Bayesian Framework for Predictive and Causal Modeling Using BART

• Activity Number: 80 - Graphical Models and Causal Inference
• Type: Contributed
• Date/Time: Sunday, July 28, 2019 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Yizhen Xu* and Tao Liu and Rami Kantor and Ann Mwangi and Michael Daniels and Joseph Hogan
• Companies: Brown University and Brown University and Brown University and Moi University and University of Florida and Brown University
• Keywords: BART; Causal Inference; Bayesian; Decision Making; Electronic Health Records

Abstract

We developed a Bayesian approach to statistical modeling of longitudinal data for event predictions and causal comparisons of policies in decision making, where we (1) formulate a sequence of predictive models over time using additive regression trees, enabling a machine learning based prediction of multiple correlated events such as engagement and death in the context of HIV care, and (2) use the predictive models as generative components of a large Bayesian functional causal model to calculate causal effects over time via the g computation algorithm (GCA). Conceptually, our method is similar to mathematical modeling in comparing policies such that sequential generative models are constructed as the basis for simulating trajectories of counterfactual outcomes over time using Monte Carlo simulation; however, in our method, generative components are predictive and data-driven, and the causal inference of counterfactual outcomes are based on posterior predictive distributions. We applied the proposed method to electronic health records collected through AMPATH to evaluate the impact of different HIV treatment initiation policies on patients' survival and engagement in care over time.