"Likelihood Evaluates to Missing" Error on Multilevel Growth Model

[salvear]

I'm working on a multilevel growth model that predicts reading score increases across grades 6 to 8. Students in the sample follow a typical grade progression, and are enrolled in the same school across grades 6 to 8. I've included pweights in my mixed model, and I received the following error message, "likelihood evaluates to missing r(430)."

mixed readingscore [pw=iptw] || school: , weight(sweight) || id: , cov(un)

Does anyone have a sense of what could be going wrong here?

Thank you in advance.

[Clyde Schechter]

The likelihood functions of multilevel models are not always well-behaved and convergence difficulties are fairly common.

The usual first step is to try simplifying the model. Your model is already pretty simple, but it could be simplified a bit, mostly as a diagnostic step. Here are some things I would try, in order:

1. Check your data to make sure that all of your variables are in their expected ranges. If readingscore is supposed to be in the range of 0-100, say, and if missing values were left coded as 999, or some entries were mistyped with extra digits, this can cause no end of problems. So fix any out-of-range data values and try again..

2. Rerun the whole thing, specifying the -trace- option. This way Stata will print out the interim results at each iteration step and you can see if there is a particular parameter that is going wild (marching to infinity, or a variance component crashing to zero.) If so, removing that parameter from the model may solve your problem (at the expense of leaving you without the full model you were hoping for. But a model with an inestimable parameter isn't very helpful anyway.

3. If the -trace- output doesn't make the problem obvious, first try removing the cov(un) option. Then try leaving out the school level, and then the id level and see if you can get convergence that way. If any of these works, you could try re-estimating the full model using the results from the converged model as starting values (see the -from()- option).

4. If none of that gets you what you want, you can consider using the -difficult- option (unlikely to help for this particular error--it's really for when you are stuck in a non-concave region), or a different estimator (see the -technique()- option).

There's a decent chance you can get convergence with some of these tricks, but no guarantee.