Comparing Causal Inference Methods for Point Exposures with Missing Confounders: A Simulation Study

TL;DR

This study compares various causal inference methods for estimating treatment effects in electronic health records with missing confounders, highlighting the strengths and limitations of each approach through simulation.

Contribution

It evaluates recently proposed estimators for combined handling of missing data and confounding, providing practical recommendations based on simulation results.

Findings

No single estimator is best across all scenarios.

Ad hoc methods perform well in some cases but lack consistency.

The study offers guidance for choosing causal inference methods with missing confounders.

Abstract

Causal inference methods based on electronic health record (EHR) databases must simultaneously handle confounding and missing data. Vast scholarship exists aimed at addressing these two issues separately, but surprisingly few papers attempt to address them simultaneously. In practice, when faced with simultaneous missing data and confounding, analysts may proceed by first imputing missing data and subsequently using outcome regression or inverse-probability weighting (IPW) to address confounding. However, little is known about the theoretical performance of such $\textit{ad hoc}$ methods. In a recent paper Levis $\textit{et al.}$ outline a robust framework for tackling these problems together under certain identifying conditions, and introduce a pair of estimators for the average treatment effect (ATE), one of which is non-parametric efficient. In this work we present a series of simulations, motivated by a published EHR based study of the long-term effects of bariatric surgery on weight outcomes, to investigate these new estimators and compare them to existing $\textit{ad hoc}$ methods. While the latter perform well in certain scenarios, no single estimator is uniformly best. We conclude with recommendations for good practice in the face of partially missing confounders.