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

Presentation

I. Overview

FOCUS

Assessing

confounding

and interaction

Valid estimate

of E–D

relationship

Three stages:

(1) Variable specification

(2) Interaction

(3) Confounding/precision

Initial model: HWF

EViVj

in initial

model

!

EVi;EVj;

Vi;Vj;ViVj

also in model

Hierarchical backward elimination:

 Can test for interaction,butnot
confounding

 Can eliminate lower order term
if corresponding higher order
term is not significant

This presentation describes a strategy for

assessing interaction and confounding when

carrying out mathematical modeling using

logistic regression. The goal of the strategy is

to obtain a valid estimate of an exposure–

disease relationship that accounts for con-

founding and effect modification.

In the previous presentation on modeling strat-

egy guidelines, we recommended a modeling

strategy with three stages: (1)variable specifi-

cation, (2)interaction assessment, and (3)con-

founding assessmentfollowed by consideration

ofprecision.

The initial model is required to behierarchi-

cally well formulated, which we denote as

HWF. This means that the initial model must

contain all lower order components of any

term in the model.

Thus, for example, if the model contains an

interaction term of the formEViVj, this will

require the lower order termsEVi,EVj,Vi,Vj,

andViVjalso to be in the initial model.

Given an initial model that is HWF, the recom-

mended strategy then involves ahierarchical

backward elimination procedurefor removing

variables. In carrying out this strategy, statisti-

cal testing is allowed for interaction terms but

not for confounding terms. Note that although

any lower order component of a higher order

term must belong to the initial HWF model,

such a component might be dropped from the

model eventually if its corresponding higher

order term is found to be nonsignificant during

the backward elimination process.

206 7. Modeling Strategy for Assessing Interaction and Confounding