Administrative data and inverse occurrence of cancer and Alzheimer's disease in elderly people II: a misclassification problem for the diagnosis of AD.

Introduction In a cohort of 115,524 person-years we observed a 40% risk reduction for the occurrence of Alzheimer's disease (AD) in persons with cancer and vice-versa. A relevant concern about these findings is about the potential effect of misclassification in biasing the results. Materials and Methods All cancer cases were identified from a population Tumor Registry. AD cases were traced from registries of drug prescriptions, hospitalizations and payment exemptions by a classification algorithm that had high specificity but lacked in sensitivity. In the co-occurrence study misclassification of "exposure" (cancer occurrence in the AD cohort) and of disease (AD in cancer cohort), might be alternatively differential (DM) or non-differential (NDM). A sensitivity analysis was performed in this cohort. Results An exposure or disease NDM had the only effect to overestimate the relative risks (RR), weakening the strength of the inverse association. The effect of a DM, unlikely to have occurred, was dependent from levels of sensitivity/specificity in the subgroups of exposure and of disease, and led to biased RR estimates only when classification performances were highly differential. Discussion Large population-based studies on administrative data allow to obtain epidemiological results accounting for possible misclassifications.