The authors examined agreement between drug treatment data and a discharge diagnosis of diabetes, considered whether agreement was modified by demographic variables and measures of comorbidity, and evaluated construct validity through consideration of relations with subsequent mortality. The study sample comprised 81,700 residents of New Jersey aged 65-99 years who had prescription drug coverage either through Medicaid or that state's Pharmacy Assistance for the Aged and Disabled program and had at least one hospitalization between July 1, 1989, and June 30, 1991. In this population, 16.4% filled a prescription for insulin or an oral hypoglycemic agent during the 120 days before admission, and 16.3% had a discharge diagnosis of diabetes. Overall agreement between these two indicators was modest (kappa = 0.67, 95% confidence interval 0.66-0.67) and was weaker in those aged 85 years and above (kappa = 0.58, 95% confidence interval 0.56-0.60), those in nursing homes (kappa = 0.42, 95% confidence interval 0.39-0.44), and those with a high level of comorbidity (modified Charlson index > or =5; kappa = 0.59, 95% confidence interval 0.56-0.62). Presence of a diagnosis of diabetes was associated with an apparent 24% reduction in the risk of death during the study interval (p<0.001), while prior treatment for diabetes had little relation to mortality (p = 0.15). These paradoxical associations with mortality and the lower agreement between discharge diagnoses and drug treatments associated with older age, nursing home residence, and comorbidity suggest limitations in the use of claims data to identify diabetes in the elderly.