Primary Submission Category: Matching, Weighting
A Deep Learning Approach to Nonparametric Propensity Score Estimation with Optimized Covariate Balance
Authors: Liang Li, Maosen Peng, Chong Wu, Yan Li,
Presenting Author: Liang Li*
This paper addresses some key challenges in causal inference using a case study on the effect of erythrocyte-to-platelet ratio (EPR) changes on sepsis outcomes based on the MIMIC-IV electronic health records database. Observed inconsistencies across existing propensity score methods highlight issues such as model misspecification, poor overlap, and inadequate covariate balance. To overcome these limitations, we propose a novel propensity score weighting method based on two sufficient and necessary conditions: “local balance,” ensuring conditional independence of covariates and treatment assignment across a dense grid of balancing scores, and “local calibration,” guaranteeing that the balancing scores correspond to the true propensity scores. Using a neural network, we develop a nonparametric propensity score model that satisfies these conditions, effectively optimizing covariate balance, minimizing bias, and stabilizing inverse probability of treatment weights. Extensive numerical studies demonstrate that the proposed method successfully addresses these challenges, providing robust treatment effect estimation. In the case study, high EPR changes were associated with a significant 16% increased risk of 28-day mortality (HR: 1.16, 95% CI: 1.04–1.29). Our method offers a practical solution for causal inference in complex observational studies, with broad implications for improving clinical decision-making in critical care settings.