I am using mplus to fit a latent curve model to an attitude measure observed over time. I wish to use the random intercept and slope relating to this variable as predictors of a consumer behavior measured at the last occasion (i.e., the last time the attitude measure was observed). There are missing data, however, for the consumer behavior. First, should I use all available data to estimate the model, or should I delete those cases with missing data on the consumer behavior? It seems to me that I should use whatever data are available. Second, given that the consumer behavior was measured at the last occasion when the repeated measure was taken, should the consumer behavior measure and the last measure of the attitude variable be allowed to correlate?
I would use all available data. Bengt did a small study of this several years ago and found that this is preferable to using listwise deletion. If you know the variables that are related to the missingness, you can include them in your model and covary them with all of the other dependent and independent variables.
Regarding the distal outcome, perhaps you could center the intercept growth factor at the last timepoint and then regress the distal outcome on i and s also if i and s are not too highly correlated.
One follow-up question: if i include measures related to the missingness, you suggest i allow them to covary with all the other DVs and IVs. Does this include both the observed repeated measures AND the latent growth factors? or only one and not the other?