I fitted two models that had autoregressive relationships for the factor scores. The first model I called full model has 35 parameters. The second model (reduced model) has 33 parameters. I had expected that the chisq for the full model was smaller than the reduced model. But it turned out the chisq for full model was bigger. Is that possible to get such kind of results? How should I interpret them? Thanks a lot.
Dear Dr. Muthen: I was not able to locate an answer in the forum so apologies if this has already been asked. The analyses I describe below are being conducted in Version 7.2.
In regard to conducting a difference test between an unrestricted model and a restricted model when TYPE=RANDOM and ESTIMATOR=ML: how is this done? I'm testing a LGM with an ordinal variable (5 categories) over 7 years. Because it is pubertal status and children were tested at different ages, I'm using t-scores. I have reviewed the instructions for difference testing when estimator=ML but my output does not include a scaling correction factor. It only provides the loglikelihood as shown below. I know this has been done in Mplus by others (Mendle, Harden, Brooks-Gunn, & Graber, 2010) using pubertal data like ours, but I am not clear on how to do it - can you provide any advice? Thanks!
The ML estimator does not require a scaling correction factor. Just take the difference in the two chi-square values and the difference in the number of parameters. See pages 486-487 of the user's guide for more information.