State-of-the-art weighting methods when analyzing multiple surveys at the same time

[Jonathan Horowitz]

Dear Statalist:

This is a bit of a longshot, but I was wondering if anyone had good references for state-of-the-art weighting methodology when you're analyzing multiple surveys in the same regression model. For example, if three different organizations ran the exact same survey questions. Or if you had three different longitudinal cohorts (like different iterations of the NLS, or ELS, or something like that) with measures that were comparable. The closest thing I've found so far is a working paper by Anna-Carolina Haensch and Bernd Weiss here: https://osf.io/preprints/socarxiv/edm3v/download. If you've got different ideas, I'd love to hear them.

If anyone has examples of how one might do this in Stata, that would make it even better. I'm guessing that the -svy- command is going to have what I need, potentially with poststratification, but the example given is for calculating means so it's hard for me to judge. Or perhaps some sort of method for composite weighting, as described here? https://www.stat.fi/isi99/proceeding...o/kalt0185.pdf

Thanks so much in advance your thoughts, ideas, and help,
Jonathan

[Carlo Lazzaro]

Jonathan:
mybe you've already taken a look at West, Brady T., Patricia Berglund and Steven G. Heeringa. 2008. A Closer Examination of Subpopulation Analysis of Complex-Sample Survey Data. The Stata Journal, 8:520–31.

[Jonathan Horowitz]

This seems like it's relevant, although I don't need subpopulation analysis. Not being too familiar with NHAMCs it's not clear whether the stratum variables were designed by the National Center for Health Statistics specifically for this purpose or not. I don't think I have that. It also looks like the design is the same for each of them, but that is not the case in the data I am analyzing.

For the surveys I am combining, if I were to analyze them separately and used the weights with the -pweight- option in a regression equation, I would not have to worry about design effects. If I were to do that when they are combined I don't think it would do what I want them to do (I tried doing it anyway just to see what it would happen, and it doesn't estimate).

I still think this is relevant because it does seem like what I'm doing would be best addressed using -svy-, especially since I'm already running a mixed model.

[Carlo Lazzaro]

Jonathan:
I can only add to take a look at -help svy estimation- and see whether your mixed model is supported by the -svy:- prefix.