We compare estimates of the effects of education on health and health behaviour using two different instrumental variables in the UK Biobank data. One is based on a conventional natural experiment while the other, known as Mendelian randomization (MR), is based on genetic variants. The natural experiment exploits a compulsory schooling reform in the UK in 1972 which involved raising the minimum school leaving age (RoSLA). MR exploits perturbations of germline genetic variation associated with educational attainment, which occur at conception. It has been widely used in epidemiology and clinical sciences. Under monotonicity, each IV identifies a LATE, with potentially different sets of compliers. The RoSLA affected the amount of education for those at the lower end of the ability distribution whereas MR affects individuals across the entire distribution. We find that estimates using each approach are remarkably congruent for a wide range of health outcomes. Effect sizes of additional years of education thus seem to be similar across the education distribution. Our study corroborates the usefulness of MR as a source of instrumental variation in education.