Detailed Outline
Abbreviated Outline
I. Overview(page 168)
Focus:
Guidelines for “best” model
3-stage strategy
Valid estimate ofE–Drelationship
II. Rationale for a modeling strategy(pages
168–169)
A. Insufficient explanation provided about
strategy in published research; typically only
final results are provided.
B. Too many ad hoc strategies in practice; need for
some guidelines.
C. Need to consider a general strategy that applies
to different kinds of modeling procedures.
D. Goal of strategy in etiologic research is to get a
valid estimate ofE–Drelationship; this
contrasts with goal of obtaining good
prediction, which is built into computer
packages for different kinds of models.
III. Overview of recommended strategy(pages
169–173)
A. Three stages: variable specification, interaction
assessment, and confounding assessment
followed by considerations of precision.
B. Reason why interaction stage precedes
confounding stage: confounding is irrelevant in
the presence of strong interaction.
C. Reason why confounding stage considers
precision after confounding is assessed: validity
takes precedence over precision.
D. Statistical concerns needing attention but
beyond scope of this presentation: collinearity,
controlling the significance level, and
influential observations.
E. The model must be hierarchically well
formulated.
F. The strategy is a hierarchical backward
elimination strategy that considers the roles
that different variables play in the model and
cannot be directly carried out using standard
computer algorithms.
G. Confounding is not assessed by statistical
testing.
H. If interaction is present, confounding
assessment is difficult in practice.
IV. Variable specification stage(pages 173–175)
A. Start withD, E,andC 1 ,C 2 ,...,Cp.
B. ChooseVs fromCs based on prior research or
theory and considering potential statistical194 6. Modeling Strategy Guidelines