I am attempting to run Tech 14 for an LCA with 8 categorical indicators (4 are 2-level, 4 are 7-level). I followed these guidelines suggested in the manual:
1. Run without TECH14 using the STARTS option of the ANALYSIS command to find a stable solution if the default starts are not sufficient. 2. Run with TECH14 using the OPTSEED option of the ANALYSIS command to specify the seed of the stable solution from Step 1. 3. Run with LRTSTARTS = 0 0 40 10; to check if the results are sensitive to the number of random starts for the k class model.
At step 3 I am getting the following Warning message:
WARNING: 1000 OUT OF 1000 BOOTSTRAP DRAWS DID NOT CONVERGE. THE P-VALUE MAY NOT BE TRUSTWORTHY. INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION.
I then successively increased the LRTSTARTS, and I am up to the point where I have LRTSTARTS = 0 0 1000 50, but I still receive the same error message. How far do I go before I give up? What does it even mean that the draws are not converging?
If number 1 has many replications of the bestloglikelihood and number 2 works, try number 3 with the default LRTSTARTS option. It may be that you need to use BIC and TECH11. TECH14 may have problems due to something about your model and data.
Sung Kim posted on Friday, February 24, 2012 - 8:55 am
I am doing an LPA using 16 continuous indicators using 2,100 subjects. I'd like to get BLRT statistics (k =5) and keep getting warning despite large random starts. However, with smaller number of random starts the best likelihood was replicated in my LPA solutions. What is your recommendation? May I see the loglikelihood values of individual bootstrap draws?
WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED IN 77 OUT OF 78 BOOTSTRAP DRAWS. THE P-VALUE MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION.
The 20 replications you have (for the K-class estimation in BLRT) is not enough.
Sung Kim posted on Tuesday, February 28, 2012 - 1:33 pm
Thanks. So, I don't need to set the number of random starts and optimizations for the k-1 model in the LRTSTARTS option. The Mplus users guide says that one optimazation is carried out for the unperturbed set of starting values (the default for the k-1 class model). My question is how much we can trust the loglikelihood with unperturbed starting values? I specified like the following:
Sung Kim posted on Thursday, March 01, 2012 - 9:14 am
Almost the same result (see below). What do you think?
WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED IN 95 OUT OF 96 BOOTSTRAP DRAWS. THE P-VALUE MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION.