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

Presentation

I. Overview

Modeling

outcomes with

dichotomous

correlated

responses

FOCUS

Examples of correlated responses:

1. Different members of the same
   house-hold.


2. Each eye of the same person.


3. Several bypass grafts on the
   same subject.


4. Monthly measurements on the
   same subject.

Observations can be grouped into
clusters:

Example
No.

Cluster Source of

observation

1 Household Household

members

2 Subject Eyes

3 Subject Bypass

grafts

4 Subject Monthly

repeats

In this chapter, we provide an introduction to

modeling techniques for use with dichotomous

outcomes in which the responses are corre-

lated. We focus on one of the most commonly

used modeling techniques for this type of anal-

ysis, known as generalized estimating equa-

tions or GEE, and we describe how the GEE

approach is used to carry out logistic regres-

sion for correlated dichotomous responses.

For the modeling techniques discussed previ-

ously, we have made an assumption that the

responses are independent. In many research

scenarios, this is not a reasonable assumption.

Examples of correlated responses include

(1) observations on different members of the

same household, (2) observations on each eye

of the same person, (3) results (e.g., success/

failure) of several bypass grafts on the same

subject, and (4) measurements repeated each

month over the course of a year on the same

subject. The last is an example of a longitudi-

nal study, since individuals’ responses are

measured repeatedly over time.

For the above-mentioned examples, the obser-

vations can be grouped into clusters. In exam-

ple 1, the clusters are households, whereas

the observations are the individual members

of each household. In example 4, the clusters

are individual subjects, whereas the observa-

tions are the monthly measurements taken on

the subject.

492 14. Logistic Regression for Correlated Data: GEE