Validation of Patients’ Race, Ethnicity, and Language Data in a Pediatric Emergency Department

Disaggregating performance metrics by patients' race, ethnicity, and language (REaL) can identify health care disparities. However, electronic health record (EHR) data are often inaccurate and not representative of patients' self-identified REaL, the gold standard.1–3 We assessed REaL data validity in our pediatric emergency department (ED) by comparing EHR-documented to survey-reported REaL data.As part of a quality improvement project, we surveyed a convenience sample of 92 patients in our pediatric ED between December 2022 and April 2023. We excluded patients with critical and severe illness, who were ≥18 years of age, or who were lacking any EHR-documented REaL data. We obtained EHR-documented REaL data through chart review. We collected self-reported REaL data via face-to-face surveys of patients' parents (race and ethnicity reflected patient identities and language reflected parent identities). Survey questions mirrored standardized language intended for use by the registration staff when soliciting REaL data.4The Supplemental Information details additional survey and institutional REaL data procedures.We reported the number and proportion of patients in each REaL group by EHR-documented and survey-report. Consistent with the practices of our health system and multiple government agencies, we reported a combined race and ethnicity category, grouping Hispanic and Latino patients, regardless of race.5,6 To assess sample representativeness, we compared our sample's EHR-documented REaL data to 1 year of ED visits. We compared EHR-documented and survey-reported REaL data using χ-square tests and concordance using Cohen's κ. We cross-tabulated EHR-documented versus survey-reported REaL data to identify patterns of discordance. Columbia University's Institutional Review Board approved this study.In comparing our sample's EHR data to the overall ED population, racial and ethnic distributions were similar but differed by language (P < .001; Table 1).EHR-documented and survey-reported data were statistically different across each data type (P < .0001). There was 53% (κ 0.30), 85% (κ 0.66), 84% (κ 0.66), and 84% (κ 0.55) agreement between EHR-documented and survey-reported race, ethnicity, combined race and ethnicity, and language data, respectively (Table 2). Most cases of discordance between EHR-documented and survey-reported race, ethnicity, and race and ethnicity included ambiguous data documented in the EHR ("other," "declined/unknown"), as follows: race (27/43), ethnicity (11/15) and combined race and ethnicity (11/14). Most language discordance occurred between EHR-documented Spanish and survey-reported English (9/15). A sensitivity analysis excluding patients with ambiguous EHR data improved concordance in all categories.Identifying and reducing health care disparities requires valid REaL data. In our study, EHR-documented and survey-reported REaL data concordance varied; concordance was lowest for race (53%) and improved with a combined race and ethnicity category (85%). Most race and ethnicity discordance occurred when ambiguous (unknown, declined, or other) data in the EHR conflicted with a specific identity listed in the survey. Most language discordances included EHR-documented Spanish and survey-reported English.We cannot infer the etiologies of REaL data discordance from our study. However, for our hospital's majority Hispanic and Latino population, discordance may represent malalignment between REaL categories and patients' identities; specifically, Latino patients may feel a stronger affiliation to the "Hispanic/Latino" ethnicity than racial categories, prompting a choice of "other/declined" race. Our data and the literature suggest that combined race and ethnicity categories may better represent the identities of Latino patients,7 despite the sacrifice of specificity and nuance.8Beyond the Latino population, discordant data may reflect inadequate data collection practices amenable to future quality improvement. For example, reducing ambiguous data in the EHR represents a key target to improve discordance. However, discordance may also reflect natural variation in REaL identities as children age and variation in caregivers' preferred language by clinical scenario or the caregiver present.No standard or benchmark rate of EHR versus self-reported REaL data concordance exists.3 However, our results reveal similar or improved concordance compared with the existing literature.8 Best practices for REaL data collection are a top disparities research priority and a topic of national discussion with practices currently in flux.9 National surveys now include combined, granular race and ethnicity categories focused on ancestry (eg, Hispanic Mexican, Cuban, Filipino)10; the Oregon Health Authority permits multiracial constituents to indicate all applicable races and ethnicities.10Study limitations include constraints related to data practices (eg, do not report multiracial identities, cannot account for discrepancies between parent-reported and patient-reported REaL) and nonresponse bias. We used a convenience sample with an unknown response rate (estimate ≤5 patients declined). Institutional patient demographics and REaL data collection practices vary, limiting generalizability.Ensuring REaL data validity is a foundational step toward reliably identifying and addressing institutional health care disparities. Our institution's REaL data can be strengthened further through efforts to reduce the use of ambiguous data.The authors would like to acknowledge Dr Sarah Schechter, who contributed to study design and power calculations, and the New York-Presbyterian Dalio Center for Health Justice for implementing the "We Ask Because We Care" campaign, an initiative to train front-line staff in a standardized approach to asking patients about their self-identified race, ethnicity, and language data.