Title: Practical Bayesian non-parametric inference in clinical development: Single arm dose escalation studies
Authors: Fabio Rigat - Janssen R&D (United Kingdom) [presenting]
Abstract: Dose escalation decisions in early phase clinical trials are sensitive to the functional form of the dose-response relation, with monotonicity being a key assumption. When little background information is available to inform these assumptions on biological and clinical grounds, it is unclear whether dose escalation decisions can be usefully based on regression models or whether regression-free methods should be preferred. The validity and efficiency of decisions based on a weighted average of regression estimates and observed event rates has not yet been fully assessed. These estimators are attractive in practice when the weight assigned to their regression component is estimated from the data, reflecting the models goodness of fit. Lack of widespread application of these methods for clinical trial decision making is surprising, given that they were first published as an application of Bayesian non-parametric (BNP) inference using the Dirichlet process over forty years ago. To show the potential of BNP inference for dose escalation trial design and analysis, we compare the operating characteristics of the modified continual reassessment method (mCRM) informed by either parametric or BNP estimates when the parametric model assumptions respectively do or do not hold. BNP mCRM, implemented by standard Markov chain Monte Carlo packages, is shown to offer important practical advantages when the parametric inference model is too wrong to be useful.