Katie,
A couple of years ago I hacked away at overidxt so that it handled
random effects. It's not entirely satisfactory in its handling of
finite sample corrections etc, which is partly why I didn't make it
more widely available, but it's probably usable for your purposes.
If you contact me off-list, I'll send it to you and we can see if it
does what you want.
As for Sargan vs. Hansen's J, this is partly just a terminological
issue, and conventions vary. Usually, Sargan's stat refers to non-
heteroskedasticity-robust statistic, even if the var-cov matrix isn't
diagonal, and Hansen's J refers to some sort of robust version. But
this isn't always the case. For example, the reporting used by
xtabond follows Arellano-Bond (1991) and refers to a "Sargan
statistic" throughout; in fact, the "one-step Sargan" is a non-het-
robust stat based on a var-cov matrix with some off-diagonal
correlations, and the "two-step Sargan" is a heteroskedasticity-
robust version that many (including -xtabond2-) would call a "J"
statistic. Others now refer to a "Sargan-Hansen" or a "Hansen-
Sargan" statistic to cover all cases.
Cheers,
Mark
Date sent: Tue, 07 Sep 2004 10:43:21 -0400
From: Katie Winder <[email protected]>
Subject: st: overidxt with random effects
To: [email protected]
Send reply to: [email protected]
> Dear Stata list,
> I am estimating a wage equation with some endogenous regressors using a
> random effects model (because a correction I am using later on for selection
> bias requires it rather than fe), so I happily went about estimating
> xtivreg, re. But, as I'm sure you all know, overidxt only works after fe.
> The Sargan stat returned after fitting the fixed effects model is good, and
> my instruments are on okay grounds theoretically, but I'm still concerned
> that I can't test them directly in the random effects model (plus, it looks
> strange if I report a statistic in one table and not another).
>
> So, my question is: Does anyone have a suggestion about how I can test the
> overidentifying restrictions in a random effects IV model?
>
> As I've been thinking about this, I've been getting a little confused about
> how I would do this. I think I need to use Hansen's J statistic instead of
> Sargan's because there will be intra-person correlation in the error terms
> using re. So, I can get the clustered weighting matrix from the xtivreg, re
> residuals, and then use the instruments Z after putting them through the GLS
> transform to construct J. I think I can modify the xtivreg ado file to do
> this, but I am not even sure if this is what I want to do (and plus, given
> my programming skills, it will not be pretty).
>
> I'd appreciate anyone's feedback on this, particularly Steve Stillman, since
> I know he is an expert on this stuff and is often on the list.
>
> Thanks!
> Katie Winder
> Johns Hopkins U.
> Dept. of Economics
>
> *
> * 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/
Prof. Mark E. Schaffer
Director
Centre for Economic Reform and Transformation
Department of Economics
School of Management & Languages
Heriot-Watt University, Edinburgh EH14 4AS UK
44-131-451-3494 direct
44-131-451-3008 fax
44-131-451-3485 CERT administrator
http://www.som.hw.ac.uk/cert
*
* 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/
