[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: 3 Problems in Panel Data Analysis

From   "Nils Braakmann" <>
Subject   Re: st: 3 Problems in Panel Data Analysis
Date   Tue, 7 Oct 2008 16:08:40 +0200

Hello Fardad,

some more comments in addition to Martin's:

> 1) FE, RE, or BE?
> ***What should I do? What is the valid approach to pursue? How should
> I justify using RE or BE? Is there any alternative tests or methods
> that can be used? What specific conditions should I check (and how?)
> to be sure about using RE for my estimations?

Essentially it depends on whether you believe that you have unobserved
heterogeneity that is correlated with your variables of interest. If
that is the case BE and RE will be inconsistent and you should rely on
the FE estimates (your Hausman tests seem to suggest that). The
dropped variables are most likely time-constant within firms so there
is no (within) variance that could be used for estimation. Similarly,
the insignificance of the remaining variables you refer to is most
likely caused by too few variation within firms so these effects are
estimated poorly. There is in fact no simple solution to that. Some
things that come to my mind: You could either (a) use BE or RE (or
pooled OLS) (with a lot of control variables to control for as much of
the unobservables as possible) and acknowledge that your results may
in fact be caused by unoberserved heterogeneity rather than by your
variable of interest (and, if possible, include a statement in your
paper why you do not believe that unobserved heterogeneity is a
problem in this estimation or provide some explanation on the likely
direction of the bias) or (b) you could try to find some outside
instruments that are uncorrelated with the outcome and your unobserved
heterogeneity but correlated with your variables of interest and apply
some sort of instrumental variable estimator.

> 2) Robust standard errors?

> ***What would you suggest? How would you correct for
> heteroskedasticity? Is there any other important characteristics that
> I need to check before I can be sure about the validity and
> reliability of my results? What pre- or post-tests do you suggest?

Stata now provided clustered standard errors (on the panel id
variable) when you request the usual robust errors as the latter are
inconsistent in a panel context (see Stock, James H. und Mark Watson,
2008: "Heteroskedasticity-Robust Standard
Errors for Fixed Effects Panel Data Regression", Econometrica 76(1):
155-174). You should use these (for a discussion of standard errors in
a panel context see e.g. chapter 21.2.3 and the example in chapter
21.3.2 in  Cameron, A. Colin and Prvain K. Trivedi, 2005
"Microeconometrics - Methods and Applications", Cambridge University

>3) SYS-GMM method?
>***How can I successfully implement this method in Stata? Is there any
>alternatives that you would suggest? In general, how would you correct
>for simultaneity problem, if you don't have access to good

System GMM is implemented in -xtabond2- by  David Roodman who has also
written two(?) "pedagogical" papers on the practical implementation
(available on the web, don't have the links right now). Stata also has
several commands: -xtabond- and from 10.0 onwards -xtdpdsys- and
-xtdpd-. For your general question: In a panel context you might want
to consider using first differences to get rid of the unobserved
heterogeneity and then use lags as instruments to get rid of any
remaining (contemporaneous) correlation between your RHS variables and
the error. However, this does not solve your problem of too few within

Hope this helped.
Best regards,
*   For searches and help try:

© Copyright 1996–2015 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index