(195) Methods for imputing race and ethnicity in real-world data: a scoping review

Background: : Real-world data (RWD) are increasingly used to inform healthcare research, but missing race and ethnicity data remain a major challenge. These gaps hinder equity-focused analysis, and while imputation methods offer a way to estimate missing values, the lack of a gold standard raises concerns about bias and accuracy.

Objective: : To examine methods used to impute missing race and ethnicity in RWD. Secondary objectives included identifying populations and settings where imputation has been applied, categorizing methods used, and evaluating their accuracy and validity.

Methods: : Following PRISMA-P guidance, we conducted a scoping review of peer-reviewed studies published in English from 2005 to 2025. Eligible studies used or evaluated imputation methods for race and ethnicity in healthcare RWD. PubMed, EMBASE, and Web of Science were searched. Two reviewers screened records and extracted data on method type, data source, validation, and population characteristics.

Results: : Fifty studies met inclusion criteria; 28 (56%) were published between 2021-2025. Thirty-eight applied imputation methods in research, while 12 evaluated them. Of the 12, 4 (33%) evaluated Bayesian surname-geocoding methods (e.g., BISG, BIFSG), 2 (17%) assessed Multiple Imputation by Chained Equations (MICE), and others evaluated reallocation, deep learning, calibrated-δ adjustment, multinomial regression, or compared methods.

Among the 38 application studies, MICE was used in 25 (66%), followed by other multiple imputation (6; 16%), Bayesian methods (4; 11%), hot deck (2; 5%), and census-based imputation (1; 3%). Among 27 studies reporting missingness percentages, levels ranged from < 5% to >75%. Most (15) assumed Missing At Random (MAR), 2 assumed Missing Not At Random (MNAR), and 21 did not specify. Validation or sensitivity analyses were reported in 20 studies (53%).

Race and ethnicity classification varied. Only 4 (11%) used full Office of Management and Budget (OMB) categories; 34 (89%) used non-standard terms (e.g., South Asian, Latinx). “Other” was used in 31 (82%) of studies, and 3 (8%) dichotomized race as White vs. non-White.

Conclusions: : Race and ethnicity imputation is increasingly used in healthcare RWD. MICE was the most common method, but assumptions and validation practices varied. In the absence of a gold standard, greater transparency and standardization are needed to support equity-focused research.

Sponsorship:: This work is supported by the Biologics and Biosimilars Collective Intelligence Consortium.