Bias Reduction Methods for Propensity Scores Estimated from Error-Prone EHR-Derived Covariates.

Area of Science:

• Health Informatics
• Epidemiology
• Biostatistics

Background:

• Electronic health records (EHR) are increasingly used for treatment effect estimation.
• Measurement error in EHR-derived covariates can introduce bias into propensity score models.
• Limited research exists on handling measurement error in propensity scores constructed from mixed-accuracy covariates.

Purpose of the Study:

• To review and compare methods for accounting for measurement error in propensity scores.
• To evaluate the performance of these methods in simulation studies across various scenarios.
• To apply and compare these methods in a real-world EHR-based comparative effectiveness study.

Main Methods:

• Literature review of approaches to correct for measurement error in propensity scores.
• Simulation studies varying outcome type, sample sizes, confounding strength, and error structure.
• Application of correction methods to an EHR study on metastatic bladder cancer treatments.

Main Results:

• Multiple imputation for propensity scores demonstrated superior performance for continuous outcomes.
• Regression calibration-based methods showed better performance for binary outcomes.
• Performance varied based on simulation parameters, including validation sample size and error structure.

Conclusions:

• The choice of measurement error correction method for propensity scores depends on the outcome type.
• Multiple imputation is recommended for continuous outcomes, while regression calibration is preferred for binary outcomes in EHR analyses.
• These findings are crucial for improving the validity of causal inference from EHR data.