I am having issues replicating the H0 loglikelihood value for the K-1 solution in the TECH11 output when comparing 5- and 6-class solutions in a LCA.
I have read "Using Mplus TECH11 and TECH14 to test the number of latent classes" and generally had good luck with the described method.
I have no problems replicating the loglikelihood values for either the c(5) or c(6) solutions themselves (with a modest number of starts, such as STARTS = 400 80), yet no matter how high I put the K-1STARTS for c(6) (I have gone up to 5000 1000), I cannot replicate the c(5) loglikelihood value in the TECH11 output (the loglikelihood for the c(5) = -5974.33, but in TECH11 it consistently = -5989.43).
Should I worry about this, and is there anything else I might try?
I assume this is a regular "exploratory" LCA so that the first class - which is dropped when doing the k-1 class model in the k-class run - doesn't have any parameter restrictions.
You can try stscale=10 or 20 or 30.
Mplus drops the first class of the 6-th class solution – this gives the starting values for the 5-th class solution. If there is a some kind of perfect indicator in the 6-class solution that is not there for the 5-class solution this can happen.
If you can match the classes between the 5-class and the 6-class solution, put the extra class as class one (you can do that by using stvalues or different optseed for the 6 class solution) so that the extra class come up as class 1 in the 6-class solution.
The loglikelihood is replicated > 10 times. However, the loglikelihood I get for this model slightly differs from the k-1 loglikelihood obtained with TECH11 in the next step (not only for 3 vs. 2 classes, but also for 4 vs. 3 and 2 vs.1). The differences are quite small (i.e. -3409.149 vs. -3409.151). Are these differences negligible?