1. I am running CFAs on a scale measured with three subtests. These subtests have 24 items each, thus I have 72 indicators. These item-level indicators are binary: the respondent was either correct or incorrect. According to the MPlus userís guide and material Iíve found online, WLSMV is the preferred estimation method. These are my questions: a. Some people say MLE is possible. I have tried it, and it appears I have too many indicators. Is there a way to run MLE on binary data? The reason I would want to use MLE is I would like to be able to see AIC and BIC model fit statistics. Thus I would be able to compare models. b. Am I correct in that WLSMV does not produce AIC and BIC? Are there equivalent statistics that it does produce? c. Am I correct in that with WLSMV can only compare models using the DIFFTEST function, and that can only be used for nested models? Un-nested models cannot be compared with WLSMV? d. Does anyone have recommendations of how to work with these models, such as alternative fit indices?
2. I will be attempting CFAs on a model which includes both binary and continuous item-level indicators. I will have 109 indicators. Would I have the same issues as above? Are there workarounds?
1a ML is quite possible with many binary items as long as you don't have a lot of factors. Each factor implies one dimension of numerical integration so computations get slower and less precise with increasing number of dimensions.
1b Right. Chi-square/CFI fit is given.
1c Right. AIC and BIC draws on ML so are not available for WLSMV.
1d. For non-nested models you can compare say CFI values.