Hi Mark,
Thanks again.
The problem with FGLS would be that my T(21, or 3yr avg=7) is less than
N(48). Xtpcse is an alternative, but then again, it assumes exogeneity.
You are right that allowing for arbitrary within panel correlation when
actually there is no autocorrelation imposes unnecessary requirements on
the data (using cluster option). Moreover, cluster assumes E(sigma_i,
sigma_j)=0 for all t.
So, if I understand it correctly, in the presence of endogeneity , it
may not be possible to get the right standard errors if there is
heteroscd. and cross panel correlation. To extend it further, if the
model is dynamic in nature, xtabond2 , twostep may also not give correct
std. errors because it also corrects only for arbitrary heteroscd. How
would one proceed in this case? Bootstrap? Could anyone shed some light
on it.
One thing which is clear though is that in the case of not cross-panel
correlation, -cluster- is the best option.
Thanks for your time.
Jeffry
-----Original Message-----
From: Mark Schaffer [mailto:[email protected]]
Sent: Saturday, August 21, 2004 3:39 PM
To: [email protected]; Jeffry Jacob
Cc: Mark Schaffer
Subject: re: st: ivreg2 and panel group heteroscedasticity
Jeffry,
> -----Original Message-----
> From: Jeffry Jacob [mailto:[email protected]]
> Sent: Saturday, August 21, 2004 2:36 PM
> To: 'Mark Schaffer'
> Subject: RE: st: ivreg2 and panel group heteroscedasticity
>
> Hi Mark,
> Thanks for your response. It is very helpful indeed.
> Since I have a fixed effect model, I need to mean difference the
> data
> (using dummies results in #of parameters > # groups/clusters).
> My thought of _cluster_ with _bw(1)_ was motivated by the fact
> that
> since _abar_ test indicates no autocorrelation, cluster with robust
> and
> bandwidth(1) might give me sigma_i's for each group where _bw(1)_
> will
> mean no autocorrelation. But now it is clear that _cluster_ by
> itself
> will do all this.
That's right, with a small caveat. -cluster- gives you SEs that are
valid
under a very wide set of circumstances (arbitrary within-group
correlation), but you might be paying a price for this robustness. You
could be better off if you can be more specific about the type of
autocorrelation you face, and then use the appropriate technique (GLS or
kernel-based HAC). The asymptotics to make the latter approach (-bw-)
work need t to go to infinity, though, and if you have a short panel
then
that's probably not going to work.
> I have one more related question though. If there is evidence of
> cross
> panel correlation, E(sigma_i,sigma_j)~=0, and there is
> endogeneity,
> xtpcse can not be used. Will ivreg2 with robust give the right
> standard
> errors in this case?
I don't think so, if I understand you correctly. -ivreg2- with -robust-
is just like -ivreg- with -robust- (if you are instrumenting) or
-regress-
with -robust- (if you aren't). To get SEs that are right with any of
these estimators, I think you would need to use -cluster-. (Somebody
out
there correct me if I'm wrong about this!)
Cheers,
Mark
> Thanks again,
> Jeffry
>
>
>
> -----Original Message-----
> From: Mark Schaffer [mailto:[email protected]]
> Sent: Saturday, August 21, 2004 12:35 PM
> To: [email protected]; Jeffry Jacob
> Cc: Mark Schaffer
> Subject: Re: st: ivreg2 and panel group heteroscedasticity
>
> snip......
*
* 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/
