Some things that you might want to consider:   It is hard for me to understand how the same variable could be both binomially distributed and Poisson distributed, unless there is some aggregation in the Poisson distributed case. I would expect that the level-1 variable, patient_id is not the subject for the level-2 variable hospital.  Try:   random intercept/subject=hospital; Let the level-1 variable be "absorbed" by the distribution.   Unless outcome is multinomial, you don't need to include it in the CLASS statement.   These should lead to a reduction in model complexity that should allow GLIMMIX to proceed.   SteveDenham ... View more

An adjustment for multiple comparisons can be added as an option to the LSMEANS statement in GLIMMIX, which would be my choice. @StatDave may recommend GENMOD or GEE, as there doesn't appear to be additional random effects in play, only repeated measures.   SteveDenham ... View more

For at 0 and 100 both of the imputed variables truncated at 0 and 100, do you feel comfortable imputing using the FCS method, and then post-processing the two variables to have a max or min of 0 or 100, as those are what could be actually measured?   SteveDenham ... View more

One analysis would just be to compare the cumulative values for the fifth period. All that is needed there is a t test.  If you want to compare the cumulative profiles, then a MANOVA using PROC GLM could be used. Once you do the repeated measures analysis (PROC GLIMMIX with a Poisson distribution), you could look at partial areas under the curve using LSMESTIMATE statements.   SteveDenham ... View more

If the Hessian is full rank with a negative eigenvalue, you have two options.  One is to use a simpler model, the other is to find out which line of data (or lines) lead to the problem.  You could use a pseudo-jackknife approach to the latter, fitting the model while successively deleting a single observation.  If all of the reduced datasets lead to the same issue, then it is almost certainly a misspecification of the model issue.   SteveDenham ... View more

What does your output look like?  Are the estimates in the Covariance Parameters table all missing?  Are there warnings in the log, or included as part of the output?  More info will definitely help address your issue.   SteveDenham ... View more

If you have millions of IDs, an unstructured option is trying to calculate (N^2 + N)/2 covariance parameters, which is in the trillions or quadrillions.  By far a better choice would be a single variance component due to ID.   Try this instead:   proc glimmix data=mydata; class exposure(ref="0") sex (ref="0") age(ref="0") id //param=ref; model outcome=exposure sex exposure*sex age_2class /dist=poisson link=log; random intercept/ subject=id; *estimate "exposure" exposure 1 0/exp; *estimate "sex" sex 1 0/exp; *estimate "age" age 1 0/exp; /* Since age is not in the model, I have commented this out. Either change this to age_2class, or change the model statement */ lsmeans exposure sex exposure*sex/ilink; run; The LSMEANS or LSMESTIMATE statement will be preferable to trying to write ESTMATE statements.  In this case, I have commented them out as they are not the correct statements to obtain a single level estimate (no intercept term).   SteveDenham   ... View more