"robust specifies that the Huber/White/sandwich estimator of variance
be used in place of the traditional calculation; see [U] 20.14
Obtaining robust variance estimates"
which is not the same as estimating a negative binomial model, where
the variance and mean of Y conditional on X are not given by the same
functional form (unlike poisson models). That said, poisson models
have the very nice feature of being consistent under *very* weak
assumptions.
On the other hand, the fixed-effects poisson you have estimated
suffers from the incidental parameters problem--you might try:
. xi:i.year
. xtpoisson depvar indepvar _Iyear*, fe i(id) cluster(id) robust
. xtpoisson depvar indepvar _Iyear*, fe i(id)
. xtnbreg depvar indepvar _Iyear*, fe i(id)
and maybe you can calculate deviance statistics by hand to measure
goodness-of-fit, though I'm not sure about that part. If someone else
on the list has an idea on how to generalize the techniques (mentioned
at -help poisson postestimation-), please chime in.
On 3/7/06, Scott Cunningham <[email protected]> wrote:
> I am estimating a Poisson with fixed effects model using:
>
> xi:poisson depvar indepvar i.year i.id, robust
>
> I use the "robust" to correct the standard errors because of
> overdispersion.
>
> Moments ago, I was looking at various poisson postestimation
> commands. I ran:
>
> estat gof
>
> The chi-squared result shows the poisson is not appropriate. But
> reading on, I couldn't be sure whether correcting the standard errors
> for overdispersion was the correct response to the problem the
> postestimation command was revealing. If the chi-squared result from
> a goodness of fit result is as large as I said, but I've corrected
> the standard errors for overdispersion, then should I be skeptical of
> using Poisson in this situation? If so, can someone help me see the
> intuition? Thanks.
*
* 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/
