Mixed/MLM model convergence with sampling weights

[Wes King]

Hello,

I am attempting to run a multilevel model using data from the CDC's Household Pulse Survey. The sampling frame for this survey was US households enumerated using the Census Bureau's database. So, each participant is given a person weight (pweight) and a household weight (hweight). I am interested in the relationship between state policies and mental health in a subsample (N=1749) of participants. Since state policies are at a higher level than individual mental health scores, I am using the mixed command.

I've standardized my outcome variable and started by running a null model that. The problem that I am running into is that this model does not converge when I specify the survey weights for individuals.

Here's my code:

*This converges*
mixed zphq4 if instudy==1 || est_st:

*This does not converge*
mixed zphq4 if instudy == 1 [pweight = pweight] || est_st:

There's a "backed up" message after each iteration in the second model. Also, Stata gives me a warning that I only specified sampling weights at the first level (person-level). There are no sampling weights in this dataset at the state level since the sampling frame was national, so my instinct is that this isn't the problem. There are 51 states (the data includes DC), but some states have low numbers of people in the subsample I am interested in (n = 3 in North Dakota is the smallest). I've tried only including observations from states with at least 10 people in this subsample, but that didn't solve the problem either. I don't think I need to worry about the household weights because I am not planning on analyzing any household-level variables.

Any thoughts on how to move forward would be very welcome!

[Gayanee Tashjian]

Hello, I know this message was posted a long time ago, but I'm currently having the same problem with a linear mixed model using weights, repeated data, and multiple levels. My model with weights doesn't work like yours, but it does without the weights. Were you able to resolve the issue?

[Erik Reinbergs]

If probability weights are large, sometimes rescaling them (i.e., to a mean of 1) helps convergence.