Mortality from diabetes is underestimated four- to fivefold by methods of analysis of death certification data which use only underlying cause of death. This problem is partially overcome by coding all conditions mentioned on death certificates. For a sample of deaths in England and Wales over the years 1972-1977, the observed proportion of certificates with specific underlying causes of death for certificates mentioning diabetes was compared with the expected proportion for all certificates. These observed/expected ratios were significantly increased in each sex for circulatory diseases and were significantly reduced for neoplasms. For 'nephritis' they were also increased, especially below 45 years of age. These results were confirmed by an analysis of underlying cause of death in a cohort of nearly 6,000 members of the British Diabetic Association. Of the 2,134 deaths in this cohort, diabetes was not mentioned on 33% of the death certificates. For the period 1972-1977, death rates for circulatory diseases associated with diabetes increased by 6% for males but remained constant for females.