Replicating Bootstrap result in Python / how does xtreg bootstrap with many fixed effects

[Ben Smith]

Hi Everyone -

I've been trying to replicate a result with handwritten python and I'm running into a bit of problems replicating the bootstrapped p-value. I've tried pair-wise bootstrapping (bootstrapping the rows), and cross sectional bootstrapping (bootstrapping the people -- it is panel data with a bunch of ids and six observations per person).

Obviously it isn't going to be exact, but it isn't close (my p-value is about .08 and stata's is about .06). Here's what I know:

1. My value is much closer to the uncorrected value
2. Stata's is closer to the corrected value

So my question is this: what procedure is stata running when you run:

xtset ID
xtreg Dep Var1 i.Category, fe vce(bootstrap, reps(1000))

And there are many ID values?

Thanks! Any help would be appreciated!

Ben

[Carlo Lazzaro]

ben:
welcome to the list.
The following Stata thread might be useful: http://www.stata.com/support/faqs/st...th-panel-data/

[Ben Smith]

Hey Carlo:

Thanks for replying. Yeah, I've seen that post before. From what I can tell that corresponds to what I described in my post as a "cross sectional bootstrap". But, perhaps there is a more correct term. Anyhow, my cross sectional approach is about the same as the pair-wise approach and does not match stata's results.

Ben

[Carlo Lazzaro]

Ben:
As per your post#1, it does not seem that a difference of 0.02 in p-value is such a relevant result.
Perhaps it may be explained by the bootstrap machinery.
Anyway, it is difficult (for me, at least) to comment on any further without seeing your results.