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

VII. SUMMARY

Five issues on model strategy
guidelines:

1. Modeling strategy when there
   are two or more exposure
   variables


2. Screening variables when
   modeling


3. Collinearity diagnostics


4. Multiple testing


5. Influential observations

Issue 1: SeveralEs

Logit PðXÞ¼aþ~

q

i¼ 1

biEiþ~

p 1

j¼ 1

gjVj

þ~

q

i¼ 1

~

p 2

k¼ 1

dikEiWk

þ~

q

i¼ 1

~

q

i^0 ¼ 1

i 6 ¼i^0

d*ii 0 EiEi^0

Modeling Strategy Summary:
SeveralEs

Step 1:Define initial model (above
formula)
Step 2:Assess interaction: overall
chunk test (?), thenEWs
and then (?)EEs
Step 3:Assess confounding and
precision (Vs) (prior to
EEs?)
Step 4:Test for nonsignifEsifnot
components of significant
EEs

Issue 2: Screening Variables
Method 0: Consider predictors
one-at-a-time
Screen-out those
Xi not significantly
associated withD

Does not consider confounding or
interaction.
Questionable if model contains
bothEs andCs

This presentation is now complete.

We have described the five issues (shown at

the left) on model strategy guidelines not cov-

ered in the previous two chapters on this topic.

Each of these issues represent important fea-

tures of any regression analysis that typically

require attention when determining a “best”

model.

Regarding issue 1, we recommend that the

initial model have the general form shown at

the left. This model involvesEs,Vs,EWs, and

EEs, so there are two types of interaction

terms to consider.

We then recommend assessing interaction,

first by deciding whether to do an overall

chunk test, then testing for theEWs, after

which a choice has to be made as to whether

to test for theEEterms prior to or subsequent

to assessing confounding and precision. The

resulting model is then further assessed to see

whether any of theEterms are nonsignificant.

Regarding issue 2, we described an approach

(calledMethod 0)inwhichthosevariablesthat

are not individually significantly associated

with the (binary) outcome are screened-out

(i.e., removed from one’s initial model).

Method 0 does not consider confounding and/

or interaction for predictors treated one-at-a-

time. Thus, Method 0 makes most sense when

the model only involvesEs but is questionable

with bothEs andCs being considered.

Presentation: VII. Summary 283