Bayesian Feature Selection for Multi-valued Treatment Comparisons: An Electronic Health Records Study of Vasopressor Effectiveness

Analyzing treatment effectiveness from electronic health records (EHR) presents unique challenges in causal inference, particularly when comparing multiple treatment options with high-dimensional covariates. We propose a novel framework combining instrumental variable (IV) analysis with advanced Bayesian feature selection methods and neural networks to estimate causal effects in multi-valued treatment settings. Our approach addresses three key methodological challenges: handling multiple treatment comparisons simultaneously, comparing Bayesian feature selection methods, and selecting relevant features while capturing complex nonlinear relationships in outcome models. Through extensive simulation studies, we demonstrate that spike-and-slab priors achieve superior performance in treatment effect estimation with the lowest mean absolute bias (0.071) compared to ALL (0.074), LASSO (0.080), and Bayesian LASSO (0.083) methods. The consistency of bias control across treatment pairs demonstrates the robustness of our Bayesian feature selection approach, particularly in identifying clinically relevant predictors. We apply this framework to compare three commonly used vasopressors (norepinephrine, vasopressin, and phenylephrine) using MIMIC-IV data[1]. Using physician prescribing preferences as instruments[2, 3, 4], our analysis reveals a clear hierarchical pattern in treatment effectiveness. Vasopressin demonstrated superior effectiveness compared to both norepinephrine (ATE = 0.134, 95% CI [0.115, 0.152]) and phenylephrine (ATE = 0.173, 95% CI [0.156, 0.191]), while phenylephrine showed inferior outcomes compared to norepinephrine (ATE = -0.040, 95% CI [-0.048, -0.031]). Our methodological framework provides a robust approach for analyzing multi-valued treatments in high-dimensional observational data, with broad applications beyond vessopressors in critical care. The integration of instrumental variable analysis, Bayesian feature selection, and advanced modeling techniques offers a promising direction for using EHR data to inform treatment decisions while addressing key challenges in causal inference.