Comparing performances of several response-adaptive designs in dose finding studies

Bruno Lecoutre (1) , Derzko Gérard (2)

1 LMRS - Laboratoire de Mathématiques Raphaël Salem
2 Sanofi-Aventis R&D

The main objectives of dose-finding studies are (i) to select the best dose(s) for future phase 3 and/or (ii) to evaluate the dose-response relationship. At that phase 2B stage of the drug development, little is known about the drug's efficacy, and only a small- to moderate-size sample can be experimented. Response-Adaptive (RA) designs, where the dose allocation to new patients depends on the information got from previous allocations and responses, do match dose-finding study conditions well. Therefore one can wonder whether they are valuable alternative designs to Balanced Randomization (BR) designs to meet the above-mentionned dose-finding objectives. Here the operating characteristics and the capability of selecting the best dose(s) and of detecting pre-specified dose contrasts are compared in four multiple dose RA designs namely two Play-the-Winner type, Generalized Drop-the-Loser and Doubly-adaptive Biased Coin designs. The restriction of these designs to two treatments have been shown to be the best choices indeed, as well as the Bayesian procedure used here for inference. The critical concept of target allocation rule is elaborated, and made explicit to match the specific dose-finding objectives. It is shown that the studied RA designs are efficient competitors to BR designs, but none of them outperforms the others on all purposes. Directions for RA design selection are given to experimenters, depending on the primary objective of their planned dose-finding study.

Mots clés: Bayesian inference; Dose selection; Power; response-Adaptive design; Treatment failures

Domaines: Méthodologie [stat.ME]

Bruno Lecoutre, Derzko Gérard. Comparing performances of several response-adaptive designs in dose finding studies. 2011. ⟨hal-00911294⟩