New Designs for Clinical Trials

Abstract

Designs for clinical trials have undergone little change in the past 30 years. Nonetheless, conventional designs have numerous inadequacies. Because it fails to use all available information accumulated during the trial and typically has an insufficient sample size, the 3+3 phase 1 design is often unlikely to select the correct dose implied in the 3+3 algorithm. The design fails to account for covariates, thus assuming, despite evidence to the contrary, that dose is the only determinant of toxicity. It does not monitor response, thus failing to recognize the phase 2 aspects of phase 1 trials and the phase 1 aspects of phase 2 trials; rather these features are often addressed in an ad hoc fashion. Phase 2 trials similarly ignore the realities of medical practice. Thus, although patients are interested in frequent data analyses, standard designs discourage these because of the connection between p-value and trial design and the desire to maintain a specified final p-value. In addition to being inadequately adaptive, phase 2 designs are frequently concerned with only a single outcome. Here we propose remedies for the shortcomings noted above. Such solutions are based on the Bayesian paradigm.