Puzzling over a multigroup lca with gender as the known class and 16 dichotomous measures of chronic health conditions as latent class indicators.
Running the model without any covariates, except having the medical conditions latent class regressed on gender, model estimation proceeds without problems and a 3 class solution for the medical classes appears to be best.
However, subsequently adding covariates to the model and regressing the latent classes on gender and the other covariate(s) yields the message that the best loglikelihood value was not replicated as when I add age:
Model: %Overall% c#1-c#2 on gender age;
I tried different covariates one at a time and I pushed Starts up very high with no luck. Any covariate I add in addition to gender seems to cause non-replication. The only thing that seems to work is including a direct effect between the indicator for arthritis and age like this:
Model: %Overall% c#1-c#2 on gender age; CC1A on age;
But I found this out just by luck and am not sure why this is happening or if there is another solution. Does it mean I essentially have a poorly fitting model (although the ICs and LMR and entropy all appear to support the model without covariates)? Should I constrain the threshold values and then include the covariates? Or does it just mean I need this direct effect in the model?
The need for direct effects of covariates on latent class indicators, such as your CC1A on age, can account for this behavior. You would see it in the class percentages changing when you add covariates and don't include any direct effects.
To investigate direct effects, you take one latent class indicator at a time and regress it on all the covariates to see if any direct effects are significant.