Scott Cunningham <[email protected]> is interested in bootstrap standard
errors from the -xtpoisson- command:
> Recently, I queried the list about the proper method of calculating
> robust standard errors in -xtpoisson- to correct for overdispersion.
> I think I may have found a method, which is on Colin Cameron's
> website (UC-Davis).
>
> http://cameron.econ.ucdavis.edu/mmabook/mmaprograms.html
>
> Scroll down to mma23p1pannonlin.do, and you'll find in that file the
> following:
>
> bootstrap "xtpoisson PAT LOGR $xextra, fe i(id)" "_b[LOGR]", cluster
> (id) reps($nreps) level(95)
>
> This clusters on id, and is a luster bootstrap. Discussion of panel
> robust standard errors can be found in Cameron and Trivedi's recent
> book, MICROECONOMETRICS: METHODS AND APPLICATIONS, setion 23.2.6,
> pp. 789-791. The .do file in question contains the code used to
> generate the various bias-corrected standard errors in table 23.1 on
> page 794.
>
> Using -xtpoisson depvar indepvar, fe i(id) vce(boostrap) - did not
> provide me with the same standard error calculation, either, as what
> Cameron recommended. Does anyone know what the -vce(bootstrap)- is
> doing in the -xtpoisson-?
Unfortunately, there is a problem with the following call to the Stata 8
-bootstrap-:
bootstrap "xtpoisson PAT LOGR $xextra, fe i(id)" ///
"_b[LOGR]", cluster(id) reps($nreps) level(95)
It is in the use of the -id- variable. When the -cluster()- option is used,
-bootstrap- generates a new bootstrap dataset by randomly sampling the groups
identified by the -id- variable. Suppose -id==1- is sampled twice, then the
new bootstrap sample contains all the observations for which -id==1- twice.
When -xtpoisson- is called to fit the model, it treats these two independently
sampled copies of the panel as a single panel.
This is the reason for the -idcluster()- option. When specified, -bootstrap-
recreates a variable that uniquely identifies the independently sampled
panels. Here is how to bootstrap panel data using the Stata 8 -bootstrap-
(with the above example):
tempvar newid
gen `newid' = id
bootstrap "xtpoisson PAT LOGR $xextra, fe i(`newid')" ///
"_b[LOGR]", cluster(id) idcluster(`newid') reps($nreps)
level(95)
Now when a panel is sampled multiple times, `newid' will identify each sampled
panel uniquely so that -xtpoisson- can treat them as seperate entities.
Scott's use of the -vce(bootstrap)- option in Stata 9's -xtpoisson- does not
have this problem, because -xtpoisson- handles communicating the panel
information to -bootstrap- for you:
xtpoisson depvar indepvar, fe i(id) vce(bootstrap)
If you want to change the number of reps from the default to 99, use the
-reps()- suboption of -vce(bootstrap)-
xtpoisson depvar indepvar, fe i(id) vce(bootstrap, reps(99))
Also, don't forget to set the random number seed so that you can reproduce
your results later.
--Jeff
[email protected]
*
* 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/
