The validity of death certificate diagnosis of stroke and its type as the underlying cause of death was investigated in a sample of in-hospital deaths of possible stroke cases from the Minnesota Heart Survey. The 228 in-hospital deaths in 1970 and the 180 deaths in 1980 had a stroke diagnosis either on hospital discharge records or as the underlying cause of death on the death certificate. Relative to a standardized physician diagnosis, positive predictive values for the death certificate diagnosis in 1970 were 96% for all types of stroke, 59% for intracranial hemorrhage, and 87% for nonhemorrhagic stroke. The respective values in 1980 were 100%, 82%, and 97%. An increase in positive predictive values, particularly for intracranial hemorrhage between 1970 and 1980, was attributed to the increased use of computerized tomography. Sensitivity for the death certificate diagnosis in 1970 was 63% for all types of stroke, 66% for intracranial hemorrhage, and 45% for nonhemorrhagic stroke. The respective sensitivities in 1980 were 70%, 76%, and 58%. The lower sensitivity for nonhemorrhagic stroke as compared with hemorrhagic stroke was due in part to 1) frequent reporting of nonhemorrhagic stroke as a contributing cause of death rather than the underlying cause of death and 2) time from stroke onset to death. Specificity among these possible strokes was high in both years. The low sensitivity of death certificate diagnosis of stroke may reduce estimated relative risks in epidemiologic studies. Nevertheless, since the advent of widespread use of computerized tomography, a death certificate diagnosis of intracranial hemorrhage versus nonhemorrhagic stroke appears to be sufficiently accurate for use in epidemiologic studies of stroke etiology.