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

Assumption:

Responses

correlated within

clusters

independent

between clusters :

8

>>

>><

>>

>>

:

Ignoring within-cluster correlation

+

Incorrect inferences

II. An Example (Infant
Care Study)

GEE vs. standard logistic regres-
sion (ignores correlation)

 Statistical inferences may differ

 Similar use of output

Data source: Infant Care Study in
Brazil

Subjects: 168 infants
136 with complete data

Response (D): weight-for-height
standardized (z) score

D¼
1 ifz< 1 ð‘‘Wasting’’Þ
0 otherwise



Independent variables:
BIRTHWGT (in grams)
GENDER

DIARRHEA¼

1 if symptoms

present

in past month

0 otherwise

8

>>

><

>>>

:

A common assumption for correlated analyses

is that the responses are correlated within the

same cluster but are independent between dif-

ferent clusters.

In analyses of correlated data, the correlations

between subject responses often are ignored in

the modeling process. An analysis that ignores

the correlation structure may lead to incorrect

inferences.

We begin by illustrating how statistical infer-

ences may differ depending on the type of

analysis performed. We shall compare a gene-

ralized estimating equations (GEE) approach

with a standard logistic regression that ignores

the correlation structure. We also show the

similarities of these approaches in utilizing

the output to obtain and interpret odds ratio

estimates, their corresponding confidence

intervals, and tests of significance.

The data were obtained from an infant care

health intervention study in Brazil (Cannon

et al., 2001). As a part of that study, height

and weight measurements were taken each

month from 168 infants over a 9-month period.

Data from 136 infants with complete data on

the independent variables of interest are used

for this example.

The response (D) is derived from a weight-for-

height standardized score (i.e.,z-score) based

on the weight-for-height distribution of a refer-

ence population. A weight-for-height measure

of more than one standard deviation below

the mean (i.e., z<1) indicates “wasting”.

The dichotomous outcome for this study is

coded 1 if thez-score is less than negative 1

and 0 otherwise. The independent variables

are BIRTHWGT (the weight in grams at

birth), GENDER, and DIARRHEA (a dichoto-

mous variable indicating whether the infant

had symptoms of diarrhea that month).

Presentation: II. An Example (Infant Care Study) 493