Xu, Man posted on Tuesday, September 07, 2010 - 7:49 am
Dear Dr. Muthen,
I am using Mplus for a SEM model of categorical variables. A binary obesrved variable is the dependent variable. It is predicted by three latent variables, each has several ordered categorical indicators.
I found it makes a diference in the results wehther i declare the dependent vraiable to be categorical in the input file, even though it is a binary variable. Is declaring it as categorical variable the right way to it? Thanks!
If you use maximum likelihood where the default is logistic regression, you will obtain odds ratios.
Xu, Man posted on Tuesday, September 14, 2010 - 1:31 am
Dear Dr.s Muthen,
Many thanks. I have put estimator=ml to the model. But the path coeff results are not comparable across probit and logit models for the same specification, although the factor loading patterns are pretty similar. The logit results showed that some previously signficant path coefficients in probit model became insignificant, and some previously insigificant path coefficients became signficant. A warning from the logit model said that THE CHI-SQUARE TEST CANNOT BE COMPUTED BECAUSE THE FREQUENCY TABLE FOR THE LATENT CLASS INDICATOR MODEL PART IS TOO LARGE. It also took a very long time to run ( 40 min compared to 6 seconds under probit). Is this normal?
Xu, Man posted on Tuesday, September 14, 2010 - 6:41 am
A colleague just reminded me that I was using different estimators for the models with different links (ml for logit, wlsmv for probit). i guess this might be what caused the difference in results. thanks!
The time difference between ML and WLSMV is due to the fact that factors with categorical factor indicators require numerical integration when ML is used. And the differences you see in the results are because one is probit and one is logit.