10-07-2011 03:22 PM
My glimmix code is:
proc glimmix data=mydata method=quad;
class group time subject;
model y = group time group*time / s link=logit dist=binary;
lsmeans group*time / ilink;
random int / subject=subject;
Suppose there are 5 groups and 10 time points (5X10 design). For one of the groups (i.e., group 2), there are no data for 4 of the 10 time points. Does this pose a problem for model convergence and/or validity of results? For example, is the estimate and standard error produced by the LSMEANS statement for group 1 at time 2 valid and interpreted in the same way as if one had a fully balanced design?
While I have read that mixed models in general can handle unbalanced data, I have not seen any literature on missing cells.
Any references or thoughts on the matter would be appreciated.
10-10-2011 07:54 AM
Take a look at the documentation regarding estimable functions and especially Type III estimable functions. These marginal estimators are valid, but with this caveat. The missing values need to be at least "missing at random". My problem is always "How do I tell if the missing data is missing at random, missing completely at random, or systematically missing?" You'll have to look at the processes that generated the data at hand.
Regarding problems for convergence--missingness can lead to problems in this area, both in non-convergence or convergence to a local extremum rather than the global extremum. It seems to be even more of a problem for binary endpoints for me, and I don't know why.
The bottom line, in my opinion, is that in most cases the unbalanced situation is by far best handled by mixed model methodology.