Evaluating multiple treatment courses in clinical trials

In oncology, a patient's treatment often involves multiple courses of chemotherapy. The most common medical practice in choosing treatments for successive courses is to repeat a treatment that is successful in a given course and otherwise switch to a different treatment. Patient outcome thus consists of a sequence of dependent response variables and corresponding treatments. Despite the widespread use of such adaptive 'play-the-winner-and-drop-the-loser' algorithms in medical settings involving multiple treatment courses, most statistical methods for treatment evaluation characterize early patient outcome as a single response to a single treatment, resulting in a substantial loss of information. In this paper, we provide a statistical framework for multi-course clinical trials involving some variant of the play-the-winner-and-drop-the-loser strategy. The aim is to design and conduct the trial to more closely reflect actual clinical practice, and thus increase the amount of information per patient. The proposed design is similar to a multi-stage cross-over trial, with the essential difference that here all treatments after the first course are assigned adaptively. We illustrate the method by application to a randomized phase II trial for androgen independent prostate cancer. We consider the goals of selecting one best treatment, or selecting a best ordered pair of treatments with the second given if the first fails to achieve a patient success. A simulation study is reported, and extensions to trials involving toxicity or regimen-related death are discussed.