A framework for understanding selection bias in real-world healthcare data

Area of Science:

• Epidemiology
• Biostatistics
• Health Informatics

Background:

• Administrative patient-care data, including Electronic Health Records (EHR) and claims data, are increasingly utilized in population-based research.
• Large sample sizes in such research can lead to minimal standard errors, necessitating attention to biases that persist regardless of sample size, such as selection bias.

Purpose of the Study:

• To present an analytic framework using directed acyclic graphs to dissect selection bias in research using administrative health data.
• To introduce and evaluate four weighting approaches for mitigating selection bias in the estimation of associations between outcomes and exposures.
• To provide practical guidance and tools for applied researchers working with complex health datasets.

Main Methods:

• Utilized directed acyclic graphs (DAGs) to conceptualize and analyze sources of selection bias.
• Developed and described four distinct weighting methodologies designed to reduce selection bias.
• Conducted a simulation study to assess the performance of the proposed weighting methods.
• Applied the methods to real-world Electronic Health Records (EHR) data from a longitudinal biorepository to examine the association between cancer and biological sex.

Main Results:

• The directed acyclic graph (DAG) framework effectively guides the dissection of selection bias in EHR research.
• Simulation studies demonstrated the practical utility of the proposed weighting approaches under specific conditions.
• Comparison of methods using real-world EHR data provided insights into their performance in practice.
• Annotated R code is provided for implementing the discussed weighting methods and associated statistical inference.

Conclusions:

• Selection bias is a critical consideration in population-based research using administrative health data.
• The proposed analytic framework and weighting methods offer practical solutions for addressing selection bias.
• The availability of R code facilitates the application of these techniques in future research, enhancing the reliability of findings from EHR data.