Im new to the SEM field and struggling to identify the best way to model an observed categorical outcome with two time points. Ideally, I would like to model the outcome at time2 as a function of time1 predictors (both latent and observed), controlling for the outcome at time1. With only two time points, growth modelling is out of the picture. In other threads, I've seen that a random intercept model is suggested. Any other guidance? Where can I learn more about how best to model this?
If you want to regress time 2 on time 1 you already account for their correlation and don't need random intercepts. The question is if you want to regress on the time 1 observed outcome or on its underlying continuous latent response variable. WLSMV and Bayes can do either in Mplus, but ML can do only the former.