Designing for cancer clinical trials: selection of prognostic factors

This paper reviews the pros and cons for stratifying on a number of variables when randomizing patients to treatments in cancer clinical trials. Arguments in favor of randomization focus on the increased precision achieved. Arguments against stratification focus on the complexity of randomization procedures and the fact that post hoc statistical adjustment can achieve nearly the same precision as stratification. Although arguments on both sides have merit, newer methods of adaptive randomization would seem to shift the balance toward the use of predictive factors in achieving balance among treatment groups. In the Northern California Oncology Group, the Efron-biased coin method of randomization is being used to balance treatment groups on as many as four to six prognostic variables. Treatment assignment is made by telephone to the Statistical Center, where the assignment is determined by computer, taking into account previous assignments and the prognostic characteristics of the patient to be assigned.