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

Guidelines applicable to:

Logistic regression

Multiple linear regression

Cox PH regression

Two modeling goals:

(1) To obtain a validE–Destimate

(2) To obtain a good predictive
model

(different strategies for different
goals)

Prediction goal:

Use computer algorithms

Validity goal:
 Our focus

 For etiologic research

 Standard computer algorithms
not appropriate

III. Overview of
Recommended Strategy

Three stages:

(1) Variable specification
(2) Interaction assessment
(3) Confounding assessment
followed by precision
Variable specification:

 Restricts attention to clinically
or biologically meaningful
variables

 Provides largest possible initial
model

Modeling strategy guidelines are also important

for modeling procedures other than logistic

regression. In particular, classical multiple lin-

ear regression and Cox proportional hazards

regression, although having differing model

forms,allhaveincommonwithlogisticregres-

sion the goal of describing exposure–disease

relationships when used in epidemiologic

research. The strategy offered here, although

describedinthecontextoflogisticregression,

is applicable to a variety of modeling procedures.

There are typically two goals of mathematical

modeling: One is to obtain a valid estimate of

an exposure–disease relationship and the other

is to obtain a good predictive model. Depend-

ing on which of these is the primary goal of the

researcher, different strategies for obtaining

the “best” model are required.

When the goal is “prediction”, it may be more

appropriate to use computer algorithms, such as

backward elimination or all possible regressions,

which are built into computer packages for dif-

ferent models. [See Kleinbaum et al. (2008)]

Our focus in this presentation is on the goal of

obtaining a valid measure of effect. This goal is

characteristic of most etiologic research in epi-

demiology. For this goal, standard computer

algorithms do not apply because the roles that

variables – such as confounders and effect

modifiers – play in the model must be given

special attention.

The modeling strategy we recommend involves

three stages: (1)variable specification,(2)inter-

action assessment, and (3)confounding assess-

ment followed by consideration of precision.We

have listed these stages in the order that they

should be addressed.

Variable specification is addressed first because

this step allows the investigator to use the

research literature to restrict attention to clini-

cally or biologically meaningful independent

variables of interest. These variables can then be

defined in the model to provide the largest possi-

ble meaningful model to be initially considered.

Presentation: III. Overview of Recommended Strategy 169