Using existing administrative data to look at issues of ethnic disparities in rehabilitation-related outcomes may lead to misleading results. Problems can emerge from apparently small issues of reliability that are magnified by reclassification of ethnic designation and missing data in complete-subject analyses. We compared the reliability of ethnic assignment in Department of Veterans Affairs (VA) medical rehabilitation records for stroke patients with administrative records; reclassified the racial identifier from the administrative data in two ways; and examined the different sources of ethnic information in relation to severity, length-of-stay, disability assessment, and discharge disposition. Our results show how small changes increase the potential for Type II error when describing ethnic differences in outcomes or using ethnicity as a predictor with dichotomous response variables. We discuss our results with reference to the literature on ethnic classification and underline the importance of initiatives for improved data collection on ethnicity in VA data sources and in rehabilitation research.