New Contributor
Posts: 3

Questions on Glimmix convergence and random level units

I have been using GLIMMIX to get the random effects estimates for second level units (counties). The data has over 40,000 level-1 records nested within 8 level-2 units (counties). The dependent variable is binary event indicator "event_pm" (1=event, 0=no event). There are no covariates except for county level random effect. The GLIMMIX program is shown below as well as the event frequencies for each level 2 county.

However, to my surprise, the model did not converge. Is it because 8 level-2 units is too few for the GLIMMIX model? Or is it because the number of events (event=1) for some counties are too few? One county has only over 30 events. What options can I use to overcome the non-convergence problem?

Any suggestions are appreciated.

PROC GLIMMIX data = ppage4 /*method=RSPL*/ noclprint pconv=1e-4;

class county;

model event_pm (event="1") =

/dist=binary  solution;

random intercept /subject=county solution cl;

run;

 Cumulative county  event_pmFrequency Frequency

------------------------------------------------

 1 0 6445 6445 1 1 212 6657 3 0 12436 19093 3 1 376 19469 5 0 1677 21146 5 1 59 21205 7 0 1216 22421 7 1 36 22457 9 0 12055 34512 9 1 377 34889 11 0 3855 38744 11 1 116 38860 13 0 1240 40100 13 1 32 40132 15 0 1673 41805 15 1 50 41855
Valued Guide
Posts: 684

Re: Questions on Glimmix convergence and random level units

There can be many reasons for nonconvergence, and it would be not for the reason you are worried about. It is not clear if it is getting close to convergence (just needing more steps) or is stuck far from the solution locus. It could be stuck on the inner or outer iteration of the doubly iterative RSPL method. I suggest you increase the maximum number of iterations and try different optimization techniques. Add

nloptions maxiter=1000 maxfunc=1000 technique=NRRIDG;

Or try TRUREG or CONGRA for technique.

For binary data, you could also change your estimation to method=LAPLACE on the procedure statement. But this might be v-e-r-y slow with the number of levels you have.

Discussion stats