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

ALR: dichotomous outcomes only
GEE: dichotomous and other out-
comes are allowed

For many health scientists, an odds ratio mea-

sure, such as that provided with an ALR model,

is more familiar and easier to interpret than a

correlation measure. However, ALR models

can only be used if the outcome is a dichoto-

mous variable. In contrast, GEE models are

not so restrictive.

The ALR model is illustrated by returning to the

Aspirin–Heart Bypass Study example, which

was first presented in Chap. 15. Recall that in

that study, researchers examined the efficacy of

aspirin for prevention of thrombotic graft

occlusion after coronary bypass grafting in a

sample of 214 patients (Gavaghan et al., 1991).

Patients were given a variable number of artery

bypasses (up to six) and randomly assigned to

take either aspirin (ASPIRIN¼1) or a placebo

(ASPIRIN¼0) every day. One year later, each

bypass was checked for occlusion and the out-

come was coded as blocked (D¼1) or unblocked

(D¼0). Additional covariates included AGE

(in years), GENDER (1¼male, 2¼female),

WEIGHT (in kilograms), and HEIGHT (in

centimeters).

Consider the model presented at left, with

ASPIRIN, AGE, GENDER, WEIGHT, and

HEIGHT as covariates.

EXAMPLE

GEE vs. ALR

Aspirin–Heart Bypass Study

(Gavaghan et al., 1991)

Subjects: received up to six coronary

bypass grafts

Randomly assigned to treatment

group:

ASPIRIN¼

1 if daily aspirin

0 if daily placebo

8

<

:

Response (D): occlusion of a bypass

graft 1 year later.

D¼

1 if blocked

0 if unblocked

(

Additional covariates:

AGE (in years)

GENDER (1¼male, 2¼female)

WEIGHT (in kilograms)

HEIGHT (in centimeters)

Model:

logit PðXÞ¼b 0 þb 1 ASPIRINþb 2 AGE

þb 3 GENDERþb 4 WEIGHT

þb 5 HEIGHT

572 16. Other Approaches for Analysis of Correlated Data