Determining the duration of comparative clinical trials while allowing for cure

Rubinstein et al. provide a technique for estimating the required length of accrual in a two treatment group randomized survival trial. An important parameter in this approach is the expected number of events as estimated by assuming exponential failure. When the underlying distribution of failure times displays a distinct plateau, as in many children's cancers (i.e. there is a "cure"), the assumption of exponential failure could be misleading. In this situation we propose the use of the general formulation in [1], but with the expected number of failures based on failure models with hazard functions which may decrease to zero. We suggest two such models and show that they provide a good fit in an example from Childrens Cancer Study Group (CCSG) trials, and that the determination of required trial duration depends strongly upon which model is assumed.