R McDowell posted on Monday, March 02, 2015 - 8:54 am
I'm modelling a dataset using factor analysis, and have a number of models I wish to fit and compare. There is missing data. I was wondering if you had any advice on which of the following is better practice:
A. Generating a separate imputed dataset for each factor model in turn, using the given factor model as part of the imputation specification.
B. Create one dataset at the very beginning without reference to any modelling, and hence using the same data to fit each factor model in turn.
Any comments you have would be appreciated. Many thanks.
For imputation I would use B. But why not use ML under MAR ("FIML") instead of imputation?
R McDowell posted on Wednesday, March 11, 2015 - 9:20 am
Thank you for your guidance. My outcome is categorical and I was keen to compare the pairwise method with multiple imputation (MI). With MI output, can you calculate 95% CI for the RMSEA using the oft used estimate+- 1.96*standard error? Also, is it possible to compare nested models obtained from MI using the means from the chisquare test of model fit?
We don't compute RMSEA confidence limits with multiple imputation. As far as I know the Browne & Cudeck (1993) method has not been generalized for multiple imputations yet.
The means from the chi-square test of model fit should not be used for model testing with the WLS estimator. Currently Mplus computes multiple imputation based test of fit only with the ML estimator and all continuous variables. Details of that computation can be found in http://statmodel.com/download/MI7.pdf