This paper uses a unique survey of Roma and non-Roma in South Eastern Europe to evaluate competing explanations for the poor performance of Roma in the labour market. The analysis seeks to identify the determinants of educational achievement, employment and wages for Roma and non-Roma. LIML methods are employed to control for endogenous schooling and two sources of sample selection bias in the estimates. Nonlinear and linear decomposition techniques are applied in order to identify the extent of discrimination. The key results are that: the employment returns to education are lower for Roma than for non-Roma whilst the wage returns are broadly similar for the two groups; the similar wage gains translate into a smaller absolute wage gain for Roma than for non-Roma given their lower average wages; the marginal absolute gains from education for Roma are only a little over one-third of the marginal absolute gains to education for majority populations; and, there is evidence to support the idea that a substantial part of the differential in labour market outcomes is due to discrimination. Explanations of why Roma fare so badly tend to fall into one of two camps: 'low education' vs. 'discrimination'. The analysis suggests that both of these explanations have some basis in fact. Moreover, a direct implication of the lower absolute returns to education accruing to Roma is that their lower educational participation is, at least partially, due to rational economic calculus. Consequently, policy needs to address both low educational participation and labour market discrimination contemporaneously.