Validity of data-derived algorithms for ascertaining causes of adult deathin two African sites using verbal autopsy

BACKGROUND Verbal autopsy (VA) is used to ascertain causes of death using i nformation obtained from bereaved relatives. Causes of death can be ascerta ined from VA questionnaires by a panel of physicians or from predefined alg orithms. In a previous study, we developed data-derived algorithms using VA data from 796 adult deaths in hospitals in Tanzania, Ethiopia, and Ghana ( primary sites). These computerized algorithms accurately estimated the caus e-specific mortality fractions (CSMFs) for deaths due to injuries, meningit is, TB/AIDS and diarrhoeal diseases in the primary sites. Since the same da ta were used to generate and to validate the algorithms, the accuracy of ou r algorithms may have been overestimated. We report here on the validity of the algorithms when they were applied to VA data from two secondary sites in Ghana and Tanzania. Here, 'validity' is taken to mean the degree to whic h the algorithms replicated the physician-generated CSMF for major causes o f death, when applied to the same VA data. METHODS VA interviews were conducted in two secondary sites: in Navrongo, G hana, on 406 adult deaths, where three local physicians independently revie wed the questionnaires and assigned a cause of death. In Morogoro, Tanzania , VA interviews were conducted on 209 adult deaths, and a panel of physicia ns independently reviewed the VA questionnaires together with the hospital death certificates or hospital records to determine the cause of death. The CSMF obtained using each algorithm was compared with the CSMF obtained usi ng physician review. RESULTS For injuries and meningitis, the algorithms and physician review es timated a similar CSMF in the Morogoro and Navrongo data. For TB/AIDS, the algorithm estimated a similar CSMF as the physicians in Morogoro. The algor ithm for diarrhoeal diseases did not agree closely with the physicians in M orogoro or Navrongo. CONCLUSIONS In general, our data-derived algorithms for assigning causes of death due to injuries, meningitis, and TB/AIDS estimated a similar CSMF as the physicians in the secondary sites. Recommendations for further validat ion and refinement are discussed. Computerized algorithms offer a potential ly quick, affordable, and feasible method for assigning causes of death in mortality surveillance or studies using VA.