How to perform bootstrap resampling for logit and fracreg models?

[Laiy Kho]

Hello,
I have 3 different research questions, and I use regress, logit, and fracreg respectively. After I estimated my models, I also performed bootstrap resampling as a robustness check. I used the command below:
. set seed 3957574 . bootstrap _b[foreign], size(37) reps(2000) dots: regress mpg weight foreign

This worked well with regress command. However, when I run it with fracreg, it takes too long. It has been 3 days, but the output is not yet estimated.
. bootstrap _b[X1], size(1404090) reps(1000) dots: fracreg logit Y X1 ....
(running fracreg on estimation sample)



Bootstrap replications (1,000)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
.................................................. 50
.................................................. 100
.................................................. 150
.................................................. 200
.................................................. 250
.................................................. 300
.................................................. 350
.................................................. 400
.................................................. 450
.................................................. 500
............................


Meanwhile, when I estimated with logit, it provides no output.
(running logit on estimation sample)

Bootstrap replications (100)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
.xxxx....x--Break--

I didnt press Break, but same continues to happen even when I re-estimate it .

Can someone advise where I am going wrong? Is there a different command when performing bootstrapping procedure for logit & fracreg models?

Thank you

[Leonardo Guizzetti]

Bootstrap standard errors are obtained through -vce(bootstrap)- so you don’t have to use the bootstrap prefix. I would do that instead.

for example

regress … , vce(bootstrap) …

[Joseph Coveney]

Two questions on your fracreg case and two suggestions for the logit.

Why are you specifying the size() option when bootstrapping the fracreg fit?

What kind of "robustness check" are you hoping to do for your coefficients' standard errors by bootstrapping when you have nearly a million and a half observations?

I've never encountered a situation in which Stata pressed the break key for me, but in the case of your logit model you might want to look into whether your red-x bootstrap samples are due to complete or quasicomplete separation. If that's the case, then you can try using the asis option of the estimation command in order to see whether that helps get you through the bootstrap replications.

And if you've found that the problem is separation, then you can also look into using a penalized maximum likelihood method in order to assure convergence.

search penlogit
search firthlogit

for further information.