I am writing an article about the comparison of LCA and CFA. I have several questions about the computation of the parameter estimates in the different analyses.
- to my understanding the optimization algorithm used to estimate parameters in LCA is (based on) iterative proportional fitting; Mplus, however, uses the more broadly applicable accelerated expectation maximization algorithm in conjunction with Fisher Scoring and Quasi-Newton. For my five-group model with 18 indicators with each 3 categories, once I get to >4 class models (which the output suggests me to explore) the analysis becomes too computationally demanding (days), and even with very high start values (starts = 6000 1500; for 5-class model) Mplus cannot verify to have converged to the global maximum. Might I better switch to a faster algorithm, and if so, which one would that be in Mplus? - although I was not able to verify this from Mplus output or User Manual, I assume continuous-variable CFA also uses EMA; is it then safe to say that CFA and LCA use the same optimization algorithm for ML estimation despite the differences in variable distribution?