Interplay of priors and skeletons in two‐stage continual reassessment method

Abstract

Model-based dose-finding designs such as the continual reassessment method (CRM) rely on some basic working model. In the Bayesian setting, these take the form of 'guess estimates' of the probabilities of toxicity at each level. In the likelihood setting, these estimates simply take the form of a model as operational characteristics are unaffected by arbitrary positive power transformations. These initial estimates are often referred to as the model skeleton. The impact of any prior distribution on the model parameter that describes the dose-toxicity curve will itself depend on the skeleton being used. We study the interplay between prior assumptions and skeleton choice in the context of two-stage CRM designs. We consider the behavior of a two-stage design at the point of transition from a 3 + 3 design to CRM. We study how use can be made of stage 1 data to construct a more efficient skeleton in conjunction with any prior information through an example of a clinical trial. We evaluate to what extent stage 1 data might be down weighted when the maximum tolerated dose (MTD) is far from the starting level and stage 1 data is strongly informative. The results show no improvement in accuracy; thus, weighting is not necessary unless the investigators feel strongly about the location of the MTD and wish to accelerate into the vicinity of the MTD. In general, because this information is not available, we recommend that the design of two-stage trials utilize stage 1 data to establish a skeleton.