Epidemiological studies over the last 10 years have greatly increased our understanding about the role vitamin D in health. During this time, a tension has arisen between the results from observational studies and randomised controlled trials of vitamin D supplementation. Several cohort studies have reported non-linear inverse associations between vitamin D status, as measured by circulating 25-hydroxyvitamin D concentrations (25(OH)D), and risk of some outcomes such as cardiovascular disease and all-cause mortality, with the increased disease risk being observed in people with 25(OH)D less than 50 nmol/L. In contrast, recent large-scale trials have reported no effect of vitamin D supplementation on major disease outcomes such as cardiovascular disease, cancer, diabetes and mortality, even in subgroup analyses of participants with baseline 25(OH)D levels less than 50 nmol/L. The latter results have been confirmed by mendelian randomisation studies which mostly have not found an association between genetically predicted 25(OH)D and disease outcome. Until now, the inconsistency between observational studies and trials has been assumed to be due to residual confounding leading to spurious inverse associations in the former. However, a major limitation of most mendelian randomisation studies is that they have assumed a linear association between genetically predicted vitamin D status and disease. Recent mendelian studies which have investigated possible non-linear associations have found that disease risk is increased in people with 25(OH)D less than 25-40 nmol/L, confirming the findings of the earlier cohort studies. These new findings and their implications will be discussed.