surveys for a piano tuner service. They note that individuals with stronger ties are more likely to activate the flow of referral. However, they do not explicitly try to document the existence and effectiveness of opinion leaders. Other researchers such as Summers (1970) have tried to examine the characteristics of opinion leaders. They found that opinion leaders are typically more knowledgeable about a product category, but do not explore the quantification of the effect of opinion leaders for outcomes. A few other papers from the medical literature have used surveys and/or field

For information on firms providing tools for identifying and targeting physician opinion-leaders see for e.g., www.estcomedical.com/thoughtleader/ and mattsonjack.com/keymd.asp, and also the Medical Science Liaison Quarterly, http://www.mslquarterly.com/. These reports indicate that most pharmaceutical companies rely on sociometric approaches (described in the next section) to identify opinion leaders. There are some attempts to identify opinion leaders using behavioral (i.e., secondary) data. However, these are subject to the Manski (1993) critique that estimates of social interactions derived via post-hoc identification of peers from outcome data are likely biased upward. 2 A related stream of work in the diffusion of innovations literature has modeled asymmetric contagion effects using a macro-level modeling approach. The general approach is to build a mathematical model with an ex-ante assumption on how two (usually) groups of consumers respond to peer behavior specifically, imitators are affected by the actions of all other previous adopters while innovators are not affected by peer actions. The resulting model is then applied to aggregate sales data to infer the size of the two groups. Some representative publications that follow this approach are Tanny and Derzko (1988), Steffens and Murthy (1992) and Van den Bulte and Joshi (2007)). With only aggregate data, this literature cannot test for (asymmetric) peer effects at the individual level. Thus, our work must be seen as complementary to this stream of work.

experiments to test for opinion leader effects. (e.g., Valente et al. 2003, Lomas et al. 1991 and Celentano et al. 2000). In economics, researchers have investigated social interaction effects more generally in the context of crop-technology adoption (Bandiera and Rasul 2006; Conley and Udry 2000), welfare participation (Bertrand, Luttmer, and Mullainathan 2000), healthplan choices (Sorensen 2005); and retirement plan choices (Duflo and Saez 2002), to name a few. A small, but growing number of recent papers in the marketing literature has also investigated the potential role of peer-effects in new product adoption (e.g. Van den Bulte and Lilien 2001, Manchanda, Xie and Youn 2004 and Iyengar et al. 2008 on new drug adoption; Bell and Song 2007 on Internet grocer adoption; Nam, Manchanda and Chintagunta 2006 on video-on-demand adoption). We refer the interested reader to Hartmann et al. 2007 for a recent and broad overview of the social interactions literature, which also discusses approaches from several related fields. Broadly speaking, relative to the previous literature cited above, our approach has several distinguishing characteristics. These include documenting the asymmetric nature of peer interactions, distinguishing causal peer effects as opposed to correlated outcomes that do not rely on peer effects, and the determination of peer effects in mature product categories i.e., using post-adoption behavior. In terms of the causal effect determination, we believe this paper is one of the first to comprehensively outline and address the identification issues related to endogenous network formation, correlated unobservables and simultaneity, and to include specific controls for targeted marketing activity in the analysis of social interactions in the presence of marketing.

3.

Model

We now discuss our model framework and empirical strategy. Our empirical framework is a descriptive linear model of prescription behavior, which we interpret as the reduced form of the behavioral process generating prescriptions for physicians and their opinion leaders (for structural approaches see Brock and Durlaf 2001; and more recently, Hartmann 2008). In the robustness section later, we discuss some extensions of this linear model that accommodate alternative specifications of the effect of peers as well as relax the linearity assumption (via the use of a count model). In the context of this model, we clarify and discuss how we address the main identification issues inherent in measuring peer effects using micro-level data. We index physicians by i, is opinion leader by j(i), and time by t. Let D denote detailing, and y and x denote new prescriptions for physicians and opinion leaders

respectively. The starting point of our empirical specification for physician prescriptions is a linear regression:

yit = Dit + x j ( i ),t + it

(1)

Here vit denotes unobserved factors that shift prescriptions of physician i over time. While ideally we would like to include the actual opinions of the opinion leader as a covariate to capture the social interaction, these are unavailable in our data. Here, we think of the prescriptions of the opinion leader as a proxy for these opinions (later, we present extensive sensitivity checks to different proxies for leader opinions). Formally, our test for the asymmetric peer effect in prescription behavior is whether is statistically significantly different from zero. An alternative model that uses the share, rather than levels of prescriptions is equivalent to (1), since the overall volume of prescriptions written for the disease condition remained roughly constant across the months in our data. Identification of peer effects in this model requires us to resolve five issues described earlier - reference-group determination, change in the external environment, existence of a communication mechanism between the physician and the opinion leader, an outcome variable that can be measured and the ability to rule out correlation in observed behavior between the physician and the opinion leader arising from endogenous group formation, targeted marketing activity, correlated unobservables and simultaneity. We discuss these in sequence below. Reference group/peer determination First, we need to identify the proper reference group or reference peer for each agent, such that the effect of the group/peers behavior on the agents actions can be measured. Manski (1993 & 2000) discusses in detail the need for exogenously defined social network information to identify peer effects from behavioral data. Intuitively, one cannot use behavior itself to define reference groups, if the goal is to obtain the effect of a reference groups behavior on an agents actions. By grouping agents with ex-post similar actions together, a researcher attempting this approach essentially produces an upward bias in any peer effects unearthed through subsequent analysis. Similarly, geographic or location specific proxies for reference groups cannot sort between peer effects and common unobservables that affect the actions of all agents in the location similarly. We overcome these challenges in our application by using a new dataset that contains detailed social network information obtained via a sociometric approach (e.g., Coleman et al. 1996; Valente et al. 2003; Valente and Pumpuang 2007). In the sociometric approach, individuals units are directly surveyed to

obtain information about other individuals who exert a peer effect on their behavior.3 Each physician in the survey self-reports the doctor whose opinions he incorporates in his prescription decisions, thus identifying his social network. This provides us an exogenous measure of the physicians reference group or peer, circumventing the need to rely on behavior, location or geography-based proxies. Thus, in the setup of equation (1), j(i) is known exogenously. Our use of the term opinion leader is to be interpreted in this sense as referring to doctors nominated by physicians in this survey (described in detail in a later section). Change in the External Environment In stable product categories with well-established brands, agents tend to have little uncertainty about product quality, and may need rely on others actions to make decisions. In stable drug categories, general practitioners may have little uncertainty about the usage and efficacy of the drugs they prescribe. Peer effects may be hard to uncover in such settings. Changes in the environment add exogenous variation that assist in unearthing the peer effect. An advantage of our data is that it covers a time-period where there was a significant change in the recommended usage of drugs in the therapeutic category. For the therapeutic category that we study, this environmental change relates to new treatment guidelines issued by the National Institutes of Health (NIH) regarding appropriate treatment for specific disease indications (we describe the new guidelines later in the paper.) The guidelines were issued by the NIH following fresh evidence available from post-release drug trials. This environmental change occurs around the mid-point of the data, and is exogenous to behavior as it arises from the behavior of a third party that is not affected by the actions of physicians and their opinion leaders. Interaction between the Physician and the Opinion Leader In the survey, physicians also report their mode of interactions with their opinion leaders. Hence, our data also allow us to provide some insights into the mechanism through which the opinion leader effect manifests itself.

In contrast, some studies follow the key informant approach, where a few individuals are polled to determine the identity of individuals with social influence (e.g., Celentano et al. 2000). Interestingly, Iyengar et al. (2008) find that the set of self-reported opinion leaders are different from those identified via a sociometric approach. Others in development economics who have adopted sociometric approaches include, Conley and Udry (2000) and Kremer and Miguel (2004). In the absence of such data, researchers have often defined networks based on geographical location (Bell and Song 2007; Manchanda et. at. 2004); dorm/work location (Sorensen 2005; Dufflo & Saez 2003; Sacerdote 2001); and ethnic/cultural proximity (Munshi and Myaux, 2002; Bertrand et al., 2000).

Outcome Variable The external change in the treatment environment generates higher uncertainty regarding drug efficacy among family and general practitioners, making them more likely to value the opinion of specialists in the category (we discuss this in detail later). The exogenous variation aids in identification. Given the nature of the exogenous change, all physicians are likely to adapt their prescription behavior to reflect the new guidelines. Hence, our identification strategy focuses on how changes in prescription behavior of nominating physicians (y in equation 1) before and after the issuance of the guidelines, is related to changes in prescription behavior of their corresponding opinion leaders. Distinguishing Causality from Correlation As mentioned above, peer effects imply that the behavior of agents in the same reference group would tend to be correlated. However, correlation in the behavior of agents per se does not imply that any one agents action has a causal effect on the actions of others in the group. In addition to peer effects, such correlation in behavior could arise due to three other factors, viz. endogenous group formation, correlated unobservables and simultaneity (see Moffitt 2001 for a discussion.) Only a causal peer effect implies a social multiplier; hence it is important to sort out causal effects from each of these sources of correlation. In our application, another factor that could lead to correlation is targeted marketing. Endogenous group formation Endogenous group formation arises in our context if physicians choose doctors with similar tastes for prescriptions as their opinion leaders. For instance, physicians who face patient bases requiring treatments using a specific class of therapeutic drugs may meet experts in that therapeutic category at conferences organized by drug companies. If physicians choose these experts as opinion leaders, it is likely that such physician-opinion leader pairs tend to prescribe more in the therapeutic category than average. In this case, the observed correlation in the behavior of the physician and his opinion leader could arise from omitted individual characteristics that are correlated within the group. In equation (1), such endogenous group formation implies that physician is unobserved tastes (vit) and opinion leader js tastes for prescriptions could be correlated if opinion leader js tastes also drive his prescriptions, xjt, this generates a correlation between xjt and vit leading to a upward bias in the estimates of . The solution to the group formation problem is facilitated by the availability of panel data (as noted by Manski 1993, the prospect for identification of peer effects in cross-

sectional data are poor). Panel data enables us to include physician-specific fixed effects in the regression (1). In terms of our model, we write,

it = i + it

(2)

where i is a fixed effect specific to physician i, which controls for unobserved (to the econometrician) time invariant tastes for prescriptions. By controlling for physician is tastes, we control for the portion of vit that is correlated with xjt via correlation with opinion leader js tastes, thus accommodating the endogenous group formation problem. The identifying assumption is that, we assume that group selection is fixed over time, and that physician group formation is not influenced by changes in the external environment. Correlated unobservables A second concern is whether there exist correlated unobservables that drive prescriptions of both the physician and the opinion leader similarly. If uncorrected, these manifest themselves as peer effects. An obvious source of correlation is sales-force activity (i.e. detailing) directed at physicians and opinion leaders by drug companies. One can partly control for this source of correlation by including time-period fixed effects that pick up common trends in marketing activity to physicians (e.g. Van den Bulte and Lilien 2001). In our setting, we fully control for such marketing activity by obtaining direct data on detailing to physicians, which we include as explanatory variables in the regression. A potential concern with using this variation arises because detailing may be targeted to physicians. As documented in the recent literature (e.g. Manchanda, Rossi and Chintagunta 2004) many pharmaceutical companies in the US, including our firm, decide detailing allocations based on a volume-based rule, whereby physicians are allocated detailing levels corresponding to their position in deciles of past prescription volume in the focal category (we find evidence for this detailing pattern in our data.) Then, the volume-based detailing rule implies that D may be correlated with vit. Our control for this potential endogeneity derives from the nature of the targeting rule. In effect, on account of stable patient bases, physicians rarely move across deciles (we find this in our data as well). Thus, the inclusion of physician fixed effects pick out the across-physician variation in detailing, and controls for the endogeneity concern. Thus, only within-physician, across-time variation in detailing is used for identification. Fixed effects do not fully absorb all detailing variation however, as in practice, actual detailing levels are centered around, but not exactly equal to top-down allocated levels due to several unanticipated factors that affect visits. These include physicians time-constraints (e.g. the physician is not in his office during a detailing visit), or unanticipated detailer time-

10

constraints (a patient is taking too long, requiring postponement of the visit). This deviation from pre-allocated levels is orthogonal to physician unobservables, and is used for identification. Thus, the underlying identifying assumption is that after controlling for i, within-physician detailing is independent of other physician and time-period specific unobservables, it. We also consider the possibility that there may be additional correlated unobservables that generate co-movement in prescriptions. Candidates for such

unobservables include trends in overall prescriptions across all physicians in the category, as well as any spatially correlated region/location specific shocks to prescription behavior that are captured by jt. We address these as follows. First, we include a full set of time-period fixed effects. These control for any time-trends common across all physicians and opinion leaders. Second, recall we include physician fixed effects. As none of the surveyed physicians in our data share a zip code, physician fixed effects are equivalent to including a full set of zip code fixed effects. Hence, time-invariant spatially correlated unobservables are also fully controlled for. A final issue is whether there are unobservables that are correlated at the level of the zip code and time. To consider this issue, we discuss a potential difference-in-difference approach. We have access to the prescription behavior of all physicians in the country. We use this data to compute the mean prescription of all other physicians in physician is zip-code, denoted by zi,t,

which we include as a covariate in the regression. Essentially z-i,t proxies for all

unobserved time-period and location-specific shocks to prescriptions that are common to all physicians in is location.4 By including these in the regression, we essentially use the prescription behavior of other physicians in is location as a control. Thus, we further decompose (2) as,

it = zit + it

(3)

where it is a mean zero error term. Note this strategy is subject to the implicit caveat that a given physicians opinion leader does not influence other physicians in his zip code. Unfortunately, data on the social networks of the universe of physicians is not available to test this. Hence, our approach will be to present extensive sensitivity checks in which z-i,t is included or excluded from the regression. Tests for correlated unobservables (presented in 5.1) suggest that most of the spurious correlation is along the temporal dimension, which is fully picked up by time-period fixed effects. Hence, the effect of z-i,t on our results is small.

Less than 10% of opinion leaders and nominators are in the same zip code in our data. Our results do not change if we exclude these physicians from our sample.

11

Simultaneity The final factor that must be considered here is simultaneity. Simultaneity implies that physician is actions and opinion leader j(i)s actions may be contemporaneously interdependent. If peer effects exist, the fact that opinion leaders affect physicians while physicians simultaneously affect them leads to an upward bias in the estimation of the interactions. In the context of our model, if physician i and opinion leader js prescriptions are simultaneously determined, high values of it would tend to induce high values of xj(i),t, thus leading to an upward bias. We control for the simultaneity problem via exclusion restrictions. In our context, detailing to the opinion leader, Dj(i),t, as well as the mean prescriptions of all other physicians in the opinion leaders zip-code, z-j(i),t, form excluded variables that affect the prescriptions of the opinion leader (the endogenous variable), and can be excluded from the prescription equation for physician i. Both Dj(i),t and z-j(i),t impact the opinion leaders prescriptions, and are thus correlated with the endogenous variable xj(i),t, but uncorrelated with it. Thus they serve as instruments for xj(i),t thus addressing the simultaneity concern. An alternative approach would be to assume that only past opinion leader prescriptions affect the physicians current prescriptions (i.e. there is no contemporaneous linkage in behavior.) We explore model sensitivity to such specifications in the Robustness section. Note that if we assumed that, given their expert status (details in the Data section below), opinion leaders were not affected by physicians, there would be no simultaneity problem by construction. Rather than assume away simultaneity concerns a priori in this manner, we use the data to check whether peer effects are truly asymmetric. We run the analogous regressions for opinion leader j(i) (i.e. xj(i),t regressed on j(i), Dj(i),t, yit and z-j(i),t) to check whether physician-prescriptions have a significant effect on the prescription behavior of their opinion leaders. Analogously, Dit and z-it are excluded variables for the opinion leaders prescription equation, and serve as instruments for yit in the opinion leaders prescription equation. Final Specification Based on the above discussion, our final specification for physician is prescriptions is,

yit = i + t + Dit + x j (i ),t + zi ,t + it , i = 1,.., N ; t = 1,.., T

The corresponding specification for is opinion leader is:

(3)

x j (i ),t = j (i ) + t + D j (i ),t + yit + z j (i ),t + j (i ),t , t = 1,.., T ;

(4)

12

We estimate both specifications via fixed-effects panel data linear instrumental variables regression. Note that the issues we identify and try to control for above are relevant for any analysis of behavioral data involving social interactions, whether using structural or reduced-form models. We now describe our data.

4.

Data Description

Our data pertain to physician prescription behavior in a large therapeutic class (we cannot reveal the name of this class due to confidentiality concerns). The drugs in this class address a serious chronic disease condition that affects about a quarter of all adults in the United States. We consider a combination drug category in this therapeutic class of drugs that treats the chronic disease. The dataset we use in our analysis is a combination of primary and secondary data. The primary data come from a market survey data of 1500 physicians chosen randomly from a set of 56,000 regularly prescribing physicians across the United States in this therapeutic category. The survey was commissioned by a large pharmaceutical company and carried out by a market research firm with the pharmaceutical company bearing all costs (confidentiality reasons preclude us from naming the companies). The survey was carried out in Jan-Feb 2004. The main objective of the survey was to obtain names of those doctors whose actions influence the nominating physicians approach to the treatment of the chronic disease treated by combination drugs. Nominating physicians were encouraged to name doctors who were known to them (by reputation or otherwise) and then queried about the mechanisms by which they were able to obtain information about the opinion leaders beliefs and actions.5 From this survey, we have access to information on 290 physician-opinion leader pairs. Note that our use of the term opinion leader is to be interpreted in this sense as referring to physicians nominated in this survey. The opinion leaders so identified reflect each nominating physicians subjective opinion regarding who in the field he considers an expert, and whose opinion he incorporates while making prescription decisions in this therapeutic category. We believe this individual-specific measure is the appropriate one for identifying peer effects in such settings.

The specific questions asked during the survey were (a) Whose opinions do you value most regarding the treatment and/or management of [disease condition] among [disease condition] patients?, and, (b) How do you obtain information from that Influencer about the treatment/management of [disease condition]? It is possible that the wording of this question could encourage respondents to name an opinion leader even when there isnt one. However, if a respondent made up the name of an opinion leader, the peer effect we calibrate using (independently) obtained behavioral data should be attenuated. In other words, the survey bias, if present, works against our finding of peer effects.

13

These data were then supplemented by secondary data also collected with the help of the company on the prescription behavior and the marketing activity directed at both the physicians and the opinion leaders. These data span a period of 24 months (from April 2002 to March 2004 inclusive) and contain the prescription counts for the combination-drug category and the count of details received for each month for the universe of physicians for the companys drug. Interestingly, for this drug, the pharmaceutical firm had decided to rely almost entirely on regular detailing to manage relationships with the opinion leaders. Our interviews with the managers of the firm indicated that while the firm would have liked to consider other forms of marketing, such as the setting up of an MSL team, they were not doing that during the period of our data. We also learnt from the firm that only about 50% of the doctors identified as opinion leaders in the survey were on the firms own list of opinion leaders. Descriptives: Primary data We have demographic and location information for 290 opinion leader-physician pairs (including primary affiliation, zip-code and specialty). There are 267 unique nominating physicians. There are 182 unique nominated physicians (opinion leaders). The distribution of nominations in the survey is presented in Tables 1 and 2. Interestingly, over 91% of physicians reported being influenced by only one opinion leader. About 38% of the nominated doctors were named as opinion leaders by more than one physician. We do not see any overlap between opinion leaders and nominating physicians.6 The typical opinion leader is a research-active specialist physician in the therapeutic category and is associated with a university-based hospital. 97.4% of the opinion leaders are specialists. Over 90% of the opinion leaders in the sample are associated with hospitals, and about 30% are affiliated with University hospitals. The average opinion leader has published about 7.2 refereed papers (st. dev. 9.68, min 0, max 40) in this therapeutic class, confirming his status as an expert. The survey also queried the nominating physicians about their mode of interaction with the opinion leader. The dominant mechanism of information transfer as reported by the physicians was direct contact, with about 94.5% of nominating physician mentions. This provides some support for our model formulation (equation 1) in
6

The pharmaceutical firm that commissioned the survey asked the market research firm for a list of only those physicians who nominated an opinion leader (and the opinion leader identity). Thus, we have only information on physicians who have an opinion leader. The consequence is that in essence, we measure the treatment on the treated, where the treatment is the effect of an opinion leader. Note that conditional on having an opinion leader, we do control for the selectivity of the chosen opinion leader (i.e. the endogenous group formation problem) via inclusion of physician fixed-effects that pick up common unobserved traits that could lead to group formation (details below).

14

which the nominating physician is assumed to respond to the prescriptions of the corresponding OPL. Other mechanisms of interaction included symposia/conferences (78%), meeting in clinical and/or hospital settings (67%) and via scientific articles published by the opinion leader (32%). Prescription/detailing data The secondary dataset contains information on 24 months of new prescriptions for the combination-drug category for the entire universe of physicians in the therapeutic class of the disease. The data also contains information on monthly physician-level detailing activity by the focal firm in the category. Unfortunately, detailing activity for the other competing drugs in the category is not available at the individual physician-level. The combination drug category we considered has 4 prescription drugs. Table 3 reports the market shares of the 4 drugs during the time-period of our data. Based on this, we focus on the two largest drugs viz. drug 2, the focal drug, and drug 1, the main competitor. We supplemented our data with monthly national aggregate detailing for drug 1 (detailing for drugs 3 and 4 was negligible). The distribution of aggregate detailing for the two drugs is presented in Figure 1. Descriptive statistics for the sample are provided in Table 4. The table shows that nominating physicians typically write a larger number of prescriptions (almost twice that of the opinion leaders) and also receive a higher level of detailing (about 50% higher). This is consistent with anecdotal evidence that opinion leaders tend to be focussed more on medical research and academic publication rather than practice. As can be seen from the table, opinion leaders are detailed less than the nominating physicians. This is likely to be a function of the category volume rule followed in the industry (whereby physicians who prescribe more at the category level get correspondingly higher details.)7 Change in drug-usage guidelines

Looking at z-it in Table 4, we see that the average physician in a zip code prescribes 0.75 new prescriptions per month in this disease category, which is much higher than that prescribed by physicians included in our survey. This is driven by the fact that we are averaging over many physicians in each zip code who do not prescribe often in this disease category (for instance, a nephrologist may not prescribe drugs for heart disease a disease condition but is included among the set of physicians in a focal doctors zip code). An alternative approach is to include only physicians who are active prescribers in this disease category (the population of interest from the firms point-ofview) in computing z-it. Our motivation for adopting the current approach is a) it is possible that the guidelines affected all doctors in the zip-code, not just active prescribers, and hence including the prescriptions of all doctors in computing z-it would better pick up the overall effect of the guidelines; b) since we did not have access to the database of physicians from which the company picked the random sample, we would otherwise have to make arbitrary guesses about who to include when computing z-it. Given this, we also present several robustness checks in which we drop z-it from the regressions, instead controlling for its effects via a full set of month-fixed effects. We thank an anonymous referee for suggesting this approach.

15

An important aspect of the data is that it covers a time-period where there was a significant change in the guidelines for the usage of drugs in this therapeutic category. This change in the treatment environment is important for us since it is then that family and general practitioners are most likely to seek and value the opinion of specialists in the category. Appendix A provides excerpts from published sources and the summary findings from survey of physicians attesting that, in general, a change in guidelines usually increases the uncertainty in terms of physician prescription decisions. The increased uncertainty of

nominating physicians regarding drug efficacy in the changed environment thus aids in identifying the effect of opinion leaders on prescription behavior. In our context, an exogenous change in the market occurred in May 2003, in the form of an announcement by the National Institutes of Health (NIH) affiliate releasing new treatment guidelines for the disease. Thus, we have 13 months of behavioral data before the guidelines were released and 11 months after. The guidelines suggested that, as against the prevailing norm, the

initiation of treatment for severe cases of this condition should comprise of at least two agents (or molecules). The guidelines also stated that more than one drug would be required to treat most cases of the disease. These guidelines tended to favor the so called combination drugs in this category. A combination drug typically had the two agents in the same pill and results in polytherapy. This had the obvious advantage of increasing compliance amongst patients i.e., taking two pills is much easier than taking one. However, there were additional therapeutic benefits for patients beginning therapy using these combination drugs. These (combination) drugs also had been shown to have therapeutic advantages such higher efficacy (than for two drugs taken individually), lower side effects and higher potency with lower dosages. Thus, we expect that all combination drugs, including drug 2, should show an increase in prescriptions after these guidelines were issued. Prior to the issuance of new guidelines, combination drugs were generally considered aggressive therapy. 8 We now document the changes in prescription and detailing behavior before and after the release of the guidelines. Figure 2 presents the distribution of mean monthly prescriptions in the combination-drug category for nominating physicians and opinion leaders before and after May 2003 (when the new treatment guidelines were introduced). For each physician, we compute the mean monthly new prescriptions before and after and present them in a box plot. As can be seen from the figure, both sets of physicians prescribe more of the combination drugs category. The mean increase in new prescriptions across both

The firm also showed us survey data collected from 319 physicians after the guidelines were released. These physicians noted that, based on the guidelines, they would expect polytherapy to become more prevalent, leading to an increase in the prescription levels of combination drugs.

16

groups is about 10%. Again, as before, the nominating physicians prescribe more than the opinion leaders before and after. We now turn our attention to the distribution of monthly detailing for drug 2 across all physicians and months. Figure 3 presents the distribution of mean monthly detailing for drug 2 for physicians and opinion leaders before and after. As before, we compute the mean monthly detailing before and after for each physician, and present them in a box plot. Interestingly, the firm seems to deviate from the detailing allocation rule cited above just after May 2003. This can be seen from the figure as the firm details more to opinion leaders after the change even though they write fewer prescriptions in the combination-drug category (see Figure 2). At the same time, detailing to the regular physicians remains relatively unchanged before and after the issuance of the guidelines (the difference is not statistically significant.) This suggests that the firm has some knowledge of the opinion leader-status of these specialists. As noted earlier, about half of the nominated opinion leaders had been characterized as opinion leaders by the firm prior to the survey. To summarize, we have a unique dataset that combines information from primary and secondary sources. Our data suggest that opinion leaders are less heavy prescribers in the therapeutic category relative to the average physician. We find that the NIH guidelines impact the prescription behavior of both nominating physicians and opinion leaders on average the prescription quantity goes up by 10%. This is consistent with the content of the guideline (as described above). We also find that the firm that markets drug 2 changes the allocation of its detailing resources after the release of the guidelines. Specifically, before the release of the new guidelines, the firm devotes a lesser amount of detailing to the opinion leaders. However, this pattern is reversed after the guidelines are released. This is consistent with firm feedback that detailing is the main instrument that is used by this firm in this therapeutic category in terms of managing its relationship with opinion leaders.

5.

Results

We now present the results from the analysis of the data. We first present some model-free evidence for the presence of correlated unobservables in prescriptions between the physicians and the opinion leaders. We then present results from various OLS and fixed effects instrumental variables specifications that control for the identification issues discussed in section 2. We then discuss robustness of our main results to alternative model specifications. We conclude by using our estimates to measure the incremental value of targeted sales force

17

activity to the opinion leaders in the data, which we use to estimate the social multiplier of detailing.

5.1.

Testing for correlated unobservables

Recall that unobservables correlated across physicians and the opinion leader which shift prescriptions similarly can result in spurious correlation, and bias the estimate of the peer effect parameters. One approach to this problem was to use the mean prescriptions of other physicians in the focal physicians zip code to control for such potentially correlated unobservables. Correlated unobservables could arise along both the temporal and spatial dimensions. For example, the new NIH guidelines could generate an overall incentive to prescribe the combination drugs for both physicians and opinion leaders. This manifests itself as an overall trend in prescriptions, which when ignored, could generate a spurious correlation in the physician and opinion leaders prescriptions over time. Alternatively, spurious correlation could arise from unobserved shocks to prescriptions of both physicians and opinion leaders that are contemporaneously spatially correlated. To check whether mean prescriptions of other physicians in the focal physicians zip code reveal a trend over time (i.e. a common temporal shock to prescriptions) we regress z-it, the mean prescriptions of all other physicians in is zip-code, on physician and time period specific fixed effects. The time period specific fixed effects from this regression are presented in Figure 4 (all the time-period dummies are significant). The plot reveals an upward trend in mean prescriptions over time. By controlling for z-it we do not use this potentially common component of variation in prescriptions over time to identify peer effects. For comparison, we also plot the corresponding coefficients from regressing the mean prescriptions of other physicians in the opinion leaders zip code (z-OPL,t) on month dummies. Recall that z-OPL,t contains unobserved shocks that shift the opinion leaders prescriptions over time. This plot also reveals an upward trend. The picture suggests that there exist correlated temporal shocks in the prescriptions of both physicians and Opinion leaders, which if not controlled for, could lead to spurious findings of peer effects. We also check the extent to which spurious correlation could arise from spatial dependence in unobservables that drive prescriptions of the physician and the opinion leader. Recall that we use the mean prescriptions of other physicians in the focal physicians zip code to control for the unobservables. We can get a sense of the extent of spatial correlation holding the temporal variation fixed, by looking at the dependence in the mean prescriptions of other physicians in the physician and the opinion leaders zip-codes timeperiod by time-period. Table 5 shows the correlations between z-it and z-OPL,t in the data each

18

month (Physician-opinion leader combinations that shared a common zip-code were dropped - about 9.6% of the observations.) In general, we do not find much evidence in the data of statistically significant spatially correlated shocks, suggesting that the main component of correlated unobservables is along the temporal dimension. Analogously, we check for correlation in z-it and z-OPL,t along the temporal dimension after controlling for across physician variation (i.e. we first regress z-it and z-OPL,t on physician and opinion leader fixed effects respectively, and report the correlation in the residuals from this regression). This correlation in the data was 0.0502, which is statistically significant (p < 0.0001).

5.2.

Parameter estimates

In the first subsection, we first present results from our primary specification (equation 3), in which we test whether physician prescriptions is significantly affected by the opinion leaders prescription behavior. We then present results for the reverse regressions (equation 4) in which we test whether physician prescriptions is significantly affected by the nominating physicians prescription behavior. Effect of Opinion Leader behavior on Physician Prescriptions OLS Estimation Results We start with discussing the estimates from OLS and fixed effects specifications of the model. The dependant variable in all regressions is the total new prescriptions by the nominating physician in the category. Results from OLS linear regression are presented in Table 6. From the table, we see that the effect of the OPLs prescriptions is positive and significant. Further, the magnitude of the effect is higher after month 13 (May 03), when the NIH guidelines were introduced. This is consistent with a basic pattern of correlation between the physician and opinion leaders prescriptions. The OLS estimate of the detailing coefficient is also large and strongly statistically significant. However, this specification does not take into account the fact that heavier detailing is targeted by firms towards higher volume physicians. Hence, the OLS estimate is likely biased upward. Fixed-Effects Estimation Results We present the results from the fixed effects linear regression in Table 7. Recall that fixed effects control for both potential endogenous group formation as well as the targeting of detailing to high-volume prescribers. However, these estimates do not control for potential simultaneity biases. Not surprisingly, the magnitude of the detailing coefficient drops under fixed effects. The F-statistics from the regression strongly reject the null that all the fixed

19

effects are zero. Although not reported, the hypothesis that the fixed effects are uncorrelated with the included variables is also strongly rejected in all specifications. The variable z-it controls for both temporal and spatially correlated shocks, and is statistically significant. The regressions in columns 3 & 4 in Table 7 indicate that after controlling for physician fixed effects and z-it, the effect of opinion prescriptions after May 2003 is positive and strongly significant (t = 3.09). The results indicate that the effect of the opinion leaders prescription is significant only after the release of the guidelines. In particular, the OPL has little effect on prescriptions prior to May 2003, when the category was relatively stable. Given the timing of the survey (Jan-Feb 2004), these results cannot be explained by a mere measurement effect i.e., opinion leaders and their behavior becoming more salient because the survey evoked the relationship between the nominating physician and the opinion leader. These results are consistent with the observed nature of change in category and suggest peer effects following the guideline release. Interestingly, after controlling for fixed effects, detailing is not significant in explaining physicians prescriptions over time. The marginal effect of an opinion leaders prescription is more than 100 times larger than the detailing effect in the latter time periods. This suggests that targeting opinion leaders with detailing with the aim of increasing their prescriptions is a better strategy for firms seeking to increase category volume among physicians in this market. We explore this strategy toward the end of this section. Fixed-effects Instrumental Variable Results We now discuss the results from the fixed-effects instrumental variable (IV) regressions. Recall that we use instrumental variables to accommodate potential biases arising due to concerns related to simultaneity. The endogenous variable we instrument for is xj(i),t. We use the detailing to the opinion leader, Dj(i),t, as well as the mean prescriptions of all other physicians in the opinion leaders zip-code, z-j(i),t, along with squared terms of both, as instruments. Results from the first-stage regressions of xj(i),t on the instruments are presented in Table 8. Columns 1 & 2 present results from the regressions of the endogenous variable on only the excluded instruments, while columns 3 and 4 present the results from the regressions of the endogenous variable on the entire first-stage matrix. The F-statistics from both regressions strongly reject the null that the exogenous instruments have no explanatory power. The first-stage is able to explain about 59% of the variation in the opinion leaders prescriptions. Thus, it is clear that we do not have a weak instruments problem. The signs of the parameters also make sense intuitively. In particular, detailing to the opinion leader has a positive and significant effect on the opinion leaders prescriptions,

20

and the mean prescriptions of other doctors in the opinion leaders zip-code is positively significantly correlated with the opinion leaders prescriptions. These suggest that the instruments are working correctly. Results for the fixed-effects IV regressions are presented in Table 9. The IV estimates parallel the results from Table 5. Looking at columns 3 - 4 in Table 8, we find that opinion leaders prescriptions have a statistically significant effect on the nominating physicians behavior after May 2003 (t =2.06), when the NIH guidelines were released. In Table 10, we repeat the IV regressions including a full set of month fixed effects, and continue to find a significant effect for the opinion leaders prescriptions post-guidelines. Table 10 also presents sensitivity to the inclusion of z-it in the regression. Recall that in computing z-it, we included the prescriptions of all other doctors in a physicians zip-code, which may include physicians who are not in the target population of interest to the firm (the target population for the firm includes only active prescribers in this disease category). In that case, z-it may not be correctly capturing unobservables relevant to the focal physicians prescriptions. We present robustness to z-it in the following way. From Table 5, we see that the z-it are capturing unobservables that are correlated mostly along the temporal, rather than spatial, dimension. These effects may thus be captured via time-period fixed effects. On a referees suggestion, we also explored month-fixed effects that are specific to each zip-code, but find that the data are too thin to support this specification. We present specifications with a full set of month-fixed effects, with and without including z-it, and find our results continue to be robust. We interpret these results as evidence that there exist peer effects in prescription behavior in these data. In subsequent sections below, we will examine whether the peer effects we find have economically significant consequences for firms marketing efforts in this industry. Effect of Physician behavior on Opinion Leader Prescriptions We now check whether peer effects in this category are asymmetric, by testing whether the prescription behavior of nominating physicians has statistically significant effects on the actions of their opinion leaders. Rather than assume each nominated doctor has a disproportionate impact on the physician-doctor dyad, we wish to verify whether the data supports the notion that the nominated doctors are experts, and are less affected by the behavior of the nominators. The specifications we estimate correspond to equation 4. For brevity, we present the fixed effect IV regressions that incorporate a full set of fixed effects for the opinion leaders, and uses the detailing to the nominating physician and the mean prescriptions of all other doctors in the nominating physicians zip-code (along with squared

21

terms)

as instruments

for

the

nominating

physicians

prescriptions. We

present

specifications with and without the mean prescription behavior of all other doctors in the opinion leaders zip-code as controls for unobservables; and with and without a full set of month-fixed effects. Table 11 presents the results. We find that the nominating physicians behavior does not have a statistically significant effect on the opinion leaders prescriptions. These results hold after allowing for a post-guideline interaction effect (columns 3 & 4). Detailing to the opinion leader is found to have a strong significant effect on prescriptions.9 We take the results in Table 11 as evidence that peer effects tend to be asymmetric, at least in this particular context. Robustness We now present several specification checks to ascertain the robustness of our findings. First, we present regressions in which peer effects are interacted with month-fixed effects. Second, recall that in our empirical model, we use the number of prescriptions by the opinion leader as a proxy for the opinion leaders opinions regarding drug use in this therapeutic category, which are transmitted to the nominating physician via interactions, and which subsequently affect the nominating physicians behavior. As noted before, we believe that this is a reasonable measure since 94.5% of physicians in the survey report direct contact with the opinion leader, thus justifying the assumption maintained in (1) that the opinion leaders prescriptions, xj(i),t, could be known to the physician. Rather than take a stand on a specific parametric model of interactions, our approach tests for these measures using a reducedform approach. In this section, we verify that the qualitative nature of our findings are robust to alternative specifications of the manner in which the opinion leaders behavior affects the nominating physician. Robustness to varying specifications of peer effects We now examine robustness to alternative specifications of the peer effect. We explore five different specifications that vary in the timing as well as the nature of the effect of the opinion leaders behavior on the nominating physicians. Let fOPL(t) denote the variable summarizing the opinion leaders behavior on the physicians prescriptions in month t. The various specifications of fOPL(t) we consider are:

One potential reason why detailing effects may be stronger for opinion leaders may be that pharmaceutical firms assign their better sales-people to make calls to these physicians. Separating out such detailer effects require data on the identity of the detailer in addition to the number of details made. Unfortunately, these data are not available.

22

1. fOPL(t) =

x j(i ),t 1 Lagged Opinion Leader prescriptions.

j i ,t

2. fOPL(t) = I 3. fOPL

( x ( ) ) Indicator for whether Opinion Leader prescribed this month. (t) = I ( x ( ) ) Lagged indicator for whether Opinion Leader prescribed.
j i ,t 1
t t 1

4. fOPL(t) = =1 x j(i ), Cumulative prescriptions by Opinion Leader as of this month. 5. fOPL(t) = =1 x j(i ), Lagged cumulative prescriptions.

Table 12 presents results from linear fixed effects IV regressions in which each of the five fOPL(t) functions are included as regressors. As before, we use Dj(i),t, as well as the mean prescriptions of all other physicians in the opinion leaders zip-code, z-j(i),t, along with squared terms of both, transformed analogously to fOPL(t) as instruments (see last 2 rows of Table 10 for precise definitions.) Looking at Table 12, we see that the main message from our early results remains robust to these alternative definitions: Opinion leaders have a statistically significant effect on prescriptions after the guidelines were issued. We also checked robustness to our assumption of linearity by estimating a fixed effects negative binomial (NBD) regression model (Hausman, Hall, Griliches 1984). This model accounts for the count nature of prescription data, allows for over dispersion, and also accommodates potential nonlinearities in the prescription response function. The results from the NBD regression (available from the authors on request) support the basic finding that opinion leader prescriptions have a positive and significant effect on physician prescriptions post-guidelines remains unchanged. We interpret this as providing some robustness checks on the results from our linear model. Finally, to check whether the peer effects varied by physician and opinion leader characteristics, we also estimated versions of the linear fixed effects model (results not reported here) in which the opinion leaders prescriptions were interacted with opinion leaders and physician characteristics (specialty & number of published papers) as well as the channel of social interaction. We found that these interactions were not statistically significant. A goal for future research is to further identify the mechanism through which social effects occur this provides a motivation for collecting more data of the form we have. An alternate approach may also be to design interventions with randomized treatments (Sacerdote, 2001; Duflo and Saez, 2003; Kremer and Miguel, 2004), or through field studies (Godes and Mayzlin 2004).

23

In sum, our results provide significant evidence for asymmetric peer effects. These effects persist after controlling for problems of endogenous group formation, targeted marketing activity, correlated unobservables and simultaneity, and are also robust to functional form. We find that opinion leader behavior significantly affects physician behavior after an exogenous change in the market that resulted in a change in the therapeutic environment. Targeting sales force activity to the Opinion Leaders We now use our results to explore the implications of targeting detailing at the opinion leaders. In particular, we wish to estimate the social multiplier associated with detailing in this industry. As was discussed previously, the focal firm seemed to be aware of the relevance of opinion leaders and did indeed shift detailing toward opinion leaders after the guideline release. An incremental detail to an opinion leader has two effects: first, it increases the prescriptions by the opinion leader; and second, the increase in the prescriptions by the opinion leader increases prescriptions by the corresponding physician. We use our model estimates to measure both effects. We first estimate a prescription response function for the opinion leaders (given our results from the previous section, we do not include the nominating physicians prescriptions as a covariate). We regress opinion leader prescriptions on opinion leader detailing while controlling for the common shocks using the mean prescriptions of all other physicians in the opinion leaders zip-code. Table 13 presents results from both OLS and fixed effects specifications. As noted before, the effect of detailing is strongly significant for opinion leaders. The effect persists even after using fixed effects to control for potential targeting by firms of opinion leaders. As expected, the coefficient of detailing drops in absolute magnitude when moving from the OLS to the fixed effects model. However, the marginal effect of detailing of drug 2 is 0.136 this is significantly larger than the corresponding effect for nominating physicians. We now use these estimates to compute the incremental revenue of an additional detail by both drugs 1 and 2 to an opinion leader. Note that total marginal effect of drug 1 and 2 detailing is 0.127 (0.136-0.009). In the aggregate data for the category, we find that one new prescription generates on average two renewals (obtained by subtracting one - the new prescription - from the number of total prescriptions generated per prescriptions for our therapeutic category). Further, the revenue from one additional dose of a combination drug is around $ 50.00. Thus, the incremental revenue from an additional prescriptions is about $ 150.00. Given this, the incremental revenue from one additional detail in the category is

24

$19.05 ($150.00*0.127). We then compute the incremental revenue arising from the increase in opinion leader prescriptions for a nominating physician for each month after the guideline change. We use the results from the main regression in the last column in Table 10. Post guidelines, the additional revenue from a nominating physician as a result of the increase in opinion leader prescriptions is 0.127*0.032*$ 150 = $ 0.61. From Table 2, the average opinion leader in our sample influences 1.56 physicians. Hence, the total effect arising out of a social interaction is 1.56*$0.61 = $0.95. For the average opinion-leader, this accounts for about 4.76% (=0.95/(19.05+0.95)) of the total revenue effect of detailing; implying a social multiplier of detailing in the category of 1.05. For the top opinion leader, who influences 17 physicians, we obtain a social-multiplier of about 1.35% (Figure 5 provides the distribution of social multipliers for the physicians in our data). Note these are likely the lower bound on the peer effect since z-it could contain some prescriptions generated via the opinion leader effect. Even so, we find that peer effects alone provide a 5%- 35% lift in the return-on-investment from targeting marketing at opinion leaders. A take-away from the analysis is that opinion leader identification is of key importance to the company.

6.

Conclusions

Our paper adds to the small but growing literature that documents peer effects using individual consumer level data. The unique and novel features of our application include the asymmetric nature of the interaction and the presence of marketing activity whose effects are moderated by the interactions among agents. The existence of the peer effect generates spillovers that multiply the returns to targeting influential agents within a group. Our key contributions can therefore be seen as providing evidence for asymmetric peer effects in physician prescription decisions via the use of a novel dataset from the pharmaceutical industry. Previous academic and industry literature has provided little support for these effects in this context, in spite of strong industry beliefs in the existence of these effects. We do this while addressing the identification challenge separating out causality from multiple sources that give rise to correlations inherent in this problem. The detected effects are robust to model specification and functional form. Finally, we use the estimated causal effects to derive implications for marketing resource allocation for firms in the industry and present estimates of the social multipliers effects of marketing activity (detailing). Our estimates indicate important peer effects in prescription choice, and economically significant social multiplier effects in detailing.

25

Our analysis has some limitations. First, the network in our approach is thin relative to networks used in previous research. Richer networks could allow researchers to investigate effect of network structure on the peer effect. Second, given that we only have the treatment effect on the treated, we cannot compute the social multiplier across all types of physicians. Our conversations with the firm suggest that most general practitioner doctors in this category may be subject to some form of peer influence; hence treatment on the treated is likely the right policy-relevant treatment effect. Third, we do not have direct measurements on the information or opinions shared by the opinion leader with the nominating physician. While this is a limitation, our results are robust to many functions of the opinion leaders prescription activity. An interesting extension of our research would be to investigate how peer effects are moderated by the nature of nominator and opinion leader interactions (e.g., in person, via e-mail etc.) over time. Fourth, our data also come from one specific therapeutic category and from one survey. We also do not have any data on the identities of the sales force and patients seen by the physicians in our sample. Fifth, we also do not have access to marketing activity for all drugs in this category at the individual physician level e.g., we only have aggregate detailing for drug 1 (the market leader). Finally, we have limited demographics on opinion leaders and nominating physicians. We are also unable to accommodate unobserved heterogeneity in peer effects; on the other hand, given we found limited evidence for observed heterogeneity in peer effects, this may be not be unreasonable for these data. Our hope is that with access to richer data, these limitations will be addressed in future work.

References
Bandiera, O. and Imran Rasul (2006), Social Networks and Technology Adoption in Northern Mozambique, Economic Journal, 116, 869-902. Becker, G. and K. Murphy (2000), Social Economics. Cambridge: Harvard University Press. Becker, Marshall H. (1970), Sociometric location and innovativeness: Reformulation and extension of the diffusion model, American Sociological Review, 35, 267-283. Bell, David R. and Sangyoung Song. (2007), Social Contagion and Trial on the Internet: Evidence from Online Grocery Retailing, Quantitative Marketing and Economics, 5(4), 361-400. Bertrand, M., E. Luttmer, and S. Mullainathan (2000). Network Effects and Welfare Cultures, Quarterly Journal of Economics, 115(3), 10191055. Brock, William A. and Steven N. Durlauf. (2001). Discrete Choice With Social Interactions, Review of Economic Studies, 68(235,Apr), 235-260. Burt, Ronald S. (1987), Social Contagion and Innovation: Cohesion versus Structural equivalence, American Journal of Sociology, 92, 1287-1335.

26

Celentano, David D., Katherine C. Bond, Cynthia M. Lyles, Sakol Eium Trakul, Vivian F.-L Go, Chris Beyrer, Chainarong na Chiangmai, Kenrad E. Nelson, Chirasak Khamboonruang and Chayan Vaddhanaphuti. (2000), Preventive Intevention to Reduce Sexually Transmitted Infections, Archives of Internal Medicine, 160, Feb 28, 535-540. Coleman, James S., Elihu Katz, and Herbert Menzel. (1966), Medical Innovation: A Diffusion Study, The Bobbs-Merrill Company, Indianapolis, IN. Conley, T. G. and C. R. Udry. (2000), Learning About a New Technology: Pineapple in Ghana, Working Paper, University of Chicago. Cutting Edge Information (CIE). (2004), Pharmaceutical Thought Leaders: Brand Strategies and Product Positioning, Report PH64. Dufflo, E. and Emmanuel Saez. (2003), The Role of Information and Social Interactions in Retirement Plan Decisions: Evidence from a Randomized Experiment, Quarterly Journal of Economics 118(3), pp. 815-842. Godes, David, and Dina Mayzlin. (2004). Firm-Created Word-of-Mouth Communication: A Field-Based Quasi-Experiment, Marketing Science, forthcoming. Granovetter, M.S. (1985). Economic action and social structure: the problem of embeddedness, American Journal of Sociology, vol. 91, pp. 481510. Hartmann, Wesley (2008), Demand Estimation with Social Interactions and the Implications for Targeted Marketing, Working paper, Stanford University. Hartmann, Wesley R., Puneet Manchanda, Harikesh Nair, Matthew Bothner, Peter Dodds, David Godes, Kartik Hosanagar and Catherine Tucker (2007), Modeling Social Interactions: Identification, Empirical Methods and Policy Implications, Marketing Letters, forthcoming. Hausman, Jerry, Bronwyn H. Hall and Zvi Griliches. (1984), Econometric Models for Count Data with an Application to the Patents-R&D Relationship, Econometrica 52: 909-938. Iyengar, R., Valente, T. and Van den Bulte, C. (2008), Opinion Leadership and Social Contagion in New Product Diffusion, working paper, University of Pennsylvaia. Kremer, M. and Miguel, E. (2004). Worms: identifying impacts on education and health in the presence of treatment externalities, Econometrica, vol. 72, pp. 159218. Lomas, Jonathan, Murray Enkin, Geoffrey M. Anderson, Walter J. Hannah, Eugene Vadya and Joel Singer. (1991), Opinion Leaders vs Audit and Feedback to Implement Practice Guidelines, Journal of the American Medical Association, 265 (17), 2202-2207. Manchanda, Puneet, Peter E. Rossi and Pradeep K. Chintagunta (2003), Response modeling with non-random marketing mix variables, Journal of Marketing Research, 41, 467-478. Manchanda, Puneet, Ying Xie and Nara Youn (2004), The Role of Targeted Communication and Contagion in New Product Adoption, Marketing Science, forthcoming. Manski, C. F., (1993), Identification of Endogenous Social Effects: The Reflection Problem, Review of Economic Studies 60, pp. 531-542. Manski, C. F., (2000), Economic Analysis of Social Interactions, Journal of Economic Perspectives 14(3), 115136. Moffitt, R., (2001). Policy Interventions, Low-Level Equilibria, and Social Interactions, in Durlauf, S. and Young, P. (Ed.) Social Dynamics, Brookings Institution Press and MIT Press 45-82

27

Munshi, K. and Myaux, J. (2002). Development as a process of social change: an application to the fertility Transition, mimeo, Brown University. Nam, Sungjoon, Puneet Manchanda and Pradeep K. Chintagunta (2006), The Effects of Service Quality and Word-of-Mouth on Customer Acquisition, Retention and Usage, working paper, University of Michigan. Reingen, Peter H. and Jerome B. Kernan. (1986), Analysis of Referral Networks in Marketing: Methods and Illustration, Journal of Marketing Research, 23 (November), 370378. Rogers, Everett M. (2003), Diffusion of Innovations, 5th Ed., The Free Press, New York. Sacerdote, B. (2001). Peer effects with random assignment: results for Dartmouth roommates, Quarterly Journal of Economics, vol. 116, pp. 681704. Sorensen, A. (2005), Social Learning and Health Plan Choice, forthcoming, RAND Journal of Economics. Steffens, P.R., and D.N.P. Murthy (1992), A mathematical model for new product diffusion: The influence of innovators and imitators, Mathematical and Computer Modeling, 16 (4), 11-26. Summers, John O. (1971), The Identity of Womens Clothing Fashion Opinion Leaders, Journal of Marketing Research, 7 (May), 178-185. Tan, Robert S. (2003), Physician executives as opinion leaders in biotechnology and pharmaceuticals, Physician Executive, 29 (3), 26. Tanny, S.M., and N.A. Derzko (1988), Innovators and imitators in innovation diffusion modeling, Journal of Forecasting, 7, 225-234. Valente, Thomas and Pumpuang, Patchareeya (2007), Identifying Opinion Leaders to Promote Behavior Change, Health Education and Behavior, 34, 881-896. Valente, Thomas H., Beth R. Hoffman, Annamara Ritt-Olson, Kara Lichtman and C. Anderson Johnson. (2003), Effects of a Social-Network Method for Group Assignment Strategies on Peer-Led Tobacco Prevention Programs in Schools, American Journal of Public Health, 93 (11), 1837-1843. Van den Bulte, Christophe and Yogesh V.Joshi (2007), New Product Diffusion with Influentials and Imitators, Marketing Science, forthcoming. Van den Bulte, Christopher and Gary L. Lilien. (2001), Medical Innovation Revisited: Social Contagion versus Marketing Effort, American Journal of Sociology, 106(5), 1409-35. Watts, D.J. and P.S. Dodds, (2007) Influentials, Networks, and Public Opinion Formation, Journal of Consumer Research., Vol. 34.

Appendix A: The Effect of New Guidelines

In general, guidelines are released to advocate new courses of therapy, to report the efficacy of the drug to alleviate symptoms not considered in the past, to report interactions with existing drugs etc. Thus, the primary outcome of these new guidelines are changes in prescription behavior, especially prescriptions. Note, however, that the release of new guidelines usually results in a period of confusion about the implications and the intended use of those guidelines. This confusion arises as guidelines are phrased in general terms and

28

not for each individual patient. In addition, many times, the guidelines can give rise to questions that need to be answered once the guidelines have been followed. We illustrate the increased uncertainty post guideline issuance using three cases of new guidelines (all text in italics in the excerpts below is ours). First, in the case of the American Heart Associations guidelines for womens use of aspirin, the issued guidelines were not precise, leading to confusion as is evident from this excerpt below (from www.health.yahoo.com): Women and Heart Disease, Part 1: Aspirin Confusion Posted by Simeon Margolis, M.D., Ph.D. on Thu, Mar 08, 2007, 4:27 am PST The 2007 update of the American Heart Association's guidelines for women has led to some confusion regarding the use of aspirin. The guidelines do indeed state that women 65 years or older should consider taking low-dose aspirin (81 mg daily or 100 mg every other day), and there is no mention that the presence of risk factors should affect this decision. The second case is that of screening guidelines for cervical cancer using Pap tests. We include excerpts from www.aafp.org and Market Wire below: Although the American Cancer Society, American College of Obstetricians and Gynecologists, and U.S. Preventive Services Task Force have released new guidelines for screening, differing language, requirements and timelines in these guidelines can confuse family physicians and their female patients. www.aafp.org, Oct 11, 2005. Confusion Over Pap Test Guidelines There's a wealth of misleading information about Pap tests and cervical cancer. Headlines are filled with reports of new guidelines and the possibility of a vaccine that could prevent cervical cancer. But this all creates confusion for women and may discourage them from getting a Pap test. Pathologists, physicians who care for patients through laboratory medicine, say that even with the new guidelines, every woman needs to have a regular Pap test. For some women, that means every year. For other women, that may mean every other year. Market Wire, December, 2002. The final illustration is that of the FDA guidelines for pharmacist guidelines on drug compounding. Despite FDA's new compounding guidelines, confusion reigns Reaction to the new guidelines was mixed. Susan Bishop, manager of regulatory affairs and political action for the American Pharmaceutical Association, said pharmacists who were hoping for guidance from the FDA would find more confusion than clarification. Rather than a strict distinction between compounding and manufacturing, the agency is laying out broad directions for enforcement and reserving the right to change course without warning. "I don't know if I would call the document useful," Bishop said. It doesn't give pharmacists any comfort to hear that they may be breaking the law, but that FDA might also decide to change its mind on whether or not they're breaking the law. www.drugtopics.com, July 1, 2002 We also polled a convenience sample of physicians (n=4, details on physician demographics below) and asked them the following question In your opinion, when new guidelines are

29

released by bodies such as the FDA and NIH, do they in general, lead to lower or higher uncertainty in treatment especially in the first few months following the guideline release? In addition to your answer, please provide the context in which we should interpret your answer. We summarize the findings below: 1. All four physicians felt that guidelines lead to increased uncertainty. 2. This is because the clinical presentation of each patient is different and the physician needs to individualize the treatment for the patient. Guidelines do not provide a clear indication of when they would be applicable and when not leading to confusion about whether the treatment is relevant for a given patient. For example, an issue with moving to a combination drug is that the use of combination drug dramatically increases the possibility of negative interactions with other courses of therapy that the patient may be undergoing. The guidelines do not explicitly detail this and therefore the physician needs to verify, for each patient, the possibility of such negative interactions. 3. In general, the role of guidelines is to increase awareness of treatment options. It is not to provide rules of thumb that would lead to lower treatment uncertainty. 4. One of the physicians offered up a specific example of her dilemma. A recent release of clinical data suggests that Vytorin/Zetia does not help patients trying to lower cholesterol. If guidelines based on this result are released, then this physician said that the guidelines would have to be interpreted on a patient-by-patient basis causing her more uncertainty. Specifically, her options would be to take all her patients off the drug, leave patients who are responding positively to the drug on the drug or follow a phased withdrawal. Physicians Consulted (full names not revealed due to confidentiality reasons): Dr. P.K. Specialist, Houston Northwest medical center, Houston, TX Dr. R.K. Former Medical director, Robert Wood Johnson hospital rehabilitation Center, Cranbury, NJ Dr. P.S. Internist, LaPeer Regional Hospital, Detroit, MI Dr. R.T. Family Practice, UMDNJ, New Jersey Taken together, these data illustrate why and how new guidelines lead to increased uncertainty in most cases. Table 1: Distribution of Nominations Number of Nominations Number of Nominators 1 245 2 21 3 1 Total 267 Notes: To be read as: there were 245 physicians who nominated 1 opinion leader, 21 who nominated 2 opinion leaders, and 1 who nominated 3 opinion leaders.

30

Table 2: Distribution of Opinion Leader Nominations Number of Number of nominations Opinion Leaders 1 112 2 56 3 7 4 5 8 1 17 1 Total 182 Notes: to be read as, there are 112 doctors who were nominated as opinion leaders by exactly 1 physician, 56 doctors who were nominated as opinion leaders by exactly 2 physicians, etc, and 1 doctor who was nominated as an opinion leader by 17 physicians. Table 3: In-sample market-shares of combination drugs
Physicians Drug 1 Drug 2 Drug 3 Drug 4 0.924 0.073 0.003 0.000 Opinion Leaders 0.861 0.138 0.002 0.000

Table 4: Sample descriptives

Variable Physician prescriptions Opinion Leader prescriptions Physician details (drug 2) Opinion Leader details (drug 2) Z-it Z-OPL,t Mean 4.16 2.23 0.75 0.52 0.75 0.40 Std. Dev. 4.40 4.85 1.35 1.15 0.94 0.38 Min 0 0 0 0 0 0 Max 39 54 11 10 13.7 3.3

Notes: Number of observations in sample = 6960. Z-it refers to the mean prescriptions of all other physicians in nominator is zip-code; Z-OPL,t refers to the mean prescriptions of all other physicians in nominator OPLs zip-code;

31

Table 5: Correlation between mean prescriptions of other physicians in the physicians and the opinion leaders zip-codes
Month Correlation p-value Month Correlation p-value Apr-02 0.103 0.097 Apr-03 -0.020 0.751 May-02 0.083 0.180 May-03 0.054 0.381 Jun-02 0.012 0.851 Jun-03 -0.025 0.687 Jul-02 0.043 0.493 Jul-03 0.009 0.888 Aug-02 0.029 0.637 Aug-03 0.021 0.730 Sep-02 0.037 0.548 Sep-03 0.038 0.545 Oct-02 0.062 0.318 Oct-03 -0.017 0.783 Nov-02 0.027 0.662 Nov-03 0.021 0.734 Dec-02 0.020 0.745 Dec-03 0.052 0.402 Jan-03 0.044 0.483 Jan-04 0.005 0.937 Feb-03 0.031 0.614 Feb-04 0.012 0.846 Mar-03 -0.004 0.949 Mar-04 0.074 0.235

Table 6: OLS Regressions of Physician prescriptions on Opinion Leaders prescriptions1

Param Constant Drug 2 Detailing Drug 1 Detailing (aggregate) OPL_Nrx Z-it I(t>=May 03) OPL_Nrx* I(t>=May 03) F R2 4.079 0.825 -0.012 0.071 t-stat 8.03 21.80 -1.38 6.77 Param 4.039 0.826 -0.012 0.070 0.059 t-stat 7.93 21.82 -1.39 6.59 1.09 Param 4.173 0.821 -0.017 0.048 0.050 0.298 0.039 t-stat 8.19 21.68 -1.91 0.88 3.36 2.63 1.86 90.76 0.0726

174.91 0.0701

131.48 0.0703

Table 7: Fixed-Effect Regressions of Physician prescriptions on Opinion Leaders prescriptions1

Param 4.243 0.014 -0.002 0.002 t-stat 12.61 0.40 -0.29 0.21 Param 4.104 0.012 -0.002 0.002 0.190 t-stat 12.07 0.35 -0.30 0.20 2.73 Param 4.297 0.001 -0.007 -0.022 0.138 0.369 0.043 t-stat 12.65 0.02 -1.27 -1.79 1.99 4.92 3.09 47.4700 0.6087

Constant Drug 2 Detailing Drug 1 Detailing (aggregate) OPL_Nrx Z-it I(t>=May 03) OPL_Nrx* I(t>=May 03) F R2
1Fixed

46.9500 0.6050

47.020 0.6054

effects for nominating physicians included, but not reported. Nobs =6960.

32

Table 8: First stage regressions of OPL prescriptions on instruments

On only excluded instruments Param t-stat Drug 2 detailing to nominating physician Aggregate Drug 1 detailing (1000-s)
I(t>May 03) Z-it

On first-stage instrument matrix1 Param t-stat -0.041 -1.02 -0.014 -2.11 0.002 0.03 -0.057 -0.72 1.296 14.37 -0.110 -6.30 1.765 4.04 -0.243 -1.06 -0.816 -1.00 47.97 0.5904

Drug 2 detailing to OPL (Drug 2 detailing to OPL)2

Z-OPL,t (Z-OPL,t)2

Constant F R2 N 6960 1Fixed effects for nominating physicians included, but not reported. Nobs =6960

2.311 -0.167 2.173 -0.993 0.724

20.89 -7.29 6.72 -5.40 7.42 343.88 0.1651

Table 9: Fixed-Effect Instrumental Variable Regressions of Physician prescriptions on Opinion Leaders prescriptions1
Param 0.078 0.011 -0.001 t-stat 1.95 0.32 -0.18 Param 0.073 0.010 -0.001 0.188 t-stat 1.83 0.27 -0.19 2.70 Param 0.000 -0.007 0.011 0.139 0.386 0.032 t-stat 0.01 -1.21 0.28 1.99 3.80 2.06 9.476 0.1485 (6)

Drug 2 Detailing Drug 1 Detailing (aggregate) OPL_Nrx Z-it I(t>=May 03) OPL_Nrx* I(t>=May 03) J-stat (Sargen) Chi2 pvalue (degrees of freedom)
1Fixed

10.572 0.0143 (3)

10.354 0.0158 (3)

effects for nominating physicians included, but not reported. Nobs =6960.

Table 10: Regressions of Physician prescriptions on Opinion Leaders prescriptions with full set of physician and time-period fixed effects1
Fixed effects Param Constant Drug 2 Detailing Drug 1 Detailing (aggregate) OPL_Nrx Z-it OPL_Nrx* I(t>=May 03) Time-period fixed effects F R2 J-stat (Sargen) Chi2 pvalue (degrees of freedom)
1Nobs

Fixed-Effects IV t-stat 2.10 -0.27 3.22 -1.82 1.71 3.01 Y 47.730 0.6127 7.299 0.2941 (6) 7.302 0.2938 (6) 0.031 Y 2.06 -0.009 0.139 0.001 -0.25 3.95 0.02 -0.010 0.137 0.000 0.120 0.031 Y -0.26 3.88 -0.01 1.72 2.05 Param t-stat Param t-stat

t-stat 2.17 -0.25 3.26 -1.83 3.02 Y

Param 1.627 -0.010 0.042 -0.023 0.119 0.042

1.678 -0.009 0.042 -0.023 0.042 47.710 0.6125

= 6960.

33

Table 11: Fixed effects instrumental variable regressions of OPL prescriptions

Param 0.001 t-stat 0.02 Param -0.037 0.044 0.136 -0.009 0.752 N 6.81 0.0782 (3) 4.35 -2.06 3.22 0.115 -0.012 t-stat -0.48 0.90 3.61 -2.70 Param -0.058 0.063 0.110 0.000 0.480 Y 4.85 0.3027 (4) t-stat -0.76 1.31 3.41 0.01 2.03 Y 2.14 0.3437 (2) Param -0.041 0.058 0.115 0.008 t-stat -0.46 1.13 3.59 0.21

Nominator_NRx Nominator_NRx*I(t >=May 03) Drug 2 detailing Drug 1 detailing (aggregate: 1000-s) Z-it I(t >=May 03) All month fixed effects? J-stat (Sargen) Chi2 pvalue (degrees of freedom)

0.538 2.28 0.066 0.34 N 5.23 0.2645 (4)

Notes: 1New treatment guidelines issued in May 03. Fixed effects for each opinion leader estimated, but not reported. Robust t-stats reported.

Table 12: Robustness checks: Fixed effects IV regressions of Physician prescriptions on functions of OPL prescriptions
[1] [2] [3] [4] [5]
t 1 x j(i ), =1

x j(i ),t 1
f(OPL_NRx):

OPL prescriptions last month

Param t-stat -0.56 3.37 -0.43 3.04 1.57

I x j(i ),t

I x j(i ),t 1

t =1 x j(i ),

Did OPL prescribe?

Param -0.018 0.347 0.582 0.303 0.104 t-stat -0.47 2.20 0.72 3.64 1.45

Did OPL prescribe last month?

Param -0.023 0.315 1.339 0.802 0.133 t-stat -0.61 1.80 1.33 3.30 1.77

How much has OPL prescribed as of this month?

Param -0.019 0.526 -0.007 0.005 0.126 t-stat -0.50 3.95 -1.91 1.94 1.75

How much has OPL prescribed as of last month?

Param -0.019 0.525 -0.007 0.005 0.126 t-stat -0.51 3.95 -1.19 1.93 1.75

Drug 2 detailing Drug 1 detailing (aggregate: 1000-s) f(OPL_NRx) f(OPL_NRx)*I(t >=May Z-it Month fixed effects included? Instruments2
Chi2 pvalue (degrees of freedom)

-0.021 0.412 -0.018 03)1 0.036 0.112

w j(i ),t1
0.262 (4)

w j(i ),t
0.316 (4)

w j(i ),t1
0.1197 (4)

=1 w j(i ),
0.7642 (4)

=1 w j(i ), 1
0.7523 (4)

Notes: Each column presents results from fixed-effects IV regressions of a physicians prescriptions on a function of his opinion leaders prescriptions. For example, column [1] presents results from fixed-effects IV regressions of a physicians prescriptions on a lagged values of of his opinion leaders prescriptions; the effect of lagged opinion-leader prescriptions pre-guidelines is -0.018, and postguidelines is -0.018+0.036 = 0.018. Dependant variable in all IV regressions is Physician NRx = yit. Physician and month fixed-effects estimated in all specifications, but not reported. 1Guidelines released in May 2003. 2w denotes the vector of available instrumental variables: wit = x j( i ),t , x j ( i ),t , z j( i ),t , z j ( i ),t .

34

Table 13: Responsiveness of OPL prescriptions to detailing

Variable Constant Drug 2 Detailing Drug 1 detailing (aggregate) Z-OPL,t F R2 N OLS Param 2.167 1.614 -0.018 0.679 426.33 0.155 6670 t-stat 4.05 34.09 -1.98 4.75 Fixed Effects1 Param 2.358 0.136 -0.009 0.753 12.79 0.826 t-stat 9.02 4.35 -2.06 3.44

Notes: Fixed effects for each OPL estimated, but not reported. Robust t-stats reported.

Figure 1: Aggregate detailing: Combination drugs

80 1000-s of details 70
65 61 60 58 54 50 53 59 54 51 65 68 65 60 58 53 57 64 62

Drug 1 details Drug 2 details

60
55 50

57

57

50

43

40

30
26 24 24 20 21 19 17 15 12 10 13 12 19 24 20 17 15 13 16 15 15 16 25

20

19

10

0
Apr- May- Jun- Jul- Aug- Sep- Oct- Nov- Dec- Jan- Feb- Mar- Apr- May- Jun- Jul- Aug- Sep- Oct- Nov- Dec- Jan- Feb- Mar02 02 02 02 02 02 02 02 02 03 03 03 03 03 03 03 03 03 03 03 03 04 04 04

Month

35

Figure 2: Distribution of Category prescriptions before and after issuance of new NIH guidelines

Physicians
5 2.8 2.7 2.6

OPL-s

Mean NRx per month

Mean NRx per month

1 2

4.5

2.5

2.4 2.3 2.2 2.1

2 1.9 3.5 1 2

Before

After

Before

After

Figure 3: Distribution of detailing before and after issuance of new NIH guidelines

Physicians
1 1.4 0.9

OPL-s

1.2

0.8

Mean details per month

Mean details per month

1 2

0.7

0.6

0.8

0.5

0.6

0.4

0.3 0.4 0.2 0.2 1

Before

After

Before

After

36

Figure 4: Variation in mean prescriptions of other physicians in the physicians zip-code over time, controlling for across physician variation

0.3

Mean NRx of other physicians in zip-code

Z-i Z-OPL 0.25

0.2

0.15

0.1

0.05

-0.05

10

15

20

25

Month (1 = May 2002)

Figure 5: Distribution of social-multipliers in data
0.7

0.6

0.5

Proportion

0.4

0.3

0.2

0.1

1.05

1.1

1.15

1.2

1.25

1.3

1.35

1.4

Social Multipliers

37