I presently have data on 158 categorical (1/0) items which were administered to 1,000 individuals. The data were collected in a planned missing data design, where a total of five different forms were used. Each form had a set of unique and common items (appx 20% of the 158 are common across the five forms). When running an EFA on these data, there are a number of missing cells due to nature of the design. The output I get from the analysis using WLSMV or ULS produces the WARNING: BIVARIATE TABLE OF X_ AND X_ HAS AN EMPTY CELL. Although I can get AIC/BIC using ML, I'm interested in the other ancillary statistics for model evaluation. Is there a way to control for this type of design and use the WLSMV or ULS estimators?
Dear Linda/Bengt, I am doing an exploratory factor analysis on binary items measuring safety culture in firms. Some items are not applicable to smaller firms and are recorded as missing. What is the best method to deal with this missingness? I have a variable measuring firm size.
Using maximum likelihood estimation would be the best, but with binary items each factor is one dimension of integration which makes EFA difficult. You could use DATA IMPUTATION to obtain imputed data sets and then use TYPE=IMPUTATION and weighted least squares to do the EFA.
Thanks, i do have the book. When data is MNAR he discusses about selection and pattern mixture models. I did not see about the performance of ML and Multiple imputation when we have 'not applicable items'. if i understand it correctly I read a section in the paper by Schafer and Graham(Missing data: state of the art) where they say when the missing values are out of scope - we can assume MAR.
The question is what an (implicitly) imputed value means for a subject for whom the question is not applicable. It's similar to imputing values for someone who drops out of a study due to death. There may be other approaches for this case. You may want to ask on SEMNET to hear if someone has a good reference.