I intend to use Latent Class Analysis on a large dataset with 12 response categories and approximately 50,000 observations. I am getting an "ALERT: Iterations finished, maximum likelihood not found" warning when I fit models with more than two latent classes. Can I conclude that the models beyond two classes are poorly identified as the log-likelihood is not exactly same? The model is fit with
nrep = 20 and
maxiter = 150 and most repetitions across all classes greater than two result in
llik values which are in the range of [..,-243389.9, -243358.4, -243269.5, -243259.4, -243269.7,...].
I found contradicting information online and the reference textbook: In this tutorial, they choose the models with 3 classes despite the said alert and in Collins, L. & Lanza, S. (2010) (page 95) they do not pick the 6 class model as the ML solution did not converge. How seriously should I take this alert?
Any insights will help.