Effect Modification and Interaction

Interaction is when a treatment effect is different across different subgroups of the population defined by a baseline covariate. Commonly examined effect modifiers include: demographic variables, study location, or baseline prognostic factors.

If an interaction does exist:

image.png

Types of Statistical Interactions

A potential issue that arises with assessing interaction is that the sample size within a subgroup will be smaller than for the whole trial population and there will be less power to detect an effect. To account for this we can adjust the sample size requirement for interactions of interest, or define a limited number of pre-specified exploratory subgroup analyses (multiple testing needs to be accounted for).

Statistical Assessment of Interaction

Interaction F is testing:
H0: No interaction vs H1: Interaction

*Adjusted Treatment Comparison;
proc glm;
class trt sex;
model change=trt sex/solution clparm;
run;

image.png

The FDA rule is to assess interactions at .15 level of significance. This is conservative and circumvents low power issues; If the interaction exists it may have a relatively low probability of obtaining a significant interaction result.

Other options:

*Cochran-Mantel-Haenszel Approach;
proc freq;
*Order: Factor*Treatment*Outcome;
table center*arm*corabn/cmh;
weight count;
run;

* Assessing treatment-by-center interaction with logistic regression;
proc logistic data=kawasaki;
class arm(param=ref ref='2-ASA') center;
model corabn(event='ABNORMAL')=arm|center /rl;
freq count;
oddsratio arm / at (center=all);
run;
Interpretation

If the interaction is qualitative it would be meaningless to talk about an "overall" treatment difference. We would need to investigate the cause of interaction and if it invalidates the study through subgroup analysis.

If non-significant or quantitative we can "ignore" interaction (remove from model assessing treatment effect)

When an interaction term is present, the main effects alone cannot be interpreted without being relative to the state of other variables.

Dummy Variable Approach to Interaction

image.png

In SAS we represent interaction in a regression model using binary variables.

*Code Binary Variable;
data dbp;
set dbp;
*Dummy variables;
if sex='F' then new_sex=1;
else new_sex=0;
if trt='A' then new_trt=1;
else new_trt=0;
*Interaction term(s);
trt_sex=new_trt*new_sex;
run;

*Model with interaction;
proc reg data=dbp;
model change=new_trt new_sex trt_sex;
*Test for interaction;
INTERACTION: test trt_sex=0;
run;

Revision #3
Created 24 March 2023 12:59:19 by Elkip
Updated 24 March 2023 16:07:41 by Elkip