Bootstrap: factual data analysis

[Felix Bittmann]

Dear all,
when we compute a bootstrap analysis, Stata always uses the factual data to compute point estimates before starting with the random bootstrap resamples for the variance estimation. I wonder: is there a way to tell Stata to treat this first, factual data analysis differently? Something like a flag, variable, or local that I could use in a program for a different treatment? I think this might be relevant for combining bootstrapping with imputation.

[Clyde Schechter]

I'm not sure what you have in mind when you speak of flagging the factual estimates. But in the context of writing programs, -bootstrap- leaves the factual estimates behind in e(b), whereas it leaves the bootstrapped estimates in e(b_bs). Does that help for your purposes?

[Felix Bittmann]

Dear Clyde,

my program first imputes the dataset and then estimates the commands of interest. Obviously, when bootstrapped, this process is slow due to the additional time for the imputation in each resample. To keep times feasible, I only use a rather small number of imputations per resample. However, the point estimate is also rather variable due to this aspect. My idea is to use a larger number of M for the point estimate and a lower number for the actual bootstrap resamples. I could adjust my program like that:

if `firstrun' == 1 {
    local M = 50
}
else {
    local M = 10
}

Note that this issue is only a problem for BC (and BCa) bootstrap CIs. As a workaround, I can estimate the point estimates in a first, separate run and then compute the bootstrap CIs (and manually compute the BC version). However, I just wonder if there is a way to automate this.