Laina Isler posted on Tuesday, April 03, 2018 - 3:00 pm
Hello Dr Muthens,
I have developed a latent profile model, and want to assess whether the inclusion of an additional variable improves model fit significantly. Am I able to compare BIC/AIC scores, as well as entropy, across these two models? Additionally, is there a way to assess improvement in fit using the LMR LRT or BLRT?
Thank you for this reply. To clarify, this suggests that we cannot compare model fit in non-nested models in LPA, in which 2 models differ slightly in regards to which variables are used to identify group membership? If I am incorrect, is there a resource or article that you can recommend to clarify this?
Model 1 and Model 2 are nested within each other and you can use LRT to test if variable Y5 contributes significantly to the latent class identification. In Model 2 variable Y5 is independent of C as the mean and the variance are held equal across classes, and thus Y5 doesn't contribute to the latent class identification. The results for model 2 should agree with the results when you run the LPA without Y5 included.
With the LRT you are essentially testing the Y5 on C significance.
You will have to do this variable by variable and always make sure that the comparison is done on the same set of variables. To double check you can use model test to test if the distribution of Y5 is the same or different across the classes.