Data on self-reported cancer from a health interview survey carried out in 1991 in the southeastern Netherlands by means of a postal questionnaire (n = 17,940) were validated against records from a population-based cancer registry. The sensitivity of the questionnaire was 0.552 (95% confidence interval (CI) 0.507-0.597), and the specificity was 0.995 (95% CI 0.994-0.996). The survey underestimated cancer prevalence in the population by 25%. Of the 212 false negative cases, 46% were registered with non-melanoma skin cancer. After the exclusion of nonmelanoma skin cancer from cancer registry records, cancer prevalence was overestimated by the survey by a negligible 2%. The misclassification of cancer by the postal survey was differential according to age, sex, education, and degree of urbanization. The survey overestimated cancer prevalence ratios for men versus women, old respondents versus young respondents, and urban residents versus rural residents. The prevalence ratios for respondents with a low educational level versus those with a high level were underestimated using survey data. These patterns remained essentially the same after exclusion of nonmelanoma skin cancer from the cancer registry records. This study shows that both overall cancer prevalence and differences in cancer prevalence between subgroups of the population may be biased when health interview survey data are used. If explicit attention is paid to nonmelanoma skin cancer in survey questions, this might improve the validity of overall cancer prevalence estimates, but not that of comparisons between subgroups of the population.