A Bayesian nonparametric approach for causal inference in EHR data in the presence of nonignorable missingness
Conference
65th ISI World Statistics Congress 2025
Format: IPS Abstract - WSC 2025
Keywords: "bayesian, causal inference, missing not at random
Session: IPS 691 - Modern Approaches for Causal Analysis Amidst Complex Data Challenges
Wednesday 8 October 2 p.m. - 3:40 p.m. (Europe/Amsterdam)
Abstract
We propose an approach for missingness in EHRs using Bayesian nonparametric (BNP) models. We show how to introduce sensitivity parameters corresponding to nonignorable missingness in the outcome and confounders by extracting unidentified distributions from the BNP model and reconstructing the distribution of interest. We also flexibly include auxiliary covariates to move closer to MAR. We use G-computation based on the reconstructed distribution to compute causal estimands of interest. We use our approach to assess the comparative effectiveness of two bariatic surgeries on BMI 18 months after surgery.