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

Interaction prior to confounding:

 If strong interaction, then
confounding irrelevant

Assess interaction before confounding

Interaction may not be of interest:

 Skip interaction stage

 Proceed directly to
confounding

Interaction assessment is carried out next,

prior to the assessment of confounding. The

reason for this ordering is that if there is strong

evidence of interaction involving certain vari-

ables, then the assessment of confounding

involving these variables becomes irrelevant.

For example, suppose we are assessing the effect

of an exposure variableEon some diseaseD,and

we find strong evidence thatgenderis an effect

modifier of theE–Drelationship. In particular,

suppose that the odds ratio for the effect ofEon

Dis 5.4 for males but only 1.2 for females. In

other words, the data indicate that theE–Drela-

tionship is different for males than for females,

that is, there is interaction due to gender.

For this situation, it wouldnotbe appropriate

to combine the two odds ratio estimates for

males and females into a single overall adjusted

estimate, say 3.5, that represents an “average”

of the male and female odds ratios. Such an

overall “average” is used to control for the con-

founding effect of gender in the absence of

interaction; however, if interaction is present,

the use of a single adjusted estimate is a mis-

leading statistic because it masks the finding of

a separate effect for males and females.

Thus, we recommend that if one wishes to assess

interaction and also consider confounding, then

the assessment of interaction comes first.

However, the circumstances of the study may

indicate that the assessment of interaction is

not of interest or is biologically unimportant.

In such situations, the interaction stage of the

strategy can then be skipped, and one proceeds

directly to the assessment of confounding.

For example, the goal of a study may be to

obtain asingleoverall estimate of the effect of

an exposure adjusted for several factors,

regardless of whether or not there is interaction

involving these factors. In such a case, then,

interaction assessment is not appropriate.

EXAMPLE

Supposegenderis effect modifier for

E–Drelationship:

OR males = 5.4, OR females = 1.2

interaction

Overall average¼ 3 : 5

not appropriate

Misleading because of separate effects

for males and females

EXAMPLE

Study goal: single overall estimate.

Then interaction not appropriate

170 6. Modeling Strategy Guidelines