Performance of multinomial designs in comparison with response-based designs in non-randomized phase II trials of...

In phase II trials of cytotoxic agents, a multinomial phase II design incorporating early progression and response end points was shown to perform more efficiently than designs based only on response. We undertook a study to evaluate the performance of these designs in trials of targeted agents using the actual phase II data.

Using best response data from sequentially enrolled patients in 15 NCIC Clinical Trials Group and 7 European Organization for Research and Treatment of Cancer trials of targeted agents, we determined that trials would have been stopped at the end of stage I of accrual by applying rules generated by the multinomial and Fleming designs. Two variants of the multinomial design were studied: to stop accrual after stage I of enrolment, Variant A required either response or progression criteria to be met, whereas Variant B required that both response and progression criteria to be met.

Using early progression, null/alternate hypotheses of 60% and 40% (60/40), the multinomial A variant recommended early stopping more often than the Fleming design. In most of the cases, this recommendation was correct given the final trial outcome. In contrast, the multinomial B variant never led to recommendations for early stopping and changing progression hypotheses did not improve the performance of this design.

The multinomial A design using 60/40 hypotheses carried out better than the Fleming design in appropriately stopping trials of inactive targeted agents early. The multinomial B design was not useful for early stopping decisions. The multinomial A design may be favored over response-based designs in phase II trials of targeted agents.