[Roaming through methodology. XV. Smoking grandfathers and other selection biases]

That some cigarette smokers live until 90 years of age does not allow conclusions about the life expectancy of smokers in general as, after all, the dead are invisible. In the same way, a case-control study of factor F (e.g. smoking) as a determinant of disease D (e.g. Alzheimer's disease) may be biased by excess mortality in subjects with the combination of disease D and factor F. The result is an underestimation of the combination F-D in survivors (survival bias). A different form of bias may occur in hospital-based observational studies, e.g. of patients with disease Y (brain infarction), with or without previous treatment with drug A (aspirin). If previous treatment is associated with relatively mild deficits, while only patients with severe deficits are admitted, the difference in stroke severity cannot be observed in hospital (referral bias). In comparison with a population-based follow-up study important groups of patients may, in either example, have been lost in the study.