Does -melogit- use full-information maximum-likelihood estimation?

[Tiago Pereira]

Dear Statalisters,

In [1], we are told that melogit uses full-information maximum likelihood for estimation. However, I have been unable to confirm that information in the manuals. Has anyone any piece of information to solve that puzzle? Perhaps I have missed some detail or alternative nomenclature?

All the best,

Tiago

Reference

[1] Liu, X. (2015). Applied ordinal logistic regression using Stata: From single-level to multilevel modeling. Sage publications.

[Leonardo Guizzetti]

I do not think that FIML is the same meaning as full ML estimation. Rather, I think the author being cited is getting at using the full likelihood equation, rather than some scaled version for a pseudo-likelihood equation (as is sometimes done for Poisson regression), but this is based on my reading and not on domain knowledge of FIML.

There a presentation from a Stata meeting linked from this page that discusses that Stata implements FIML (also called MLMV) as part of the SEM suite. In that sense, it seems reasonable to parameterize a multilevel logistic regression within -gsem- and request the FIML estimator. You might find that a useful starting point. Someone with a deeper understanding may care to enlighten us.

[Rich Goldstein]

note that -gsem- does not offer FIML (MLMV) as an option - only -sem- allows this option; note that the ml methods used for multi-level models, and for -gsem- if I understand it correctly, is robust to "missing at random" but this is not the same as the MLMV option for -sem-

[Tiago Pereira]

Thank you so much, Leonardo and Rich, for the insightful and very helpful comments. I agree that full ML may not mean FIML.