Keywords: dose finding, maximum tolerated dose, toxicity probability interval, interval design, model-assisted design
A number of novel model-based and model-assisted designs have been proposed to find the maximum tolerated dose (MTD) in phase I clinical trials, but their differences and relative pros and cons are not clear to many practitioners. We review three model-based designs, including the continual reassessment method (CRM), dose escalation with overdose control (EWOC), and Bayesian logistic regression model (BLRM), and three model-assisted designs, including the modified toxicity probability interval (mTPI), Bayesian optimal interval (BOIN), and keyboard designs. We conduct numerical studies to assess their accuracy, safety and reliability, and the practical implications of various empirical rules used in some designs, such as skipping a dose and imposing overdose control. Our results show that the CRM outperforms EWOC and BLRM with higher accuracy of identifying the MTD. For the CRM, skipping a dose is not recommended as it substantially increases the chance of overdosing patients, while providing limited gain for identifying the MTD. EWOC and BLRM appear excessively conservative. They are safe, but have relatively poor accuracy of finding the MTD. The BOIN and keyboard designs have similar operating characteristics, outperforming the mTPI, but the BOIN is more intuitive and transparent. The BOIN yields competitive performance comparable to the CRM, but is simpler to implement and free of the issue of irrational dose assignment caused by model misspecification, thereby providing an attractive approach for designing phase I trials.