Detecting treatment-by-centre interaction in multi-centre clinical trials.

This paper considers several permutation tests for treatment-by-centre interaction in multi-centre clinical trials in which the endpoint is survival time subject to censoring. Some of the tests are based on existing tests and some are new. To evaluate and compare the tests with respect to power under different conditions, we generated survival times and censoring times through simulation. We used special methodology to handle the unusual problems that arise in power simulations when the tests under study are permutation tests. Different conditions yielded different interaction tests as the best performers. Although one test gave comparatively good power under almost all conditions, some of the other tests also appear to be useful. For the sample sizes and configurations in the simulations, power is generally low; thus it may not be possible to detect interaction reliably when it exists, a finding in agreement with the known low power for interaction tests in general. Copyright 2001 John Wiley & Sons, Ltd.