CFE 2019: Start Registration
View Submission - CMStatistics
Title: Dempster-Shafer adaptive clinical trials and A/B tests Authors:  Paul Edlefsen - Fred Hutchinson Cancer Research Center (United States) [presenting]
Raabya Rossenkhan - Fred Hutchinson Cancer Research Center (United States)
Abstract: Dempster-Shafer is a statistical framework that extends Bayesian analysis by incorporating a set-valued representation of probabilities, yielding ``imprecise probability'' measures called Dempster-Shafer PQR values. These are posterior assessments of a hypothesis that allow for a new dimension of uncertainty: ``don't know'' ($R>0$). In this framework, the outcome of an analysis might be ``there is insufficient evidence to make a conclusion; collect more data''. Recent advances in employing binomial Dempster-Shafer inferences sequentially in an ``online'' adaptive experiment are presented, where decisions to draw a conclusion or to gather further data are made sequentially. We demonstrate the power of the Dempster-Shafer adaptive binomial clinical trial design through simulation studies under various cost (loss) function scenarios, with applications to early-phase clinical testing of (for example) HIV vaccines, as well as to A/B testing in the (other sense of the) ``online'' setting.