Matt Keough posted on Thursday, April 09, 2015 - 11:38 am
I am running a multi-group CFA with categorical data (48 yes/no items) using WLSMV. Given low frequencies on some variables, I used the add frequency command. I still get an error message about empty cells, but I vaguely remember reading a post on the forum stating that we can ignore these errors if we use the add frequency command and if the model fit etc. is supported. Just to verify, is this the case?
I would not ignore this if there are many such messages. The correlations may become less reliable which ultimately influences the model estimates. I would also try ML as a comparison in such cases. See also our FAQ
Estimator choices with categorical outcomes
Matt Keough posted on Thursday, April 09, 2015 - 12:07 pm
Thanks very much. I have read through this previously and it was really helpful. Do you have any ideas on how to potentially fix the low frequency problem other than adding a frequency of .50. As it stands, I am getting quite a few empty cell errors.
For such a pair, combine the 2 binary variables into 1 3-category variable.
Or, don't use WLSMV.
Matt Keough posted on Thursday, April 09, 2015 - 1:30 pm
Combining items doesn't seem like the best option. From my understanding, ML estimation is not appropriate for categorical data? This may be an uninformed question, but does adding frequency change this? I should not that WLS and USLMV also do not work.
ML not being appropriate for categorical data is a common misunderstanding. It comes from the mistake of thinking that ML means treating the variables as continuous. ML is used for many models with different kinds of non-continuous outcomes, including IRT. This is also why ML is included in the FAQ I referred to.