Background: Phase I studies in oncology using the continual reassessment method (CRM), which is a Bayesian approach to utilize prior information about the clinical question, are increasing. If a drug has already been tested in another country, some information should be available before the study. Based on the available informative, the prior distribution of model parameters, which reflects (un) certainty, can be reasonably informative. However, if dose-limiting toxicity (DLT) is observed in the initial cohort given the starting dose that is considered “sufficiently safe,” investigators would query the validity of the prior information, but it is difficult at this stage to differentiate whether the DLT is an incidental event or caused by underestimating the true DLT probability. In the latter case, the incorrect prior distribution will lead to an overestimation of the maximum tolerated dose (MTD) or increase the number of patients treated with higher dose levels.
Methods: We propose a CRM that addresses the possibility of underestimation of the true DLT probability. Our method allows adaptive change of the prior distribution according to the state of DLT observed in the initial cohort, thereby minimizing the effect of an incorrect prior distribution.
Results: We compared the proposed method with the conventional CRM in some simulation studies. According to the results of simulation study, our method offers robustness against incorrect prior information, although efficiency is slightly decreased.
Conclusions: Considering that the true dose-response relationship is not known prior to a study, the proposed method shows reasonable operation characteristics compared to the conventional CRM, based on our simulation study, and has considerable flexibility for further applications.