A comparative study of dose-finding models with late-onset toxicities

Clinical trials in which the subject's results are not immediately known make it challenging to use conventional dose-finding methods. Several methods have been developed to accommodate the characteristics of these trials, and this study compares these dose-finding models with late-onset toxicities through extensive simulation studies. The impact of the combinations of dose-response relationship, time-to-toxicity model, cohort size, and inter-arrival time on the performance of the approaches was compared. Our simulation study demonstrated that dose-response relationships had the greatest impact on the performance, followed by cohort size, inter-arrival time, and time-to-toxicity models. Overall, the rule-based methods had the advantage of being easy to use with a short test period, but it had the problem of inaccurately estimating the maximum tolerated dose (MTD). Also, the model-based methods had the advantage of accurately estimating the MTD with relatively long test durations and high dose limiting toxicity numbers, and the model-assisted methods accurately predicted the MTD compared to the rule-based methods and the performances were robust to the testing scenarios. This study aims to provide the guidelines for choosing the most appropriate method under a given circumstance.