A0911
Title: Bayesian methods for estimating optimal treatment strategies from real-world data
Authors: Jason Roy - Rutgers University (United States) [presenting]
Abstract: In many clinical studies, interest is in learning about the best treatment strategies for individual patients. The optimal treatment strategy might be a function of history data, including how effective previous treatments have been. Challenges in learning about treatment strategies from real-world data include the possibility that waiting times between treatments are informative, that not all variables are measured at all times, and that there might be a large number of confounders to control for. To address these issues, an approach for modeling repeated measures data is developed in continuous time while allowing for informative observation times and missing data. Several Bayesian nonparametric models are used to avoid making strong parametric assumptions. G-computation was used to obtain posterior distributions for the effectiveness of different treatment strategies. The methods were applied to a study of pediatric acute myeloid leukemia.
