Thank you for the suggestion.  This occurred to me too but while there are nearly 500 facilities, the median sample size per facility is 150 and an average prevalence of about 10%.  My intuition is that except for a handful of the largest facilities the sample size isn't really large enough to support variable selection methods, but might be worth a try.  ... View more

Hi Rick, Can you say a little more about which specific termination criteria might be adjusted?  Maxiter and maxfunc are not a problem since I have "GCONV convergence criterion satisfied"  well before the maximum number of iterations and functions are reached.  I have set gconv=1E-9[2] but with no effect.  NLMIXED is like a machine with many nobs to adjust, and I am not sure where to really start. Greg ... View more

Yes, indeed!  Thanks for the correction. ... View more

It sounds like you just need to use events/trials syntax and then specify institutions as the SUBJECT= effect indicating that the sets of trials are correlated within institutions. The data would contain one observation for each quarter. For example, if you assume an AR correlation structure across quarters: proc genmod class instID quarter group; model NumPos / NumSubj = group; repeated subject=instID / type=AR within=quarter; run; ... View more