intervals for the gold standard odds ratio and the odds
ratio for the model that drops NSAS. If the latter
confidence interval is meaningfully narrower, then
precision is gained by dropping NSAS, so that this
variable should, therefore, be dropped. Otherwise, one
should control for NSAS because no meaningful gain
in precision is obtained by dropping this variable. Note
that in assessing both confounding and precision,
tables of odds ratios and confidence intervals obtained
by specifying values of NS need to be compared
because the odds ratio expression involves an effect
modifier.- If NSAS is dropped, the onlyVvariable eligible to be
dropped is AS. As in the answer to Question 6,
confounding of AS is assessed by comparing odds ratio
tables for the gold standard model and reduced model
obtained by dropping AS. The same odds ratio
expression as given in Question 5 applies here, where,
again, the coefficients for the reduced model (without
AS and NSAS) may be different from the coefficient
for the gold standard model. Similarly, precision is
assessed similarly to that in Question 6 by comparing
tables of confidence intervals for the gold standard
model and the reduced model. - The odds ratio expression is given by exp(1.9381
1.1128NS). A table of odds ratios for different values of
NS can be obtained from this expression and the results
interpreted. Also, using the estimated
variance–covariance matrix (not provided here), a table
of confidence intervals (CIs) can be calculated and
interpreted in conjunction with corresponding odds
ratio estimates. Finally, the CIs can be used to carry out
two-tailed tests of significance for the effect of SMK
at different levels of NS.
Chapter 8 1. a. The screening approach described does not
individually assess whether any of the control (C)
variables are either potential confounders or
potential effect modifiers.
b. Screening appears necessary because the number
of variables being considered (including possible
interaction terms) for modeling is large enough to
expect that either a model containing all main
effects and interactions of interest won’t run, or will
yield unreliable estimated regression coefficients.
In particular, if such a model does not run,
collinearity assessment becomes difficult or even
impossible, so the only way to get a “stable” model
requires dropping some variables.
676 Test Answers