|
Activity Number:
|
202
- Meta-Analysis, Mediation, and Causal Inference from a Bayesian Perspective
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 8, 2022 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
|
Abstract #320888
|
|
|
Title:
|
Uncertainty Calibration and Exemplar Identification for Heterogeneous Treatment Effects with Individualized Bayesian Causal Forests
|
|
Author(s):
|
Jennifer Starling* and Lauren Vollmer and Erin Lipman and Peter Mariani and Daniel Thal and Irina Degtiar and Mariel McKenzie Finucane
|
|
Companies:
|
Mathematica Policy Research and Mathematica Policy Research and University of Washington and Mathematica Policy Research and Mathematica Policy Research and Mathematica Policy Research and Mathematica
|
|
Keywords:
|
heterogeneous treatment effects;
causal forest;
regression tree;
BART;
random effects
|
|
Abstract:
|
Bayesian Causal Forests (BCF) has proven to be an effective framework for estimating heterogeneous treatment effects, particularly where standard methods yield biased estimates due to residual confounding. BCF parameterizes BART models to allow for separate regularization of treatment and prognostic effects, making it possible to shrink towards homogeneity. However, BCF’s credible intervals are known to under-cover, and treatment effect estimates are purely a function of measured covariates, without the ability to identify “exemplar” observations that respond to treatment more strongly than covariates predict. We introduce iBCF, a new version of the BCF prior that incorporates observation-level random effects to allow for robust estimation of individual causal impacts. We demonstrate the utility of this approach in calibrating uncertainty and identifying exemplar observations. We showcase the benefits of this approach in a simulation based on real data from a Medicare primary care program.
|
Authors who are presenting talks have a * after their name.
