VI. SUMMARY
Chap. 6
Overall guidelines for three
stages
Focus: variable specification
HWF model
Chap. 7
Focus: interaction and confound-
ing assessment
Interaction: use hierarchical back-
ward elimination
Usehierarchy principleto identify
lower order components that
cannot be deleted (EVs,Vis, and
ViVjs)
Confounding:nostatistical testing:
Compare whetherdOR meaning-
fully changes whenVs are deleted
Drop nonconfounders if precision
is gained by examining CIs
No interaction: assess confounding
by monitoring changes in^b, the
coefficient ofE
A brief summary of this presentation is now
given. This has been the second of two chap-
ters on modeling strategy when there is a sin-
gle E. In Chap. 6, we gave overall guidelines
for three stages, namely, variable specifica-
tion, interaction assessment, and confo-
unding assessment, with consideration of
precision. Our primary focus was the variable
specification stage, and an important require-
ment was that the initial model be hierarchi-
cally well formulated (HWF).In this chapter, we have focused on the inter-
action and confounding assessment stages of
our modeling strategy. We have described
how interaction assessment follows a hierar-
chical backward elimination procedure, start-
ing with assessing higher order interaction
terms followed by assessing lower order inter-
action terms using statistical testing methods.If certain interaction terms are significant,
we use the hierarchy principle to identify all
lowerordercomponentsofsuchterms,which
cannot be deleted from any further model
considered. This applies to lower order inter-
action terms (i.e., terms of the formEV)and
to lower order terms involving potential con-
founders of the formViorViVj.Confounding is assessed without the use of
statistical testing. The procedure involves
determining whether the estimated odds
ratio meaningfully changes when eligibleV
variables are deleted from the model.If some variables can be identified as noncon-
founders, they may be dropped from the
model provided their deletion leads to a gain
in precision from examining confidence
intervals.If there is no interaction, the assessment of
confounding is carried out by monitoring
changes in the estimated coefficient of the
exposure variable.Presentation: V. The Evans County Example Continued 231