Mark posted on Wednesday, December 11, 2013 - 5:02 pm
Hello Dr. Muthen,
I have been working with both Geoff McLachlan's EMMIX program and your MPlus program to conduct some finite mixture modeling analyses on the same data. The dataset contains 2 continuous observed indicator variables across 150 cases.
I have run all of the EMMIX analyses using the "unrestricted" option, which creates a completely free model in terms of means, variances, and covariances.
I have also written the MPlus code to run an unrestricted model as well, i.e., free means, variances, & covariances.
I have stumbled upon something odd. For the one class and two class models, both EMMIX and MPlus generate exactly the same results across all relevant model fit statistics/indexes (log likelihood, AIC, BIC, mixing proportions), to the second decimal place.
However, when I move to a three class model, the results generated by EMMIX and MPlus diverge (different values of log likelihood, AIC, BIC, mixing proportions).
I don't know if you have used EMMIX, but might you have any thoughts as to what might be different computationally across the two programs when moving from a two class to a three class (and more) model? What might be responsible for the programs diverging in their estimations in moving to a three class model, whereas their estimations for the one and two class solutions are identical?
You may be comparing a local and a global max, or two local max. With a higher number of classes the occurrence of local solutions increases so that you want to use more random starts (STARTS option in Mplus).
Mark posted on Thursday, December 12, 2013 - 4:33 pm
Thank you --- I will give that a try and compare results.