Bootstrap replicate weights?

[Adam Mayer]

Hello,

I am trying to bootstrap a coefficient from a logistic regression model which uses complex survey data. I have a working knowledge of bootstrapping and complex survey data, but I've come up with something of a blind spot when the two are combined.

As I understand it something called "bootstrap replicate weights" might be really important when bootstrapping with complex survey data. However, I have yet to find an intuitive explanation of bootstrap replicate weights; rather, the literature mentions them but rarely defines them. At this point I'm not sure if I need to worry about bootstrap replicate weights or if I can just start bootstrapping with the svy bootstrap command but I am very hesitant to proceed given my lack of knowledge.

Can anyone point me to a reasonably intuitive, conceptual introduction to bootstrap replicate weights or, more generally, bootstrapping with survey data?

Thanks for any help! I know requesting readings is a bit unusual for this forum but I find that Statalist provides some of the most practical, useful knowledge of any of the methodological forums out there.

Adam

[Steve Samuels]

See Stas Kolenikov's Stata Journal article at http://www.stata-journal.com/article...article=st0187, especially page 172.

The logic of the replicate weights is simple and it applies to all resampling methods, not just to the bootstrap. The total of sampling weights for a sample is an estimate of the total size of the population, \(N\), say. This will not be true of a resampling replicate, because some observations are omitted and others may be duplicated. The weights of the included observations for each replicate must therefore be altered so that the replicate total is equal to the total of the original weights.

Survey organization often create replicates to disguise stratum and PSU information. You don't say if your data already contains replicate weights. If not, then the default linearized standard errors may be fine. See page 180 of Stas's article (pdf page 18) where the different methods are compared.

[Adam Mayer]

Steve,

Thank you very much for the prompt and very clear reply.

My survey data does not use replicates to disguise stratum and/ or PSU information. It is proprietary, novel survey data and not public use.

[Sayoree Gooptu]

I got the concept of replicate weights, but I do not understand, how to create them in stata. I did not find a clear exposition about the creation of such weights while bootstrapping. Any help would be very helpful for me.

[Aiban Nongspung]

Is it advisable to use " bootstrap replicate weights " for survey data with no replicate weights ?