This paper examines the impact of a program that subsidizes children to remain in school for up to two years beyond the statutory age. The programme was first piloted in a number of areas in England from September 1999. Evaluating such interventions is of course critical to the shaping of education policy and the effectiveness or otherwise of a conditional cash transfer to 16 and 17 year olds on school dropout rates is of general policy interest to policy makers worldwide. The authors find that the impact of the subsidy is quite substantial, especially for those who receive the maximum payment. The subsidy increases the initial education participation of eligible males by 4.8 percentage points and eligible females by 4.2 percentage points. In the second year this increases to 7.6 percentage points for eligible males and 5.3 percentage points for females, suggesting that the effect of the policy is not only to increase participation, but also retention in full-time education. The initial effects are largest for those who receive the maximum payment although the retention effects are concentrated among individuals who are only partially eligible. The authors estimate that just over half of individuals who stayed in education were drawn from inactivity rather than work. The overall impact of the Education Maintenance Allowance (EMA) was not diminished when it was paid to the mother rather than to the child, though there is some weak evidence that paying to the child is more effective for those fully eligible whereas the opposite is true for those who are partially eligible. The authors also find that the effect of EMA is largest for children coming from a poorer socio-economic background. Both girls and boys coming from low-income families who qualify for the full EMA payment have very high drop out rates and the EMA has proved especially effective in plugging the dropout gap for this vulnerable group. Appended are: (1) Indicators used in each deprivation score; (2) Sample characteristics; (3) Covariate balancing indicators (best specification): before and after matching; (4) Attrition between wave 1 and wave 2; and (5) Identifying assumptions and Estimation method. (Contains 9 tables and 38 footnotes.)