ABSTRACT
Historically, the primary objective of a dose-finding clinical trial in oncology has been to identify the maximum tolerated dose (MTD) of the agent being investigated, from a discrete set of available doses D = {d 1, ... , dK }. Numerous phase I clinical trial designs [1–5] have been proposed for identifying the MTD for situations in which toxicity is described as a binary random variable (dose-limiting toxicity, DLT; yes/no). Since the works of Storer [1] and O’Quigley, Pepe, and Fisher [2], there has been considerable development in the statistical design and analysis of dose-finding methods, both for phase I and phase I–II trials, a recent review of which is given in Iasonos and O’Quigley [6]. Many of these approaches appeal to Bayesian techniques and involve the use of prior information. Such information is typically not informative, although sometimes attempts are made to elicit relevant information from the clinicians involved in the studies. One common criticism of the standard 3 + 3 design in this context is that it behaves like a random walk [1] so that no real learning is taking place as more and more patients are included into a study. For model-based designs, on the other hand, there is learning taking place, and, generally, if we allow sample size to increase without bound (as a theoretical construction), then we will find the correct MTD with a probability of 1.
