Melogit or xtlogit?

[Maria Ventura]

Hi,

I need to estimate a logistic regression with fixed effects on panel data. I am using probability weights though, so I was going for melogit as I can't use them with xtlogit.

I tried running a simple regression (without weights) with xtlogit y x, fe and melogit y x, but the results are quite different? I assume there is something I am missing and running melogit with a fe equations only is not equivalent to a xtlogit, fe?

Thanks

Maria

[Clyde Schechter]

-xtlogit, fe- is a fixed effects model; -melogit- is a random effects model and running it without any re equations is not equivalent to a fixed-effects logit; it is equivalent to running -logit- with no hierarchy at all.

[Maria Ventura]

Thanks!! So is there any other command that would allow me to use probability weights but with e FE model? I see clogit wouldn't work as prob weights must be the same within the group...

[Clyde Schechter]

I don't know of anything that will support that. But others might. If so, I hope they will chime in.

[Richard Williams]

So, is it like you have 5 records each from different points in time for each person? Sometimes panel data sets have a single weight you are supposed to use when, say, you are using waves 1 through 5 or waves 1 and 2 or whatever. See what the documentation says about weighting when using multiple waves.

[Maria Ventura]

Thanks Richard! So the documentation distinguishes between longitudinal and cross sectional weights, depending on the type of analysis, but both would change over time

[Richard Williams]

Yes, but you are probably not the first one to want to use more than one wave. Hopefully it gives a single weight for, say, using all the waves or whatever waves you want to use.

Some people actually argue against weighting, at least under certain conditions. I summarize the arguments on pp. 11-12 of

https://www3.nd.edu/~rwilliam/xsoc73...yCautionsX.pdf

Much of that discussion is adapted from

https://www.statalist.org/forums/for...-be-considered