The accuracy of self-reported cancer diagnoses in a prospective study was compared with population-based cancer registry data in four states. The study cohort included 65,582 men and women aged 39-96 years who were participants in the Cancer Prevention Study II Nutrition Survey, begun by the American Cancer Society in 1992. Estimates of sensitivity (the proportion of study participants with a registry-documented cancer who self-reported the cancer) ranged from 0.79 for an exact match of cancer site and year of diagnosis (+/- 1 year) to 0.93 for a match of any reported cancer. The sensitivity of exact matches varied considerably by cancer site and was highest for breast, prostate, and lung cancers (0.91, 0.90, and 0.90, respectively) and lowest for rectal cancer and melanoma (0.16 and 0.53, respectively). Sensitivity also varied slightly by the age, education, and smoking status of study participants. Estimates of sensitivity were virtually identical for each of the four states. The positive predictive value (the proportion of self-reported cancers that were confirmed by the registries) was 0.75 overall and also varied by cancer site. Unlike sensitivity, however, this proportion varied considerably by state. All self-reports of cancer that were not confirmed by the registries were further investigated by repeat questionnaires and acquisition of medical records. Low positive predictive values were due to underascertainment of true cancer cases by the registries, inaccurate reporting on the part of study participants, and problems with the algorithm used by the state to link the study population to the registry data. In conclusion, the ability of members of this cohort to report a past diagnosis of cancer accurately is quite high, especially for cancers of the breast, prostate, lung, and colon, or for the occurrence of any cancer.