59: Comparing Proxy Variable Methods for Reducing Omitted Variable Bias in EHR-Based Causal Inference

Music City Center 

Propensity score methods are frequently used to estimate causal effects but rely on a no-unobserved-confounder assumption, which may be a concern when using electronic health record (EHR) data to estimate causal effects. We compare three methods that use proxy variables to reduce omitted-variable bias: 1) a naïve approach using proxies directly in estimation, 2) proximal causal inference with inverse probability weighting (PCI) (Cui et al., 2024), and 3) propensity score weighting with an inclusive factor score (IFS) (Nguyen and Stuart, 2020). In simulations with two proxy variables and an unobserved confounder, the naïve approach generally reduced but did not eliminate bias, while the PCI and IFS methods produced unbiased estimates if assumptions were met. Under violations of the conditional independence assumptions which define the role of the proxies in PCI and/or IFS, bias and variability of PCI and IFS estimates could be higher than the naïve approach. Finally, we use these methods to estimate the probability of hospitalization one year after the prescription of one of two antidepressants, using proxies for a potential unobserved confounder, financial assets.

propensity score methods

proximal causal inference

electronic health records

measurement error

unobserved confounding

comparative effectiveness 

Health Policy Statistics Section