Multilevel mixed-effects linear regression of longitudinal data with subpopulation and survey weights

[Sabine Liebenehm]

Good afternoon everyone,

I'm attempting to run a 3-level model, where I observe repeated observations of individuals (qid)'s log consumption (lconspcm) over time within subdistricts (subdistr). My data is in long format. I am currently using Stata 13 with xtmixed command as follows:

xtmixed lconspcm || subdistr: || qid:, mle

I have 2 questions:

1) How to incorporate sample weights in this case: Level 1: time; Level 2: individual; Level 3: subdistrict? Do I need to weight the repeated observations at all?
2) How to run the model for a specific subpopulation? I have seen svy commands using subpop for many multilevel models in Stata 14 - however not for mixed (xtmixed respectively).

Has someone any suggestions how to proceed with these 2 questions?

Thanks a lot in advance!
Sabine

[Steve Samuels]

The Manual entries for meglm an mixed both have a section on Survey Data. You specify a weight for each level either in svyset or in the multilevel command itself.

You an use subpop in any command that accepts a svy prefix.

[Sabine Liebenehm]

Hi Steve,

thanks a lot for your quick response. I will follow your suggestion. However, I am not sure which weight to use at level 1, i.e. repeated observations of consumption of an individual over time (2 waves panel data)?

Best
Sabine

[Steve Samuels]

Since there was no sampling at level 1, just give those observations a weight equal to 1. Page 399 of the manual entry for mixed discusses scaling the weights, but with the individual observation weight equal to 1, no scaling is needed.

[Sabine Liebenehm]

Thanks a lot Steve! That helps! Best, Sabine