New Epidemiologic Methods for Reducing Measurement Error and Misclassification Bias in Cancer Epidemiology

Project Summary/Abstract Uncertainty in exposure and outcome measurements poses substantial challenges to the identification and quantification of the causes of cancer. For example, although difficult to measure well, physical activity patterns form the basis of many etiologic hypotheses concerning cancer risk. Cancer cases identified in electronic health records (EHR) and other administrative ‘big data’ sources, such as Medicare claims data, are also subject to misclassification. This exposure and outcome uncertainty leads to considerable bias in estimated health effects, masking our ability to detect true associations, which are likely underestimated if detected at all. It is the role of measurement error and misclassification correction methods to validly and efficiently estimate the relationship between exposures and cancer outcomes. To accomplish this, a validation study is required for estimating key features of the error process. Although much has been accomplished in this domain over the years, the current aims address unsolved problems of high scientific significance that would otherwise remain unanswered without this additional work. We will drill down into the multi-faceted themes that arise in cancer research, tackling several seminal new directions of critical importance for the translation of the results of population-based research to practice and policy. These methods will include estimation of the effects of within-individual change in lifestyle behaviors on cancer risk corrected for measurement error in the change variables, utilizing complex, currently under-accessed validation studies of diet and physical activity comprised of repeated paper and online questionnaire self-reports and repeated concentration and recovery biomarkers to obtain relative risk estimates unbiased by general measurement error structures which may include correlated and biased errors, and estimating effects of exposures, including medications, other clinical treatments, and health behaviors, on cancer incidence in EHR data. The new methods will be applied to studies of the impact of within-participant change in alcohol intake on breast cancer incidence in the American Cancer Society’s CPS-II cohort and in Harvard’s Nurses’ Health Study, and to a study disentangling the impacts of diabetes and diabetes medications on colorectal cancer risk in Yale New Haven’s Epic EHRs. Dissemination is a central feature of this research. User-friendly publicly available software will accompany all new methods to be developed. The new methods will be disseminated through short courses and lectures at national and international epidemiologic and statistical conferences, and through the development of a massive online open course (MOOC). We have assembled an outstanding team of experts in measurement error methods and statistical theory, along with an exceptional team of cancer epidemiologists with much prior collaborative experience with the methods team, to guide the developments and their applications to the scientific problems at hand. With the talented junior faculty and trainees to be recruited for this project, we will solve the challenging problems that have been identified.