milan lee posted on Saturday, December 07, 2013 - 12:54 pm
Hello everyone, I wanted to specify a GMM model with covariates in my dataset. I read Jung and Wickrama (2008)and Dr. Muthen's chpt 19. Both of them suggest that specifying a LCGA model without within-class variance is an initial modeling step prior to specifying a GMM model. However, I think the best number of latent classes found in a LCGA may not suit a GMM, particularly with covariates, since LCGA disregards the within-class variance. Jung's paper does not say much about whether the difference between LCGA and GMM will lead to the validity of class number. Should we decide the class number based on a unconditional GMM instead? Thanks a lot for any suggestions in advance! Milan
Re conflicting indices, again! I'm wondering if you have any updates or suggestions about how to resolve conflicts between the BIC and the VLMR in choosing the best fitting solution for an unconditional GMM? For three different sets of scales measured over 7 assessments, the BIC indicates that three classes fit the data better than two with as much as a 10point difference, but the VLMR is clearly not significant for the 3 vs 2-class solution. Given the Nylund finding that the VLMR tends to overfit classes, so that when it is nonsignificant it strongly indicates that too many classes have been extracted, it seems that the two class solution would be the better choice. Any thoughts would be greatly appreciated!
No new simulation or theoretical results as far as I have seen. These days I tend to make it simple and use BIC. But in any given choice situation I would compare the 2 alternatives (your 2 and 3 class solutions) and see if the extra class teaches you anything substantively interesting that you don't see with 2 classes.
Thank you, Bengt - your advice is welcome and very helpful! In the past, you've suggested that I might do a Monte Carlo study of my own to explore some of the problems I've written about. I'd like to do that now that I've learned a little about MC analysis. In doing one, do you have some way of extracting the results from series of analyses to collect them into a single (or a few) tables? Or does it have to be done the old fashioned way -- by hand!? Thanks, Bruce
Thank you for both suggestions -- runall and MplusAutomation! I'll see whether I can learn how to use those at the same time I'm learning more about your terrific Monte Carlo modeling options before I bother Karen. -bac