Austin,
Another rationale for marking singletons as not in the estimation sample
is that it follows the treatment of perfectly predicted observations in
-logit- and -probit-; when this happens the observations get dropped
from the estimation.
What worries me slightly, though, is whether postestimation tests like
-hausman- or others will work properly when the two estimation samples
appear different even though they are originally the same. I think they
will, but maybe I am missing something.
--Mark
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]] On Behalf Of
> Austin Nichols
> Sent: 07 November 2008 13:07
> To: [email protected]
> Subject: Re: st: RE: e(sample) not working after XTIVREG2
>
> Mark--
> I think the singletons should be marked as `touse'=0 as this kind of
> thing is just too odd:
>
> webuse grunfeld, clear
> keep if t==1 | company==1
> xtreg mvalue invest, fe
> reg mvalue invest if company==1
>
> where the only thing the extra 9 singleton obs are used for is the
> calculation of the "constant"
> ...but others may disagree.
>
> I suppose the resolution of this question is related to whether you
> believe the individual-specific means are estimated consistently (see
> also -fese- on SSC) but it seems to me that since -xtivreg2- does not
> even report a constant, you can take the stand that singletons are not
> in the estimation sample.
>
> It certainly seems odd to me to report a coef identified by 20 obs on
> one company and a "constant" (mean FE) identified by 29 obs, and to
> have the same df for tests on each, but I suppose there can be
> situations where that is the behavior you want.
>
> On Fri, Nov 7, 2008 at 7:27 AM, Schaffer, Mark E
> <[email protected]> wrote:
> > Erasmo,
> >
> >> -----Original Message-----
> >> From: [email protected]
> >> [mailto:[email protected]] On Behalf Of
> >> Erasmo Giambona
> >> Sent: 07 November 2008 11:23
> >> To: statalist
> >> Subject: st: e(sample) not working after XTIVREG2
> >>
> >> Dear Prof. Schaffer and Statalisters,
> >>
> >> For some reasons, e(sample) does not seem to work after
> XTIVREG2. In
> >> particular, if I run XTIVREG2 some observations are not
> used because
> >> they are singleton (likely due to the fact that some of my RHS
> >> variables are lagged 1 period relative to the dependent variable).
> >> However, If I try to calculate summary statistics with the
> e(sample)
> >> option (i.e., sum ...... if e(sample)), the singleton
> observations are
> >> now used again. I contacted STATA Support about it. They cannot
> >> explain why this happens and suggests that I contact you.
> >
> > This is a good question. -xtivreg2- works this way mostly because
> > that's how Stata's official FE estimators work. If you use official
> > -xtivreg- or -xtreg,fe- and you have singleton groups, they aren't
> > explicitly dropped from the sample and e(sample) will include them.
> >
> > Probably -xtivreg2- should either (a) drop the singletons from the
> > estimation sample, or (b) add an option that allows you to
> recover the
> > singletons. But I am not sure what is best; (a) is a significant
> > variation from official Stata and should not be adopted lightly.
> >
> > If anyone on the list has views on this, I'd be interest to
> hear them!
> >
> > --Mark
> *
> * 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/
>
--
Heriot-Watt University is a Scottish charity
registered under charity number SC000278.
*
* 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/
