Uncertainty is inherent in the conception and measurement of ethnicity, by both individuals themselves and those who seek to gather evidence of discrimination or inequalities in social and economic outcomes. These issues have received attention in the literature, yet rather little research has been carried out on the uncertainty subsequently created through the analysis of such measurements. We argue that, while general-purpose ethnicity classifications offer a method of standardising results, such groupings are inherently unstable, both in their upward aggregation and in their downward granulation. As such, the results of ethnicity analysis may possess no validity independent of the ethnicity classes upon which it is based. While this conclusion is intuitive, it nevertheless seems to pass unnoticed in the interpretation of research conducted in public policy applications such as education, health and residential segregation. In this paper we use examples based on the standard Census classification of ethnicity, alongside new rich ethnicity datasets from the education domain, in order to evaluate the sensitivity of results to the particular aggregation that is chosen. We use a case study to empirically illustrate the far-reaching consequences of this commonly overlooked source of uncertainty.