D. Control for the subset that gives largest gain in
precision, i.e., tighter confidence interval around
odds ratio.
E. Example.
IV. Confounding assessment with interaction(pages
215–223)
A. Flow diagram representation.
B. Use hierarchy principle to identify allVs that
cannot be eliminated from the model; the
remainingVs are eligible to be dropped.
C. EligibleVs can be dropped as nonconfounders if
odds ratio does not change when dropped; then
control for subset of remainingVs that gives
largest gain in precision.
D. Alternative ways to determine whether odds ratio
changes when different subsets ofVs are
dropped.
E. In practice, it is difficult to evaluate changes in
odds ratio when eligibleVs are dropped;
consequently, safest strategy is to control for allVs.
F. Example.
V. Evans County example continued(pages 223–230)
A. Evans County CHD data descriptions.
B. Variable specification stage.
C. Confounding assessment stage.
D. Final model results.Practice Exercises
Practice Exercises
A prevalence study of predictors of surgical wound infec-
tion in 265 hospitals throughout Australia collected data
on 12,742 surgical patients (McLaws et al., 1988). For each
patient, the following independent variables were deter-
mined: type of hospital (public or private), size of hospital
(large or small), degree of contamination of surgical site
(clean or contaminated), and age and sex of the patient. A
logistic model was fit to this data to predict whether or not
the patient developed a surgical wound infection during
hospitalization. The abbreviated variable names and the
manner in which the variables were coded in the model are
described as follows:Variable Abbreviation Coding
Type of hospital HT 1 ¼public, 0¼private
Size of hospital HS 1 ¼large, 0¼small
Degree of
contaminationCT 1 ¼contaminated,
0 ¼clean
Age AGE Continuous
Sex SEX 1 ¼female, 0¼male234 7. Modeling Strategy for Assessing Interaction and Confounding