Dose-finding in phase I clinical trials based on toxicity probability intervals

Background: Most phase I clinical trials conducted at the M. D. Anderson Cancer Center use the algorithmic 3 + 3 design, despite the availability of more advanced model-based designs such as the continual reassessment method. Purpose: Through simple statistical modeling and computing, we develop a dose-finding design that can be easily understood and implemented by non-statisticians. Methods: We propose a beta/binomial Bayesian model and a probabilistic up-and-down rule that allow all possible dose-assignment actions to be tabulated in a spreadsheet. We have developed an Excel macro (available at http://odin.mdacc.tmc.edu/~yuanj) that generates trial monitoring tables, which contain the dose-assignment actions corresponding to various toxicity outcomes. Results: The new design outperforms the 3 + 3 design and performs comparably to other model-based methods in the literature. Limitations: The proposed method assumes that the observed toxicity is a binary variable and that toxicity increases with dose level. Conclusion: The new dose-finding design enables physicians to readily determine dose assignments for new patients by referencing a trial monitoring table.