Convergence Problem in GENMOD

[oakHILLS68]

I'm having problems with the following GENMOD code:

proc genmod descending data=dat1;
where DV ne .;
class subjectid timepoint / param=glm;
model vitalstatus = DV  timepoint DV*timepoint / d=b type3 maxiter=100;
repeated subject=subjectid / type=cs corrw ;
estimate 'at time 0' intercept 0 &v 1 timepoint 0 0 0 0 0 0 0 0 &v*timepoint 1 0 0 0 0 0 0 0 / exp;
...
estimate 'at time 72' intercept 0 &v 1 timepoint 0 0 0 0 0 0 0 0 &v*timepoint 0 0 0 0 0 0 0 1 / exp;run;

My continuous dependent variable is measured at 8 time points per subject, and is highly correlated across time.  The Pearson coefficients for timepoint pairs are all in the range 0.6-0.8.

Even with a type = cs correlation structure, I cannot get this to work.

If I ignore the repeated measure aspects, this works just fine.  But what's the problem with even a simple CS structure?

Anyone have ideas?

[SteveDenham · Accepted solution]

Wild guess: You have missing values at different timepoints.  If that is the case, add withinsubject=timepoint as an option to your REPEATED statement. You probably ought to analyze using PROC GEE as it is better equipped for handling missingness.

Wild guess 2: You have multiple measurements at the same timepoints.  Actually, this should cause a full stop of GENMOD.

Not so wild guess: Something other than this that we don't have enough information to go on. Things like NOTES, WARNINGS, ERRORS from the log file.  Dataset size (number of records per DV). Messages in the output file. Possible complete or quasi-separation (always something to look out for with a binomial distribution).

Something that might help: Get the parameter estimates from the model without the repeated measures aspect, and use those as starting values by specifying them in an INITIAL= and INTERCEPT= option.

SteveDenham