Testing for interaction in studies of noninferiority.

We consider the role of interaction tests in the context of active-controlled clinical trials that aim to demonstrate the noninferiority of an experimental treatment compared to a standard (control) treatment. When the subjects can be grouped into strata (e.g., study sites, gender, race, etc.), there may be a desire to determine whether the experimental treatment is noninferior to the standard in each of the strata. We present five possible analysis strategies to test for heterogeneity of relative treatment effects among strata. These strategies are either identical to or straightforward modifications of strategies that can be used to test for interaction when the objective of the study is to show differences rather than noninferiority. The various analysis strategies implicitly depend on different definitions of interaction. Power of the various tests will be low, a phenomenon that often occurs when testing for interaction. We present simulation results to quantify the power and type I error rates under different scenarios and an example to demonstrate the proposed tests. None of the analysis strategies is best under every parameter configuration. The tests may be best used in a descriptive or exploratory manner. Extensions to two-sided equivalence testing are also discussed.