Abstract
BACKGROUND: Electronic health records (EHRs) are increasingly used to investigate health inequalities across ethnic groups. While there are some studies showing that the recording of ethnicity in EHR is imperfect, there is no robust evidence on the accuracy between the ethnicity information recorded in various real-world sources and census data. METHODS AND FINDINGS: We linked primary and secondary care NHS England data sources with Census 2021 data and compared individual-level agreement of ethnicity recording in General Practice Extraction Service (GPES) Data for Pandemic Planning and Research (GDPPR), Hospital Episode Statistics (HES), Ethnic Category Information Asset (ECIA), and Talking Therapies for anxiety and depression (TT) with ethnicity reported in the census. Census ethnicity is self-reported and, therefore, regarded as the most reliable population-level source of ethnicity recording. We further assessed the impact of multiple approaches to assigning a person an ethnic category. The number of people that could be linked to census from ECIA, GDPPR, HES, and TT were 47.4m, 43.5m, 47.8m, and 6.3m, respectively. Across all 4 data sources, the White British category had the highest level of agreement with census (≥96%), followed by the Bangladeshi category (≥93%). Levels of agreement for Pakistani, Indian, and Chinese categories were ≥87%, ≥83%, and ≥80% across all sources. Agreement was lower for Mixed (≤75%) and Other (≤71%) categories across all data sources. The categories with the lowest agreement were Gypsy or Irish Traveller (≤6%), Other Black (≤19%), and Any Other Ethnic Group (≤25%) categories. CONCLUSIONS: Certain ethnic categories across all data sources have high discordance with census ethnic categories. These differences may lead to biased estimates of differences in health outcomes between ethnic groups, a critical data point used when making health policy and planning decisions.