HT variable. Such a test is also inappropriate because
AGE is a potential confounder, and confounding
should not be assessed by statistical testing.- A hierarchical backward elimination procedure for
the model in Exercise 3 would involve first assessing
interaction involving the four interaction terms and
then considering confounding involving the four
potential confounders. The interaction assessment
could be done using statistical testing, whereas the
confounding assessment should not use statistical
testing. When considering confounding, any
Vvariable that is a lower order component of a
significant interaction term must remain in all further
models and is not eligible for deletion as a
nonconfounder. A test for any of these latterVsis
inappropriate because such a test would be dependent
on the coding of any variable in the model. - If HTCT and HTSEX are found significant, then
theVvariables CT and SEX cannot be removed from
the model and must, therefore, be retained in all
further models considered. The HT variable remains
in all further models considered because it is the
exposure variable of interest. CT and SEX are lower
order components of higher order interaction terms.
Therefore, it is not apropriate to test for their
inclusion in the model. - At the end of the interaction assessment stage, the
remaining model is given by
logit PðXÞ¼aþbHTþg 1 HSþg 2 CTþg 3 AGEþg 4 SEX
þd 2 HTCTþd 4 HTSEX:
202 6. Modeling Strategy Guidelines