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

EXAMPLE (continued)

Diagram 2)PREVHOSP and PAMU

independent risk factors;

AGE and GENDER

confounders

Diagram 2 appropriate)initial

model containing both

E 1 andE 2 is justified

Diagram 1 appropriate)initial

model should not

contain both

PREVHOSP and

PAMU

i.e., PAMU intervening variable

Logit PðXÞ¼

+

aþb 1 E 1 þg 1 V 1 þg 2 V 2

þd 11 E 1 V 1 þd 12 E 1 V 2

The moral: Causal diagram can
influence choice of initial model

Diagram 2 indicates that PREVHOSP and

PAMU are independent risk factors for MRSA

outcome, and that AGE and GENDER are con-

founders of both PREVHOSP and PAMU.

The initial model that we considered in our

analysis of the MRSA data, containing both

PREVHOSP and PAMU in the model asEvari-

ables, can be justified if we decide that Dia-

gram 2 is a correct representation of the

causal pathways involved.

In contrast, if we decide that Diagram 1 is more

appropriate than Diagram 2, we should not put

both PREVHOSP and PAMU in the same

model to assess their joint or separate effects.

In other words, if PAMU is an intervening vari-

able, we should consider a model involving

only oneEvariable, preferably PREVHOSP.

An example of such a model, which controls

for AGE and GENDER and allows for interac-

tion effects, is shown at the left.

Thus, as mentioned previously in Chap. 6, the

choice of the initial model can be influenced

by the causal diagram considered most appro-

priate for one’s data.

262 8. Additional Modeling Strategy Issues