Using verbal autopsy to track epidemic dynamics: the case of HIV-related mortality in South Africa

Background: Verbal autopsy (VA) has often been used for point estimates of cause-specific mortality, but seldom to characterize long-term changes in epidemic patterns. Monitoring emerging causes of death involves practitioners' developing perceptions of diseases and demands consistent methods and practices. Here we retrospectively analyze HIV-related mortality in South Africa, using physician and modeled interpretation. Methods: Between 1992 and 2005, 94% of 6,153 deaths which occurred in the Agincourt subdistrict had VAs completed, and coded by two physicians and the InterVA model. The physician causes of death were consolidated into a single consensus underlying cause per case, with an additional physician arbitrating where different diagnoses persisted. HIV-related mortality rates and proportions of deaths coded as HIV-related by individual physicians, physician consensus, and the InterVA model were compared over time. Results: Approximately 20% of deaths were HIV-related, ranging from early low levels to tenfold-higher later population rates (2.5 per 1,000 person-years). Rates were higher among children under 5 years and adults 20 to 64 years. Adult mortality shifted to older ages as the epidemic progressed, with a noticeable number of HIV-related deaths in the over-65 year age group latterly. Early InterVA results suggested slightly higher initial HIV-related mortality than physician consensus found. Overall, physician consensus and InterVA results characterized the epidemic very similarly. Individual physicians showed marked interobserver variation, with consensus findings generally reflecting slightly lower proportions of HIV-related deaths. Aggregated findings for first versus second physician did not differ appreciably. Conclusions: VA effectively detected a very significant epidemic of HIV-related mortality. Using either physicians or InterVA gave closely comparable findings regarding the epidemic. The consistency between two physician coders per case (from a pool of 14) suggests that double coding may be unnecessary, although the consensus rate of HIV-related mortality was approximately 8% lower than by individual physicians. Consistency within and between individual physicians, individual perceptions of epidemic dynamics, and the inherent consistency of models are important considerations here. The ability of the InterVA model to track a more than tenfold increase in HIV-related mortality over time suggests that finely tuned "local" versions of models for VA interpretation are not necessary.