Some epidemiologic studies have selected subjects from death registries. This method is, however, subject to substantial systematic bias that arises from the fact that different segments of society have different risks of death in any year. Different occupational groups will thus have unequal probabilities of entering the pool of potential controls, and biased estimates of the odds ratios that relate occupation to disease will result. Failure to recognize the bias may lead to invalid conclusions. This selection bias is discussed, with particular reference to studies of malignancies among farmers. The bias is demonstrated by using data from studies of Parkinson's disease and sinonasal cancer.