I have plenty of missing data in my longitudinal sample with 5 measurement points, both on dependent variables (5% first, 40% last measurement point) and covariates (20-30% missing). Biggest problem are my time-varying covariates, for which missings are always problematic.
Sample size = 800.
(1) Multiple Imputation, even with constraints, gives pretty weird results. Example: categorical dependent variable, observed distribution 50% 0, 30% 1, 15 2, 5% 3 leads to estimated distributions with >80% in category 3, in all 5 imputations.
Maybe the reason is that my categorical dependent variable is skewed to the left? Any ideas? Read the tutorials about MI but they didn't help me in my case.
(2) I heard that using FIML (full information maximum likelihood) is as good as using multiple imputation. I'm running mixed model analyses in SPSS, but I don't think FIML in SPSS is possible for that, or is it?