Programming the statistical procedures from SAS

Proc Genmod GEE model lsmeans difference and its CI on the proportion scale

Reply
New User
Posts: 1

Proc Genmod GEE model lsmeans difference and its CI on the proportion scale

Dear All,

I am fitting a GEE model using proc genmod, the syntax is below. Dependent variable 'agree' has 2 levels (Yes/No). Independent variable 'method' has 2 levels as well (A vs. B). Variable caseid has 200 levels for the 200 cases involved in the study. There are repeated measurements on each case.  Each case was read by 4 investigators, we are only interested in the population average estimates, hence the GEE model. We are not interested in fitting the investigators as a random factor.

proc genmod data=pop descending;

  class caseid method;

  model agree=method /dist=bin link=logit type3;

  repeated subject=caseid/corr=cs;

  lsmeans method/diff cl ilink;

run;

I need to have the difference in proportion between the 2 methods (A - B) in the probability scale, not the logit scale.

This is not available, per SAS doc

ILINK

Computes and displays estimates and standard errors of LS-means (but not differences) on the inverse linked scale

The difference and its CI are all in the logit scale, ilink is not applicable for the difference (log(p1/(1-p1)) - log(p2/1-p2)), of course.

Does anyone know a way (maybe a SAS macro someone implemented) that can calculate the lsmeans difference and its CI in the probability scale (the inverse linked scale)?

I asked SAS technical support this question and was pointed to a support note which recommends 2 methods (either use the link=identity in proc genmend or use proc nlmixed). None of those 2 solutions is satisfactory.

Thanks

Isaac

Respected Advisor
Posts: 2,655

Re: Proc Genmod GEE model lsmeans difference and its CI on the proportion scale

See this discussion: https://communities.sas.com/thread/58661 for an approximate method for proportions from PROC LOGISTIC.  Since you have specified corr=cs, any comparison of different timepoints will have the same covariance, so that should be incorporated when calculating the approximate standard error of the difference..

Still, the NLMIXED approach is preferred.

Steve Denham

Ask a Question
Discussion stats
  • 1 reply
  • 662 views
  • 0 likes
  • 2 in conversation