A Bayesian model averaging method for dose ranging studies in oncology

Dose finding and optimization studies are important in oncology drug development for making new drugs available to patients at pace and for reducing the risk of toxicity. The requirement by Project Optimus to conduct a randomized dose optimization study necessitates a change in the oncology drug development paradigm, and standard dose-response modeling approaches (e.g., Multiple Comparisons Procedure-Modeling, MCP-Mod) are not always applicable due to small sample sizes and a small number of doses that are typical of dose optimization studies. An innovative Bayesian model averaging method for dose ranging studies (BAMADOS) is proposed for the design and analysis of oncology trials with binary endpoints. The method assumes a monotonic relationship between response and doses of an investigational drug. It does not require pre-specification of candidate models but instead evaluates all possible models in a constrained model space. To minimize the potential impact of the Occam's razor property, non-conjugate moderately informative priors from the family of generalized normal priors are implemented. We show via simulation studies that BAMADOS correctly identifies the optimal biological dose in a dose optimization setting. It further estimates response rates with little bias, even in the presence of discordance between the priors and observed data. Compared to MCP-Mod, BAMADOS exhibited higher power for small sample sizes (30 or fewer participants per dose) and comparable power for larger sample sizes in several scenarios considered.