Introduction We begin this chapter by giving the rationale for having a
strategy to determine a “best” model. Focus is on a logistic
model containing a single dichotomous exposure variable
that adjusts for potential confounding and potential inter-
action effects of covariates considered for control. A strat-
egy is recommended, which has three stages: (1) variable
specification, (2) interaction assessment, and (3) con-
founding assessment followed by consideration of preci-
sion. Causal diagrams are introduced as a component of
the variable specification stage. The initial model must be
“hierarchically well-formulated”, a term to be defined and
illustrated. Given an initial model, we recommend a strat-
egy involving a “hierarchical backward elimination proce-
dure” for removing variables. In carrying out this strategy,
statistical testing is allowed for assessing interaction terms
but is not allowed for assessing confounding. Further
description of interaction and confounding assessment is
given in the next chapter (Chap. 7).
Abbreviated
Outline
The outline below gives the user a preview of the material
in this chapter. A detailed outline for review purposes
follows the presentation.I. Overview (page 168)
II. Rationale for a modeling strategy (pages
168–169)
III. Overview of recommended strategy (pages
169–173)
IV. Variable specification stage (pages 173–175)
V. Causal diagrams (pages 175–179)
VI. Other considerations for variable specification
(pages 180–181)
VII. Hierarchically well-formulated models (pages
181–184)
VIII. The hierarchical backward elimination
approach (page 184–185)
IX. The hierarchy principle for retaining variables
(pages 185–187)
X. An example (pages 188–192)
XI. Summary (pages 192–193)166 6. Modeling Strategy Guidelines