To incorporate the monotonic toxicity order, a dose-toxicity curve or a parametric model is often assumed. In contrast to algorithm-based designs, such as the 3+3 design, where dose escalation or de-escalation is solely based on the data collected at the current dose level, model-based designs can pool all the information across different doses for decision making. Among a variety of model-based dose-finding methods, the continual reassessment method (CRM) is a very popular approach which is often cast in the Bayesian framework [1]. The CRM may take a single-parameter power model, a two-parameter logistic regression model, or a probit model. Different parametric models represent different patterns of the dose-toxicity ...