This paper considers the problems which arise in seeking to measure socioeconomic inequalities in health when the health indicator is a categorical variable, such as self-assessed health. It shows that the standard approach--which involves dichotomizing the categorical variable--is unreliable. The degree of measured inequality is found to depend on the cut-off point chosen and the choice of cut-off point to affect the conclusions one can reach about trends in or differences in health inequality. The paper goes on to propose an alternative approach which involves constructing a latent health variable and then measuring inequalities in this latent variable by means of a variant of the health concentration curve.