Stratified split sampling of electronic health records

Area of Science:

• Health Informatics
• Biostatistics
• Epidemiology

Background:

• Electronic health record (EHR) data, while resembling clinical research data, necessitates distinct analytical strategies.
• EHR data's primary design for clinical care, not scientific inquiry, requires explicit definition of outcome and predictor variables.
• Iterative variable definition and analysis in EHR studies can inflate Type I errors, compromising replicability.

Purpose of the Study:

• To propose a stratified split sample approach for enhancing the replicability and generalizability of findings from electronic health record studies.
• To address the challenges of variable definition, subgroup analysis, and heterogeneity in EHR data.
• To provide a methodological framework for robust scientific inference using EHR data.

Main Methods:

• A split sample approach divides EHR data into exploratory and confirmatory sets for iterative analysis and validation.
• Stratified sampling oversamples rare subgroups within the exploratory set to ensure adequate sample size for heterogeneity assessment.
• Testing for effect modification by group membership allows for the examination of heterogeneous associations within subgroups.

Main Results:

• The recommended approach aims to increase the chances of obtaining consistent results across studies.
• It helps to avoid masking heterogeneous associations by accounting for subgroup differences.
• The method enhances the generalizability of findings by ensuring subgroup representation.

Conclusions:

• A stratified split sample approach is recommended for electronic health record research to improve replicability and generalizability.
• This methodology facilitates the assessment of subgroup heterogeneity and effect modification.
• The approach was illustrated using a study on socio-demographic factors and hepatic cancer screening uptake.