Data on adult mortality are very limited in sub-Saharan Africa where only small proportions of deaths occur in health facilities. In such settings, ascertainment of causes of death from data obtained from relatives or associates of the deceased through interviews in surveys or longitudinal surveillance systems appears to be an attractive option. This technique, known as verbal autopsy (VA) is based on the assumption that important causes of death have distinctive symptoms and signs, and these can be recognised, remembered and reported by lay respondents, and that based on the reported information causes of death can be reached. The existing experience of VA for adult death is limited mainly to maternal deaths and the validity of VA for adult death is unknown. We developed a VA questionnaire, mortality classification system and "expert opinion" based algorithms for reaching diagnoses for adult deaths and tested their validity on deaths occurring at hospitals in Tanzania (n=315), Ethiopia (n=249) and Ghana (n=232). Hospital records of adult deaths occurring at study hospitals from June 1993 to April 1995 were collected prospectively. VA interviews were conducted by trained non-medical interviewers. Caused of death from VA data were reached by a panel of three physicians and by a computerised algorithm. The validity of VA was assessed by comparing the VA diagnoses with hospital diagnoses. Specificity of VA fell below 95% only for few common causes of adult death. Sensitivity and kappa of VA for all common causes of adult death were low and this suggests that the accuracy of VA at the individual level is low. However, the misclassification of causes of death was bi-directional and the number of false positive and false negative diagnosis for most common causes of adult death tend to be similar. Thus there was robust agreement between the true and VA estimates of cause specific mortality fractions of common causes of adult death and VA is useful for assessing cause specific mortality fractions of common causes of adult death.