Log likelihood (not concave) message - melogit /meqrlogit

[Anna Petrova]

Hi,

I am running in Stata 15 the logistic model with wave fixed effect. As I was previously recommended Statalist I can use both -meqrlogit- and -melogit- and if one one fails to converge, the other often will.

I have run both models and received in both models for some log likelihood the message that it is not concave:

melogit charity_participation i.generation sex i.income_class i.educ_level happiness char_conf i.marital_status i.religion i.country(S002) || wave(S003):

meqrlogit charity_participation i.generation sex i.income_class i.educ_level happiness char_conf i.marital_status i.religion i.country(S002) || wave(S003):

Can you advise please if it is a problem and which model is better to use?

BR,
Anna

[Weiwen Ng]

Please don't show results in screenshots. You can show them in code delimiters. Screenshots often don't display properly, and indeed, your first one is barely legible!

That said, I can sort of squint and see that in the iteration log for melogit, you have one not concave message in the first few iterations. Similar story with meqrlogit. Per Stata's manual on likelihood maximization, this isn't a problem unless you have not concave at the last iteration. If you had that sort of situation, my experience is that Stata won't declare convergence in that situation unless you had changed one of the maximization defaults (specifically, you used the nonrtolerance option, which is generally not recommended).

Both models look like they converged at a very similar log likelihood value. It's hard to compare coefficients between the models, but I bet they are similar. If you had substantially different results between them, then it would be a problem. Chances are you don't, so the difference is irrelevant, so you could just use melogit so you don't have to explain what the QR decomposition is.

[Anna Petrova]

Weiwen Ng thanks!