A simple technique to evaluate model sensitivity in the continual reassessment method

Abstract

The continual reassessment method (CRM) is a sequential design used in phase I cancer trials to determine the maximal dose with acceptable toxicity. It has been established that the CRM is consistent under model misspecification but not generally. When the method does not converge to the target percentile, some dose-response models will be more sensitive than others in terms of how close the converged recommendation is to the target. In this article, we interpret the main condition under which the CRM is consistent and apply it to evaluate the sensitivity of the model used with the CRM. The technique presented is found to be a useful supplement to simulation when planning a phase I trial.