Chris Weber posted on Wednesday, January 30, 2008 - 12:46 pm
I'm trying to run a random intercept CFA. I have four factors and 34 clusters. All of my indicators are categorical. Because there are 8 integration dimensions (and 812 integration points), I was only able to get the model to run using monte carlo integration. The other integration alternatives appear to be too computationally intensive, as indicated by an error message.I was also able to get the model to run using the WLSMV estimator. My question is which approach is appropriate, MLR using monte carlo integration, or WLSMV? I was under the assumption that a random intercept confirmatory model wouldn't even run using WLSMV. Is there any literature comparing these approaches?
I would go with the new WLSM approach. With 8 dimensions, Monte Carlo integration for ML doesn't give very precise loglikelihood results, although the estimates might be reasonably approximate. The new WLSM approach is briefly described in the Technical Appendices section on the web site, last entry. It gives a limited comparison with ML. Many more studies are needed.
I am doing a cross-sectional MLM analysis with 17 clusters with 500 cases. My research question is focused more on level 1 predictors (although a couple of level 2 predictors are also in the model). I tried REML in SAS, ML and MLR in Mplus. The results of regression coefficients for level 1 predictors are different. One predictor at level 1 is not significant under REML and ML, but significant under MLR. In this case, should I trust the results from ML or MLR? What would you suggest? Thank you.