Identifying Factor Productivity by Dynamic Panel Data and Control Function Approaches: A Comparative...
-
Upload
mathias-kloss -
Category
Science
-
view
72 -
download
0
Transcript of Identifying Factor Productivity by Dynamic Panel Data and Control Function Approaches: A Comparative...
Identifying Factor Productivity by Dynamic Panel Data and Control
Function Approaches: A Comparative Evaluation for EU Agriculture
by Martin Petrick and Mathias Kloss
Mathias Kloss
Economics Cluster Seminar Wageningen UR | 3 October 2013
www.iamo.de 2
www.iamo.de 3
www.iamo.de 4
Stagnating agricultural productivity
growth in Europe
Source: Coelli & Rao 2005, p. 127.
www.iamo.de 5
Stagnating agricultural productivity
growth in the EU
Source: Piesse & Thirtle 2010, p. 171.
www.iamo.de 6
Outline
• An insight into recent innovations in production function estimation
Comparative evaluation of 2 recently proposed production function estimators
How plausible are these for the case of agriculture?
• Unique and current set of production elasticities for 8 farm-level data sets at the EU country level
• Some evidence on shadow prices
• Conclusions
www.iamo.de 7
Two problems of identification
A general production function:
𝑦𝑖𝑡 = 𝑓 𝐴𝑖𝑡 , 𝐿𝑖𝑡 , 𝐾𝑖𝑡, 𝑀𝑖𝑡 + 𝜔𝑖𝑡 + 휀𝑖𝑡
with
y Output
A Land
L Labour
K Capital (fixed)
M Materials (Working capital)
𝜔 Farm- & time-specific factor(s) known to farmer, unobserved by analyst
휀 Independent & identically distributed noise
i, t Farm & time indices
www.iamo.de 8
Two problems of identification
Collinearity problem • If variable and intermediate inputs are chosen simultaneously factor use across farms varies only with 𝜔 (Bond & Söderbom 2005; Ackerberg et al. 2007)
Production elasticities for variable inputs not identified!
Endogeneity problem • 𝜔 likely correlated with other input choices • Need to take ω into account in order to identify 𝑓, as 𝜔𝑖𝑡 + 휀𝑖𝑡
is not i.i.d No identification of 𝑓 possible if ω is not taken into account!
www.iamo.de 9
Traditional approaches to solve the
identification problems
1. Ordinary Least Squares: forget it. Assume ω is non-existent.
– Bias: elasticities of flexible inputs too high (capture ω)
2. “Within” (fixed effects): assume we can decompose ω in
– Assumption plausible?
– Bias: elasticities too low as signal-to-noise is reduced
– Collinearity problem not adressed
time-specific shock
farm-specific fixed effect
remaining farm- and time specific shock
www.iamo.de 10
Recent solutions to solve the
identification probems
3. Dynamic panel data modelling
– current (exogenous) variation in input use by lagged adjustment to
past productivity shocks (Arellano & Bond 1991; Blundell & Bond 1998)
• feasible if input modifications s.t. adjustment costs (Bond & Söderbom 2005)
• plausible for many factors (e.g. labour, land or capital ) but less so for intermediate inputs
– one way to allow costly adjustment: 𝜐𝑖𝑡 = 𝜌𝜐𝑖𝑡−1 + 𝑒𝑖𝑡, with 𝜌 < 1
– dynamic production function with lagged levels & differences of inputs
as instruments in a GMM framework (Blundell & Bond 2000)
– Bias: hopefully small. Adresses both problems if instruments induce
sufficient exogenous variation
𝜌 autoregressive parameter
𝑒𝑖𝑡 mean zero innovation
www.iamo.de 11
Recent solutions to solve the
identification probems
4. Control Function approach
– assume ω evolves along with observed firm characteristics (Olley/Pakes
1996, Econometrica)
– materials a good control candidate for ω (Levinsohn & Petrin 2003)
– further assume: (a) M is monotonically increasing in ω & (b) factor
adjustment in one period
1. Estimate “clean” A & L by controlling ω with M & K
2. Recover M & K from additional timing assumptions
– solves endogeneity problem if control function fully captures ω • productivity enhancing reaction to shocks less input use violating (a)
• some factors (e.g. soil quality) might evolve slowly violating (b)
– collinearity problem not solved • Solutions by Ackerberg et al. (2006) and Wooldridge (2009)
www.iamo.de 12
Data
FADN individual farm-level panel data made available by EC
Field crop farms (TF1) in Denmark, France, Germany East, Germany
West, Italy, Poland, Slovakia & United Kingdom
T=7 (2002-2008) (only 2006-2008 for PL & SK)
Cobb Douglas functional form (Translog examined as well)
Annual fixed effects included via year dummies
Estimation with Stata12 using xtabond2 (Roodman 2009) & levpet
estimator (Petrin et al. 2004)
www.iamo.de 13
Cobb Douglas production elasticities
Blundell/Bond
www.iamo.de 14
Cobb Douglas production elasticities
LevPet
www.iamo.de 15
Elasticity of materials LevPet
www.iamo.de 16
Elasticity of materials:
Comparison of estimators (LevPet)
www.iamo.de 17
Returns to scale LevPet
Point estimates
www.iamo.de 18
Returns to scale LevPet
Not sig. different from 1 displayed as 1
www.iamo.de 19
Examining the Translog specification
• OLS: Highly implausible results at sample means
• Within: Interaction terms not sig. in the majority of cases
• BB: No straightforward implementation, as assumption of linear
addivitity of the fixed effects is violated
• LevPet: No straightforward implementation, as M & K are assumed to
be additively separable
www.iamo.de 20
-10
0-5
00
50
10
015
020
0
Sha
dow
price o
f w
ork
ing c
apita
l (%
)
DK FR DEE DEW IT PL SK UKexcludes outside values
Shadow interest rate of materials (%):
distributions per country
www.iamo.de 21
-10
010
20
30
40
50
60
Sha
dow
wage
(E
UR
/hou
r)
DK FR DEE DEW IT PL SK UKexcludes outside values
Shadow wage (€/h): distributions per country
www.iamo.de 22
Conclusions
• Adjustment costs relevant for important inputs in agricultural production – LP and BB identification strategies a priori plausible
• LP plausible results combined with FADN data but is a second-best choice – corrected upward (downward) bias in OLS (Whithin-OLS) regressions
– conceptual problems in identifying flexible factors
• BB only performed well with regard to materials
• Materials most important production factor in EU field crop farming (prod.
elasticity of ~0.7)
• Fixed capital, land and labour usually not scarce
• Shadow price analysis reveals heterogenous picture – Credit market imperfections: Funding constraints (DEE, IT) vs. overutilisation
(DEW, DK)? Effects of financial crisis?
– Low labour remuneration (except DK)
www.iamo.de 23
Future research
• Estimated shadow prices as starting point for analysis of
drivers & impacts
• Extension to other production systems (e.g., dairy)
• Examine other identification strategies
• Wooldridge (2009) is a promising candidate
• unifies LP in a single-step efficiency gains
• solves collinearity problem
www.iamo.de 24
The END.
www.iamo.de 25
Appendix
www.iamo.de 26
Blundell/Bond in detail
• Substituting 𝑣𝑖𝑡 = 𝜌𝑣𝑖𝑡−1 + 𝑒𝑖𝑡 and 𝜔𝑖𝑡 = 𝛾𝑡 + 𝜂𝑖 + 𝑣𝑖𝑡 into the production
function implies the following dynamic production function
𝑦𝑖𝑡 = 𝛼𝑋𝑥𝑖𝑡 − 𝛼𝑋𝜌𝑥𝑖𝑡−1 + 𝜌𝑦𝑖𝑡−1 + 𝛾𝑡 − 𝜌𝛾𝑡−1𝑋
+ 1 − 𝜌 𝜂𝑖 + 휀𝑖𝑡
• Alternatively:
𝑦𝑖𝑡 = 𝜋1𝑋𝑋
𝑥𝑖𝑡 + 𝜋2𝑋𝑋
𝑥𝑖𝑡−1 + 𝜋3𝑦𝑖𝑡−1 + 𝛾𝑡∗ + 𝜂𝑖
∗ + 휀𝑖𝑡∗
subject to the common factor restrictions that 𝜋2𝑋 = −𝜋1𝑋𝜋3 for all X.
(allows recovery of input elasticities)
• Farm-specific fixed effects removed by FD, allows transmission of 𝜔 to
subsequent periods
www.iamo.de 27
Olley Pakes and Levinsohn/Petrin in
detail
• Log investment (𝑖𝑖𝑡) as an observed characteristic driven by 𝜔𝑖𝑡:
• 𝑖𝑖𝑡 = 𝑖𝑡 𝜔𝑖𝑡 , 𝑘𝑖𝑡 and 𝑘𝑖𝑡 evolves 𝑘𝑖𝑡+1 = 1 − 𝛿 𝑘𝑖𝑡 + 𝑖𝑖𝑡, with 𝛿=
depreciation rate
• Given monotonicity we can write 𝜔𝑖𝑡 = ℎ𝑡 𝑖𝑖𝑡 , 𝑘𝑖𝑡
• Assume: 𝜔𝑖𝑡 = 𝐸 𝜔𝑖𝑡|𝜔𝑖𝑡−1 + 𝜉𝑖𝑡,
– 𝜉𝑖𝑡 is an innovation uncorrelated with 𝑘𝑖𝑡 used to identify capital
coefficient in the second stage
• Idea
1. control for the influence of k and ω
2. recover the true coefficient of k as well as ω in the second stage
www.iamo.de 28
Olley/Pakes and Levinsohn/Petrin
continued
• Plugging 𝜔𝑖𝑡 = ℎ𝑡 𝑖𝑖𝑡 , 𝑘𝑖𝑡 into production function gives
𝑦𝑖𝑡 = 𝛼𝐴𝑎𝑖𝑡 + 𝛼𝐿𝑙𝑖𝑡 + 𝛼𝑀𝑚𝑖𝑡 + 𝜙𝑡 𝑖𝑖𝑡 , 𝑘𝑖𝑡 + 휀𝑖𝑡
• 𝜙 is approximated by 2nd and 3rd order polynomials of i and k in the first stage
• Here parameters of variable factors are obtained by OLS
• Second stage:
1. using 𝜙𝑡 and candidate value for 𝛼𝐾, 𝜔 𝑖𝑡 is computed for all t
2. Regress 𝜔 𝑖𝑡 on its lagged values to obtain a consistent predictor of that part of ω that is free of the innovation ξ (“clean” 𝜔𝑖𝑡)
3. using first stage parameters together with prediction of the “clean” 𝜔𝑖𝑡 and 𝐸 𝑘𝑖𝑡𝜉𝑖𝑡 = 0 consistent estimate of 𝛼𝐾 by minimum distance
www.iamo.de 29
Elasticity of materials:
Comparison of estimators (BB)
www.iamo.de 30
Comparison of estimators - East German field
crop farms: marginal return on materials
www.iamo.de 31
-10
0-9
5-9
0-8
5-8
0-7
5-7
0-6
5-6
0-5
5
Sha
dow
price o
f fixe
d c
apita
l (%
)
DK FR DEE DEW IT PL SK UKexcludes outside values
Shadow interest rate of fixed capital (%):
distributions per country
www.iamo.de 32
-.000
50
.000
5
Sha
dow
price o
f la
nd
(E
UR
/ha)
DK FR DEE DEW IT PL SK UKexcludes outside values
Shadow land rent (€/ha): distributions per
country
www.iamo.de 33
The Wooldridge-Levinsohn-Petrin
approach
• Unifies the Olley/Pakes and Levinsohn/Petrin procedure within a
IV/GMM framework
– Estimation in a single step
– Analytic standard errors
– Implementation of translog is straightforward
• Suppose for parsimony:
𝑦𝑖𝑡 = 𝛼 + 𝛽1𝑙𝑖𝑡 + 𝛽2𝑘𝑖𝑡 +𝜔𝑖𝑡 + 𝑒𝑖𝑡, and remember
𝜔𝑖𝑡 = ℎ 𝑘𝑖𝑡 , 𝑚𝑖𝑡 ,
– Now assume:
𝐸 𝑒𝑖𝑡|𝑙𝑖𝑡 , 𝑘𝑖𝑡 , 𝑚𝑖𝑡 , 𝑙𝑖,𝑡−1, 𝑘𝑖,𝑡−1 , 𝑚𝑖,𝑡−1, … , 𝑙𝑖1, 𝑘𝑖1 , 𝑚𝑖1 = 0
www.iamo.de 34
The Wooldridge-Levinsohn-Petrin
approach
• Again, assume: 𝜔𝑖𝑡 = 𝐸 𝜔𝑖𝑡|𝜔𝑖𝑡−1 + 𝜉𝑖𝑡 and
𝐸 𝜔𝑖𝑡|𝑘𝑖𝑡 , 𝑙𝑖,𝑡−1, 𝑘𝑖,𝑡−1 , 𝑚𝑖,𝑡−1, … , 𝑙𝑖1, 𝑘𝑖1 , 𝑚𝑖1
= 𝐸 𝜔𝑖𝑡|𝜔𝑖𝑡−1 = 𝑓 𝜔𝑖𝑡−1 = 𝑓 ℎ 𝑘𝑖,𝑡−1, 𝑚𝑖,𝑡−1,
• Plugging into the production function gives
𝑦𝑖𝑡 = 𝛼 + 𝛽1𝑙𝑖𝑡 + 𝛽2𝑘𝑖𝑡 + 𝑓 ℎ 𝑘𝑖,𝑡−1, 𝑚𝑖,𝑡−1, + 휀𝑖𝑡
where 휀𝑖𝑡 = 𝜉𝑖𝑡 + 𝑒𝑖𝑡.
• Now, we have two equations to identify the parameters
𝑦𝑖𝑡 = 𝛼 + 𝛽1𝑙𝑖𝑡 + 𝛽2𝑘𝑖𝑡 + ℎ 𝑘𝑖𝑡 , 𝑚𝑖𝑡 + 𝑒𝑖𝑡
𝑦𝑖𝑡 = 𝛼 + 𝛽1𝑙𝑖𝑡 + 𝛽2𝑘𝑖𝑡 + 𝑓 ℎ 𝑘𝑖,𝑡−1, 𝑚𝑖,𝑡−1, + 휀𝑖𝑡
www.iamo.de 35
The Wooldridge-Levinsohn-Petrin
approach
• And
𝐸 𝑒𝑖𝑡|𝑙𝑖𝑡 , 𝑘𝑖𝑡 , 𝑚𝑖𝑡 , 𝑙𝑖,𝑡−1, 𝑘𝑖,𝑡−1 , 𝑚𝑖,𝑡−1, … , 𝑙𝑖1, 𝑘𝑖1 , 𝑚𝑖1 = 0
𝐸 휀𝑖𝑡|𝑘𝑖𝑡 , 𝑙𝑖,𝑡−1, 𝑘𝑖,𝑡−1 , 𝑚𝑖,𝑡−1, … , 𝑙𝑖1, 𝑘𝑖1 , 𝑚𝑖1 = 0.
• Unknown function ℎ approximated by low-order polynomial and 𝑓
might be a random walk with drift.
• Estimation:
– Both equations within a GMM framework, or
– Second equation by IV-estimation and instrument for 𝑙 (Petrin and
Levinsohn 2012)