Adaptive Designs

Abstract

The designs described in this article involve treatment allocation probabilities that change according to prior data. Such designs are useful to protect from extreme results that might result from unrestricted randomization, to improve power (e.g., when group variances are unknown and unequal), when prior information is too limited to fix the sampling space a priori (e.g., regions of desirable efficacy and unacceptable toxicity are unknown). Procedures are described for studying K‐groups, but emphasis is placed on dose‐finding experiments. Some procedures involve randomization and some are deterministic given past data. Some are model based and others are nonparametric. Some procedures have a specific optimization criterion as a goal (e.g., maximize power or some other utility) and others are heuristically formulated, with large variation in ease of application and subsequent data analysis.