This paper considers the identification of the effect of tobacco on mortality. If individuals select into smoking according to some unobserved health characteristic, then estimates of the effect of tobacco on health that do not account for this are biased.We show that using information on mortality, morbidity and smoking, it is possible tocontrol for this selection effect and obtain consistent estimates of the effect of smokingon mortality. We implement our method on Swedish data. We show that there is selection into smoking, and considerable dispersion around the average effect, so that health policies that aim at decreasing smoking prevalence and quantities smoked might have less effect in terms of average number of years of life gained than previouslyestimated. We also empirically show that selection into smoking has increased overthe last fifty years with the availability of information on the dangers of smoking,so that future studies comparing smokers and non smokers will spuriously reveal aworsening effect of tobacco on health if they fail to control for selection.