If you are missing only the outcome or dependent variable then there is no need to impute the missing values. You may be able to use the maximum likelihood or the pseudo maximum likelihood (to account for survey design)using the observed data. Two exceptions are:
(1) you have missing data in covariates (2) You have variables that are not part of your model but are highly predictive of the outcome. In both these cases imputations might help. For imputations, you might want to consider this as a one record per person (that is, variables from different waves are horizontally concatenated) and then apply Proc MI. This way you are assuming that the vector of variables from all the waves has a multivariate normal distribution and modeling the joint correlation. Raghu --On Tuesday, September 09, 2008 11:05 AM -0400 Anne Stephenson <[email protected]> wrote: > Hi. > > I am using MI to deal with missing data. My dataset contains longitudinal > data over a 10 yr period (not all subjects are followed for the full 10 > years however). In the data model stage, I am wondering if I should > include the subject ID as a sample design variable to denote "clustering" > of the data within subjects (ie. repeated measures on individuals). Is > this appropriate? > > Thanks for any input. > Anne > > Anne Stephenson MD, FRCPC > Divison of Respirology > St. Michael's Hospital > 30 Bond Street, Room 6-040 > Toronto, Ontario > M5B 1W8 > > Tel: 416-864-6060 x 4103 > Fax: 416-864-5651 > > _______________________________________________ > Impute mailing list > [email protected] > http://lists.utsouthwestern.edu/mailman/listinfo/impute > >
