An issue I've run into for two separate sets of LCA models is a very large (e.g., over 1000) odds ratio resulting from the multinomial logistic model comparing a covariate effect on class membership.
Specifically, the effect has been observed in two separate datasets when comparing effects of peer substance use on drug use class memberships when non-users were treated as the reference condition.
Conceptually, this makes sense (since non-users typically do not have drug using friends in either data set). However, I'm wondering if I need to factor this issue into my model (e.g., constraining the relationship of peer use for non-users, etc.)? Are there any examples where such constraints are used in the manual?
This typically happens when a covariate has zero variance within one of the latent classes. Typically for a binary covariate. In line with regular logistic regression with an observed categorical dependent variable, this means that the slope for this covariate goes toward infinity. Mplus typically fixes such slopes at high values automatically. This does not complicate the interpretation because all this means is that the probability is 1 to be in a certain class when the covariate has one of its values and zero otherwise.