[Joseph Coveney]

Remove the quietly and take a look at the messages that Stata gives just before the iterations begin. It seems that Stata is excluding observations for some reason, and you'll need to know why. Also, I don't understand why you have fractional observations. wtssall ought to be only integers with frequency weights.

If the expectation (mean) for the year effect is zero, as it is for the individual-respondent effect, then you would put the effect into the random effects side only. But, I would suspect some kind of systematic (average) difference across years, and so in order to account for that (and make the conditional expectation to be zero for year on the random effects side), I would put year into the fixed effects equation, too.

[Marie Eisenstein]

Joseph, thank you so much for the insight about including year as a fixed effect.

I did take a look at the messages Stata gives just before the iterations begin, and they are not useful in determining why some observations are excluded when I have two random effects versus one. All Stata says is:
Obtaining starting values by EM:

Performing gradient-based optimization:

And, I also don't understand why I have fractional observations. Looks like I have some things to investigate here.

[Joseph Coveney]

Looks like I have some things to investigate here.

A couple of things that you might start with:

1. Based upon the error message from the likelihood ratio test command, I would start with

assert !missing(cohort5)

2. For the weights, you could

generate byte fractional = wtssall != floor(wtsall)
preserve
keep if fractional
list in 1/2
restore

or something similar to start exploring what's going on there.

[Wayne Kay]

I have the same problem with Marie when trying to get p values for the random effect parameters. I switched from SAS to Stata. When SAS reports the results, SAS will report a p values for each random-effect parameter. If my understanding of this thread is correct based on the feedback from tenured members, p value for random effect parameters should not be reported or calculated at all because these p values make no sense anyways?