Objective: To describe the accuracy of ethnicity coding in contemporary National Health Service (NHS) hospital records compared with the 'gold standard' of self-reported ethnicity. Design: Secondary analysis of data from a crosssectional survey (2011). Setting: All NHS hospitals in England providing cancer treatment. Participants: 58 721 patients with cancer for whom ethnicity information (Office for National Statistics 2001 16-group classification) was available from self-reports (considered to represent the 'gold standard') and their hospital record. Methods: We calculated the sensitivity and positive predictive value (PPV) of hospital record ethnicity. Further, we used a logistic regression model to explore independent predictors of discordance between recorded and self-reported ethnicity. Results: Overall, 4.9% (4.7-5.1%) of people had their self-reported ethnic group incorrectly recorded in their hospital records. Recorded White British ethnicity had high sensitivity (97.8% (97.7-98.0%)) and PPV (98.1% (98.0-98.2%)) for self-reported White British ethnicity. Recorded ethnicity information for the 15 other ethnic groups was substantially less accurate with 41.2% (39.7-42.7%) incorrect. Recorded 'Mixed' ethnicity had low sensitivity (12-31%) and PPVs (12- 42%). Recorded 'Indian', 'Chinese', 'Black-Caribbean' and 'Black African' ethnic groups had intermediate levels of sensitivity (65-80%) and PPV (80-89%, respectively). In multivariable analysis, belonging to an ethnic minority group was the only independent predictor of discordant ethnicity information. There was strong evidence that the degree of discordance of ethnicity information varied substantially between different hospitals (p