Predict after meqrlogit produces missing values for all random effects

[Heiko Giebler]

For a paper, I am estimating a three-level random-intercept model with a binary DV with Stata 13. I have ~150,000 observations nested in ~27,000 respondents nested in 28 countries and there are IVs on all three levels including a cross-level interaction. The simplified command looks like this:

meqrlogit y x1 x2##x3 x4 x5, intp(3) variance ///
|| z1: , covariance(independent) || z2: , covariance(independent)

The model runs smoothly without problems of convergence or in regard to adaptive quadrature. However, the lines

predict yhat
predict u*, reffects

produce missing values for all observations. I get probabilities for the fixed part (predict yhat_fixed, fixed) without problems. Moreover, if I run the model deleting a large number of observations randomly (~2,000 observations, ~400 repsondents and 28 countries left), Stata calculates the predicted random effects. I tried different number of integration points (including LaPlace) as well as different specifications of the covariance structure. The problem keeps appearing under all conditions. In other words, it seems that the problem is caused by the large number of observations which seems rather odd.
I am probably missing something.

Any idea, what could possibly cause this?
Thanks!

[Elizabeth Colantuoni]

Heiko, I am having a similar problem, did you figure this out?

[Olli Sagi]

Hi guys,

I figured it out, its very simple. You predict random effects for variables which are usually quite broad, for example cities. And you are ofcourse using two datasets, say 1 and 2. Dataset 1 you use to get estimates to predict random effects for Dataset 2. However, if Dataset 1 does not contain a city which IS present in Dataset 2 stata will not be able to predict any results. Fairly simple because it does not have any information about that particular city to predict random effects.

Hope this helps! Took me quite long to figure out tho.

Greetings,
Olcay