Double Sampling for Informatively Missing Data in Electronic Health Record-Based Comparative Effectiveness Research.

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Missing data are common in electronic health records (EHR), particularly when data are missing not at random (MNAR).
• Existing sensitivity analyses for MNAR data often lack actionable conclusions.
• Bariatric surgery outcome studies frequently encounter MNAR data.

Purpose of the Study:

• To introduce and evaluate double sampling as a method to address MNAR outcome data in EHR.
• To enable accurate estimation and inference of causal effects despite missing data.
• To provide robust statistical tools for health outcomes research.

Main Methods:

• Developed assumptions for identifying joint distributions under double sampling.
• Derived efficient and robust estimators for average causal treatment effect (ACTE).
• Compared estimators under nonparametric and missing at random (MAR) models via simulations.

Main Results:

• Double sampling provides a framework for causal inference with MNAR data.
• Proposed estimators demonstrate efficiency and robustness.
• The method extends to handle arbitrary data coarsening mechanisms.

Conclusions:

• Double sampling is a viable strategy for mitigating MNAR data in EHR studies.
• The derived estimators offer improved causal effect estimation.
• This approach enhances the reliability of findings in health research with incomplete data.