I'm running a series of latent variable mixture models that include classes and factors. One of the models is returning the error:
WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED. THE SOLUTION MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS.
The error occurs with the maximum number of random starts (5k) specified, and I have no a priori justification for starting values. Given the nature of the study, changing the model substantively is not desirable. The only other convergence related parameters specified are "iterations" (set at 1k) and "convergence" (set at .00005), which I believe are primarily relevant to the factor portion of the model. Would changing either of these parameters potentially help? If so, which one and what would be reasonable alternative specifications?
I don't have an active support license. Would changing "iterations" or "convergence" alter the model's ability to replicate the best loglikelihood or are these unrelated? Does it depend on the model? Any general advice you could provide would be appreciated.