R(322): could not compute empirical Bayes' means: missing values were returned by the evaluator

[Sander Spoelstra]

I'm unsure why or how this happens, but it's in a postestimation test after a meologit command. The help file says that the estimation command returned an unexpected result and doesn't know how to proceed. Any clue why this happens and how to fix it?

[Marcos Almeida]

Hello Sander,

Unfortunatelty you didn't present your commands, neither you presented the details of the model.

That said, there has been some recent discussion on this matter:http://www.statalist.org/forums/foru...ion-prediction

Hopefully it helps.

Best,

Marcos

[Sander Spoelstra]

Hi, sorry about not posting the commands, they're pretty straightforward though.

meologit trustep i.education agea female || cntry: || region:, or

with trustep being a 4 category variable, education a categorical and agea a continuous. I've tried running the female as i.female since it is a dummy variable, but I've read that dummy variables do not need the i. addition. Also if I run this with normal coefficients instead of odds ratios, the problem persists.

There are roughly 50.000 observations, which makes the me command a bit difficult according to this post http://hsphsun3.harvard.edu/cgi-bin/...rticle-73.html
However, I dont see any alternative for a Multilevel Ordered Logistic Regression.

I've tried the vce(robust), intmethod (mvaghermite) and evaltype(gf0) (as was suggested by stata developers) options: none work. The same error code 322 keeps popping up:

. predict randint3
(predictions based on fixed effects and posterior means of random effects)
(option mu assumed)
(using 7 quadrature points)
could not compute empirical Bayes' means;
missing values were returned by the evaluator
r(322);

This same error also persists with the margins command and pretty much every other postestimation test.

[Marcos Almeida]

Hello, Sander,

I wonder if you may want to try the intmethod - laplace -

According to the Stata manual (http://www.stata.com/manuals13/memeologit.pdf),

The Laplacian approximation has been known to produce biased parameter estimates; however, the bias tends to be more prominent in the estimates of the variance components rather than in the estimates of the fixed effects.

[Sander Spoelstra]

I wonder if you may want to try the intmethod - laplace -

I just tried this, but it returns instantly with a r(1400); initial values not feasible. If I understand correctly, there are too many observations for this method.