Summary
Electronic health record (EHR) data provides an essential resource for comparative effectiveness research, but missing data is a persistent challenge, particularly for confounder variables.
Propensity score-based inverse probability of treatment weights (IPTWs) is a promising approach to control for confounders, but there needs to be more guidance on how to handle missing data. Most investigators ignore missing data or use crude methods, with only 45% describing a missing data approach.
Multiple imputation (MI) and propensity score calibration (PSC) are popular methods, but their performance has not been directly compared in EHR-based comparative effectiveness analyses. In this study, researchers from the University of Pennsylvania and Flatiron Health compared the performance of complete case analysis (CC), MI, and PSC in an EHR-based comparative effectiveness analysis, considering potential misclassification of outcome variables.
Why this matters
The increasing use of EHR data for research is hindered by the issue of missing data, particularly in the confounding variables such as disease severity, symptoms, and social determinants of health. This missing data can lead to biased results and incorrect conclusions about treatment effectiveness, which can have serious consequences for patient care and healthcare policy. Therefore, it is important to develop robust methods to handle missing data in EHR-based comparative effectiveness research and understand which are reasonable to apply in RWE analyses.