04.22.2025 Seminar Series: A Deep Learning Approach to Nonparametric Propensity Score Estimation with Optimized Covariate Balance

Seminar Series

Tuesday, April 22, 2025

12:00-1:00PM

Join Zoom Webinar: https://www.zoomgov.com/j/1606031888?pwd=9jP5evH9G0vnd7Lweto0V2HU5XYzZv.1

Deep Learning Approach to Nonparametric Propensity Score Estimation with Optimized Covariate Balance

Presented by: Moasen Peng, MS, MD Anderson Graduate School of Biomedical Sciences

Abstract: Causal inference in observational studies presents significant challenges, particularly when estimating treatment effects under model misspecification, poor covariate overlap, and inadequate balance. This talk introduces a novel propensity score weighting method designed to address these issues, motivated by a case study on the effect of erythrocyte-to-platelet ratio (EPR) changes on sepsis outcomes using the MIMIC-IV electronic health records database.

The proposed approach is grounded in two essential conditions: local balance, which ensures conditional independence of covariates and treatment assignment across a dense grid of balancing scores, and local calibration, which guarantees that balancing scores accurately reflect true propensity scores. A neural network is employed to construct a nonparametric propensity model that optimizes covariate balance, minimizes bias, and stabilizes inverse probability weights.

Extensive numerical studies validate the effectiveness of this method, and findings from the case study indicate that high EPR changes are associated with a 16% increased risk of 28-day mortality (HR: 1.16, 95% CI: 1.04–1.29). Additionally, a live demonstration of the R package implementation will be provided, illustrating how this method can be seamlessly applied to real-world causal inference problems.

Virtual Seminar: Contact Anirudh Babu (asbabu@utmb.edu) for additional information.