Why the enhanced module of lclogit (lclogit2) does not converge?

[Federica Di Marcantonio]

Dears

I am estimating latent class models and i have some problems with the command lclogit2. Apparently what can be estimated with lclogit does not converge when i use lclogit2. Is there any reason? And what could be the solution?

Thanks
Federica

[Phil Bromiley]

You will increase your chances of useful answer by following the FAQ on asking questions – provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

Exactly why one user written procedure converges and another doesn't is impossible to answer from your posting. I don't use these procedures, but it's possible for example that the precise details by which one specifies models differ so you actually are specifying a different model than you think you are. There are many standard things to do with models that don't converge beginning with running simpler models and building up to see where the problem is. Alternatively, if the second estimator accepts starting values, use the ones from the first estimator as starting values. Oddly, it appears that both of those estimators were at least partially written by the same person and appear to use the same algorithm.

If nothing else works, contact the authors.

[Hong Il Yoo]

Federica Di Marcantonio: Have you checked whether your -lclogit- results really reached a local maximum by passing it to -lclogitml-? My guess is that your seemingly converged model is empirically unidentified and -lclogitml- will fail to converge. -lclogit2- carries out all calculations in Mata while -lclogit- does not, and they may shoot in different directions when the model is numerically unstable due to underidentification. The default cap on the number of iterations is also larger for -lclogit2- than -lclogit- so unless you have changed the default settings, you may have a situation -lclogit2- correctly fails to converge whereas -lclogit- is terminated prematurely as the maximum number of iterations is reached.