Evaluating Methods for Imputing Race and Ethnicity in Electronic Health Record Data.

Area of Science:

• Health Informatics
• Epidemiology
• Biostatistics

Background:

• Accurate race and ethnicity data are crucial for understanding chronic disease burden.
• Electronic Health Records (EHRs) are valuable data sources but often have missing demographic information.
• Imputation methods aim to fill in missing data, but their performance can vary.

Purpose of the Study:

• To compare anonymized and non-anonymized imputation methods for race and ethnicity in EHR data.
• To assess the accuracy and agreement of different imputation techniques.
• To evaluate the impact of imputation methods on chronic disease burden estimates.

Main Methods:

• Simulation analyses were conducted using various missing data mechanisms.
• Imputation methods included Bayesian Improved Surname Geocoding (BISG), single imputation, random forest, and multiple imputation with chained equations (MICE).
• Performance was measured by sensitivity, precision, accuracy, and Cohen's kappa (κ) agreement with self-reported data. Methods were applied to two EHR datasets.

Main Results:

• Non-anonymized BISG showed the highest accuracy (66-73%) in simulations.
• Anonymized methods' agreement decreased when data was missing not at random (MNAR).
• Despite simulation differences, racial/ethnic distributions and disease burden estimates were consistent across methods in EHR data. All imputation methods improved estimate precision compared to complete case analysis.

Conclusions:

• Non-anonymized BISG may offer superior racial and ethnic classification, especially under MNAR conditions.
• Descriptive studies on chronic disease burden may be robust to the choice of imputation method for race and ethnicity.
• Further research can explore the nuances of imputation in diverse EHR populations.