Dennis Li posted on Thursday, July 28, 2016 - 2:06 am
I am running an analysis with 8-10 binary milestone variables measured over 5 time periods. Once a milestone becomes 1, there's no transitioning backwards to 0. Also, the proportion of 1s starts high >80% for most milestones. Without adding covariates yet, I put all 50 variables into a RMLCA to see if it would run, and, to my surprise, the models converged. I was able to calculate VLMRLRT and BLRT for all models, and entropy values for k=1-5 were >.96, with fairly substantive interpretations.
However, for models k>2, I received a NON-POSITIVE DEFINITE FIRST-ORDER DERIVATIVE PRODUCT MATRIX error, with the issue being a threshold parameter and a small condition number. Scouring the message boards, I found an old discussion (URL copied below) in which Bengt says "Thresholds that go large like that are harmless causes of the non-pos def message" but also "You don't want a very small condition number and [-0.783D-15 in this case] is very small."
Should I just take that my models are non-identified and not even look at the results? Is it even a good idea to run multiple indicators with multiple time points in the same model? I am quite new to LCA and am stumbling with some of the basic ins and outs. Thanks.
If you have very large thresholds that weren't fixed by Mplus, that can give you a small condition number. This is not necessarily harmful, however. Your model may well be identified.
A bigger issue is perhaps that your model does not reflect the no transitioning backwards from 1 to 0. That feature is characteristic of discrete-time survival analysis (although that is typically done for a single variable).