Comparing causal inference methods for point exposures with missing confounders: a simulation study | JoVE Visualize

Area of Science:

Biostatistics
Health Informatics
Causal Inference

Background:

Electronic health record (EHR) databases present challenges for causal inference, requiring simultaneous handling of confounding and missing data.
Existing methods often address these issues sequentially using imputation followed by outcome regression or inverse-probability weighting (IPW), with limited understanding of their combined performance.
Few studies formally integrate missing data and confounding adjustments within a unified causal inference framework.

Purpose of the Study:

To investigate the performance of novel causal inference estimators designed to simultaneously address confounding and missing data in EHRs.
To compare these new estimators against traditional ad hoc methods using simulations motivated by a real-world study on bariatric surgery outcomes.
To provide recommendations for best practices when dealing with partially missing confounders in causal inference from EHR data.

Main Methods:

Simulation study based on a published EHR study examining bariatric surgery's long-term weight outcomes.
Evaluation of newly proposed non-parametric efficient and other estimators for average treatment effect (ATE) under missing confounder scenarios.
Comparison with established ad hoc methods combining imputation and confounding adjustment techniques like outcome regression and IPW.

Main Results:

Ad hoc methods combining imputation and confounding adjustment demonstrate good performance in specific scenarios but lack universal superiority.
No single causal inference estimator consistently outperformed others across all simulated conditions for handling partially missing confounders.
The study highlights the complexity of simultaneously addressing missing data and confounding in EHR-based causal inference.

Conclusions:

The choice of causal inference method for EHR data with missing confounders is context-dependent.
While ad hoc methods can be effective, their performance is not uniform, necessitating careful consideration.
Recommendations for best practices are provided to guide analysts in selecting appropriate methods for handling partially missing confounders in EHR studies.