You probably want to use -xtreg, fe- to control for unobservable panel
effects in the first place, right? You suspect your regressors are
correlated with the panel effect, yada-yada, biased estimates, etc. If
you can twist -mi-'s arms to generate the completed data sets that
incorporate those unobservables, you might be able to justify the
multiple imputation approach. But this is not feasible since by
definition those panel effects are not observed. Ergo, I would not
trust any application that would do -mi estimate: xtreg, fe-.
Conceptually, if you ran -mi- within each panel, that might produce
something you could claim makes sense. But say 10 observations you
would probably have in a typical panel is hardly enough to fit a good
imputation model, even more so if your regression will have 20
explanatory variables. No free lunch, sorry; -mi- is not a panacea,
and should not be thought of as such. I personally think about it as a
last resort, and only if I can convincingly incoroporate all the data
generating process features into the imputation model.
On Thu, Aug 27, 2009 at 11:49 AM, Woolton Lee<[email protected]> wrote:
> Does anyone know of simulation studies that have looked at statistical
> properties of simple fixed effects regression with multiple
> imputation? I would like to know whether there is evidence justifying
> the use of this approach. Thanks.
>
> PS if you know some cites please send them to me. Thanks.
>
> Woolton
> *
> * For searches and help try:
> * http://www.stata.com/help.cgi?search
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
--
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
