Epidemiologic studies of cancer and exposure to herbicides have shown puzzling inconsistencies. Exposure-response gradients have been reported for non-Hodgkin's lymphoma in Sweden and Kansas, but no significant associations were seen in New Zealand or Washington State. Subjects in these studies were categorized by exposure using information obtained primarily by interview. A number of questions can be raised regarding the reliability and validity of such an exposure assessment. We examined procedures used to assess pesticide exposures in case-control studies of cancer to evaluate their limitations and their probable effects on risk estimates. Except for case recall bias, problems of misclassification in these studies would tend to bias risk estimates toward the null and dilute exposure-response gradients. These problems are, therefore, unlikely explanations for the positive associations between cancer and herbicide use noted in some investigations. A tendency for false-negative findings, however, is not reassuring, and improvements in exposure assessment are needed if epidemiologic investigations are to continue to provide reliable information on the relationships of cancer and pesticide exposure.