Trying out different seeds is kinda cheating -- it is like you run a
hundred regressions and pick the one you liked best :)). The actual
problems might be twofold:
1. the problem is with -mfx-, as the marginal effects in Poisson
regression may have come out as the difference of two discrete
variables, and as long as you have a large sample size, the difference
between two bootstrap samples might be too small for those discrete
values to change much, so you are getting the differences in outcomes
equal to zero leading to zero or missing standard errors. I am not
really sure if that is applicable, but it might be worth looking into.
2. there are problems with the bootstrap procedure itself: I am
personally well convinced of the use of the bootstrap to estimate the
variance/get the CIs of a sample mean of i.i.d. data. Otherwise,
please show me the proof of consistency of the bootstrap (I am in the
"show-me" state of the US :)). Again, I am not sure if that really is
the culprit, but the bootstrap is not a magic wand, either, and things
may go off with it, too, in surprisingly simple circumstances. It may
not come out as a software failure, but you may run into one of those
cases where the bootstrap is inconsistent.
As for the -cluster- option, it is a variant of the general sandwich
variance estimator, which is likely to give you somewhat more
conservative standard errors, but it is not designed so as to correct
for overdispersion directly. Yes, it will correct the standard errors
somewhat for it, but it won't model overdispersion in the way -zip-
does compared to -poisson- (think of -regress, robust- with
White/Huber/linearization estimator of the standard errors vs. [F]GLS
estimator).
On 9/12/06, Richard Williams <[email protected]> wrote:
At 07:12 AM 9/12/2006, Scott Cunningham wrote:
>I'm estimating a model of sexual partners using -xtpoisson- and -
>poisson-. The data suffers from overdispersion, and so I'm trying to
>correct for that using bootstrapping within -xtpoisson-. But as I
>posted the other day, I'm having trouble recovering the marginal
>effects in post-estimation. I have a memory of someone telling me
>that the cluster() option within -poisson- can correct for
>overdispersion. Does anyone with experience in count data have
>recommendations? This is micro-level data from the National
>Longitudinal Survey of Youth (1997). The problem with -mfx, dydx-
>appears to be that bootstrapping within -xtpoisson- had numerous
>failures in calculating the standard errors. That's at least what I
>think is going wrong.
Could you try setting different seeds and see if you can get one that
does not produce any failures?
--
Stas Kolenikov
http://stas.kolenikov.name
*
* 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/