Is there an alternative to Firth Penalized Likelihood for GLIMMIX?

[VXB]

Hello - 

I would like to know if there is an alternative to the Firth's regression (as available in PROC LOGISTIC) but for PROC GLIMMIX ?

I have clustered data (students within schools) and a rare binary outcome (referrals to an intervention with rates of about 1%) with quite some schools at zero, so any suggestion of an alternative approach will be welcome! 

Many thanks,

Van

[sbxkoenk]

Hello,

The GLIMMIX procedure does not include a FIRTH option. 

Are you aware of a GLMM version of the Firth approach? If you are aware of a statistical reference for this type of bias reduction for GLMM model, please post it here !

There are no special features for dealing with rare events built into GLIMMIX.  The algorithm is applied to fit any generalized linear mixed model and if there are no numerical problems encountered along the way, then the solution is found and reported.  If the data becomes extremely sparse, the numerical problems often result in nonconvergence.

If you run into non-convergence, there are several things you can try to avoid non-convergence / to get convergence.

Let us know what happens in your case!

Koen

[VXB]

Hello Koen,

Thank you for your answer!

I don't know of a statistical reference for this type of bias reduction for GLMM model, but will of course post if I find something.

I can get convergence if I stay reasonable with the number of predictors (less than 10-15 estimated coefficients) but I find my estimated predicted probabilities suspiciously low...and some increasing as I decrease the number of estimated coefficients. 

I will post as I learn more,

Thanks,

Van