What are general guidelines for deciding on the number of classes? I know the model with the larger Loglikelihood Ho is preferable. Is it possible to obtain Ho values that are positive? If this happens, should you still prefer the larger Ho? I know you also want a model with lower information criteria. What about the Pearson chi-square and Likelihood ratio chi-square, should these be lower or higher for the preferred model and what do the p-values indicate? I am testing a one- versus two-class model of a construct but I have not been able to obtain similar Ho values from TECH 11 and the previous model even though I increased the number of starting values (500 50). Is this a red flag? Also, do you have any general advice for deciding on the number of starting values and final stage optimization? I have a dataset of over 20,000 cases and 7, 18, and 25 binary variables for three investigations.
These are big topics - the November Mplus short courses are recommended. Brief answers follow. See also the Nylund et al SEM article on our web site.
H0 loglikelihood values can be positive. Yes, higher LL is better.
Pearson and LR chi-square testing for frequency tables should have small values with p not too low - see literature on freq table testing.
Difficulty in replicating the LL can be a sign that this many classes is not needed.
Sung Kim posted on Thursday, October 14, 2010 - 6:21 pm
I want to fit an FMM on a sample of 3,200 (Women = 2,800; Men = 400) with 14 continuous indicators (3 factors). I want to include gender and age as covariates. My concern is the fact that I don't know how the sample size discrepancy between women and men will affect the result. I have thought about a few plans:
1. Combine all of them and fit an FMM (N = 3,200). 2. Randomly select 400 women out of 2,800 and combine them with the men (N = 800). 3. Fit FMMs separately for each gender
#3 is least favorable for me, though. What would be your recommendation?
Also, my overall plan is to fit an unconditional FMM first without the two covariates and choose the best fitting model (e.g., 3-factor 4-class model). Next, I add the covariates to the best FMM and fit an conditional model to see the effects of the covariates. Is this an appropriate plan?