Observational analyses of the association between body mass index (BMI) and all-cause mortality often suggest that overweight is neutral or beneficial, but such analyses are potentially confounded by smoking or by reverse causation. The use of BMI measured in early adulthood offers one means of reducing the latter problem. We used a cohort who were first measured while 16-24 year old students at Glasgow University in 1948-1968 and subsequently re-measured in 2000-2003, offering a rare opportunity to compare BMI measured at different ages as a predictor of mortality. Analysis of the later BMI measurements suggested that overweight was beneficial to survival, while analysis of BMI measured in early adulthood suggested that overweight was harmful and that the optimum BMI lay towards the lower end of the recommended range of 18.5-25 kg m-2. We interpret the association with later BMI as being probably distorted by reverse causality, although it remains possible instead that the optimum BMI increases with age. Differences when analyses were restricted to healthy non-smokers also suggested some residual confounding by smoking. These results suggest that analyses of BMI recorded in middle or old age probably over-estimate the optimum BMI for survival and should be treated with caution.