DEEQA,Ecole Doctorale MPSE
Academic year 2003-2004
Advanced Econometrics
Panel data econometrics
and GMM estimation
Alban Thomas
MF 102,
[email protected]
2
Purpose of the course
Present recent developments in econometrics, that allow for
a consistent treatment of the impact of unobserved heterogeneity
on model predictions:
Panel data analysis.
Present a convenient econometric framework for dealing with
restrictions imposed by theory:
Method of Moments estimation.
Deal with discrete-choice models with unobserved hetero-
geneity.
Two keywords: unobserved heterogeneity and endogeneity.
Methods:
- Fixed Eects Least Squares
- Generalized Least Squares
- Instrumental Variables
- Maximum Likelihood estimation for Panel Data models
- Generalized Method of Moments for Times Series
- Generalized Method of Moments for Panel Data
- Heteroskedasticity-consistent estimation
- Dynamic Panel Data models
- Logit and Probit models for Panel Data
- Simulation-based inference
- Nonparametric and Semiparametric estimation
Statistical software: SAS, GAUSS, STATA (?)
3
4
Contents
I
Panel Data Models
7
1
Introduction
9
1.1
Gains in pooling cross section and time series . . .
9
1.1.1
Discrimination between alternative models .
9
1.1.2
Examples . . . . . . . . . . . . . . . . . . .
10
1.1.3
Less colinearity between explanatory variables 11
1.1.4
May reduce bias due to missing or unobserved variables
2
. . . . . . . . . . . . . . .
11
1.2
Analysis of variance . . . . . . . . . . . . . . . . .
12
1.3
Some denitions . . . . . . . . . . . . . . . . . . .
15
The linear model
17
2.1
Notation . . . . . . . . . . . . . . . . . . . . . . .
17
2.1.1
Model notation
. . . . . . . . . . . . . . .
18
2.1.2
Standard matrices and operators . . . . . .
19
2.1.3
Important properties of operators
. . . . .
20
The One-Way Fixed Eects model . . . . . . . . .
21
2.2
2.2.1
The estimator in terms of the Frisch-WaughLovell theorem . . . . . . . . . . . . . . . .
21
2.2.2
Interpretation as a covariance estimator
. .
23
2.2.3
Comments . . . . . . . . . . . . . . . . . .
24
2.2.4
Testing for poolability and individual eects
25
5
6
CONTENTS
2.3
The Random Eects model . . . . . . . . . . . . .
26
2.3.1
Notation and assumptions
. . . . . . . . .
26
2.3.2
GLS estimation of the Random-eect model
27
2.3.3
Comparison between GLS, OLS and Within
29
2.3.4
Fixed individual eects or error components? 29
2.3.5
Example: Wage equation, Hausman (1978)
2.3.6
Best Quadratic Unbiased Estimators (BQU)
of variances
3
31
Extensions
33
3.1
The Two-way panel data model . . . . . . . . . . .
33
3.1.1
The Two-way xed-eect model
33
3.1.2
Example: Production function (Hoch 1962)
3.2
3.3
4
. . . . . . . . . . . . . . . . .
30
More on non-spherical disturbances
. . . . . .
. . . . . . . .
36
37
3.2.1
Heteroskedasticity in individual eect
. . .
37
3.2.2
`Typical heteroskedasticity . . . . . . . . .
38
Unbalanced panel data models
. . . . . . . . . . .
39
3.3.1
Introduction . . . . . . . . . . . . . . . . .
39
3.3.2
Fixed eect models for unbalanced panels .
40
Augmented panel data models
47
4.1
Introduction . . . . . . . . . . . . . . . . . . . . .
47
4.2
Choice between Within and GLS . . . . . . . . . .
48
4.3
An important test for endogeneity
49
4.4
Instrumental Variable estimation: Hausman-Taylor
. . . . . . . . .
GLS estimator . . . . . . . . . . . . . . . . . . . .
51
4.4.1
Instrumental Variable estimation . . . . . .
51
4.4.2
IV in a panel-data context
51
4.4.3
Exogeneity assumptions and a rst instru-
. . . . . . . . .
ment matrix . . . . . . . . . . . . . . . . .
52
7
CONTENTS
4.4.4
More ecient procedures: Amemiya-MaCurdy
and Breusch-Mizon-Schmidt
4.5
4.5.1
. . . . . . . . . . . . . . . . . . . . . .
Full IV-GLS estimation procedure
Example: Wage equation
4.6.1
4.7
55
. . . . .
56
. . . . . . . . . . . . . .
56
. . . . . . . . . . . . .
56
Model specication
Application: returns to education
. . . . . . . . .
4.7.1
Variables related to job status
4.7.2
Variables related to characteristics of households heads
5
53
Computation of variance-covariance matrix for IV
estimators
4.6
. . . . . . . .
. . . . . . .
. . . . . . . . . . . . . . . . .
58
58
58
Dynamic panel data models
63
5.1
63
Motivation . . . . . . . . . . . . . . . . . . . . . .
5.1.1
5.2
5.3
Dynamic formulations from dynamic programming problems . . . . . . . . . . . . .
63
5.1.2
Euler equations and consumption . . . . . .
65
5.1.3
Long-run relationships in economics
. . . .
67
The dynamic xed-eect model . . . . . . . . . . .
69
5.2.1
Bias in the Fixed-Eects estimator . . . . .
70
5.2.2
Instrumental-variable estimation . . . . . .
73
The Random-eects model
. . . . . . . . . . . . .
75
5.3.1
Bias in the ML estimator . . . . . . . . . .
75
5.3.2
An equivalent representation
. . . . . . . .
76
5.3.3
The role of initial conditions
. . . . . . . .
77
5.3.4
Possible inconsistency of GLS . . . . . . . .
78
5.3.5
Example: The Balestra-Nerlove study
78
. . .
8
II
6
CONTENTS
Generalized Method of Moments estimation
The GMM estimator
6.1
6.2
6.3
85
Moment conditions and the method of moments
.
85
. . . . . . . . . . . . .
85
6.1.1
Moment conditions
6.1.2
Example: Linear regression model
6.1.3
Example: Gamma distribution
. . . . .
86
. . . . . . .
87
6.1.4
Method of moments estimation . . . . . . .
87
6.1.5
Example: Poisson counting model
. . . . .
88
6.1.6
Comments . . . . . . . . . . . . . . . . . .
89
The Generalized Method of Moments (GMM) . . .
91
6.2.1
Introduction . . . . . . . . . . . . . . . . .
91
6.2.2
Example: Just-identied IV model . . . . .
91
6.2.3
A denition
92
6.2.4
Example: The IV estimator again
. . . . . . . . . . . . . . . . .
. . . . .
92
Asymptotic properties of the GMM estimator . . .
93
6.3.1
Consistency
. . . . . . . . . . . . . . . . .
94
6.3.2
Asymptotic normality . . . . . . . . . . . .
95
6.4
Optimal and two-step GMM
. . . . . . . . . . . .
97
6.5
Inference with GMM
. . . . . . . . . . . . . . . .
99
6.6
Extension: optimal instruments for GMM . . . . .
102
6.6.1
Conditional moment restrictions
. . . . . .
102
6.6.2
A rst feasible estimator
. . . . . . . . . .
104
6.6.3
Nearest-neighbor estimation of optimal instruments
6.6.4
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . .
GMM estimators for time series models
7.1
GMM and Euler equation models
7.1.1
106
Generalizing the approach: other nonparametric estimators
7
83
109
115
. . . . . . . . .
115
Hansen and Singleton framework . . . . . .
115
9
CONTENTS
7.1.2
7.2
7.3
7.4
8
GMM estimation
. . . . . . . . . . . . . .
117
GMM Estimation of MA models . . . . . . . . . .
118
7.2.1
A simple estimator
. . . . . . . . . . . . .
118
7.2.2
A more ecient estimator . . . . . . . . . .
120
7.2.3
Example: The Durbin estimator
. . . . . .
121
. . . . . . . .
122
. . . . . . . . . . .
122
. . . . . . . . . . . . . . . .
123
Covariance matrix estimation . . . . . . . . . . . .
125
7.4.1
Example 1: Conditional homoskedasticity .
126
7.4.2
Example 2: Conditional heteroskedasticity .
126
7.4.3
Example 3: Covariance stationary process .
127
7.4.4
The Newey-West estimator . . . . . . . . .
128
7.4.5
Weighted autocovariance estimators
. . . .
130
7.4.6
Weighted periodogram estimators
. . . . .
133
GMM Estimation of ARMA models
7.3.1
The ARMA(1,1) model
7.3.2
IV estimation
GMM estimators for dynamic panel data
135
8.1
Introduction . . . . . . . . . . . . . . . . . . . . .
135
8.2
The Arellano-Bond estimator . . . . . . . . . . . .
136
8.2.1
Model assumptions
136
8.2.2
Implementation of the GMM estimator
. . . . . . . . . . . . .
. .
137
More ecient procedures (Ahn-Schmidt) . . . . . .
139
8.3.1
Additional assumptions . . . . . . . . . . .
139
8.4
The Blundell-Bond estimator . . . . . . . . . . . .
140
8.5
Dynamic models with Multiplicative eects
. . . .
141
8.5.1
Multiplicative individual eects . . . . . . .
141
8.5.2
Mixed structure
143
8.3
8.6
. . . . . . . . . . . . . . .
Example: Wage equation
. . . . . . . . . . . . . .
145
10
III
9
CONTENTS
Discrete choice models
149
Nonlinear panel data models
9.1
9.2
151
Brief review of binary discrete-choice models
. . .
151
. . . . . . . . . .
151
9.1.1
Linear Probability model
9.1.2
Logit model
. . . . . . . . . . . . . . . . .
152
9.1.3
Probit model . . . . . . . . . . . . . . . . .
152
Logit models for panel data . . . . . . . . . . . . .
153
9.2.1
Sucient statistics . . . . . . . . . . . . . .
153
9.2.2
Conditional probabilities
. . . . . . . . . .
155
9.2.3
Example:
. . . . . . . . . . . . . . .
156
. . . . . . . . . . . . . . . . . . . .
157
T =2
9.3
Probit models
9.4
Semiparametric estimation of discrete-choice models 158
9.5
9.4.1
The binary choice model
. . . . . . . . . .
159
9.4.2
The IV estimator
. . . . . . . . . . . . . .
162
SML estimation of selection models
. . . . . . . .
164
9.5.1
The GHK simulator . . . . . . . . . . . . .
164
9.5.2
Example
168
. . . . . . . . . . . . . . . . . . .
Appendix 1. Maximum-Likelihood estimation of the
Random-eect model
Appendix 2. The two-way random eects model
171
173
Appendix 3. The one-way unbalanced random eects
model
179
Appendix 4. ML estimation of dynamic panel models181
Appendix 5. GMM estimation of static panel models185
11
CONTENTS
Appendix 6. A framework for simulation-based inference
194
c Software
c
Appendix 8. A crash course in Gauss
c
Appendix 9. Example: The Gauss
software
Appendix 7. Example: the SAS
203
211
219
c 224
Appendix 10. IV and GMM estimation with Gauss
c
Appendix 11. DPD estimation with Gauss
232
References
238
12
CONTENTS
Part I
Panel Data Models
13
Chapter 1
Introduction
Panel data: Sequential observations on a number of
units (individuals, rms).
cross-sections over time, longitudinal data
cross-section time-series data.
Also called
or
pooled
1.1 Gains in pooling cross section and time series
1.1.1
Discrimination between alternative models
Many economic models in the form:
F (Y; X; Z; ) = 0;
where
Y:
individual control variables (workers, rms);
policy or principal's) variables;
:
Z:
(public
(xed) individual attributes;
parameters.
Linear model:
Y = 0 + xX + z Z + u:
15
X:
16
CHAPTER 1. INTRODUCTION
Alternative views concerning this model:
Policy variables have a signicant impact whatever individual
characteristics, or
Dierences across individuals are due to idiosyncratic individual
features,
not included in Z .
In practice, observed dierences across individuals may be due
to both inter-individual dierences
and the impact of policy vari-
ables.
1.1.2
Examples
a) W AGE = 0 + 1EDUCAT ION + 2Z .
People with higher education level have higher wages because
rms value those people more;
People have higher education because they have higher ability
(expected productivity) anyway, and rms value worker ability
more.
b)
SALES = 0 + 1ADV ERT ISEMENT + 2Z .
Advertisement expenditures boost sales;
More ecient rms enjoy more sales, and thus have more money
for advertisement expenditures.
c)
OUT P UT = 0 + 1REGULAT ION + 2Z .
Regulatory control aects rm output;
Firms with higher output are more regulated on average.
d) W AGE = 0 + 11I(UNION ) + 2Z .
Belonging to a union signicantly raises wages;
1.1.
GAINS IN POOLING CROSS SECTION AND TIME SERIES
17
Firms react to higher wages imposed by unions by hiring higherquality workers, and
1.1.3
1I(UNION ) is a proxy for worker quality.
Less colinearity between explanatory variables
In consumer or production economics, input, output or consumer
prices are dicult to use, because:
Time-series:
Aggregated macro price indexes are highly cor-
related;
Cross-sections: Not enough price variation across individuals
or rms.
With panel data, variations across individuals and across time periods are accounted for.
Time-series: no information on the impact of individual characteristics (socioeconomic variables,...);
Cross-sections: no information on adjustment dynamics. Estimates may reect inter-individual dierences inherent in comparisons of
1.1.4
dierent people or rms.
May reduce bias due to missing or unobserved
variables
With panel data, easy to control for unobserved heterogeneity
across individuals. This is critical in practice, explains why panel
data models are now so popular in micro- and macro-econometrics.
Point related to endogeneity and omitted variables issues.
18
CHAPTER 1. INTRODUCTION
Example: Output supply function under perfect competition
max = pQ C (; Q) where C (; Q) = c(Q)
(Q)
, p = @c@Q
= AQ 1 (Cobb-Douglas)
= (0 + 1Q) (Quadratic).
1
Cobb-Douglas case: log Q = 1 (log p
log A ). From
equilibrium condition to estimable equation: Observations (Qit ; pit ),
unobserved heterogeneity i , rm i, period t.
1
(log pit log i A )
log Qit =
1
Identication issue: estimable equation is
Q~ it = a0 + a1p~it + uit; i = 1; 2; : : : ; N; t = 1; 2; : : : ; T;
~ it = log Qit, p~it = log pit, a1 = 1=( 1),
where Q
a0 = ( A E log i) =( 1), Euit = 0.
Model identied if E log i = 0, i.e., Ei = 1, otherwise A is biased if i is overlooked and E log i 6= 0.
Empirical issue: possible correlation between output price
and eciency term
i.
pit
1.2 Analysis of variance
Consider the model
yit = i + xiti + "it;
where
xit
is scalar,
i
and
i = 1; 2; : : : ; N; t = 1; 2; : : : ; Ti;
i
are parameters, and
time periods available for individual
i.
Ti:
number of
1.2.
19
ANALYSIS OF VARIANCE
Useful rst-order empirical moments are
Ti
1X
y ;
yi =
T t=1 it
Sxxi =
Ti
X
t=1
x )2;
(xit
and
Syyi =
i
Ti
X
t=1
(yit
Ti
1X
x ;
xi =
T t=1 it
Sxyi =
yi)2;
Ti
X
t=1
(xit
xi)(yit
yi);
i = 1; 2; : : : ; N:
Least-square parameter estimates are computed as
^ i = Sxyi=Sxxi
and
xi^
^ i = y i
and the Residual Sum of Squares (RSS) for individual
2 =S ;
Sxyi
xxi
RSSi = Syyi
with
(Ti
i is
2) degrees of freedom:
Consider now a restricted model with constant slopes and constant intercepts:
yit = + xit + "it;
which obtains by imposing the following restrictions
1 = 2 = = N (= )
1 = 2 = = N (= ):
Under these restrictions, least-squares parameter estimates would
be
^ =
PN PTi
)(yit
i=1 t=1(xit x
PN PTi
)2
i=1 t=1 (xit x
y)
20
CHAPTER 1. INTRODUCTION
and
^ = y x^ , where
y =
N
Ti
N X
X
1
P
i Ti i=1 t=1
yit; x =
N
1
P
Ti
N X
X
i Ti i=1 t=1
xit:
The Residual Sum of Squares is
RSS =
hP
Ti
N X
X
i=1 t=1
(yit
y)2
with as number of degrees of
N PTi
i=1 t=1(yit y)(xit
PN PTi
)2
i=1 t=1(xit x
PN
freedom:
i=1 Ti 2.
i2
x)
;
For a majority of applications, the rst model is too general and
estimation would require a great number of time observations. If
unobserved heterogeneity is additive in the model, we might consider the following specication with constant slope and dierent
intercepts:
Minimizing
P P
i t (yit
yit = i + xit + "it:
i xit )2 with respect to i and , we
have
XX
t
i
(yit
i
xit ) = 0;
XX
i
t
xit(yit
i
xit ) = 0;
so that
P P
x (y y )
^ i = yi xi and ^ = P i P t it it i :
i )
i t xit (xit x
P
Residual Sum of Squares has now
i Ti (N + 1) degrees of
N + 1 parameters are estimated).
free-
dom (
This is the most popular model encountered in empirical applications.