I was running Latent class analysis with 19 indicators on problem behavior. The output showed analyses terminated normally but I did receive the following warnings in both 2 and 3 class solutions.
3-class solution: IN THE OPTIMIZATION, ONE OR MORE LOGIT THRESHOLDS APPROACHED AND WERE SET AT THE EXTREME VALUES. EXTREME VALUES ARE -15.000 AND 15.000. THE FOLLOWING THRESHOLDS WERE SET AT THESE VALUES: * THRESHOLD 1 OF CLASS INDICATOR L3A12DM FOR CLASS 1 AT ITERATION 4 * THRESHOLD 1 OF CLASS INDICATOR L3A11DM FOR CLASS 1 AT ITERATION 5 * THRESHOLD 1 OF CLASS INDICATOR L3A13DM FOR CLASS 1 AT ITERATION 6 * THRESHOLD 1 OF CLASS INDICATOR L3A10DM FOR CLASS 1 AT ITERATION 41
WARNING: WHEN ESTIMATING A MODEL WITH MORE THAN TWO CLASSES, IT MAY BE NECESSARY TO INCREASE THE NUMBER OF RANDOM STARTS USING THE STARTS OPTION TO AVOID LOCAL MAXIMA.
The sample size is about 10, 000 cases. The overall prevalence of problem behavior is ranged from 10%-38% with few indicators having small prevalence, like those indicators listed in the warning are having the smallest prevalence¡XL3A12DM (0.2%), L3A11DM (4.5), L3A13DM (1.5%). Is the warning coming from the small prevalence? I ran the analysis with 10 random values (STARTS=100 10), but still got the same warning.
You are showing two warnings. In the second one, you are reminded that you need to replicate the best loglikelihood in order to accept the solution.
In the first one, you are told certain thresholds are being fixed. This is not a problem. It simply means that in some classes there is a probability of zero or one of subscribing to the item. This actually may be helpful in interpreting the classes.
Thanks, Linda. Regarding to the second warning that one needs to replicate the likelihood-- I re-ran the model with different starting values and starts=500 20 couple times, however, I still received the same warning. "WARNING: WHEN ESTIMATING A MODEL WITH MORE THAN TWO CLASSES, IT MAY BE NECESSARY TO INCREASE THE NUMBER OF RANDOM STARTS USING THE STARTS OPTION TO AVOID LOCAL MAXIMA".
Any suggestions? Or should I assume this is the best I can have and accept the solution?
The warning is given whenever a model with more than two classes is estimated. You need to look at the output to see if you replicated the best loglikelihood. You should increase the STARTS option to STARTS= 500 50; for example if you did not. If this is not clear to you, please send your input, data, output, and license number to support at statmodel.com.
ksk posted on Wednesday, January 22, 2020 - 2:13 pm
Dear Dr. Muthen,
I'm running LCA models. I got the 5-class solution is the best. For a 6-class model, I received similar a warning shown in the first posting:
"IN THE OPTIMIZATION, ONE OR MORE LOGIT THRESHOLDS APPROACHED AND WERE SET AT THE EXTREME VALUES.EXTREME VALUES ARE -15.000 AND 15.000.THE FOLLOWING THRESHOLDS WERE SET AT THESE VALUES: * THRESHOLD 1 OF CLASS INDICATOR MARRY6 FOR CLASS 1 AT ITERATION 77"
You already said that warning can help interpret the results. My questions are 1) How can I interpret the results then? 2) With this warning, can I still trust model evaluation statictics (e.g., AIC, BIC, entropy, bootstrap) from the 6-class solution? The statistics show that the 6-class is not better than the 5-class one.