block bootstrap (cluster-robust bootstrap) with xtpoisson

[Bobby Wood]

Hi,

I have panel count data with an endogenous variable and use a first stage residual inclusion control function (cf) to overcome that.
To get correct inference I need to bootstrap the standard errors.

Is it possible to do this with clusters?

bootstrap, reps(1000) cluster(id): xtpoisson y x cf, fe i(id)

doesnt' work.

Would

bootstrap, reps(1000): xtpoisson y x cf, fe i(id) r

do the job?

Many thanks,

Bobby

[Joseph Coveney]

bootstrap, reps(1000) cluster(id): xtpoisson y x cf, fe i(id)

doesnt' work.

You might want to take a look at this for some advice on that.

Would

bootstrap, reps(1000): xtpoisson y x cf, fe i(id) r

do the job?

For xt commands, the cluster bootstrap is now built-in. Have you considered something like this?

xtset id
xtpoisson y x cf, fe vce(bootstrap, reps(1000))

On the other hand, if you need to include the "first stage residual inclusion control function" into the bootstrap procedure, then you would probably have to write a brief wrapper command that incorporates both the first-stage regression command and the Poisson follow-on regression so that the bootstrap sample is used for both. You can check the entry for bootstrap in Stata's set of user's manuals, but I believe that it would look (very roughly) like this:

program define bootem
    version 15.1
    syntax varname [if]

    marksample touse

    xtreg stuff1 i.stuff2 c.etc if `touse', i(`varlist') fe
    tempvar cf
    quietly predict double `cf', ue

    xtpoisson y x c.`cf' if `touse', i(`varlist') fe
end

bootstrap x = _b[x], reps(1000) cluster(id) idcluster(newid): bootem newid

The suggestion obviously has not been tested; moreover, I'm no econometrician and so I'm guessing as to what your commands look like, but I believe that you can take it from there.

You might want to take a look at gsem and see whether that estimation command can fit a model that addresses your research question more cleanly.