Scribd
Upload
60 views18 pages

Dynamic Panel Estimator: GMM: Dr. Elya Nabila Abdul Bahri

Uploaded by

moira142560
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
60 views18 pages

Dynamic Panel Estimator: GMM: Dr. Elya Nabila Abdul Bahri

Uploaded by

moira142560
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 18

Dynamic panel estimator:

GMM

Dr. Elya Nabila Abdul Bahri


Why Dynamic Panel?
• Some relationships are dynamic and panel
data allows us to understand adjustment.
• Y does not necessarily respond
immediately to changes in X.
• Example:
– Current demand depends on past demand
– Increased in educational expenditure today
does not decrease income gap today.
Why Dynamic Panel?
• Allowing for dynamics in the underlying
process may be crucial for recovering
consistent estimates of other parameters
(Bond, 2002).
Why can't we use standard panel
estimators
• The presence of individual-specific effects,
lagged dependent variable and potential
endogeneity of independent variables
making the traditional panel estimators
(OLS, FE and RE) inappropriate (see
Nickell, 1981).
GMM estimator designed for
situations with:
• “small T, large N” panels
• a linear functional relationship
• one left-hand-side variable that is dynamic,
depending on its own past realizations
• independent variables that are not strictly
exogenous, meaning they are correlated with past
and possibly current realizations of the error
• fixed individual effects (country effects)
• heteroscedasticity and autocorrelation within
individuals but not across them.
Linear dynamic panel-data model
• Repeated observations on the same individuals
allows us to analyze individual dynamics.
• Dynamic relationships are typically modeled by
introducing lagged dependent variable in panel
model specifications. Consider the model:

yit = yi ,t −1 + xit + i +  it (1)


where
i = 1, 2, ..., N; t = 1, 2, ..., T
Linear dynamic panel-data model
• First-differencing the equation removes the fixed-effects (ηi), thus
eliminating a potential source of omitted variable bias in estimation.

yit − yi ,t −1 =  ( yi ,t −1 − yi ,t − 2 ) +  ( xit − xi ,t −1 ) + ( it −  i ,t −1 )
(2)
Difference GMM
• To address the simultaneity bias of explanatory
variables, Arellano and Bond (1991) develop a
GMM estimator that instruments the differenced
variables that are not strictly exogenous with all
their available lags in levels.
• The following moment condition:

E yi ,t − s • ( it −  i ,t −1 ) = 0 for s  2; t = 3, ..., T (3)


E xi ,t − s • ( it −  i ,t −1 ) = 0 for s  2; t = 3, ..., T (4)
Difference GMM estimates eq. (2) with moment
conditions in (3) and (4).
Difference GMM
• However, when the independent variables are
persistent, the lagged levels of the variables
become weak instruments in difference GMM.
• The weak instruments may lead to biased
parameter estimates in small samples and
larger variance asymptotically (Alonso-Borrego
and Arellano (1999); Blundell and Bond
(1998)).

1/26/2023
System GMM
• Arellano and Bover (1995) proposed an alternative system estimator that
combines the difference and the level equations.
• Blundell and Bond (1998) show that system estimator is able to reduce
biases and imprecision associated with difference estimator.
• The additional moment conditions for the regression in levels are set as
follows:

E ( yi ,t − s − yi ,t − s −1 )• (i +  it ) = 0 for s = 1 (5)

E (xi ,t − s − xi ,t − s −1 )• (i +  it ) = 0 for s = 1 (6)

• System GMM estimates eq. (1) and (2) simultaneously with moment
conditions in (3)-(6).

1/26/2023
GMM estimations
• The GMM estimator has one- and two-step
variants.
• The one-step estimator assumes independent
error terms and homoskedastic error
variances across individuals and times.
• The second-step estimator uses the residuals
of the first-step estimation to construct a
consistent variance–covariance matrix when
the assumptions of independence and
homoscedasticity do not hold.

1/26/2023
Diagnostic Tests
• The consistency of the GMM estimator
depends on the validity of the assumption
that the error terms do not exhibit serial
correlation (AR(2)) and on the validity of the
instruments.

1/26/2023
Diagnostic Tests
• The validity of the instruments can be tested
using the Sargan or Hansen tests of over-
identifying restrictions
• The null hypothesis is that the overidentifying
restrictions are valid, that is, the instrumental
variables are not correlated with the error
term.

1/26/2023
Diagnostic Tests
• Failure to reject the null of the Sargan or
Hansen test would imply that the instruments
are valid and the model is correctly specified.
• Reconsider the model or the instruments if
the null of Sargan/Hansen is rejected (p-value
< 0.10)

1/26/2023
Diagnostic Tests
• The Arellano-Bond test for autocorrelation is
applied to the differenced residuals.
• The null hypothesis is that no second order
serial correlation.
• Outlier detection.

1/26/2023
In Practice...
• Panels with a small time dimension should
be of no concern given that all the
asymptotic properties of GMM estimator
rely on the size of the cross-sectional
dimension of the panel (Beck et al. 2000).
• GMM do not require large time dimension
in order to obtain consistent parameter
estimates.
• The number of instruments in GMM tends
to increase with T.
In Practice...
• Use all available lagged values of
independent variables as instruments to
ensure maximum efficiency (Beck et al.,
2001).
• Collapse the matrix of instruments to
minimize the number of instruments in the
regressions as suggested by Roodman
(2009).
In Practice...
• A large instrument collection can overfit
endogenous variables.
• How many instruments is “too many”?
Keep the instrument count < individual
units, N.

You might also like