using the MODELSE option in the REPEATED statement. The table titled “Con-
trast Estimate Results” contains the output requested by the ESTIMATE state-
ment. The odds ratio estimate for DIARRHEA¼1 vs. DIARRHEA¼0 is given as
1.2479. The table titled “Contrast Results for GEE Analysis” contains the output
requested by the CONTRAST statement. Thep-value for the requested Score test
is 0.3819.
Other correlation structures could be requested using the TYPE¼option in the
REPEATED statement. Examples of code requesting an independent, an exchange-
able, a stationary 4-dependent, and an unstructured correlation structure using the
variable IDNO as the cluster variable are given below.
REPEATED SUBJECT¼IDNO / TYPE¼IND;
REPEATED SUBJECT¼IDNO / TYPE¼EXCH;
REPEATED SUBJECT¼IDNO / TYPE¼MDEP(4);
REPEATED SUBJECT¼IDNO / TYPE¼UNSTR MAXITER¼1000;
The ALR approach, which was described in Chap. 16, is an alternative to the GEE
approach with dichotomous outcomes. It is requested by using the LOGOR¼option
rather than the TYPE¼option in the REPEATED statement. The code requesting the
alternating logistic regression (ALR) algorithm with an exchangeable odds ratio
structure is:
REPEATED SUBJECT¼IDNO / LOGOR¼EXCH;The MAXITER¼option in the REPEATED statement can be used when the default
number of 50 iterations is not sufficient to achieve numerical convergence of the
parameter estimates. It is important that you make sure the numerical algorithm
converged correctly to preclude reporting spurious results. In fact, the ALR model in
this example, requested by the LOGOR¼EXCH option, does not converge for the
infant care dataset no matter how many iterations are allowed for convergence. The
GEE model, using the unstructured correlation structure, also did not converge, even
with MAXITER set to 1,000 iterations.
Generalized Linear Mixed Models with Dichotomous
Outcomes
Generalized linear mixed models (GLMMs) can be run in SAS using PROC GLIMMIX
or PROC NLMIXED. Our focus here is to illustrate PROC GLIMMIX. GLMMs are a
generalization of linear mixed models in that they allow for the inclusion of fixed and
random effects with nonnormally distributed outcome data. PROC GLIMMIX is a
flexible procedure that can run relatively simple or quite complex models. We begin
our illustration of PROC GLIMMIX by demonstrating how a standard logistic regres-
sion is run using the Evans County data. Typically, we would not use PROC GLIM-
MIX to run a standard logistic regression, but for illustration we present it here as a
626 Appendix: Computer Programs for Logistic Regression