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

Option A: Overall (chunk) LR test for

interaction; then “subchunk” LR tests

for EWs and EEs; then Vs; finally Es

EXAMPLE

MRSA example:

 Overall chunk test:LRw^25 df

underH 0 :

d 11 ¼d 12 ¼d 21 ¼d 22 ¼d*¼ 0

 Subchunk tests:

LRw^24 dfunderH 01 :

d 11 ¼d 12 ¼d 21 ¼d 22 ¼ 0

LRw^21 dfunderH 02 :d*¼ 0

Option B: Assess EWs first, then EEs, prior

to Vs and Es

Reasons:
Assess interaction (EWs andEEs)
prior to confounding and preci-
sion, andAssess EWs prior toEEs

Option C: Assess EWs first, then Vs, prior

to EEs and Es

Reason:
Assess effect modification (Ws)
and confounding (Vs) before con-
sidering exposures (Es andEEs)

EXAMPLE

Initial Model Output:

2 2ln L¼275.683

Analysis of maximum likelihood estimates

Es

EWs

EEs

Vs

Reduced Model A:

Logit PðXÞ¼aþðb 1 E 1 þb 2 E 2 Þ

þðg 1 V 1 þg 2 V 2 Þ

One Option (A) begins with a “chunk” LR test

that simultaneously evaluates all product

terms. We then test separate “subchunks”

involvingEWs andEEs, after which we assess

theVs for confounding and precision. Finally,

we consider dropping nonsignificantEs.

For the initial MRSA model, since there are five

product terms, the overall chunk test would

involve a chi square statistic with 5 degrees of

freedom. The two “subchunks” would involve

the 4EWterms and the singleEEterm, as we

illustrate on the left.

Alternatively, a second Option (B) differs from

OptionAby simply skippingthe overall chunk

test. Both Options A and B make sense if we

decide that assessing interaction should always

precede assessing confounding and precision,

and thatEWs should always be assessed prior to

EEs.

As another Option (C), recall that when we

considered a model with only a singleE,we

left thisEin the model throughout the entire

process of evaluating interaction, confound-

ing, and then precision. An analogous

approach for severalEs is to evaluate effect

modifiers (Ws) and potential confounders (Vs)

before considering any terms involving Es,

including product terms (EEs).

We will apply Options A through C to the

MRSA data. First, we present, at the left, edited

results from fitting the initial model.

We now show the results from usingOptionA

for the reduced model (A) that eliminates all

five interaction terms from the initial model.

248 8. Additional Modeling Strategy Issues