Xu, Man posted on Wednesday, February 03, 2016 - 7:37 am
Dear Dr. Muthen,
I am applying ESEM analysis to a large number of bio-markers that are supposed to have some sort of homeostasis. The final chosen factor structure left quite a lot of near/zero loadings, so I suspect this is one of the reasons that fit indices such as TLI, CFI are rather poor in all tested models (althouh there is systematic improvement corresponding to model adjustment, and RMSEA can become acceptable too). I wonder if it makes any sense for me to use these traditional CFA fit indices, or is there a more suitable strategy for model comparisons in these situation?
I currently use estimator WLSMV, because some of the indicators have a censored regression link due to flooring effect, and many indicators are quite skewed as well.
I don't really know. BSEM could be a candidate if it weren't for the fact that we don't have censored for Bayes yet. You can check the Modindices for the residual correlations in ESEM to get a better fit.
Xu, Man posted on Friday, February 05, 2016 - 3:39 am
Thank you. I don't know much about BSEM, but do you mean fitting CFA type of analysis but put small priors on those with very low cross loadings? If it is CFA type of analysis, then maybe it is not easy to start everything from scratch, because choosing number of factors this way sounds challenging.
Maybe it is a good idea to use BSEM to fit the final model following ESEM, for dealing with the cross loadings. But it seems BSEM gives BIC and DIC as way of model comparison/fit. If this is the case, then I still don't know how well the model fit the data like the REMSE, TLI and CLI.
For the censoring and sknewness, it has been considered by some as acceptable to do +1 log transformation prior to data analysis.