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

If, looking at these two tables collectively, we

find thatyes, there is one or more meaningful

difference in corresponding odds ratios, we

would conclude that the variableV 3 needs to

be controlled for confounding. In contrast, if

we decide thatno, the two tables are not mean-

ingfully different, we can conclude that vari-

able V 3 does not need to be controlled for

confounding.

If the decision is made thatV 3 does not need to

be controlled for confounding reasons, we still

may wish to control forV 3 because of precision

reasons. That is, we can compare confidence

intervals for corresponding odds ratios from

each table to determine whether we gain or

lose precision depending on whether or notV 3

is in the model.

In other words, to assess whether there is a

gain in precision from droppingV 3 from the

model, we need to make an overall comparison

of two tables of confidence intervals for odds

ratio estimates obtained whenV 3 is in and out

of the model.

EXAMPLE (continued)

OR OR OR

OR OR OR

corresponding odds ratios

Gain in precision?

Gold standard CI CI* (excludes V 3 )

OR tables

meaningfully

different?

Control V 3

for con-

founding

Do not need to control

V 3 for confounding

Consider precision with and without

V 3 by comparing confidence intervals

yes

no

CI CI CI

CI CI CI CI* CI* CI*

CI* CI* CI*

Gold standard OR

OR* OR* OR*

OR* OR* OR*

OR* (excludes V 3 )

222 7. Modeling Strategy for Assessing Interaction and Confounding