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

Gold standard OR estimate controlling for SBP

and ECG: 2.07, Narrowest CI obtained when

only ECG is controlled: OR estimate is 2.08,

Overall conclusion: OR approximately 2, but not

significant.

VI. Assessing Interaction Involving Matching

Variables(pages 404–406)

A. Option 1: Add product terms of the formEV 1 i,

whereV 1 iare dummy variables for matching

strata.

Model:logit PðXÞ¼aþbEþ~g 1 iV 1 iþ~g 2 jV 2 j

þE~d 1 iV 1 iþE~dkWk;

whereV 2 jare other covariates (not matched)

andWkare effect modifiers defined from other

covariates.

Criticism of option 1:

 Difficult to identify specific effect modifiers

 Number of parameters may exceedn

B. Option 2: Add product terms of the formE

W1m, whereW1mare the matching variables in

original form.

Model:logit PðXÞ¼aþbEþ~g 1 iV 1 iþ~g 2 jV 2 j

þE~d 1 iW 1 mþE~dkWk;

whereV 2 jare other covariates (not matched)

andWkare effect modifiers defined from other

covariates.

Criticism of option 2:

 Model is not HWF (i.e.,EW1min model

but notW1m)

But, matching variables are in model in

different ways as both effect modifiers and

confounders.

C. Other alternatives:

 Do not match on any variable considered as

an effect modifier

 Do not assess interaction for any matching

variable

VII. Pooling Matching Strata(pages 407–409)

A. Example: Pair-match on SMK (0, 1), 100 cases,

60 matched pairs of smokers, 40 matched pairs

of nonsmokers.

B. Controls for two or more matched pairs that

have same SMK status are interchangeable.

Detailed Outline 417