I am trying to determine the best fit # of classes for an LPA/LCA with continuous variables, and I'm running into some problems I haven't seen before. I'm using 3 variables and the sample size is about 4,300. Here's the basic info:
a. All the IC statistics continue improving as I add more classes--I've run up to 8 classes so far.
b. The parametric boot strap test also improves with each additional group, although I get errors that the best LL was not replicated in all draws.
c. The VLMR and the LMR suggest that a 6-class model fits best.
d. For all classes >4, the log-likelihood is only replicated a small number of times. For example, I ran the 6-class model with 10,000 starts, and the best LL was replicated only 3 times, suggesting a local maximum.
One thing I'm noticing is that is some classes from models with 5 or more classes have either a mean or a SE on one of the indicators of 0. As the three indicators are counts, a substantial portion of the sample has a value of 0 for one or more indicator (26%-38%). Does it indicate a fundamentally unusable/unidentified model?
Thanks in advance for any help anyone can provide.