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

Modeling Strategy Summary: Several Es

Step 1: Define initial model (above formula)

Step 2: Assess interaction

Option A: Overall chunk test + Options B or C

Option B: Test EWs, then EEs

Option C: Test EWs, but assess Vs before EEs

Step 4: Test for nonsignificant Es if not

Step 3: Assess confounding and precision (Vs)

Option A and B (cont’d):

Vs after EWs and EEs

Option (cont’d):

Vs after EWs, but prior to EEs

components of significant EEs

Special Cases: SeveralEs

(a) AllVs are controlled as main
effects,
i.e., confounding and
precision forVs not
considered
Modeling Strategy: All Vs controlled
Step 1: Define initial model (above formula)
Step 2: Assess Interaction
Option A: Overall chunk test + Options B
Option B: Test EWs, then EEs
Step 4: Test for nonsignif Es if not components
of significant EEs

EXAMPLE

MRSA Initial Model, Special case(a)

Logit PðXÞ¼aþðb 1 E 1 þb 2 E 2 Þ

þðg 1 V 1 þg 2 V 2 Þ

þðd 11 E 1 W 1 þd 12 E 1 W 2

þd 21 E 2 W 1 þd 22 E 2 W 2 Þ

þd*E 1 E 2

ModelA: Final Model

Logit PðXÞ¼aþðb 1 E 1 þb 2 E 2 Þþðg 1 V 1

þg 2 V 2 Þ

(b) The model contains onlyEs
andEEs, but noCs
(i.e., noVsorWs)
“Hypothesis Generating”
Model

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 theEE

terms prior to or subsequent to assessing con-

founding and precision (involving theVs).

The resulting model can then be further

assessed to see whether any of theEterms in

the model can be dropped as nonsignificant.

There are two special cases that we now

address:

What if you decide to control for allVsas

main effects(without assessing confounding

and/or precision)?

In this case (a), we only need to consider

Options A and B, so that Step 3 of our previ-

ously described strategy can be omitted.

For example, using the MRSA data, the initial

model, shown again at the left contains twoEs,

twoVs, 4EWs and oneEEterm.

When previously applying Options A and B to

this model, we dropped all interaction terms,

resulting in reduced model A shown at the left.

If we decide in advance to control for bothVs,

then this is our final model, since bothEs were

significant in this model.

As a second special case,what if our model con-

tains only Es,so there are no Cs to control?This

case is often referred to as a “hypothesis gener-

ating” model, since we are essentially assuming

that we have limited knowledge on all possible

predictors and that no risk factors have been

established.

260 8. Additional Modeling Strategy Issues