Questions about how to analyze these data now follow:- In addition to the variables listed above, there were 12
other variables also listed in the dataset and identified
from the literature review and conceptualization of the
study as potential control variables. These variables
were screened out by the investigators as not being
necessary to include in the multivariable modeling ana-
lyses that were carried out.
a. Assume that screening was carried out one variable
at a time using tests of significance for the relation-
ship between each potential control variable and
the outcome variable, so that those potential con-
trol variables not found significantly associated
with the outcome variable were not included in
any modeling analysis.
How can you criticize this approach to screening?
b. Why was some kind of screening likely necessary
for this analysis?
Suppose that the logistic regression treated the variables
mattress type (F) and perceived mattress type (PF)asordi-
nalvariables. Suppose also that the variableBASEis also
considered to be anexposure variable(even though it was
not involved in the randomization) in addition to the vari-
able mattress type (F).
Suppose further that this model allows for two-way inter-
actions (i.e., products of two variables) between mattress
type (F) and each of the other independent variables
(POST, BASE, PF, OCC, AGE, andGEN) and two-way
interactions betweenBASEand each of the control vari-
ablesPF, POST, OCC, AGE, andGEN. - State the logit formula for the logistic regression model
just described. Make sure to consider bothFandBASE
as exposure variables. - For each of the following product terms, state whether
the product term is anEEvariable, anEVvariable, or
neither:
F 3 POST
F 3 BASE
POST 3 BASE
PF 3 POST
BASE 3 PF
(To answer this part, simply writeEE, EV, or neither
next to the variable regardless of whether the product
term given is contained in the revised model described
in question 2.)
294 8. Additional Modeling Strategy Issues