Effect of misclassification of causes of death in verbal autopsy: can it be adjusted?

BACKGROUND: Verbal autopsy (VA) is an indirect method of ascertaining cause of death from information about symptoms and signs obtained from bereaved relatives. This method has been used in several settings to assess cause-specific mortality. However, cause-specific mortality estimates obtained by VA are susceptible to bias due to misclassification of causes of death. One way of overcoming this limitation of VA is to adjust the crude VA estimate of cause-specific mortality fractions (CSMF) using the sensitivity and specificity of the VA tool. This paper explores the application of sensitivity and specificity of VA data obtained from a hospital-based validation study for adjusting the effect of misclassification error in VA data obtained from a demographic surveillance system. METHOD: Data from a multi-centre validation study of 796 adult VA, conducted in Tanzania, Ethiopia and Ghana, were used to explore the effect of distribution of causes of death in the validation study population and the pattern of misclassification on the sensitivity and specificity of VA. VA estimates of CSMF for six causes (acute febrile illness, diarrhoeal diseases, TB/AIDS, cardiovascular disorders, direct maternal causes and injures) were obtained from a demographic surveillance system in Morogoro Rural District in Tanzania. These were adjusted for misclassification error by using sensitivity and specificity values of VA obtained from the validation study in a model proposed for correcting the effect of misclassification error in morbidity prevalence surveys. RESULTS: Sensitivity and specificity of VA differed between the three validation study sites depending on the distribution of causes of death. These differences were explained by variations in the level and pattern of misclassification between sites. When these estimates of sensitivity and specificity were applied to data from the demographic surveillance system with a comparable structure of causes of death the difference between crude and adjusted VA estimates of CSMF ranged from 3 to 83%. CONCLUSION: Estimates of sensitivity and specificity obtained from hospital-based validation studies must be used cautiously as a de facto 'gold standard' for adjusting the misclassification error in CSMF derived from VA. It is not possible to use sensitivity and specificity estimates derived from a location-specific validation study to adjust for misclassification in VA data from populations with substantially different patterns of cause-specific mortality.