st: Re: re: how to test for the country and time effects...

["Rodrigo A. Alfaro" <raalfaroa@gmail.com>]

If you can get Kit's book in a hurry, I suggest to go directly to the 
primary sources. Hahn and Kuersteiner (2002) Econometrica Vol. 70, No. 4 
1639-1657 and Alvarez and Arellano (2003) Econometrica Vol. 74, No. 4 
1121-1159.

The first paper shows that Within Groups (WG) estimator can be used with 
lagged dependent variables if T is large and proposes a very simple 
correction for the asymptotic bias. Note that T/n could be lower than one 
(usually the case), but must be finite. Also the bias is completely 
different if the autoregressive coefficient is one.

The second paper shows the asymptotic bias for various estimators, including 
WG and Arellano and Bond (1991) estimators (GMM and LIML). It seems that the 
asymptotic bias for LIML is the lower one (programming that is not 
difficult, I think that there is not implemented in Stata, maybe Stata 10... 
I don't know), the second lower is the GMM (the usual) and the last is WG if 
T<n.

Also, I am writing my dissertation :-). I slightly modified Lemma 2 in 
Alvarez and Arellano (2003), having a truncated number of lags used in the 
estimation of AB/GMM. For large T the asymptotic bias is zero!!!, but given 
that the number of instruments is lower than AB/GMM it seems that the the 
standard errors must be higher. Similar argument applies for WG and AB/GMM, 
the first has a higher asymptotic bias (again if T<n), but lower standard 
errors.

In other words: (1) You can use AB/GMM with large T, at least this is what 
Alvarez and Arellano say. Even more, they recommended to use it. (2) It is 
fine to truncate the maximum number of lags used as instruments, 
option -maxldep()- in Stata. However, there is not theorem to determine an 
optimal number of lags (to minimize the Mean Square Error for example).

Finally, to the best of my knowledge, it is possible to use AB/GMM without 
lag of the dependent variable in the model. Example: Blundell, Bond, 
Devereux and Schiantarelli (1995) uses AB/GMM in the context of corporate 
investment, where the model does not have lag of the dependent variable, but 
the explanatory variable (Tobin's Q) is endogenous and the errors could be 
serial correlated.

Rodrigo.

References:
Arellano and Bond (1991) Review of Economic Studies 53, 277-297.
Blundell, Bond, Devereux and Schiantarelli (1995) Journal of Econometrics 
51, 233-257.

----- Original Message ----- 
From: "Kit Baum" <kitbaum@mac.com>
To: <statalist@hsphsun2.harvard.edu>
Sent: Friday, June 15, 2007 9:08 AM
Subject: st: re: how to test for the country and time effects...



> Vicente said
>
> I think you=C2=B4ll need to create dummy variables for that.
> I think that with gmm/xtabond2 and tsset command you will have the time
> effects already managed by stata but will still need dummies to deal
> with other kind of effects.
> I=C2=B4m also a beginner but I think this is the way to go on.
>
>
> BACKWARDS. xtabond(2) automatically includes firm/country effects,
> but does NOT include time effects.
>
> In any case the DPD / Arellano-Bond estimator should not be employed
> with large T (that is, T large relative to N), and is only
> appropriate if you have a lagged dependent variable in the model.
> See the Panel Data chapter of my book, cited below.
>
>
> Kit Baum, Boston College Economics and DIW Berlin
> http://ideas.repec.org/e/pba1.html
> An Introduction to Modern Econometrics Using Stata:
> http://www.stata-press.com/books/imeus.html
>
>
>
> *
> *   For searches and help try:
> *   http://www.stata.com/support/faqs/res/findit.html
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/ 
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/