Matthieu Resche-Rigon1, Sarah Zohar, Sylvie Chevret. 1. Biostatistical Department, U717 Inserm, AP-HP, Paris 7 University, Saint-Louis Hospital, 1 Avenue Claude Vellefaux, 75475 Paris Cedex 10, France. matthieu.resche-rigon@univ-paris-diderot.fr
Abstract
BACKGROUND: In non-cancer phase II trials, dose-finding trials are usually carried out using fixed designs, in which several doses including a placebo are randomly distributed to patients. However, in certain vulnerable populations, such as neonates or infants, there is an heightened requirement for safety, precluding randomization. PURPOSE: To estimate the minimum effective dose of a new drug from a non-cancer phase II trial, we propose the use of adaptive designs like the Continual Reassessment Method (CRM). This approach estimates the dose closest to some target response, and has been shown to be unbiased and efficient in cancer phase I trials. METHODS: Based on a motivating example, we point out the individual influence of first outliers in this setting. A weighted version of the CRM is proposed as a theoretical benchmark to control for these outliers. Using simulations, we illustrate how this approach provides further insight into the behavior of the CRM. RESULTS: When dealing with low targets like a 10% failure rate, the CRM appears unable to rapidly overcome an early unexpected outcome. This behavior persisted despite changing the inference (Bayesian or likelihood), underlying dose-response model (though slightly improved using the power model), and the number of patients enrolled at each dose level. LIMITATIONS: The choices for initial guesses of failure rates, the vague prior for the model parameter, and the log-log shape of weights can appear somewhat arbitrary. CONCLUSIONS: In phase II dose-finding studies in which failure targets are below 20%, the CRM appears quite sensitive to first unexpected outcomes. Using a power model for dose-response improves some behavior if the trial is started at the first dose level and includes at least three to five patients at the starting dose before applying the CRM allocation rule.
BACKGROUND: In non-cancer phase II trials, dose-finding trials are usually carried out using fixed designs, in which several doses including a placebo are randomly distributed to patients. However, in certain vulnerable populations, such as neonates or infants, there is an heightened requirement for safety, precluding randomization. PURPOSE: To estimate the minimum effective dose of a new drug from a non-cancer phase II trial, we propose the use of adaptive designs like the Continual Reassessment Method (CRM). This approach estimates the dose closest to some target response, and has been shown to be unbiased and efficient in cancer phase I trials. METHODS: Based on a motivating example, we point out the individual influence of first outliers in this setting. A weighted version of the CRM is proposed as a theoretical benchmark to control for these outliers. Using simulations, we illustrate how this approach provides further insight into the behavior of the CRM. RESULTS: When dealing with low targets like a 10% failure rate, the CRM appears unable to rapidly overcome an early unexpected outcome. This behavior persisted despite changing the inference (Bayesian or likelihood), underlying dose-response model (though slightly improved using the power model), and the number of patients enrolled at each dose level. LIMITATIONS: The choices for initial guesses of failure rates, the vague prior for the model parameter, and the log-log shape of weights can appear somewhat arbitrary. CONCLUSIONS: In phase II dose-finding studies in which failure targets are below 20%, the CRM appears quite sensitive to first unexpected outcomes. Using a power model for dose-response improves some behavior if the trial is started at the first dose level and includes at least three to five patients at the starting dose before applying the CRM allocation rule.
Authors: Alexina J Mason; Juan Gonzalez-Maffe; Killian Quinn; Nicki Doyle; Ken Legg; Peter Norsworthy; Roy Trevelion; Alan Winston; Deborah Ashby Journal: Stat Med Date: 2016-11-27 Impact factor: 2.373