Unexpected results after estimating mixed with weights

[Felix Bittmann]

Dear all,

I have longitudinal data and a binary treatment indicator. I would like to match my sample on one outcome variable at t1 to study whether the trajectories of the two groups differ, accounting for initial group differences. I generate a matching weight using entropy balancing. This works fine. When I test the balancing with a simple OLS regression on the outcome at t1 using the weights, the group difference is virtually zero. However, when I estimate the mixed model with all panel waves, I see large group differences, even at t1. The command is like follows:

mixed outcome i.wave##i.group [pweight=mygenweight] || id:

I proceed to create a marginsplot which looks as follows:


as you can see, the groups are always different, even at t1, where the weights should force them to be equal. When I estimate the same model with regress, it works as intended. I am a bit surprised by this huge differences I see between the two models. From a statistical point of view, mixed is superior to regress as the data is panel. I wonder what causes this behaviour. Am I making a mistake here or is mixed not the right way? Any suggestions?

Best wishes

[Felix Bittmann]

I think I have found the solution. Since mixed usually includes at least 2 levels, the handling of weights is different. For level 2 weights, as in this case, the option pweight() should be specified.

[Erik Ruzek]

I think you might want to revise your understanding on this issue. See help mixed##options.

There it states that both frequency and sampling weights can be specified at level 1 or at higher levels. For lower level weights, you use the syntax you have in the first post:

mixed outcome i.wave##i.group [pweight=mygenweight] || id:

But to specify the weights at level 2, you use the following syntax:

mixed outcome i.wave##i.group || id:, pweight(mygenweight)

[Felix Bittmann]

Thanks for the reply. I think your second syntax is what I need to apply since the generated weights are on level 2 (since they vary only between individuals but are constant within one person). My data are observations (L1) nested within individuals (L2).