Fewer than 25% of people diagnosed during life as being demented were found to have this diagnosis coded as the underlying cause of death. In a sample of deaths certified as due to dementia the majority were found to have occurred in long-stay institutions. This distorts the geographical pattern of mortality because the Office of Population Censuses and Surveys (OPCS) considers these institutions to be the patient's usual address six months after his admission. Analysis of all certified deaths from dementia during 1968-78 by place of residence shows that areas with a significantly high SMR usually contain a large psychiatric hospital. Changes in diagnostic fashion and in the procedure by which OPCS selects the underlying cause of death have also affected numbers of deaths coded as dementia. Death certificate data are unlikely to be useful in examining either geographical variation or time trends in rates of dementia.