Alex Pavluck <apavluck@GMAIL.COM> replied (to me): > David, > Thanks for your reply. The problem is that we want to make sure that > our main effects are not confounded by screening. So, to check this we > are running a model of only screeners and comparing it to the total > population main effects. However, screening is a time dependent > variable because we asked it three times. If they are screening in > internal 1 and 3 but not 2 then they will only be considered during > internal 1 and 3 and all person-time and covariate information will be > dropped for interval 2. So, unless I am missing the obvious, it is not > so easy to drop missing because missing in interval 1 can be present in > interval 2.

It seems to me that you would want to include screening and non-screening, so you *can* determine whether your main effects are related to screening. If you only work with 'screening' cases, I don't se how you can separate the effects.

The temporal features can complicate things, but I don't see that you can safely ignore screening/non-screening when doing a full analysis of your data. IMHO, anyway.

> Oh, it is a population based survey.

Then you ought to be looking at using your sampling weights and design effects in PROC SURVEYREG or PROC SURVEYLOGISTIC if at all possible.

David -- David Cassell, CSC Cassell.David@epa.gov Senior computing specialist mathematical statistician