Correlation Matrix Specification in PROC GENMOD Repeated Statement

[lcmichael_unc]

I'm running a repeated measures analysis with GENMOD with a binary outcome and want to specify the best correlation matrix. The following code is used but changing the TYPE does not change the parameter estimates or the predicted probabilities. This seems somewhere between ridiculous and impossible. LINK=IDENTITY produces the predicted probability differences from the LSMEANS in the correct scale but the TYPE still has no impact on the output. All clues gratefully embraced.

proc genmod data=&library..qstrans_dm descending;
where RANDREG in ('SOF/LDV' 'EBR/GZR') and HCVGT2 in ('1A' '1B');
class USUBJID RANDREG (ref='SOF/LDV') HCVGT2 (ref='1B') CIRRHOSIS (ref='N') EVENT (ref='T2') / param=ref;
model ADHERENT = RANDREG HCVGT2 CIRRHOSIS EVENT / dist=binomial link=logit;
repeated subject=USUBJID / corrb covb type=CS;
contrast 'Regimen' RANDREG 1;
lsmeans RANDREG / ilink exp diff cl;

lsmestimate RANDREG 'HVN vs ZEP' 11 / exp cl;
run;

[SteveDenham]

The type of correlation matrix affects the standard errors and significance tests, and is often referred to as a variance-covariance matrix.  With that in mind, we should expect that lsmeans would not be affected unless there is missing data, and even then, the effect should be minor.  This is because the R part of the V matrrix is what is modeled with the type= option, and it is separate from the X part (where the fixed effects reside).  At least that is how I think about it.

SteveDenham