Bayesian Nonparametric Causal Models for Point Treatments and Continuous Outcomes

• Activity Number: 601
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Kirsten Lum* and Michael J. Daniels and Jason Roy
• Companies: University of Pennsylvania and The University of Texas at Austin and University of Pennsylvania
• Keywords: Bayesian inference ; causal inference ; confounder ; Dirichlet process ; causal model

Abstract

Causal models are used to model the marginal mean of potential outcomes (marginal structural model) or the average contrast of a potential outcome at the observed treatment level and no treatment, conditional on treatment and confounders (structural mean model). Research on Bayesian methods for fitting causal models is limited, possibly due to the potential for model misspecification when parametric models are used for observed data. We propose a Bayesian nonparametric approach to fitting causal models in the setting of a point treatment, observed confounders, and a continuous outcome. We model the distribution of the outcome conditional on the treatment and confounders in a flexible way using a dependent Dirichlet process mixture model. Under causal identifying assumptions, we show how this model can be adapted to satisfy the constraints of the causal model and estimate the posterior distributions of the causal parameters. We present results from simulation studies comparing the robustness of our method with doubly robust estimators. We illustrate our approach with a study of antiretroviral therapy regimens and hepatic decompensation biomarkers among HIV/HCV coinfected patients.