In interventions it may be important to determine whether the benefits extend beyond the active treatment period. This is clearly of interest for intensive lifestyle interventions, and there are also examples in the pharmaceutical literature. We consider estimation of carryover effects on time-to-event outcomes such as incident hypertension or incident diabetes. These are defined by a noisy measurement exceeding a diagnostic threshold, and diagnosis is followed by interventions that make subsequent measurements useless for treatment comparison.
We present the results of a systematic simulation study to determine the ability of a parallel-group trial design to detect carryover. None of the designs we examined had acceptable Type I error rate; most also had low power. When a treatment is effective during the intervention period, reliable testing for a carryover effect is difficult. Parallel-group designs do not appear to be a feasible approach.
Sarah Gwynn Sturdevant and Thomas Lumley