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

All strata
concordant) model will not run

Within-subject variability for one
or more independent variable

•  Model will run

 Parameters estimated for only
those variables

Matched analysis:

 Advantage: control of
confounding factors

 Disadvantage: cannot separate
effects of time-independent
factors

Besides ASPIRIN, additional covariates

include AGE, GENDER, WEIGHT, and

HEIGHT. We restate the model from the previ-

ous section at left, which also includes 213

dummy variables for the 214 study subjects.

The output from running a conditional logistic

regression is presented on the left. Notice that

all of the coefficient estimates are zero with

their standard errors missing. This indicates

that the model did not execute. The problem

occurred because none of the independent

variables changed their values within any clus-

ter (subject). In this situation,allof the predic-

tor variables are said to beconcordantin all the

matching strata and uninformative with

respect to a matched analysis. Thus, the condi-

tional logistic regression, in effect, discards all

of the data.

If at least one variable in the model does vary

within a cluster (e.g., a time-dependent variable),

then the model will run. However, estimated

coefficients will be obtained only for those vari-

ables that have within-cluster variability.

An advantage of using a matched analysis with

subjectas the matching factor is the ability to

control for potential confounding factors that

can be difficult or impossible to measure.

When the study subject is the matched vari-

able, as in the Heartburn Relief example,

there is an implicit control of fixed genetic

and environmental factors that comprise each

subject. On the other hand, as the Aspirin–

Heart bypass example illustrates, a disadvan-

tage of this approach is that we cannot model

the separate effects of fixed time-independent

factors. In this analysis, we cannot examine the

separate effects of aspirin use, gender, and

height using a matched analysis, because the

values for these variables do not vary for a

given subject.

EXAMPLE (continued)

logit PðXÞ¼b 0 þb 1 ASPIRINþb 2 AGE

þb 3 GENDER

þb 4 WEIGHT

þb 5 HEIGHTþ~

213

i¼ 1

giVi

CLR model

Variable Coefficient

Standard

Error

Wald

p-value

AGE 0..

GENDER 0..

WEIGHT 0..

HEIGHT 0..

ASPIRIN 0..

578 16. Other Approaches for Analysis of Correlated Data