Dear all, I'm conducting a bi-factor EFA in a large dataset (n=13150) of a 21-item questionnaire with a 4-point likert scale. Whenever < 25% of the responses were missing, I've inputted the average for the available items and afterwards rounded the imputed values so the variables could become categorical. Initially, I did not specify that the variables were categorical and my syntax looked like: USEVARIABLES ARE BDIR1-BDIR21; ANALYSIS: TYPE= EFA 3 3; ROTATION= BI-CF-QUARTIMAX;
Than I declare the indicators as categorical (CATEGORICAL = BDIR1-BDIR21;), ran a new analysis and the results changed substantially in the subgroup factors. I thought the reason could be the fact that I was using a different estimator. However, I tried to use MLR and Mplus crashes before I can get an output. Therefore, I would like to ask: 1) Is it correct do use categorical items (in a bi-factor analysis) and not declare them as categorical (as in CATEGORICAL=)? I was considering to refer to Jennrich and Bentler, since in their paper they also apparently used continuous factor indicators. 2) I haven't seen any example of a bifactor analysis using the "CATEGORICAL = " argument. Am I doing the right way? 3) Do you have idea why am I getting different results (if it's in fact due to the estimator)? Thanks ! Ricardo
The comparison between treating the items as continuous or categorical depends largely on whether or not you have strong floor or ceiling effects (piling up at the end points of the Likert scale).
There should be no problem in declaring the items as categorical in bi-factor EFA. If you have a problem with MLR, you may want to send the input, output, data and license number to Support@statmodel.com.
So, if I have floor of ceiling effect it would be better to declare them as continuous (and maybe even not round them)? Do you know any paper in the literature approaching it (or descrbing) this procedure with categorical indicators? I'll send the files right away! Thanks in advance