GEE model question

[deleted_user]

Dear All;

When I run GEE model, I usually use 'Deviance' or 'log likelihood' to check the model fit or compare two models. However, when I change the different correlation-structures in the repeated statement, the 'Deviance' from the models with the two different correlation structures are the same. Do you have some suggestions for how to choose the best model for different correlation structures? Please see the models that I use.


proc genmod data=test;
class studyid allgp ;
model use=time allgp/ wald type3 d=bin;
repeated subject=studyid/ type=cs corrw ;
run;

proc genmod data=test;
class studyid allgp ;
model use=time allgp/ wald type3 d=bin;
repeated subject=studyid/ type=un corrw ;
run;

Thank.
Ruby.

[Olivier]

Hi Ruby.
You can use the QIC criterion (similar to the Akaike Information Criterion in usual models) that can be computed by this macro here (http://support.sas.com/kb/26/100.html). The smaller the better.
You can also use the ODS GRAPHICS simulated cumulative results if you are using SAS 9 : check there (http://support.sas.com/onlinedoc/913/getDoc/fr/statug.hlp/genmod_sect53.htm) for details.

Regards,
Olivier