Flexible Almost-Exact Matching for Trustworthy Causal Inference

3835 

Contributed Abstract 

Poster 

Quinn Lanners (1), Harsh Parikh (2), Cynthia Rudin (3), Alexander Volfovsky (3), David Page (3)

(1) Duke University School of Medicine, N/A, (2) Johns Hopkins University, N/A, (3) Duke University, N/A

Harsh Parikh  
Johns Hopkins University

Cynthia Rudin  
Duke University

Alexander Volfovsky  
Duke University

David Page  
Duke University

Quinn Lanners  
Duke University School of Medicine

Quinn Lanners  
Duke University School of Medicine

Our goal is to produce methods for observational causal inference that are auditable, easy to troubleshoot, yield accurate treatment effect estimates, and are adaptable to various datasets. We describe an almost-exact matching approach that achieves these goals by (i) learning a distance metric via outcome modeling, (ii) creating matched groups using the distance metric, and (iii) using the matched groups to estimate treatment effects. Our proposed method uses variable importance measurements to construct a distance metric, making it a flexible method that can be adapted to various applications. We operationalize this method into a safe and interpretable framework to identify optimal treatment regimes in a noisy ICU dataset. In this application, we face challenges including missing data, inherent stochasticity, and the critical requirements for interpretability and patient safety. Using our approach, we match patients with similar medical and pharmacological characteristics, allowing us to construct an optimal policy via interpolation. Our findings strongly support personalized treatment strategies based on a patient's medical history and pharmacological features.

causal inference|optimal treatment regime|machine learning|variable importance|matching|medicine

ENAR

Miscellaneous