Avoiding failed bootstrap replications

[Belinda Foster]

Hello. I am boostrapping a probit model which in some replications does not converge resulting in missing values in the bootstrap dataset. This is obviously related to the data that i have but i was wondering if there is a way to instruct Stata to discard the replications when they fail and instead draw another sample in their place so my boostrap dataset does not have missing values.

[Clyde Schechter]

I don't believe there is any way to do this, short of hacking the code of -bootstrap.ado- or writing your own version. But there generally is no need to do that. Bootstrap soldiers on after finding failed analyses. If you have already run a moderately long series of bootstrap replications you will have a sense of what proportion of samples lead to estimation failure, and you can just adjust the number of replications you specify in your bootstrap command so you will end up with the (approximate) number you want. The exact number of bootstrap replications isn't really important: what matters is that the number be large enough to adequately sample the bootstrap distribution. If you figured out originally that you need 1000 reps for your purposes, it will be just fine to have 975 or 1050 or something like that. To be safe, err on the side of somewhat more replications than are strictly needed. So if you have a 5% estimation failure rate, and you need 1000 replications, asking for 1000/.95 = 1053 (nearest integer) should get you there more or less. To be on the safe side, ask for 1100. You'll probably end up with a bit more than 1000 replications, but that's not a problem.

Added: if the proportion of replications that lead to estimation failure is large, then you should probably rethink your problem. Perhaps some other way of estimating the parameters you seek would work better.

[Belinda Foster]

Thanks Clyde. What you proposed is actually what i am doing at the moment. I was just curious whether there was some other way to avoid constantly playing with the number of replications.