Estimating the Magnitude of Interaction

Qualitative interaction (Peto, 1982) is a useful concept, but its practical value i~ limited because its presence is hard to detect. This difficulty arises from trying to detect a treatment effect of opposite direction to the main effect in a small number of subgroups. Simulation results are presented which reinforce this point. A new measure of interaction magnitude, I AI, is introduced and proposed as an alternative way to think about important interaction between treatments and patient subgroups. Simulation results show that comparing a lower bound for 1M to a specific cutoff provides greater power to detect important interaction than do the tests for qualitative interaction.