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

X. An Example We review the guidelines recommended to this
point through an example. We consider a car-
diovascular disease study involving the 9-year
follow-up of persons from Evans County, Geor-
gia. We focus on data involving 609 white
males on which we have measured six vari-
ables at the start of the study. These are cate-
cholamine level (CAT), AGE, cholesterol level
(CHL), smoking status (SMK), electrocardio-
gram abnormality status (ECG), and hyperten-
sion status (HPT). The outcome variable is
coronary heart disease status (CHD).

In this study, the exposure variable is CAT,

which is 1 if high and 0 if low. The other five

variables are control variables, so that these

may be considered as confounders and/or

effect modifiers. AGE and CHL are treated con-

tinuously, whereas SMK, ECG, and HPT, are

(0, 1) variables.

The question of interest is to describe the

relationship betweenE(CAT) andD(CHD),

controlling for the possible confounding and

effect-modifying effects of AGE, CHL, SMK,

ECG, and HPT. These latter five variables are

theCs that we have specified at the start of our

modeling strategy.

To follow our strategy for dealing with this

data set, we now carry out variable specifica-

tion in order to define the initial model to be

considered. We begin by specifying theVvari-

ables, which represent the potential confoun-

ders in the initial model.

In choosing theVs, we follow our earlier re-

commendation to let theVs be the same as

theCs. Thus, we will letV 1 ¼AGE,V 2 ¼CHL,

V 3 ¼SMK,V 4 ¼ECG, andV 5 ¼HPT.

We could have chosen otherVs in addition to the

fiveCs. For example, we could have considered

Vs that are products of twoCs, such asV 6 equals

AGECHL orV 7 equals AGESMK. We could

also have consideredVsthataresquaredCs,

such asV 8 equals AGE^2 orV 9 equals CHL^2.

EXAMPLE

Cardiovascular Disease Study

9-year follow-up Evans County, GA

n¼609 white males

The variables:

CAT|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl};AGE;CHL;SMK;ECG;HPT

at start

CHD¼outcome

CAT: (0, 1) exposure

AGE;CHL:continuous

SMK;ECG;HPT:ð 0 ; 1 Þ

)

control

variables

E¼CAT? D¼CHD

controlling for

AGE|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl};CHL;SMK;ECG;HPT

Cs

Variable specification stage:

Vs: potential confounders in initial

model

Here,Vs¼Cs:

V 1 ¼AGE;V 2 ¼CHL;V 3 ¼SMK;

V 4 ¼ECG;V 5 ¼HPT

Other possibleVs:

V 6 ¼AGECHL

V 7 ¼AGESMK

V 8 ¼AGE^2

V 9 ¼CHL^2

188 6. Modeling Strategy Guidelines