Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: Autocorrelation and heteroskedasticity in panel models


From   "Clive Nicholas" <Clive.Nicholas@newcastle.ac.uk>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Autocorrelation and heteroskedasticity in panel models
Date   Tue, 10 Aug 2004 23:03:34 +0100 (BST)

Cris Carambas wrote:

>   Having read that your mail, I wonder then what I could use if T < N and
> there is panel heteroskedasticity (In your mail, OLS-PCSE would be the
> best
> option if T > N and there's panel het.) I could of course resort to GLS,
> in
> particular -xtgls- but then it is also advised (read from FAQ in Stata
> http://www.stata.com/support/faqs/stat/awreg.html) that when fitting a
> variance-unconstrained model---which I am doing)  not to use -xtgls- if
> the
> number of df (N-k) is below 25 (in one of the 2 groups I have, I will
> certainly violate it). I am almost convinced to use -xtgls- given the
> arguments for T and N requierements and also my hausman test favors random
> effects and i have both autocorr and het in panel data but then the df
> constraints came in. Do you or anyone have any idea how to resolve this T,
> N
> and df constrains in -xtpcse- and -xtgls-?

First, David Greenberg was correct in his original assertion that there's
nothing wrong with FGLS when N is large, so I have to retract what I said
there. Indeed, I've done this once already. In that particular post, I
remarked that the only thing that was unresolved in my mind was the
question, "Where is the boundary between small and large, above which you
should use FGLS and below which you should use PCSE?" Is it 100? 200? 500?
I don't know, but I'm sure somebody else has a good answer to this.

Second, the df criterion is indeed important, and any estimates you obtain
after using -xtgls- if df < 25 for any of your groups may be problematic.

Third, the fact that T > N in your data allows you to consider the
possibility of using something like -arima-, does it not? I only mention
this since Kit Baum's survey lecture to the London SUG
(http://ideas.repec.org/p/boc/bocoec/598.html) notes that -arima- is quite
versatile. Stimson (1985: 928) notes that if T is (very) long, then
autocorrelation poses an ever greater threat and -arima- may well help to
overcome this. He calls this model 'GLS-ARMA' (notice there's no 'I' as it
contains no differencing mechanism in order to achieve stationarity, which
of course -arima- does have): I don't know how you would use Stata to
obtain GLS estimates from using -arima-! Somebody cleverer than me
probably has an answer to that as well.

I don't know enough about -arima- as to whether you can get it to
recognise that the data has a pooled strucutre, but - if you have a
manageable number of units - you can, of course, include dummy variables.
Moreover, -arima- is much more flexible in setting the -ar()- parameter
than -xtgls-, where you can only specify an AR(1) option (either overall
or at unit level).

Other alternatives that may help in this regard include
Baum/Schaffer/Stillman's -ivreg2- and David Roodman's -newey2-.

I hope any of this helps, but I daresay that hardly any of it probably does.

CLIVE NICHOLAS        |t: 0(044)191 222 5969
Politics              |e: clive.nicholas@ncl.ac.uk
Newcastle University  |http://www.ncl.ac.uk/geps

Stimson JA (1985) "Regression in Space and Time: A Statistical Essay" AMER
J POL SCI 29(4): 914-47.
*
*   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/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index