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

Confounding assessment when
interaction present (summary):

 Compare tables of ORs and CIs

 Subjective – debatable

 Safest decision – control for all
potential counfounders

V. The Evans County
Example Continued

If, overall, we decide thatyes, the asterisked

confidence intervals, which exclude V 3 , are

narrower than those for the gold standard

table, we would conclude that precision is

gained from excludingV 3 from the model. Oth-

erwise, if we decideno, then we conclude that

no meaningful precision is gained from

droppingV 3 , and so we retain this variable in

our final model.

Thus, we see that when there is interaction and

we want to assess both confounding and preci-

sion, we must compare tables of odds ratio

point estimates followed by tables of odds

ratio confidence intervals. Such comparisons

are quite subjective and, therefore, debatable

in practice. That is why the safest decision is to

control for all potential confounders even if

someVs are candidates to be dropped.

We now review the interaction and confound-

ing assessment recommendations by returning

to the Evans County Heart Disease Study data

that we have considered in the previous chap-

ters.

Recall that the study data involves 609 white

males followed for 9 years to determine CHD

status. The exposure variable is catecholamine

level (CAT), and theCvariables considered for

control are AGE, cholesterol (CHL), smoking

status (SMK), electrocardiogram abnormality

status (ECG), and hypertension status (HPT).

The variables AGE and CHL are treated contin-

uously, whereas SMK, ECG, and HPT are (0, 1)

variables.

EXAMPLE

Evans County Heart Disease Study

n¼609 white males

9-year follow-up

D¼CHD(0, 1)

E¼CAT(0, 1)

Cs:AGE|fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl};CHL

continuous

SMK|fflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflffl};ECG;HPT

ð 0 ; 1 Þ

EXAMPLE (continued)

CI*

narrower

than CI?

Precision

gained from

excluding V 3

No precision gained

from excluding V 3

Ye s

No

Retain V 3

Exclude V 3

Presentation: V. The Evans County Example Continued 223