Estimating

Production
Functions

Paul Schrimpf

Estimating Production Functions

Paul Schrimpf

UBC
Economics 565

January 12, 2017

 Estimating
Production
Functions

Paul Schrimpf
1 Introduction

Introduction
2 Setup
Setup

Simultaneity
Instrumental 3 Simultaneity
variables
Panel data Instrumental variables
Fixed effects
Dynamic Panel data
Control functions
Critiques and
Fixed effects
extensions
Dynamic
Selection
OP and selection
Control functions
Applications
Critiques and extensions

4 Selection
OP and selection

5 Applications
 Estimating
Production
Functions

Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables
Panel data
Section 1
Fixed effects
Dynamic
Control functions
Critiques and
extensions
Introduction
Selection
OP and selection

Applications
 Estimating
Production
Functions
Why estimate production
Paul Schrimpf
functions?
Introduction

Setup

Simultaneity
Instrumental
variables • Primitive component of economic model
Panel data
Fixed effects • Gives estimate of firm productivity — useful for
Dynamic
Control functions
understanding economic growth
Critiques and
extensions • Stylized facts to inform theory, e.g. Foster, Haltiwanger,
Selection and Krizan (2001)
OP and selection
• Effect of deregulation, e.g. Olley and Pakes (1996)
Applications
• Growth within old firms vs from entry of new firms, e.g.
Foster, Haltiwanger, and Krizan (2006)
• Effect of trade liberalization, e.g. Amiti and Konings
(2007)
• Effect of FDI Javorcik (2004)
 Estimating
Production
Functions
These slides based on:
Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables
Panel data
Fixed effects
Dynamic
Control functions
• Aguirregabiria (2017) chapter 2
Critiques and
extensions • Ackerberg et al. (2007) section 2
Selection
OP and selection • Van Beveren (2012)
Applications
 Estimating
Production
Functions

Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables
Panel data
Section 2
Fixed effects
Dynamic
Control functions
Critiques and
extensions
Setup
Selection
OP and selection

Applications
 Estimating
Production
Functions
Setup
Paul Schrimpf

• Cobb Douglas production

Introduction

Setup β β
Simultaneity
Yit = Ait Kitk Litl
Instrumental
variables
Panel data • In logs,
Fixed effects
Dynamic yit = βk kit + βl lit + ωit + εit
Control functions
Critiques and
extensions with log Ait = ωit + εit , ωit known to firm, εit not
Selection
OP and selection
• Problems:
Applications 1 Simultaneity: if firm has information about log Ait when
choosing inputs, then inputs correlated with log Ait , e.g.
price p, wage w, perfect information
(p ) 1−β1
Lit = βl Ait Kitβk l

w
2 Selection: firms with low productivity will exit sooner
3 Others: measurement error, specification
 Estimating
Production
Functions

Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables Section 3
Panel data
Fixed effects
Dynamic

Simultaneity
Control functions
Critiques and
extensions

Selection
OP and selection

Applications
 Estimating
Production
Functions
Simultaneity solutions
Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables
Panel data
Fixed effects
Dynamic
Control functions 1 IV
Critiques and
extensions
2 Panel data
Selection
OP and selection 3 Control functions
Applications
 Estimating
Production
Functions
Instrumental variables
Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
• Instrument must be
variables
Panel data • Correlated with k and l
Fixed effects
Dynamic
• Uncorrelated with ω + ε
Control functions
Critiques and • Possible instrument: input prices
extensions

Selection
• Correlated with k, l through first-order condition
OP and selection • Uncorrelated with ω if input market competitive
Applications
• Other possible instruments: output prices (more often
endogenous), input supply or output demand shifter
(hard to find)
 Estimating
Production
Functions
Problems with input prices as IV
Paul Schrimpf

Introduction

Setup
• Not available in some data sets
Simultaneity • Average input price of firm could reflect quality as well
Instrumental
variables
Panel data
as price differences
Fixed effects
Dynamic
• Need variation across observations
Control functions • If firms use homogeneous inputs, and operate in the
Critiques and
extensions
same output and input markets, we should not expect
Selection to find any significant cross-sectional variation in input
OP and selection
prices
Applications
• If firms have different input markets, maybe variation
in input prices, but different prices could be due to
different average productivity across input markets
• Variation across time is potentially endogenous because
could be driven by time series variation in average
productivity
 Estimating
Production
Functions
Fixed effects
Paul Schrimpf

Introduction

Setup • Have panel data, so should consider fixed effects

Simultaneity • FE consistent if:
Instrumental
∗
variables
1 ωit = ηi + δt + ωit
Panel data
∗ ∗
Fixed effects 2 ωit uncorrelated with lit and kit , e.g. ωit only known to
Dynamic
Control functions firm after choosing inputs
∗
Critiques and
extensions 3 ωit not serially correlated and is strictly exogenous
Selection
OP and selection
• Problems:
Applications • Fixed productivity a strong assumption
• Estimates often small in practice
• Worsens measurement error problems

βk Var(∆ε)
Bias(β̂kFE ) ≈ −
Var(∆k) + Var(∆ε)
 Estimating
Production
Functions
Dynamic panel: motivation I
Paul Schrimpf

Introduction

Setup

Simultaneity • General idea: relax fixed effects assumption, but still

Instrumental
variables exploit panel
Panel data
Fixed effects
Dynamic
• Collinearity problem: Cobb-Douglas production, flexible
Control functions
Critiques and
labor and capital implies log labor and log capital are
extensions
linear functions of prices and productivity (Bond and
Selection
OP and selection Söderbom (2005))
Applications
• If observed labor and capital are not collinear then
there must be something unobserved that varies across
firms (e.g. prices), but that would invalidate
monotonicity assumption of control function
 Estimating
Production
Functions
Dynamic panel: moment
Paul Schrimpf
conditions
Introduction

Setup
• See Blundell and Bond (2000)
Simultaneity
Instrumental
variables
• Assume ωit = γt + ηi + νit with νit = ρνi,t−1 + eit , so
Panel data
Fixed effects
Dynamic yit = βl lit + βk kit + γt + ηi + νit + εit
Control functions
Critiques and
extensions
subtract ρyi,t−1 and rearrange to get
Selection
OP and selection

Applications yit =ρyi,t−1 + βl (lit − ρli,t−1 ) + βk (kit − ρki,t−1 )+

+ γt − ργt−1 + ηi (1 − ρ) + eit + εit − ρεi,t−1
| {z } | {z }
=ηi∗ =wit

• Moment conditions:
• Difference: E[xi,t−s ∆wit ] = 0 where x = (l, k, y)
• Level: E [∆xi,t−s (ηi∗ + wit )] = 0
 Estimating
Production
Functions
Dynamic panel: economic
Paul Schrimpf
model I
Introduction

Setup
• Adjustment costs
Simultaneity
Instrumental V(Kt−1 , Lt−1 ) = max Pt Ft (Kt , Lt ) − PKt (It + Gt (It , Kt−1 )) −
variables
It ,Kt ,Ht ,Lt
Panel data
Fixed effects
Dynamic
− Wt (Lt + Ct (Ht , Lt−1 )) +
Control functions
Critiques and
extensions
ψE [V(Kt , Lt )|It ]
Selection s.t. Kt = (1 − δk )Kt−1 + It
OP and selection

Applications
Lt = (1 − δl )Lt−1 + Ht

Implies
( [ L ])
∂Ft ∂Ct λ
Pt − Wt =Wt + λLt 1 − (1 − δl )ψE t+1 |I t
∂Lt ∂Lt λLt
( [ K ])
∂Ft ∂Gt λ
Pt − PKt =λKt 1 − (1 − δk )ψE t+1 |It
∂Kt ∂Kt λKt
 Estimating
Production
Functions
Dynamic panel: economic
Paul Schrimpf
model II
Introduction

Setup

Simultaneity
Instrumental
variables
Panel data
Fixed effects
∂Ft
Dynamic
• Current productivity shifts and (if correlated with
∂Lt
Control functions
[ L ]
Critiques and λ
extensions future) the shadow value of future labor E λt+1L |It
t
Selection
OP and selection
• Past labor correlated with current because of
Applications adjustment costs
 Estimating
Production
Functions
Dynamic panel data: problems
Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables
Panel data • Problems:
Fixed effects
Dynamic • Sometimes imprecise (especially if only use difference
Control functions
Critiques and moment conditions)
extensions
• Differencing worsens measurement error
Selection
OP and selection
• Weak instrument issues if only use difference moment
Applications conditions but levels stronger (see Blundell and Bond
(2000))
 Estimating
Production
Functions
Control functions
Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables • From Olley and Pakes (1996) (OP)
Panel data
Fixed effects • Control function: function of data conditional on
Dynamic
Control functions which endogeneity problem solved
Critiques and
extensions • E.g. usual 2SLS y = xβ + ε, x = zπ + v, control function
Selection is to estimate residual of reduced form, v̂ and then
OP and selection
regress y on x and v̂. v̂ is the control function
Applications

• Main idea: model choice of inputs to find a control

function
 Estimating
Production
Functions
OP assumptions
Paul Schrimpf

Introduction yit = βk kit + βl lit + ωit + εit

Setup

Simultaneity
Instrumental
variables 1 ωit follows exogenous first order Markov process,
Panel data
Fixed effects
Dynamic p(ωit+1 |Iit ) = p(ωit+1 |ωit )
Control functions
Critiques and
extensions
2 Capital at t determined by investment at time t − 1,
Selection
OP and selection

Applications kt = (1 − δ)kit−1 + iit−1

3 Investment is a function of ω and other observed

variables
iit = It (kit , ωit ),
and is strictly increasing in ωit
4 Labor variable and non-dynamic, i.e. chosen each t,
current choice has no effect on future (can be relaxed)
 Estimating
Production
Functions
OP estimation of βl
Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables
• Invertible investment implies ωit = I−1
t (kit , iit )
Panel data
Fixed effects
Dynamic yit =βk kit + βl lit + I−1
t (kit , Iit ) + εit
Control functions
Critiques and
extensions
=βl lit + ft (kit , iit ) + εit
Selection
OP and selection
• Partially linear model
Applications
• Estimate by e.g. regress yit on lit and series functions of
t, kit , iit
• Gives β̂l , f̂it = f̂t (kit , iit )
 Estimating
Production
Functions
OP estimation of βk
Paul Schrimpf

Introduction
• Note: f̂t (kit , iit ) = ω̂it + βk kit
Setup

Simultaneity
• By assumptions, ωit = E[ωit |ωit−1 ] + ξit = g(ωit−1 ) + ξit
Instrumental
variables
with E[ξit |kit ] = 0
Panel data
Fixed effects
• Use E[ξit |kit ] = 0 as moment to estimate βk .
Dynamic
Control functions
• OP: write production function as
Critiques and
extensions

Selection
yit − βl lit =βk kit + g(ωit−1 ) + ξit + εit
OP and selection
=βk kit + g (fit−1 − βk kit−1 ) +
Applications
+ ξit + εit

Use β̂l and f̂it in equation above and estimate β̂k by e.g.
semi-parametric nonlinear least squares
• Ackerberg,
[ ]Caves, and Frazer (2015): use
E ξ̂it (βk )kit = 0
 Estimating
Production
Functions
Critiques and extensions
Paul Schrimpf

Introduction
• Levinsohn and Petrin (2003): investment often zero, so
Setup

Simultaneity
use other inputs instead of investment to form control
Instrumental
variables
function
Panel data
Fixed effects • Ackerberg, Caves, and Frazer (2015): control function
Dynamic
Control functions often collinear with lit — for it not to be must be firm
Critiques and
extensions specific unobervables affecting lit (but not investment /
Selection other input or else demand not invertible and cannot
OP and selection

Applications
form control function)
• Gandhi, Navarro, and Rivers (2013): relax scalar
unobservable in investment / other input demand
• Wooldridge (2009): more efficient joint estimation
• Maican (2006) and Doraszelski and Jaumandreu (2013):
endogenous productivity
 Estimating
Production
Functions

Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables
Panel data
Section 4
Fixed effects
Dynamic
Control functions
Critiques and
extensions
Selection
Selection
OP and selection

Applications
 Estimating
Production
Functions
Selection
Paul Schrimpf

Introduction

Setup

Simultaneity • Let dit = 1 if firm in sample.

Instrumental
variables • Standard conditions imply d = 1{ω ≥ ω∗ (k)}
Panel data
Fixed effects • Messes up moment conditions
Dynamic
Control functions • All estimators based on E[ωit Something] = 0, observed
Critiques and
extensions data really use E[ωit Something|dit = 1]
Selection • E.g. OLS okay if E[ωit |lit , kit ] = 0, but even then,
OP and selection

Applications
E[ωit |lit , kit , dit = 1] =E[ωit |lit , kit , ωit ≥ ω∗ (kit )]
=λ(kit ) ̸= 0

• Selection bias negative, larger for capital than labor

 Estimating
Production
Functions
Selection in OP model
Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental • Estimate βl as above
variables
Panel data
Fixed effects
• Write
Dynamic
Control functions
dit = 1{ξit ≤ ω∗ (kit )−ρ(fi,t−1 −βk kit−1 ) = h(kit , fit−1 , kit−1 )}
Critiques and
extensions • Propensity score Pit ≡ E[dit |kit , fit−1 , kit−1 ]
Selection
OP and selection
• Similar to before estimate βk , from
Applications
yit − βl lit =βk kit + g̃ (fit−1 − βk kit−1 , Pit ) +
+ ξit + εit
 Estimating
Production
Functions

Paul Schrimpf

Introduction

Setup

Simultaneity
Instrumental
variables Section 5
Panel data
Fixed effects
Dynamic

Applications
Control functions
Critiques and
extensions

Selection
OP and selection

Applications
 Estimating
Production
Functions
Applications
Paul Schrimpf

Introduction
• Olley and Pakes (1996): productivity in telecom after
Setup
deregulation
Simultaneity
Instrumental
variables
• Söderbom, Teal, and Harding (2006): productivity and
Panel data
Fixed effects
exit of African manufacturing firms, uses IV
Dynamic
Control functions • Levinsohn and Petrin (2003): compare estimation
Critiques and
extensions methods using Chilean data
Selection
OP and selection
• Javorcik (2004): FDI and productivity, uses OP
Applications • Amiti and Konings (2007): trade liberalization in
Indonesia, uses OP
• Aw, Chen, and Roberts (2001): productivity differentials
and firm turnover in Taiwan
• Kortum and Lerner (2000): venture capital and
innovation
 Estimating
Production
Functions

Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Part II
Relation to dynamic
panel
Empirical example

Gandhi,
Navarro, and
Rivers (2013)
Selected applications and
Identification
problem
Identification from
extensions
first order conditions
Value added vs gross
production
Empirical results

Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
6 Ackerberg, Caves, and Frazer (2015)
Paul Schrimpf
Collinearity in OP
Ackerberg, ACF estimator
Caves, and
Frazer (2015) Relation to dynamic panel
Collinearity in OP
ACF estimator Empirical example
Relation to dynamic
panel
Empirical example
7 Gandhi, Navarro, and Rivers (2013)
Gandhi,
Navarro, and Identification problem
Rivers (2013)
Identification
Identification from first order conditions
problem
Identification from Value added vs gross production
first order conditions
Value added vs gross
production
Empirical results
Empirical results

Grieco and 8 Grieco and McDevitt (2017)

McDevitt
(2017)

Amiti and 9 Amiti and Konings (2007)

Konings (2007)

Doraszelski
and 10 Doraszelski and Jaumandreu (2013)
Jaumandreu
(2013)

References
 Estimating
Production
Functions

Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Section 6
Empirical example

Gandhi,
Navarro, and
Rivers (2013) Ackerberg, Caves, and Frazer (2015)
Identification
problem
Identification from
first order conditions
Value added vs gross
production
Empirical results

Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Ackerberg, Caves, and Frazer
Paul Schrimpf
(2015): contributions
Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
• Document collinearity problem in OP and Levinsohn
Empirical example and Petrin (2003)
Gandhi,
Navarro, and
• Need lit , fit (kit , iit ) not collinear, i.e. something causes
Rivers (2013) variation in l, but not k
Identification
problem
Identification from
first order conditions
• Propose alternative estimator
Value added vs gross
production • Relates estimator to dynamic panel (Blundell and Bond,
Empirical results

Grieco and
2000) approach
McDevitt
(2017) • Illustrates estimator using Chilean data
Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013) 0∗
These slides are based on the working paper version Ackerberg,
References Caves, and Frazer (2006).
 Estimating
Production
Functions
Collinearity in OP I
Paul Schrimpf

Ackerberg,
Caves, and • OP assume iit = It (kit , ωit )
Frazer (2015)
Collinearity in OP • Symmetry, parsimony suggest lit = Lt (kit , ωit )
ACF estimator
Relation to dynamic
panel
• Then lit = Lt (kit , I−1
t (kit , iit )) = gt (kit , iit )
Empirical example

Gandhi,
Navarro, and
yit = βl lit + ft (kit , iit ) + εit
Rivers (2013)
Identification
problem
Identification from
lit collinear with ft (kit , iit )
first order conditions
Value added vs gross
• Worse in Levinsohn and Petrin (2003)
production
Empirical results • Uses other input mit to form control function
Grieco and
McDevitt
(2017)
yit =βl lit + βk kit + βm mit + ωit + εit
Amiti and mit =Mt (kit , ωit )
Konings (2007)

Doraszelski • Even less reason to treat labor demand differently than

and
Jaumandreu other input demand
(2013)

References
 Estimating
Production
Functions
Collinearity in OP II
Paul Schrimpf • Collinearity still problem with parametric input
Ackerberg,
demand
Caves, and
Frazer (2015) • Plausible models that do not solve collinearity
Collinearity in OP
ACF estimator
• Input price data
Relation to dynamic
panel
• Must include in control function to preserve scalar
Empirical example unobservable
Gandhi, • Same logic above implies m and l are functions of both
Navarro, and
Rivers (2013)
prices, so still collinear
Identification
problem
• Adjustmest costs in labor
Identification from
first order conditions
• Need to add lit−1 to control function
Value added vs gross
production • Change in timing assumptions
Empirical results
• Measurement error in l (but not m)
Grieco and
McDevitt • Solves collinearity, but makes β̂l inconsistent
(2017)

Amiti and
• Potential model change that removes collinearity
Konings (2007)
• Optimization error in l (but not m)
Doraszelski
and
• m chosen, l specific shock revealed, l chosen
Jaumandreu • OP only: lit chosen at t − 1/2, lit = Lt (ωit−1/2 , kit ), iit
(2013)
chosen at t
References
 Estimating
Production
Functions
ACF estimator
Paul Schrimpf
• Idea: like capital, labor is harder to adjust than other
Ackerberg, inputs
Caves, and
Frazer (2015) • Model: lit chosen at time t − 1/2, mit at time t
Collinearity in OP
• Implies mt = Mt (kit , lit , ωit )
ACF estimator
Relation to dynamic
panel
• Estimation:
Empirical example 1 yit = βk kit + βl lit + ft (mit , kit , lit ) +εit gives
| {z }
Gandhi,
≡Φt (mit ,kit ,lit )
Navarro, and
Rivers (2013)
Identification
problem
ω̂it (βk , βl ) = Φ̂it − βk kit − βl lit
Identification from
first order conditions
Value added vs gross
2 Moments from timing and Markov process for ωit
production
Empirical results
assumptions:
Grieco and
McDevitt
ωit = E[ωit |ωit−1 ] + ξit
(2017)

Amiti and
Konings (2007) • E[ξit |kit ] = 0 as in OP
• E[ξit |lit−1 ] = 0 from new timing assumption
Doraszelski
and • ξ̂it (βk , βl ) as residual from nonparametric regression of
Jaumandreu ω̂it on ω̂it−1
(2013)
• Can add moments based on E[εit |Iit ] = 0
References
 Estimating
Production
Functions
Relation to dynamic panel
Paul Schrimpf
estimators
Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
• Both derive moment conditions from assumptions
Empirical example
about timing and information set of firm
Gandhi,
Navarro, and • Dealing with ω
Rivers (2013)
Identification
• Dynamic panel: AR(1) assumption allows
problem
Identification from
quasi-differencing
first order conditions
Value added vs gross
• Control function: makes ω estimable function of
production
Empirical results
observables
Grieco and
McDevitt
• Dynamic panel allows fixed effects, does not make
(2017)
assumptions about input demand
Amiti and
Konings (2007) • Control function allows more flexible process for ωit
Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Empirical example
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
• Chilean plant level data
ACF estimator
Relation to dynamic • Compare OLS, FE, LP, ACF, and dynamic panel estimators
panel
Empirical example
• LP and ACF using three different inputs (materials,
Gandhi,
Navarro, and electricity, fuel) for control function
Rivers (2013)
Identification
• Results:
problem
Identification from
first order conditions
• 311=food, 321=textiles, 331=wood, 381=metal
Value added vs gross
production
• Expected biases in OLS and FE
Empirical results • ACF and LP significantly different
Grieco and • ACF less sensitive to which input used for control
McDevitt
(2017) function
Amiti and • Dynamic panel closer to ACF than LP, but still significant
Konings (2007) differences
Doraszelski
and
Jaumandreu
(2013)

References
 TABLE 1
Industry 311
Capital Labor Returns to Scale
Estimate SE Estimate SE Estimate SE
OLS 0.336 0.025 1.080 0.042 1.416 0.026
FE 0.081 0.038 0.719 0.055 0.800 0.066
ACF – M 0.371 0.037 0.842 0.048 1.212 0.034
ACF – E 0.379 0.031 0.865 0.047 1.244 0.032
ACF – F 0.395 0.033 0.884 0.046 1.279 0.028
LP – M 0.455 0.038 0.676 0.037 1.131 0.035
LP – E 0.446 0.032 0.764 0.040 1.210 0.034
LP – F 0.410 0.032 0.942 0.040 1.352 0.036
DP 0.391 0.026 0.987 0.043 1.378 0.028

Industry 321
Capital Labor Returns to Scale
Estimate SE Estimate SE Estimate SE
OLS 0.256 0.035 0.953 0.056 1.210 0.034
FE 0.204 0.068 0.724 0.087 0.927 0.108
ACF – M 0.242 0.041 0.893 0.063 1.135 0.040
ACF – E 0.272 0.037 0.832 0.060 1.104 0.039
ACF – F 0.272 0.038 0.873 0.061 1.145 0.040
LP – M 0.320 0.037 0.775 0.059 1.094 0.049
LP – E 0.241 0.037 0.978 0.065 1.219 0.047
LP – F 0.254 0.039 1.008 0.062 1.262 0.048
DP 0.320 0.042 0.837 0.064 1.157 0.041

Industry 331
Capital Labor Returns to Scale
Estimate SE Estimate SE Estimate SE
OLS 0.236 0.047 1.038 0.074 1.274 0.052
FE -0.028 0.103 0.897 0.095 0.869 0.136
ACF – M 0.196 0.064 0.923 0.085 1.119 0.076
ACF – E 0.195 0.065 0.897 0.088 1.092 0.073
ACF – F 0.212 0.062 0.915 0.086 1.127 0.075
LP – M 0.352 0.056 0.678 0.077 1.030 0.072
LP – E 0.305 0.059 0.786 0.086 1.090 0.075
LP – F 0.241 0.052 0.993 0.079 1.234 0.071
DP 0.252 0.054 0.998 0.073 1.249 0.061
 TABLE 2
Industry 311 Industry 321
M E F M E F
ACF vs OLS ACF vs OLS
K 0.111 0.040 0.010 K 0.585 0.192 0.192
L 1.000 1.000 1.000 L 0.970 1.000 0.996
RTS 1.000 1.000 1.000 RTS 0.998 1.000 0.998

ACF vs LP ACF vs LP
K 1.000 1.000 0.707 K 0.982 0.052 0.070
L 0.000 0.000 0.899 L 0.048 0.998 1.000
RTS 0.000 0.061 0.990 RTS 0.198 1.000 1.000

ACF vs DP ACF vs DP
K 0.737 0.788 0.505 K 1.000 0.992 0.992
L 1.000 1.000 1.000 L 0.052 0.511 0.084
RTS 1.000 1.000 1.000 RTS 0.820 0.996 0.669

Industry 331 Industry 381

M E F M E F
ACF vs OLS ACF vs OLS
K 0.892 0.840 0.830 K 0.060 0.058 0.054
L 0.974 0.990 0.984 L 1.000 1.000 1.000
RTS 1.000 1.000 1.000 RTS 1.000 1.000 1.000

ACF vs LP LP vs ACF
K 1.000 1.000 0.860 K 0.996 0.980 0.683
L 0.000 0.024 0.876 L 0.000 0.072 0.910
RTS 0.056 0.431 0.984 RTS 0.002 0.323 0.984

ACF vs DP ACF vs DP
K 0.962 0.922 0.884 K 0.834 0.916 0.892
L 0.940 0.986 0.962 L 0.852 0.844 0.649
RTS 1.000 1.000 0.998 RTS 0.984 0.992 0.934

Note: Value is the % of bootstrap reps where ACF coeff is less than OLS, LP, or DP coef. A value
either above 0.95 or below 0.05 indicates that coefficients are significantly different from each other.
 Estimating
Production
Functions

Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Section 7
Empirical example

Gandhi,
Navarro, and
Rivers (2013) Gandhi, Navarro, and Rivers (2013)
Identification
problem
Identification from
first order conditions
Value added vs gross
production
Empirical results

Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Gandhi, Navarro, and Rivers
Paul Schrimpf
(2013)
Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Empirical example
• Show that control function method is not
Gandhi, nonparametrically identified when there are flexible
Navarro, and
Rivers (2013) inputs
Identification
problem
Identification from
• Propose alternate estimate that uses data on input
first order conditions
Value added vs gross
shares and information from firm’s first order condtiion
production
Empirical results • Show that value-added and gross output production
Grieco and
McDevitt
functions are incompatible
(2017)
• Application to Colombia and Chile
Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Assumptions
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel 1 Hicks neutral productivity Yjt = eωjt +εjt Ft (Ljt , Kjt , Mjt )
Empirical example

Gandhi, 2 ωjt Markov, εjt i.i.d.

Navarro, and
Rivers (2013) 3 Kjt and Ljt determined at t − 1, Mjt determined flexibly at
Identification
problem t
Identification from
first order conditions • K and L play same role in the model, so after this slide I
Value added vs gross
production will drop L
Empirical results

Grieco and 4 Mjt = Mt (Ljt , Kjt , ωjt ), monotone in ωjt

McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Reduced form
Paul Schrimpf
• Let h(ωjt−1 ) = E[ωjt |ωjt−1 ], ηjt = ωjt − h(ωjt−1 )
Ackerberg,
Caves, and • log output
Frazer (2015)
Collinearity in OP
ACF estimator yjt =ft (kjt , mjt ) + ωjt + εjt
Relation to dynamic
panel
=ft (kjt , mjt ) + h(M−1 (kjt−1 , mjt−1 )) +ηjt + εjt
Empirical example
| t−1 {z }
Gandhi,
Navarro, and =ht−1 (kjt−1 ,mjt−1 )
Rivers (2013)
Identification
problem • Assumptions imply
Identification from
first order conditions
Value added vs gross E[ηjt | kjt , kjt−1 , mjt−1 , ...kj1 , mj1 ] = 0
production
Empirical results
| {z }
=Γjt
Grieco and
McDevitt
(2017) • Reduced form
Amiti and
Konings (2007)
E[yjt |Γjt ] =E[ft (kjt , mjt )|Γjt ] + ht−1 (kjt−1 , mjt−1 ) (1)
Doraszelski
and
Jaumandreu • Identification: given observed E[yjt |Γjt ] is there a unique
(2013)

References
ft , ht−1 that satisfies (3)?
 Estimating
Production
Functions
Example: Cobb-Douglas I
Paul Schrimpf

Ackerberg,
• Let ft (k, m) = βk k + βm m
Caves, and
Frazer (2015) • Assume firm is takes prices as given
Collinearity in OP
ACF estimator • First order condition for m gives
Relation to dynamic
panel
Empirical example
βk 1
Gandhi, m = constant + k+ ω
Navarro, and 1 − βm 1 − βm
Rivers (2013)
Identification
problem
Identification from
• Put into reduced form
first order conditions
Value added vs gross
production βk βm
Empirical results E[yjt |Γjt ] =C + kjt + E[ωjt |Γjt ] + ht−1 (kjt−1 , mjt−1
Grieco and
1 − βm 1 − βm
McDevitt (2)
(2017)

Amiti and
Konings (2007) • ω Markov and ωjt−1 = M−1
t−1 (kjt−1 , mjt−1 ) implies
Doraszelski
and
Jaumandreu E[ωjt |Γjt ] =E[ωjt |ωjt−1 = M−1
t−1 (kjt−1 , mjt−1 )] =
(2013)

References
=ht−1 (kjt−1 , mjt−1 )
 Estimating
Production
Functions
Example: Cobb-Douglas II
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015) • Which leaves
Collinearity in OP
ACF estimator
βk 1
Relation to dynamic
panel E[yjt |Γjt ] =constant + kjt + ht−1 (kjt−1 , mjt−1 )
Empirical example 1 − βm 1 − βm
Gandhi, (3)
Navarro, and
Rivers (2013)
Identification
problem from which βk , βm are not identified
Identification from
first order conditions
Value added vs gross
• Rank condition fails, E[mjt |Γjt ] is colinear with
production
Empirical results ht−1 (kjt−1 , mjt−1 )
Grieco and
McDevitt
• After conditioning on kjt , kjt−1 , mjt−1 , only variation in
(2017)
mjt is from ηjt , but this is uncorrelated with the
Amiti and
Konings (2007) instruments
Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Identification from first order
Paul Schrimpf
conditions I
Ackerberg,
Caves, and
Frazer (2015)
• Since m flexible, it satisfies a simple static first order
Collinearity in OP
ACF estimator
condition,
Relation to dynamic
panel
∂Ft
Empirical example
ρt =pt E[eεjt ]eωjt
Gandhi, ∂M
Navarro, and
∂Ft
Rivers (2013)
Identification
log ρt = log pt + log (kjt , mjt ) + log E[eεjt ] + ωjt
problem ∂M
Identification from
first order conditions
Value added vs gross
production
• Problem: prices often unobserved, endogenous ω
Empirical results
• Solution: difference from output equation to eliminate
Grieco and
McDevitt ω, rearrange so that it involves only the value of
(2017)
materials and the value of output (which are often
Amiti and
Konings (2007) observed)
Doraszelski
and
Jaumandreu sjt = log Gt (kjt , mjt ) + log E[eεjt ] −εjt
(2013) |{z} | ( {z ) } | {z }
ρt Mjt E
References ≡log pt Yjt
≡ Mt ∂Ft
∂M /Ft
 Estimating
Production
Functions
Identification from first order
Paul Schrimpf
conditions II
Ackerberg,
Caves, and • Identifies elasticity up to scale, Gt E and εjt which
Frazer (2015)
Collinearity in OP
identifie E
ACF estimator
Relation to dynamic • Integrating,
panel
Empirical example
∫ mjt
Gandhi,
Navarro, and Gt (kjt , m)/m = ft (kjt , mjt ) + ct (kjt )
Rivers (2013) m0
Identification
problem
Identification from
first order conditions identifies f up to location
Value added vs gross
production
Empirical results
• Output equation
Grieco and ∫ mjt
McDevitt
(2017) yjt = G̃t (kjt , m)/m − ct (kjt ) + ωjt + εjt
Amiti and m0
Konings (2007) ∫ mjt
Doraszelski −ct (kjt ) + ωjt = yjt − G̃t (kjt , m)/m − εjt
and
m0
Jaumandreu
(2013)
| {z }
≡Yjt
References
 Estimating
Production
Functions
Identification from first order
Paul Schrimpf
conditions III
Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Empirical example

Gandhi,
where the things on the right have already been
Navarro, and
Rivers (2013)
identified
Identification
problem • Identify ct from
Identification from
first order conditions
Value added vs gross
production
Empirical results
Yjt = − ct (kjt ) + h̃t (Yjt−1 , kjt−1 ) + ηjt
Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Value added vs gross production
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
• Value added:
ACF estimator
Relation to dynamic
panel
Empirical example
VAjt =pt Yjt − ρt Mjt
Gandhi, =pt Ft (Kjt , Mt (Kjt , ωjt ))eωjt +εjt − ρt Mt (Kjt , ωjt )
Navarro, and
Rivers (2013)
Identification
problem
• Envelope theorem implies
ρt Mjt
Identification from
first order conditions elasticityYeω ≈ elasticityVA
eω (1 − pt Yjt )
Value added vs gross
production
Empirical results
Problems
Grieco and
McDevitt • Production Hicks-neutral productivity does not imply
(2017)

Amiti and
value-added Hicks-neutral productivity
Konings (2007)
• Ex-post shocks εjt not accounted for in approximation
Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Empirical results
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Empirical example

Gandhi,
Navarro, and • Look at tables
Rivers (2013)
Identification • Value-added estimates imply much more productivity
problem
Identification from
first order conditions
dispersion than gross (90-10) ratio of 4 vs 2
Value added vs gross
production
Empirical results

Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions

Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Section 8
Empirical example

Gandhi,
Navarro, and
Rivers (2013) Grieco and McDevitt (2017)
Identification
problem
Identification from
first order conditions
Value added vs gross
production
Empirical results

Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Grieco and McDevitt (2017)
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Empirical example

Gandhi, https://www.ftc.gov/sites/default/files/documents/
Navarro, and
Rivers (2013) public_events/
Identification
problem fifth-annual-microeconomics-conference/grieco-p_0.
Identification from
first order conditions
Value added vs gross
pdf
production
Empirical results

Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
My thoughts while reading I
Paul Schrimpf

Ackerberg, • How they got the data

Caves, and
Frazer (2015) • Incentive shifters may be correlated with quality and
Collinearity in OP
ACF estimator productivity
Relation to dynamic
panel • Time since inspection
Empirical example

Gandhi,
+ Inspect recently ⇒ higher quality incentive
Navarro, and - Low productivity ⇒ low quality ⇒ more inspections
Rivers (2013)
Identification • Referral rate
problem
Identification from • Section 5.2: incentive shifters can be correlated with
first order conditions
Value added vs gross productivity, only need that shape of production
production
Empirical results possibilities frontier is invariant
Grieco and
McDevitt • Should hemoglobin level be controlled for when
(2017) measuring quality?
Amiti and
Konings (2007)
• Anemia (low hemoglobin) is risk-factor for infection
Doraszelski
• Anemia can be treated through diet, iron supplements
and (pills or IV), EPO, etc
Jaumandreu
(2013) • Are dialysis facilities responsible for this treatment?
References
 Estimating
Production
Functions
My thoughts while reading II
Paul Schrimpf

Ackerberg,
• In 2006-2014 data average full-time dieticiens = 0.5,
Caves, and average part-time = 0.6
Frazer (2015)
Collinearity in OP
ACF estimator • Estimation details:
Relation to dynamic
panel Step 1: Estimate αq
Empirical example

Gandhi,
Navarro, and yjt Ê[y|hjt , ijt , kjt , ℓjt , xjt ] = αq (qjtÊ[q|hjt , ijt , kjt , ℓjt , xjt ]) + εjt
Rivers (2013)
Identification
problem
Identification from
first order conditions • Drop observations with hjt = 0 (not invertibility)
Value added vs gross
production • Okay here, because selecting on ω, and residual, εjt is
Empirical results
uncorrelated with ω
Grieco and
McDevitt
• Problematic in last step? No, see footnote 49
(2017)
Step 2: Estimate βk , βℓ from
Amiti and
Konings (2007)
yjt + αˆq + βk kjt + βℓ ℓjt = g(ω̂jt−1 (β)) + ηjt + εjt
Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
My thoughts while reading III
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015) • Only have hatωjt−1 (β) when hjt−1 ̸= 0, okay because εjt
Collinearity in OP
and ηjt are uncorrelated with ωjt−1 , would be problem if
ACF estimator
Relation to dynamic using ω̂jt
panel
Empirical example • Nothing about selection — number of centers, 4270, vs
Gandhi,
Navarro, and
center-years, 18295, implies there must be entry and exit
Rivers (2013)
Identification
• Would like to see some results related to productivity
problem
Identification from dispersion e.g.
first order conditions
Value added vs gross • Decompose variation in infection rate into: productivity
production
Empirical results variation, incentive variation, quality-quantity choices,
Grieco and and random shocks
McDevitt
(2017) • Compare strengthening incentives vs closing least
Amiti and productive facilities as policies to increase quality
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions

Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Section 9
Empirical example

Gandhi,
Navarro, and
Rivers (2013) Amiti and Konings (2007)
Identification
problem
Identification from
first order conditions
Value added vs gross
production
Empirical results

Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Overview
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
• Effect of reducing input and output tariffs on
ACF estimator
Relation to dynamic
productivity
panel
Empirical example • Reducing output tariffs affects productivity by
Gandhi,
Navarro, and
increasing competition
Rivers (2013)
Identification
• Reducing input tariffs affects productivity through
problem
Identification from learning, variety, and quality effects
first order conditions
Value added vs gross
production • Previous empirical work focused on output tariffs;
Empirical results
might be estimating combined effect
Grieco and
McDevitt
(2017)
• Input tariffs hard to measure; with Indonesian data on
Amiti and
plant-level inputs can construct plant specific input
Konings (2007)
tariff
Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Methodology
Paul Schrimpf

Ackerberg, • Estimate TFP using Olley-Pakes

Caves, and
Frazer (2015) • Output measure is revenue ⇒ may confound
Collinearity in OP
ACF estimator
productivity and markups
Relation to dynamic
panel
Empirical example
• Estimate relation between TFP and tariffs
Gandhi,
Navarro, and
Rivers (2013)
log(TFPit ) =γ0 + αi + αtl(i) + γ1 (output tariff)tk(i) +
Identification
problem + γ2 (input tariff)tk(i) + εit (4)
Identification from
first order conditions
Value added vs gross
production
Empirical results
• k(i) = 5-digit (ISIC) industry of plant i
Grieco and
McDevitt
• l(i) = island of plant i
(2017)
• Explore robustness to:
Amiti and
Konings (2007) • Different productivity measure
Doraszelski • Specification of 4
and
Jaumandreu • Endogeneity of tariffs
(2013)

References
 Estimating
Production
Functions
Data and tariff measure
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
• Indonesian annual manufacturing census of 20+
panel
Empirical example employee plants 1991-2001, after cleaning 15, 000 firms
Gandhi, per year
Navarro, and
Rivers (2013) • Input tariffs:
Identification
problem
Identification from
• Data on tariffs on goods, τjt , but also need to know
first order conditions
Value added vs gross
inputs
production
Empirical results
• 1998 only: have data on inputs, use to construct input
Grieco and
weights at industry level,
∑ wjk
McDevitt • Industry input tariff = j wjk τjt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Results
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015) • Look at tables
Collinearity in OP
ACF estimator • Input tariffs have larger effect than output, γ̂1 ≈ −0.07,
Relation to dynamic
panel
Empirical example
γ̂2 ≈ −0.44
Gandhi, • Robust to:
Navarro, and
Rivers (2013) • Productivity measure
Identification
problem
• Tariff measure
Identification from
first order conditions
• Including/excluding Asian financial crisis
Value added vs gross
production
Empirical results
• Less robust to instrumenting for tariffs
Grieco and • Qualitatively similar, but larger coefficient estimates
McDevitt
(2017) • Explore channels for productivity change
Amiti and • Markups (maybe), product switching/addition (no),
Konings (2007)
foreign ownership (no), exporters (no)
Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Criticism
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel • Methodology:
Empirical example
• Could use LP, ACF, or dynamic panel methods to
Gandhi,
Navarro, and estimate TFP
Rivers (2013)
Identification
• Standard errors in productivity regression appear to
problem
Identification from
ignore uncertainty from estimating TFP
first order conditions
Value added vs gross
• Measurement error in tariffs could be taken more
production
Empirical results
seriously (is this why IV estimates are larger?)
Grieco and
McDevitt
•
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions

Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Section 10
Empirical example

Gandhi,
Navarro, and
Rivers (2013) Doraszelski and Jaumandreu (2013)
Identification
problem
Identification from
first order conditions
Value added vs gross
production
Empirical results

Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Overview
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator • Estimable model of endogenous productivity, which
Relation to dynamic
panel combines:
Empirical example
• Knowledge capital model of R&D
Gandhi,
Navarro, and • OP & LP productivity estimation
Rivers (2013)
Identification
problem
• Application to Spanish manufacturers focusing on R&D
Identification from
first order conditions • Large uncertainty (20%-60% or productivity
Value added vs gross
production unpredictable )
Empirical results
• Complementarities and increasing returns
Grieco and
McDevitt • Return to R&D larger than return to physical capital
(2017) investment
Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Model (simplified) I
Paul Schrimpf

Ackerberg, • Cobb-Douglas production:

Caves, and
Frazer (2015)
Collinearity in OP yit = βl lit + βk kit + ωit + εit
ACF estimator
Relation to dynamic
panel
Empirical example • Controlled Markov process for productivity,
Gandhi,
Navarro, and
p(ωit+1 |ωit , rit ),
Rivers (2013)
Identification
problem ωit = g(ωit−1 , rit−1 ) + ξit
Identification from
first order conditions
Value added vs gross
production • Labor flexible and non-dynamic
Empirical results

Grieco and • Value function

McDevitt
(2017)

Amiti and
V(kt , ωt , ut ) = maxΠ(kt , ωt ) − Ci (i, ut ) − Cr (r, ut )+
Konings (2007)
i,r
Doraszelski 1
and + E [V(kt+1 , ωt+1 , ut+1 )|kt , ωt , i, r, ut ]
Jaumandreu 1+ρ
(2013)

References
 Estimating
Production
Functions
Model (simplified) II
Paul Schrimpf
• u scalar or vector valued shock
Ackerberg, • u not explicitly part of model, but identification
Caves, and
Frazer (2015)
discussion (especially p10 and footnote 6) implicitly
Collinearity in OP adds it
ACF estimator
Relation to dynamic
• u independent of? k, l? across time?
panel
Empirical example
• Control function incorporating Cobb-Douglas
Gandhi,
Navarro, and assumption (and perfect competition):
Rivers (2013)
Identification
problem
Identification from
ωit = h(lit , kit , wit −pit ; β) = λ0 +(1−βl )lit −βk kit +(wit −pit )
first order conditions
Value added vs gross
production
Empirical results
• Form moments based on
Grieco and
McDevitt
(2017) yit = βl lit +βk kit +g (h(lit−1 , kit−1 , wit−1 − pit−1 ; β), rit−1 )+ξit +εi
Amiti and
Konings (2007) • No collinearity because:
Doraszelski • Parametric h
and
Jaumandreu • Variation in k, r due to u
(2013)

References • Estimated model adds

 Estimating
Production
Functions
Model (simplified) III
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Empirical example

Gandhi,
• Material input instead of labor for control function
Navarro, and • h based on imperfect competition
Rivers (2013)
Identification
problem
Identification from
• Comparison to OP, LP, ACF
first order conditions
Value added vs gross
production
Empirical results

Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Results
Paul Schrimpf

Ackerberg,
Caves, and
Frazer (2015)
Collinearity in OP
ACF estimator
Relation to dynamic
panel
Empirical example
• Look at tables and figures
Gandhi,
Navarro, and • Large uncertainty (20%-60% or productivity
Rivers (2013)
Identification
problem
unpredictable )
Identification from
first order conditions • Complementarities and increasing returns
Value added vs gross
production
Empirical results
• Return to R&D larger than return to physical capital
Grieco and
McDevitt
(2017)

Amiti and
Konings (2007)

Doraszelski
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions Ackerberg, D., K. Caves, and G. Frazer. 2006. “Structural
Paul Schrimpf identification of production functions.” URL
Ackerberg,
http://mpra.ub.uni-muenchen.de/38349/.
Caves, and
Frazer (2015) Ackerberg, D., C. Lanier Benkard, S. Berry, and A. Pakes.
Collinearity in OP
ACF estimator 2007. “Econometric tools for analyzing market
Relation to dynamic
panel outcomes.” Handbook of econometrics 6:4171–4276. URL
Empirical example

Gandhi,
http://www.sciencedirect.com/science/article/
Navarro, and pii/S1573441207060631. Ungated URL
Rivers (2013)
Identification
problem
http://people.stern.nyu.edu/acollard/Tools.pdf.
Identification from
first order conditions Ackerberg, Daniel A., Kevin Caves, and Garth Frazer. 2015.
Value added vs gross
production “Identification Properties of Recent Production Function
Empirical results

Grieco and
Estimators.” Econometrica 83 (6):2411–2451. URL
McDevitt
(2017)
http://dx.doi.org/10.3982/ECTA13408.
Amiti and Aguirregabiria, Victor. 2017. “Empirical Industrial
Konings (2007)
Organization: Models, Methods, and Applications.” URL
Doraszelski
and http://www.individual.utoronto.ca/vaguirre/
Jaumandreu
(2013) courses/eco2901/teaching_io_toronto.html.
References
 Estimating
Production
Functions Amiti, Mary and Jozef Konings. 2007. “Trade Liberalization,
Paul Schrimpf
Intermediate Inputs, and Productivity: Evidence from
Ackerberg, Indonesia.” The American Economic Review 97 (5):pp.
Caves, and
Frazer (2015) 1611–1638. URL
Collinearity in OP
ACF estimator
http://www.jstor.org/stable/30034578.
Relation to dynamic
panel
Empirical example
Aw, Bee Yan, Xiaomin Chen, and Mark J. Roberts. 2001.
Gandhi,
“Firm-level evidence on productivity differentials and
Navarro, and
Rivers (2013)
turnover in Taiwanese manufacturing.” Journal of
Identification
problem
Development Economics 66 (1):51 – 86. URL
Identification from
first order conditions http://www.sciencedirect.com/science/article/
Value added vs gross
production pii/S0304387801001559.
Empirical results

Grieco and Blundell, R. and S. Bond. 2000. “GMM estimation with

McDevitt
(2017) persistent panel data: an application to production
Amiti and functions.” Econometric Reviews 19 (3):321–340. URL
Konings (2007)
http://www.tandfonline.com/doi/pdf/10.1080/
Doraszelski
and 07474930008800475.
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Bond, Steve and Måns Söderbom. 2005. “Adjustment costs
Paul Schrimpf
and the identification of Cobb Douglas production
Ackerberg,
Caves, and
functions.” IFS Working Papers W05/04, Institute for
Frazer (2015)
Collinearity in OP
Fiscal Studies. URL
ACF estimator http://ideas.repec.org/p/ifs/ifsewp/05-04.html.
Relation to dynamic
panel
Empirical example Doraszelski, Ulrich and Jordi Jaumandreu. 2013. “R&D and
Gandhi, Productivity: Estimating Endogenous Productivity.” The
Navarro, and
Rivers (2013) Review of Economic Studies 80 (4):1338–1383. URL
Identification
problem http://restud.oxfordjournals.org/content/80/4/
Identification from
first order conditions
Value added vs gross
1338.abstract.
production
Empirical results Foster, L., J.C. Haltiwanger, and C.J. Krizan. 2001. “Aggregate
Grieco and
McDevitt
productivity growth. Lessons from microeconomic
(2017) evidence.” In New developments in productivity analysis.
Amiti and
Konings (2007)
University of Chicago Press, 303–372. URL
Doraszelski http://www.nber.org/chapters/c10129.pdf.
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Foster, Lucia, John Haltiwanger, and C. J. Krizan. 2006.
Paul Schrimpf
“Market Selection, Reallocation, and Restructuring in the
Ackerberg,
Caves, and
U.S. Retail Trade Sector in the 1990s.” The Review of
Frazer (2015)
Collinearity in OP
Economics and Statistics 88 (4):pp. 748–758. URL
ACF estimator http://www.jstor.org/stable/40043032.
Relation to dynamic
panel
Empirical example Gandhi, A., S. Navarro, and D. Rivers. 2013. “On the
Gandhi, Identification of Production Functions: How
Navarro, and
Rivers (2013) Heterogeneous is Productivity?” URL https:
Identification
problem //sites.google.com/site/econsalvador/Research/
Identification from
first order conditions
Value added vs gross
production_9_25_13_FULL.pdf?attredirects=0.
production
Empirical results Grieco, Paul LE and Ryan C McDevitt. 2017. “Productivity
Grieco and
McDevitt
and Quality in Health Care: Evidence from the Dialysis
(2017) Industry.” Review of Economic Studies (forthcoming) URL
Amiti and
Konings (2007)
http://www.restud.com/wp-content/uploads/2016/
Doraszelski 08/MS17933manuscript.pdf.
and
Jaumandreu
(2013)

References
 Estimating
Production
Functions
Javorcik, Beata Smarzynska. 2004. “Does Foreign Direct
Paul Schrimpf Investment Increase the Productivity of Domestic Firms?
In Search of Spillovers through Backward Linkages.” The
Ackerberg,
Caves, and American Economic Review 94 (3):pp. 605–627. URL
Frazer (2015)
Collinearity in OP
http://www.jstor.org/stable/3592945.
ACF estimator
Relation to dynamic
panel
Kortum, Samuel and Josh Lerner. 2000. “Assessing the
Empirical example Contribution of Venture Capital to Innovation.” The RAND
Gandhi,
Navarro, and
Journal of Economics 31 (4):pp. 674–692. URL
Rivers (2013)
Identification
http://www.jstor.org/stable/2696354.
problem
Identification from
first order conditions
Levinsohn, James and Amil Petrin. 2003. “Estimating
Value added vs gross
production
Production Functions Using Inputs to Control for
Empirical results
Unobservables.” The Review of Economic Studies 70 (2):pp.
Grieco and
McDevitt 317–341. URL http://www.jstor.org/stable/3648636.
(2017)

Amiti and
Maican, F.G. 2006. “Productivity dynamics, r&d, and
Konings (2007) competitive pressure.” ECONOMIC STUDIES DEPARTMENT
Doraszelski
and
OF ECONOMICS SCHOOL OF BUSINESS, ECONOMICS AND
Jaumandreu
(2013)
LAW UNIVERSITY OF GOTHENBURG .
References
 Estimating
Production
Olley, G.S. and A. Pakes. 1996. “The dynamics of productivity
Functions in the telecommunications equipment industry.”
Paul Schrimpf
Econometrica 64 (6):1263–1297. URL
Ackerberg, http://www.jstor.org/stable/2171831.
Caves, and
Frazer (2015) Söderbom, Måns, Francis Teal, and Alan Harding. 2006. “The
Collinearity in OP
ACF estimator Determinants of Survival among African Manufacturing
Relation to dynamic
panel
Empirical example
Firms.” Economic Development and Cultural Change
Gandhi,
54 (3):pp. 533–555. URL
Navarro, and
Rivers (2013)
http://www.jstor.org/stable/10.1086/500030.
Identification
problem Van Beveren, I. 2012. “Total factor productivity estimation: a
Identification from
first order conditions
Value added vs gross
practical review.” Journal of Economic Surveys URL
production
Empirical results
http://onlinelibrary.wiley.com/doi/10.1111/j.
Grieco and 1467-6419.2010.00631.x/full.
McDevitt
(2017) Wooldridge, Jeffrey M. 2009. “On estimating firm-level
Amiti and
Konings (2007)
production functions using proxy variables to control for
Doraszelski
unobservables.” Economics Letters 104 (3):112 – 114. URL
and
Jaumandreu
http://www.sciencedirect.com/science/article/
(2013) pii/S0165176509001487.
References