Hi, since I'am beginner with CFA, one question: I have made on my data an EFA (all variables are categorical) to determine how many factors I have (with the parallel analysis) and so on. Then I build a CFA accordingly to the EFA results and it turned out for the final model that CFI is 0.427, TLI is 0.468 and RMSEA 0.199 which indicates a poor model fit. However, since my factors in the EFA explain a little bit more than 50% of the total variance, so can I expect a good model fit at all? If no, is there some criteria in the Mplus output which allows me to judge the quality of the model?
I'm not sure how you decided on the number of factors with your EFA. The default estimator does not give fit statistics. You can change the estimator to WLSMV using the ESTIMATOR option of the ANALYSIS command and you will obtain fit statistics. If you have a lot of cross-loadings in the EFA and fix them to zero in the CFA, this could lead to poor model fit. You can also ask for modification indices in the CFA and see where the model misfit is.
Hi, the number of factors was decided with the help of parallel analysis which is a modification of "eigenvalues larger than one" criteria. When do you consider a loading as cross-loading? The absolute loadings (even after varimax) are pretty small (only 3-4 out of 67x9 are larger than 0,7). To get a reasonable interpretable model we decided to take all variables for estimating a factor where the absolute loading is larger than 0.4.
When I do an EFA, I use a strict rule. I require a factor loading to be two times as large for the factor that it should load on in comparison to other factors. I can't fully understand what you are doing, but if you go from an EFA to a CFA and get such poor model fit, you are most likely fixing factor loadings to zero and that is causing the poor model fit. You may benefit from purchasing our Day 1 course handout where we go through EFA in detail and also show how to do an EFA in a CFA framework. We use EFA to determine items that perform poorly and eliminate them if they seem not to be good measures.
my fault, I'am still thinking too much in terms of EFA. I tried to force the correlation matrix between the factors to be the identity matrix which is of course nonsense. When I gave up that then my CLI/TLI goes up to 0.7/0.8 and the RMSEA goes down to 0.1 which is much better then before.