Suppose the following initial model is specified for
assessing the effect of type of hospital (HT), consid-
ered as the exposure variable, on the prevalence of
surgical wound infection, controlling for the other
four variables on the above list:
logit PðXÞ¼aþbHTþg 1 HSþg 2 CTþg 3 AGEþg 4 SEX
þd 1 HTAGEþd 2 HTSEX:
Describe how to test for the overall significance (a
“chunk” test) of the interaction terms. In answering
this, describe the null hypothesis, the full and reduced
models, the form of the test statistic, and its distribu-
tion under the null hypothesis.
Using the model given in Exercise 1, describe briefly
how to carry out a backward elimination procedure to
assess interaction.
Briefly describe how to carry out interaction assess-
ment for the model described in Exercise 1. (In
answering this, it is suggested you make use of the
tests described in Exercises 1 and 2.)
Suppose the interaction assessment stage for the
model in Example 1 finds no significant interaction
terms. What is the formula for the odds ratio for the
effect of HT on the prevalence of surgical wound infec-
tion at the end of the interaction assessment stage?
WhatVterms remain in the model at the end of inter-
action assessment? Describe how you would evaluate
which of theseVterms should be controlled as con-
founders.
Considering the scenario described in Exercise 4 (i.e.,
no interaction terms found significant), suppose you
determine that the variables CT and AGE do not need
to be controlled for confounding. Describe how you
would consider whether dropping both variables will
improve precision.
Suppose the interaction assessment stage finds that
the interaction terms HTAGE and HTSEX are
both significant. Based on this result, what is the for-
mula for the odds ratio that describes the effect of HT
on the prevalence of surgical wound infection?
For the scenario described in Example 6, and making
use of the hierarchy principle, whatVterms are eligi-
ble to be dropped as possible nonconfounders?
Describe briefly how you would assess confounding
for the model considered in Exercises 6 and 7.
Suppose that the variable CT is determined to be a
nonconfounder, whereas all otherVvariables in the
model (of Exercise 1) need to be controlled. Describe
