On 6/9/09, Carolina Lennon <[email protected]> wrote:
> Previously, I thought that sometimes it could be not advisable to
> implement cluster errors but still to implement robust errors (then,
> the utility of having the two options separately). For instances,
> when the number of observations was higher enough but either the
> number of clusters was lower than 50 or the number of explanatory
> variables exceeded the number of clusters. In any case, the response
> might be in the article! then many thanks again!
That's cheating on degrees of freedom. With few clusters, the
-vce(cluster)- variance estimator will be unstable -- in any case, the
asymptotics of both point estimates and the standard errors is that
the number of clusters goes to infinity. You would get standard errors
that are anti-conservative, although I don't think I've seen any proof
(analytical or by simulation) that they are inconsistent. I am sure
you can find situations where they are though.
The qualitative explanation that I can offer is that -robust- assumes
the data are independent, but when we are in the panel data world, we
know we are not. The contributions to the statistical model come from
a panel as a whole -- this is the case with the likelihood with
-xtreg, mle- for instance, and this is the case with -vce(cluster)-.
--
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/