Number of quadrature points in -predict- after -gsem-

[Laura Gibbons]

I'm having a problem that I can't replicate in any of the publicly-available SEM datasets so I guess I'm looking for general explanations or solutions. When I fit a -gsem- model and then try to -predict- the latent trait, sometimes I get an error message of

could not compute empirical Bayes means;
missing values were returned by the evaluator

I usually use 30 quadrature points (intpoints). When I reduce the number to 20, all my examples run fine, and at least for those datasets, I'm comfortable with using 20 points.

But this problem came up when someone was using my ssc program -difdetect-, and sometimes he had to reduce to as low as 5 quadrature points before the latent trait was estimable.

I'll admit I know very little about what is going on under the hood here. Can anyone offer an explanation on why this might be? Bonus points for any solutions!

Many thanks,
Laura

[Erik Ruzek]

This is really weird behavior. I would suggest that you contact Stata Technical Support directly about it. They will need a copy of your data, most likely. But you are not the only one experiencing this, so it feels like it needs to be addressed by Stata directly.

[Laura Gibbons]

Yes, I always start here, but I was guessing this might be something for the tech support folks. thanks!

[Laura Gibbons]

Stata has identified a precision problem and this will be fixed in an upcoming update to Stata 18. I hope they fix it in earlier versions as well, but I don't know what they will do.

This came to my attention because someone was trying to use my -difdetect- program (on SSC) in Stata 17, so please contact me directly (my email is in the help file) if you are having trouble running that program.