George Y posted on Sunday, October 06, 2013 - 7:41 pm
I am running growth mixture models and am trying to interpret the output from the Vuong-Lo-Mendell-Rubin likelihood ratio test and bootstapped parametric likelihood ratio tests.
When running simple latent class growth models (i.e., with growth parameter variances fixed) then I can see that the loglikelihood for H0 (i.e., K-1 classes) is the same as what is estimated if I was to run the K-1 class model itself. However, if I start freeing up some of the variances (e.g., say I free up two intercept variances) then I am unsure of what model this H0 model is. Is it the same as the K-1 model also with two intercept variances freed? I cannot seem to identify what this H0 model is that it is comparing to.
I assumed that I should be able to estimate this H0 model itself. For example, if I was comparing the K=3 class model with 2 intercept variances free, then I thought that the K-1 H0 model in the VLMR and BLRT tests would have the same loglikelhood as if I just estimated a K=2 class model with the same 2 intercept variances free. But the loglikelihoods are not the same when I run it myself. I also quickly tried to increase the starts for the VLMR and BLRT tests thinking that perhaps it was a local solution but not sure whether that was the right thinking or not. I didn't get any error saying the loglikelihood was not replicated.
Any guidance here would be much appreciated. Thanks for your time.