Logistic Regression: A Self-learning Text, Third Edition (Statistics in the Health Sciences)

briefly how you would assess whether the variable CT

needs to be controlled for precision reasons.

10. What problems are associated with the assessment of
    confounding and precision described in Exercises
    8 and 9?

Test The following questions consider the use of logistic regres-
sion on data obtained from a matched case-control study
of cervical cancer in 313 women from Sydney, Australia
(Brock et al., 1988). The outcome variable is cervical can-
cer status (1¼present, 0¼absent). The matching vari-
ables are age and socioeconomic status. Additional
independent variables not matched on are smoking status,
number of lifetime sexual partners, and age at first sexual
intercourse. The independent variables are listed below
together with their computer abbreviation and coding
scheme.

Variable Abbreviation Coding

Smoking status SMK 1 ¼ever, 0¼never

Number of sexual

partners

NS 1 ¼ 4 þ,0¼0–3

Age at first intercourse AS 1 ¼ 20 þ,0¼ 19

Age of subject AGE Category matched

Socioeconomic status SES Category matched

Assume that at the end of the variable specification stage,

the followingE,V,Wmodel has been defined as the initial

model to be considered:

logit PðXÞ¼aþbSMKþ~g*iVi*þg 1 NSþg 2 AS

þg 3 NSASþd 1 SMKNSþd 2 SMKAS

þd 3 SMKNSAS;

where theV*iare dummy variables indicating matching

strata, theg*i are the coefficients of theVi*variables,

SMK is the only exposure variable of interest, and the

variables NS, AS, AGE, and SES are being considered

for control.

1. For the above model, which variables are interaction
   terms?


2. For the above model, list the steps you would take to
   assess interaction using a hierarchically backward
   elimination approach.


3. Assume that at the end of interaction assessment, the
   only interaction term found significant is the product
   term SMKNS. What variables are left in the model at

236 7. Modeling Strategy for Assessing Interaction and Confounding