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

Answers about Screening (con-
tinued)

Q2. Yes: It depends.

 Use Method 0 if severalEs
andCs? No
 Use Method 0 if oneEand
severalCs? No
 Use Method 0 if onlyEs?
Yes
Questionable when model con-
sidersCs

Q3. No clear-cut rules for
“largek.”
However, need screening if
initial model does not run.

Q4. Other options for (1-at-a-
time) screening?Yes
 Variations to assess
confounding or
interaction, e.g., use
stratified analysis instead
of logistic regression
 Such options needed if
consideringCs

5. Collinearity?
   pffiPrior to screening and/or
   after screening

Initial model does not run

+

Cannot obtain collinearity diagnostics

+

Start with screening

Screening completed

+

Model may still be unreliable

+

Consider collinearity diagnostics

Q2. Should the use of Method 0 depend on

types of predictors?

Yes, Method 0 makes most sense when the

model only involvesEs, i.e., no potential con-

founders (Cs) and no corresponding changes in

ORs are being considered. However, Method

0 is questionable whenever there are variables

being controlled (Cs).

Q3. How large doeskhave to be relative ton

in order to justify screening?

There are no clear-cut rules, but you will

become aware that screening should be consid-

ered if your initial model does not run (see next

section on collinearity).

Q4. Are there other ways to carry out (one-at-

a-time) screening and when, if at all, should

they be preferred to the typical approach?

There are several reasonable options for

screening, all of which are variations of ways

to assess possible confounding and/or effect

modification involving covariates. Such options

should be preferred whenever there is a mixture

ofEsandCstobeconsidered.

Q5. Where does collincarity assessment fit in

with this problem?

Collinearity may be considered prior to

screening or after screening is performed.

If your initial model does not run, typical col-

linearity diagnostics (e.g., condition indices, to

be described in the next section) cannot be

obtained, so screening must be considered

from the beginning.

Also, once screening has been performed, col-

linearity assessment may determine that

your reduced model (after screening) is still

unreliable.

Presentation: III. Screening Variables 267