Logistic Regression: A Self-learning Text, Third Edition (Statistics in the Health Sciences)

Modeling Correlated Dichotomous Outcomes with GEE

The programming of a GEE model with the infant care dataset is demonstrated using
PROC GENMOD. The model is stated as follows:

logit PðOUTCOME¼ 1 jXÞ¼b 0 þb 1 BIRTHWGTþb 2 GENDERþb 3 DIARRHEA

The code and output are shown for this model assuming an AR1 correlation struc-
ture. The code for specifying other correlation structures using the REPEATED
statement in PROC GENMOD is shown later in this section, although the output is
omitted.

First a PROC PRINT will be run on the infant care dataset. Again, the use of the follow-
ing LIBNAME statement assumes the permanent SAS dataset is stored on the C drive.

LIBNAME REF‘C:\’;

PROC PRINT DATA¼REF.INFANT; RUN;

The output for one infant obtained from running the PROC PRINT is presented. Each
observation represents one of the nine monthly measurements.

IDNO MONTH OUTCOME BIRTHWGT GENDER DIARRHEA

244 1 0 2850 2 0

244 2 1 2850 2 1

244 3 1 2850 2 0

244 4 0 2850 2 0

244 5 0 2850 2 0

244 6 0 2850 2 0

244 7 0 2850 2 0

244 8 0 2850 2 0

244 9 0 2850 2 0

The code for running a GEE model with an AR1 correlation structure follows:

PROC GENMOD DATA¼REF.INFANT DESCENDING;

CLASS IDNO MONTH;

MODEL OUTCOME¼BIRTHWGT GENDER DIARRHEA / DIST¼BIN LINK¼LOGIT;

REPEATED SUBJECT¼IDNO / TYPE¼AR(1) WITHIN¼MONTH CORRW;

ESTIMATE‘log odds ratio (DIARRHEA 1 vs 0)’DIARRHEA 1/EXP;

CONTRAST‘Score Test BIRTHWGT and DIARRHEA’BIRTHWGT 1, DIARRHEA 1;

RUN;

The variable defining the cluster (infant) is IDNO. The variable defining the order of
measurement within a cluster is MONTH. Both these variables must be listed in the
CLASS statement. If the user wishes to have dummy variables defined from any
nominal independent variables, these can also be listed in the CLASS statement.

The LINK and DIST options in the MODEL statement define the link function and
the distribution of the response. Actually, for a GEE model, the distribution of the

622 Appendix: Computer Programs for Logistic Regression