A REPEATED-MEASURES DESIGN WITH REPEATED RANDOMIZATION

We consider a two-phase design to evaluate a total of nu treatments. Eligible patients are randomly assigned to a set of nu1 treatments during Phase 1 and observed repeatedly. Then following a washout period of patient-specific duration, patient eligibility is redetermined for a patient willing to participate in Phase 2. The patients eligible for Phase 2 are randomly assigned to a set of nu2 treatments and observed repeatedly during Phase 2. Phase 1 dropouts are not rerandomized. Thus the design allows incomplete data only in the last observed phase. This design avoids the problem of carryovers associated with a crossover design and is more economical in terms of recruitment time than a conventional one-phase design. Assuming multivariate normality with ante-dependence structure for a vector of responses, a procedure is given for testing the significance of a linear hypothesis.