Assessing verbal autopsy as a component of vital registration: a data-driven simulation study

Abstract Background Verbal autopsy (VA) has been proposed as an alternative or supplement to mortality estimation in countries with low vital registration (VR) coverage. Low VR in countries can result in inaccurate estimations of disease burden and inefficient allocation of resources. We used VA and VR data from the states of Hidalgo and Morelos (Mexico) to explore the reliability of augmenting incomplete VR with VA data. Methods As part of the Population Health Metrics Research Consortium study, we identified death certificates from 2009 in the states of Hidalgo and Morelos, Mexico, and created a sample. We used data on location of death to simulate 100 scenarios of VR coverage ranging from 10% to 90% and augmented each with VA data to simulate full coverage, using different strategies to blend VR and VA information. We attributed underlying cause of death to the VAs using the Tariff method, an additive algorithm that predicts 34 causes of death for adults, 21 for children, and six for neonates. We measured the reliability of cause fraction predictions comparing VR only to VR and VAs. To quantify uncertainty, we replicated this 500 times for each scenario. Findings As the percent (simulated) VR coverage decreased and the missing deaths were assigned using VAs, the mean absolute difference in cause fractions for adults increased from 0·219% (95% CI 0·217–0·221) at the 90% VR coverage level to 1·84% (1·82–1.86) at the 10% VR coverage level. These differences were greater among child and neonatal deaths, and varied by cause of death. Interpretation A mean change of two percentage points in cause fractions from a 34-cause list is substantial when the true cause fractions can fall below 2%. This study shows that countries whose mortality estimates are based on incomplete VR data could see substantial changes in estimated cause composition if they supplemented their VR with VA data. Funding Bill & Melinda Gates Foundation through grant 51488 "Population Health Metrics Research Consortium—Mexico".