Searching for the Robust Method to
Estimate Total Factor Productivity
at Firm Level
Yin Heng Li Shigang Liu Di
SEBA Beijing Normal University
Email: yheng@bnu.edu.cn
Motivation

Discuss the robust TFP estimation method at
firm level, using competitive industry as an
example.
What does TFP measure?



Evaluate the input-output efficiency
Labor productivity cannot describe the true
efficiency at firm level
The core of TFP estimation is dealing with
the substitution among input factors
The importance of TFP estimation


Productivity is not everything, but approximates
everything in the long run. Krugman（1997）
The factors affecting TFP
–

Is the TFP of firms improved?
Misallocation of resources
- Can the economic environment promotes firms
with high TFP, and suppress or expel those
with low TFP?
The current situation of TFP estimation

Great differences exist even in researches
appeared in the top journals
- Young (1995)’s estimation of the growth rate of TFP
in Hong Kong and Taiwan district in China is
between 2% and 3%, the growth rate of Korea is
1.7%
- Hsieh (1999) got 3% more than Young’s.
Structure






The measurement of data and variables
Traditional methods
Firms’ decision and structure estimation
Value-added or gross output production
function
Sample selection, function form and other
robust test
summary
The measurement of data and
variables

Panel construction
–
–
Goal : identify firms across years
Problems :


–
Different firms may share the same code
Firms may change the code because of changing name
or structure etc.
Idea :



Make sure that firms with the same code is the same
one;
Match firms with the combination of relatively stable
information, such as name, head, telephone, etc.;
Correct the wrong matching.
The measurement of data and
variables

The measurement of output
–
–

The real output
deflate gross output from three dimensions:
time(year), space(province), industry(two-digit)
The measurement of input
–
Construct the input deflator from time(year),
space(province) and industry(two-digit), based on
input-output table, like Brandt (2012).
The measurement of data and
variables

The measurement of capital
–
–
–

Estimate nominal investment from the found year
with the data of original fixed capital
Deflate the nominal investment to get the real
investment
Get the real capital with perpetual inventory
method
The measurement of labor
–
Total average number of staff
The measurement of data and
variables

The choice of industries
–

Two-digit industry 18: manufacture of clothing, shoes
and hats; two-digit industry 19: manufacture of
leather, fur and feather
Data clearing
–
–
–
Delete the sample with non-positive output, capital,
labor and input
Delete the sample with less than 8 workers
Delete the sample with bigger value-added than
output
Table 1: Descriptive Statistics
Year
Obs.
1998
Output
Capital
Labor
Material
Mean
Std.
Mean
Std.
Mean
Std.
Mean
Std.
8795
29634.70
71628.43
4589.94
13460.20
331.38
787.72
19406.05
48162.89
1999
8482
31358.37
76530.40
4640.89
14266.12
331.24
638.53
20317.37
51764.68
2000
8872
33683.06
89291.19
4399.33
14301.95
332.19
663.23
21526.77
58845.14
2001
10269
33896.71
99122.08
4030.10
14122.47
320.76
616.40
21716.56
65105.94
2002
11488
34759.67
106787.20
3763.74
13890.40
315.26
592.85
22062.48
69204.64
2003
13219
37826.08
129803.20
3758.72
14743.61
321.98
642.17
23571.01
82674.73
2004
16210
36327.99
165626.10
3577.54
19389.14
316.30
616.25
21990.80
108535.00
2005
17549
43655.98
201137.20
3969.65
21885.35
319.08
657.42
26398.21
131528.40
2006
19260
47903.11
235153.90
4283.38
28367.65
313.16
676.45
28537.35
150848.60
2007
21314
52152.53
249305.10
4344.99
27960.65
301.58
671.94
30418.95
159284.20
Traditional methods



DEA (Data Envelopment Analysis)
Index method
Tradition parametric methods
–
–
–
OLS
FE
BB
Traditional methods

DEA
–
–
–
Considering the heterogeneity of firms’ TFP
Get the TFP measurement from the input and
output data with linear programming, treating the
production process as a black box
It is a determinate method which can be sensitive
to the random error or extreme values.
Traditional methods

Index method
𝜔𝑖𝑡 = 𝑦𝑖𝑡 − 𝑆𝑖𝑡𝐿 𝑙𝑖𝑡 − 𝑆𝑖𝑡𝑀 𝑚𝑖𝑡 − 1 − 𝑆𝑖𝑡𝐿 − 𝑆𝑖𝑡𝑀 𝑘𝑖𝑡
–
–
–
–
Free of function form
Need the given the parameter of return to scale
Without the consideration of random error
Based on the hypothesis that all inputs are static
input without adjustment cost
Traditional methods

Parametric method
𝑦𝑖𝑡 = 𝛽𝑘 𝑘𝑖𝑡 + 𝛽𝑙 𝑙𝑖𝑡 + 𝛽𝑚 𝑚𝑖𝑡 + 𝜔𝑖𝑡
–
–
Based on the set that all firms in the same
industries have the same elasticity of output of
capital, labor and input
Deal with random error
Traditional methods

OLS
–

FE
–

Endogenous problem
Neglect the change of TFP with time
AB/BB
–
–
𝜔𝑖𝑡 = 𝑢𝑖 + 𝑣𝑖𝑡 ，𝑣𝑖𝑡 = 𝜌𝑣𝑖𝑡−1 + 𝑒𝑖𝑡
System GMM
Table 2：Traditional Methods
DEA
INDEX
OLS
FE
BB
Coef.
Std.
Coef.
Std.
Coef.
Std.
k
-
-
-
-
0.0431
0.0017
0.0320
0.0032
0.0632
0.0247
l
-
-
-
-
0.0738
0.0024
0.1057
0.0048
0.2333
0.0497
m
-
-
-
-
0.8678
0.0018
0.7490
0.0033
0.8643
0.0363
k
-
-
-
-
0.0492
0.0013
0.0599
0.0023
0.1282
0.0215
l
-
-
-
-
0.1191
0.0020
0.1527
0.0044
0.0106
0.0349
m
-
-
-
-
0.8158
0.0017
0.6969
0.0027
0.9518
0.0204
Year
98-02
03-07
98-02
03-07
98-02
03-07
98-02
03-07
98-02
03-07
Growth %
4.2136
6.8645
2.6128
3.2202
1.5407
2.7222
2.3363
4.2434
0.4844
0.1696
90/10
4.1528
5.4557
2.2865
2.5604
1.4636
1.5532
1.7671
1.8278
1.8370
1.7388
95/5
6.9894
9.1132
3.0946
3.7640
1.7207
1.8629
2.1345
2.2355
2.2774
2.1942
16780
29704
16780
29704
16780
29704
16780
29704
16780
29704
98-02
03-07
Ratio
Obs.
Firms’ decision and structure
estimation


The more information of firm’s action and
decision we use, the more robust and
accurate result we can get.
Tradition methods neglect the information of
firm’s action and decision structure.
Firms’ decision and structure
estimation

Data generating process at firm level
–
–
Firms choose input and output to maximize the
profit based on the observed productivity 𝜔𝑖𝑡
𝑄𝑖𝑡 = 𝑄 𝐿𝑖𝑡 , 𝐾𝑖𝑡 , 𝑀𝑖𝑡 , 𝜔𝑖𝑡
Where 𝑄𝑖𝑡 is planed output, the real output is
𝑌𝑖𝑡 = 𝑄𝑖𝑡 𝑒 𝜇𝑖𝑡
Firms’ decision and structure
estimation

The decision structure of firm’s factor input:
dynamic and static
–
Two adjustment frictions make the firm’s input
decision dynamic:


Adjustment cost, such as the cost of installment, test
and dismantle
Adjustment lag, because the factor used now is decided
at the former period
Firms’ decision and structure
estimation

The decision structure of firm’s dynamic input:
take capital as an example
V  Kit , it , J it   max   Kit , it , J it   C  Rit 
Rit
1

E V  Kit 1 , it 1 , J it 1  Kit , it , J it , Rit 
1 
Firms’ decision and structure
estimation

The decision structure of firm’s static input :
materials
max   Kit , it , J it   PQ
 Kit , M it , it   M it PMt
t
M it
Firms’ decision and structure
estimation

The decision structure of firm’s labor input
( may change with industry)
–
Treated as dynamic if the adjustment cost cannot
be neglected


–
Adjustment cost : training cost when employing new
staff and the cost of layoff
Adjustment lag : new staff can only get to work after the
training
Treated as static if the adjustment cost can be
neglected
Firms’ decision and structure
estimation

Model
–
–
–
C-D production function
Hicks-neutral techniques
Static labor input
𝛽
𝛽
𝛽
𝑄 𝐿𝑖𝑡 , 𝐾𝑖𝑡 , 𝑀𝑖𝑡 , 𝜔𝑖𝑡 = 𝐿𝑖𝑡𝑙 𝐾𝑖𝑡𝑘 𝑀𝑖𝑡𝑚 𝑒 𝜔𝑖𝑡
𝑦𝑖𝑡 = 𝛽𝑘 𝑘𝑖𝑡 + 𝛽𝑙 𝑙𝑖𝑡 + 𝛽𝑚 𝑚𝑖𝑡 + 𝜔𝑖𝑡 + 𝜇𝑖𝑡
Firms’ decision and structure
estimation

Olley & Pakes（1996）
–
–
–
Get productivity 𝜔𝑖𝑡 = 𝜔 𝑘𝑖𝑡 , 𝑟𝑖𝑡 from the
investment function 𝑟𝑖𝑡 = 𝑟 𝑘𝑖𝑡 , 𝜔𝑖𝑡 , and then take
it into the production function
𝑦𝑖𝑡 = 𝛽𝑙 𝑙𝑖𝑡 + 𝛽𝑚 𝑚𝑖𝑡 + 𝜑 𝑟𝑖𝑡 , 𝑘𝑖𝑡 + 𝜇𝑖𝑡
Step 1. get 𝛽𝑙 , 𝛽𝑚 , 𝜑 and 𝜇𝑖𝑡 with nonparametric
method, and then the productivity can be
expressed as 𝜔𝑖𝑡 =𝜑 − 𝛽𝑘 𝑘𝑖𝑡 ；
Step 2. let productivity follows the Markov
process𝜔𝑖𝑡 = 𝐸 𝜔𝑖𝑡 |𝜔𝑖𝑡−1 + 𝜉𝑖𝑡 = 𝑔 𝜔𝑖𝑡−1 +
𝜉𝑖𝑡 ,get the estimation of 𝛽𝑘 with the moment
condition 𝐸 𝜉 𝑘 = 0
Firms’ decision and structure
estimation

Levinsohn & Petrin (2003)
–
–
A great loss of investment information
Use materials as proxy variables:
mit  m  kit , it 
it    kit , mit 
Firms’ decision and structure
estimation

Bond & Söderbom (2005)and Ackerberg et al.
(2006): Collinearity problem
lit  l  kit , it  ; mit  m  kit , it 
–
–
Robinson (1988): “The variables in the parametric
part cannot be predicted by those in the
nonparametric part in the sense of OLS.”
Newey et al. (1999): There should exist no function
between parametric part and nonparametric part in
semi-parametric model.
Firms’ decision and structure
estimation

Ackerberg et al.(2006)
–
–
Capital is decided before TFP
Labor decision is before materials
lit  l  kit , it 
mit  m lit , kit ,it 
Firms’ decision and structure
estimation
–
–
Step 1. the production function is 𝑦𝑖𝑡 = 𝜑 𝑚𝑖𝑡 , 𝑘𝑖𝑡 , 𝑙𝑖𝑡 +
𝜇𝑖𝑡 , get 𝜑、 𝜇𝑖𝑡 with nonparametric method, and the
productivity is 𝜔𝑖𝑡 = 𝑦𝑖𝑡 − 𝛽𝑘 𝑘𝑖𝑡 − 𝛽𝑙 𝑙𝑖𝑡 − 𝛽𝑚 𝑚𝑖𝑡 − 𝜇𝑖𝑡
Step 2.the productivity follows Markov process𝜔𝑖𝑡 =
𝐸 𝜔𝑖𝑡 |𝜔𝑖𝑡−1 + 𝜉𝑖𝑡 = 𝑔 𝜔𝑖𝑡−1 + 𝜉𝑖𝑡 , get the other
parameters with the moment condition 𝐸 𝜉𝑖𝑡 ⋅
Firms’ decision and structure
estimation

The idea of the new structural estimation of
TFP at firm level
–
Review the index method about estimating static
input



–
Solow (1957)；
Caves et al. (1982)；
Hall (1989)
Separate the estimation of static input and dynamic
input

Gandhi et al. (2011)
Firms’ decision and structure
estimation

new structural estimation of TFP
Get the following formula according to the optimal
condition of static input
𝑄𝑀 𝐿𝑖𝑡 , 𝐾𝑖𝑡 , 𝑀𝑖𝑡 , 𝜔𝑖𝑡 𝑀𝑖𝑡 𝑃𝑀𝑡 𝑀𝑖𝑡
=
𝜇
𝑖𝑡
𝑄 𝐿𝑖𝑡 , 𝐾𝑖𝑡 , 𝑀𝑖𝑡 , 𝜔𝑖𝑡 𝑒
𝑃𝑡 𝑌𝑖𝑡
– the Hicks-neutral technique allows
𝑀
𝑀
𝑠𝑖𝑡
= 𝑓 𝐿𝑖𝑡 , 𝐾𝑖𝑡 , 𝑀𝑖𝑡 − 𝜇𝑖𝑡 Where 𝑠𝑖𝑡
is the share of
materials to nominal output
–
–
Get 𝑓𝑖𝑡 、𝜇𝑖𝑡 、𝛽𝑚𝑖𝑡 with nonparametric regression,
and in the situation of C-D production function,
the mean of 𝛽𝑚𝑖𝑡 is 𝛽𝑚
Firms’ decision and structure
estimation
–
–
If labor is static input, then get 𝛽𝑙 with the method
above, if not, get the estimation of 𝛽𝑙 at the next
step
The productivity follows Markov process same as
OP/LP/ACF 𝜔𝑖𝑡 = 𝐸 𝜔𝑖𝑡 |𝜔𝑖𝑡−1 + 𝜉𝑖𝑡 = 𝑔 𝜔𝑖𝑡−1 +
𝜉𝑖𝑡 , and get 𝛽𝑘 with the moment
condition 𝐸 𝜉𝑖𝑡 𝑘𝑖𝑡 = 0
Firms’ decision and structure
estimation

New structural estimation of TFP
–
–
–
Step 1. estimate the parameter of static input
following the idea of index method
Step 2. estimate the parameter of dynamic input
following the idea of structural estimation
The advantages



Avoid the assumptions in the proxy variables method such
as the reversible proxy function and the measurement error
Make full use of firms’ decision
Solve the endogenous problem and the collinearity problem
Table 3：Structural Estimation for Aggregate Output
OP
LP
ACF
NEW-S1
NEW-S2
Coef.
Std.
Coef.
Std.
Coef.
Std.
Coef.
Std.
Coef.
Std.
k
0.0459
0.0005
0.0084
0.0037
0.0115
0.0021
0.0537
0.0014
0.0621
0.0014
l
0.0936
0.0015
0.0773
0.0010
0.0501
0.0054
0.1358
0.0021
0.1090
0.0002
m
0.8309
0.0012
0.9144
0.0102
0.9372
0.0092
0.7302
0.0002
0.7302
0.0002
Year
98-02
03-07
98-02
03-07
98-02
03-07
98-02
03-07
98-02
03-07
Growth%
0.3810
0.5745
0.2169
0.1836
0.2574
0.3691
1.4941
3.9874
1.5763
4.0379
90/10
1.0562
1.0514
1.1133
1.0822
1.1816
1.1270
1.4517
1.4990
1.4676
1.5141
95/5
1.0856
1.0809
1.2142
1.1271
1.3234
1.2080
1.6544
1.7231
1.6803
1.7427
Ratio
Obs.
67483
97196
97196
97196
97196
Gross output or Value-added?

Gross output (sales) is the real observable
variable by firms who experience the
production and management process, while
value-added is just a statistical concept.
Gross output or Value-added?
Value-added can be proper only if the
theoretic definition is agreed with empirical
measurement, which needs the following
assumptions
Assumption 1. Labor and capital produce
value-added following 𝑉 = 𝜓 𝐿, 𝐾 𝑒 𝜔 , and
combine with materials according to 𝑌 =
𝐹 𝜓 𝐿, 𝐾 𝑒 𝜔 , 𝑀 to form output

Gross output or Value-added?

The core in TFP estimation is to control the
substitution among factors
–
Make the following choices to maximize profit



–
Labor intensive
Capital intensive
Outsource and material intensive
Value-added production function only consider the
substitution between labor and capital and neglect
the efficiency from materials
Gross output or Value-added?
Gross output or Value-added?

The result misusing value-added
𝑉𝑖𝑡 = 𝑌𝑖𝑡 1 − 𝑆𝑖𝑡𝑀 = 𝑄 𝐿𝑖𝑡 , 𝐾𝑖𝑡 , 𝑀𝑖𝑡 𝑒 𝜔𝑖𝑡 +𝜇𝑖𝑡 1 − 𝑆𝑖𝑡𝑀
𝑣𝑖𝑡 = 𝛽′ 0 + 𝛽𝑘 𝑘𝑖𝑡 + 𝛽𝑙 𝑙𝑖𝑡 + 𝛽𝑚 𝑚𝑖𝑡 + 𝜔𝑖𝑡 + 𝜇𝑖𝑡
–
–
New endogenous problem appears because
𝛽𝑚 𝑚𝑖𝑡 is put into the error term
TFP heterogeneity will be exaggerated because
the heterogeneity coming from materials is put
into TFP difference
Table 4：Structural Estimation for Value-added
OP
LP
ACF
NEW-S
Coef.
Std.
Coef.
Std.
Coef.
Std.
Coef.
Std.
k
0.3417
0.0028
0.2691
0.0054
0.1689
0.0115
0.3869
0.0044
l
0.4959
0.0040
0.2169
0.0024
0.0909
0.0133
0.0930
0.0001
Year
98-02
03-07
98-02
03-07
98-02
03-07
98-02
03-07
Growth%
0.9916
1.3299
4.4379
9.4516
5.8774
12.0461
11.4778
17.9319
90/10
1.6245
1.5912
5.1236
5.1989
6.3752
6.6307
8.8184
7.1292
95/5
1.9829
1.9762
9.0538
8.6154
12.0695
11.7247
19.7805
14.1120
Ratio
Obs.
67489
97196
97196
96920
Sample selection, function form
and other robust test

Sample selection problem
–

There is a great number of entry and exit in the
data, and we can only observe the existed ones
The structure estimation method don’t have to
deal with sample selection problem because
of the proxy of 𝜔𝑖𝑡 in the first step
𝐸 𝜇𝑖𝑡 |𝑙𝑖𝑡 , 𝑟𝑖𝑡 , 𝑘𝑖𝑡 , 𝑚𝑖𝑡 , 𝜔𝑖𝑡 = 𝐸 𝜇𝑖𝑡 |𝑙𝑖𝑡 , 𝑟𝑖𝑡 , 𝑘𝑖𝑡 , 𝑚𝑖𝑡 = 0
Sample selection, function form
and other robust test


We can only observe the existed samples
with𝜗𝑖𝑡 = 1, and 𝜔𝑖𝑡 = 𝐸 𝜔𝑖𝑡 |𝜔𝑖𝑡−1 , 𝜗𝑖𝑡 = 1 +
𝜉𝑖𝑡 ,so there is endogenous problem in the
second step
How to deal with it?
1 𝜔𝑖𝑡 ≥ 𝜔𝑖𝑡
–
Rules of entry and exit：𝜗𝑖𝑡 =
–
Conditional expectation：𝐸 𝜔𝑖𝑡 |𝜔𝑖𝑡−1 , 𝜗𝑖𝑡 = 1
=𝐸 𝜔𝑖𝑡 |𝜔𝑖𝑡−1 ,𝜔𝑖𝑡 ≥ 𝜔 =𝜑 𝜔𝑖𝑡−1 , 𝜔
0
𝜔𝑖𝑡 < 𝜔𝑖𝑡
Sample selection, function form
and other robust test
The probability that a firm i stay in period t
𝑃𝑖𝑡|𝑡−1 = Pr𝑜𝑏 𝜗𝑖𝑡 = 1|𝜔𝑖𝑡−1 , 𝐼𝑡−1
= Pr𝑜𝑏 𝜔𝑖𝑡 ≥ 𝜔𝑖𝑡 |𝜔𝑖𝑡−1 , 𝐼𝑡−1 = 𝜓𝑡−1 𝜔𝑖𝑡 , 𝜔𝑖𝑡−1
–
–
−1
Get 𝜔𝑖𝑡 = 𝜓𝑡−1
𝑃𝑖𝑡|𝑡−1 , 𝜔𝑖𝑡−1
Put into the conditional expectation of productivity
𝐸 𝜔𝑖𝑡 |𝜔𝑖𝑡−1 , 𝜗𝑖𝑡 = 1 =𝜑 𝜔𝑖𝑡−1 , 𝜔𝑖𝑡
–
−1
= 𝜑 𝜔𝑖𝑡−1 , 𝜓𝑡−1
𝑃𝑖𝑡|𝑡−1 , 𝜔𝑖𝑡−1
= 𝑔 𝜔𝑖𝑡−1 , 𝑃𝑖𝑡|𝑡−1
Table 5：Structural Estimation for Aggregate Output:
Sample Selection Considered
OP
LP
ACF
NEW-S1
NEW-S2
Coef.
Std.
Coef.
Std.
Coef.
Std.
Coef.
Std.
Coef.
Std.
k
0.0457
0.0005
0.0259
0.0103
0.0304
0.0118
0.0343
0.0014
0.0341
0.0015
l
0.0936
0.0015
0.0773
0.0010
-0.0277
0.0380
0.1065
0.0028
0.1090
0.0002
m
0.8309
0.0012
0.7968
0.0783
0.8170
0.0665
0.7302
0.0002
0.7302
0.0002
Year
98-02
03-07
98-02
03-07
98-02
03-07
98-02
03-07
98-02
03-07
Growth%
0.3848
0.5796
0.9931
1.8084
1.4416
2.4926
1.7720
4.4306
1.7608
4.4184
90/10
1.0562
1.0514
1.2926
1.3153
1.5180
1.5157
1.5034
1.5482
1.4996
1.5456
95/5
1.0858
1.0810
1.4062
1.4371
1.7527
1.7450
1.7337
1.7862
1.7296
1.7821
Ratio
Obs.
67483
97196
97196
97196
97196
Sample selection, function form
and other robust test

Trans-log production function
2
2
2
𝑦𝑖𝑡 = 𝛽𝑘 𝑘𝑖𝑡 + 𝛽𝑙 𝑙𝑖𝑡 + 𝛽𝑚 𝑚𝑖𝑡 + 𝛽𝑘𝑘 𝑘𝑖𝑡
+ 𝛽𝑙𝑙 𝑙𝑖𝑡
+ 𝛽𝑚𝑚 𝑚𝑖𝑡
+
𝛽𝑘𝑙 𝑘𝑖𝑡 𝑙𝑖𝑡 + 𝛽𝑘𝑚 𝑘𝑖𝑡 𝑚𝑖𝑡 + 𝛽𝑙𝑚 𝑙𝑖𝑡 𝑚𝑖𝑡 + 𝜔𝑖𝑡 + 𝜇𝑖𝑡

Cobb-Douglas production function is a special
situation of trans-log production function.
Table 6：Sensitivity Analysis
BASE
CHECK1
CHECK 2
CHECK 3
CHECK 4
Coef.
Std.
Coef.
Std.
Coef.
Std.
Coef.
Std.
Coef.
Std.
k
0.0717
0.0054
0.0771
0.0047
0.0587
0.0088
0.0821
0.0043
0.0752
0.0062
l
0.2944
0.0113
0.2679
0.0000
0.2588
0.0120
0.2752
0.0086
0.1887
0.0094
m
0.3925
0.0010
0.3925
0.0010
0.4049
0.0010
0.3984
0.0013
0.4719
0.0013
 kk
0.0089
0.0004
0.0091
0.0004
0.0104
0.0006
0.0080
0.0003
0.0075
0.0004
ll
0.0274
0.0012
0.0307
0.0000
0.0261
0.0011
0.0290
0.0010
0.0263
0.0008
 mm
0.0392
0.0001
0.0392
0.0001
0.0390
0.0001
0.0387
0.0001
0.0426
0.0001
 kl
0.0035
0.0010
0.0019
0.0000
0.0053
0.0012
0.0038
0.0008
0.0033
0.0009
 lm
-0.0499
0.0002
-0.0499
0.0002
-0.0464
0.0002
-0.0494
0.0002
-0.0516
0.0002
 mk
-0.0185
0.0001
-0.0185
0.0001
-0.0222
0.0001
-0.0182
0.0001
-0.0188
0.0001
t
-
-
-
-
-
-
0.0342
0.0005
Year
98-02
03-07
98-02
03-07
98-02
03-07
98-02
03-07
98-02
03-07
Growth%
1.0973
3.3210
1.0765
3.3036
0.5199
2.3967
1.4325
2.6047
0.8027
2.5936
90/10
1.4631
1.4616
1.4631
1.4616
1.4167
1.4232
1.6414
1.3594
1.4308
1.3966
95/5
1.6683
1.6467
1.6683
1.6467
1.5899
1.5791
1.9345
1.5050
1.6283
1.5619
Ratio
Obs.
97196
97196
97196
97196
97196
Summary

The problems of tradition methods
–
DEA method tries to measure TFP by construct a
set of substitution of factors by linear
programming, but determinate method cannot get
the robust estimation with the data at firm level,
because the measurement error cannot be
neglected.
Summary
–
–
Index method is also not satisfactory because all
the inputs are assumed to be static and the
parameter of return to scale should be given.
Traditional methods, such as FE,IV and dynamic
panel, will not get the robust result because the
disturbance should be given before the estimation.
Summary

Structural estimation method, which is becoming
the most potential approach, tries to open the
black box of the firms’ production process by
making full use of the information of their
behavior and decision-making.
–
–
Olley and Pakes (1996), Levinsohn and Petrin
(2003),Ackerberg et al.(2006) all face the “collineraity”
problem.
The new structural estimation, which combines the
structural estimation with the traditional index method,
may get the most robust estimation of TFP at firm level.
Summary

The definition of variables affects the robustness
of TFP estimation
-measuring firms’ output with value-added will
exaggerate TFP heterogeneity seriously

Sample selection and the production function
form also affect the TFP estimation
Summary

The most robust estimation of TFP for clothing
and leather industry in China
Summary

Unsolved problem:
–
–
The use of proxy variable in structural method
and the index method need new foundation if
firms have market power.
More information is needed to separate the effect
of demand and price from TFP
Thank you!