Background: In sub-Saharan Africa and elsewhere, many people frequently still do not have access to health services that allow medical certification of the cause of their death. Health systems need estimates of cause-specific mortality for planning, and the only way to realistically obtain these in sub-Saharan Africa in the medium term will be through verbal autopsy (VA). Methods: This thesis investigates three methods for assigning HIV/AIDS as a cause of death – physician review, InterVA-4 and the Lopman algorithm – using VA data on 15–59 year-olds from two demographic surveillance systems in Tanzania (Kisesa) and Zimbabwe (Manicaland). The performance of the methods is assessed against the reference standard of known HIV status, allowing the calculation of performance metrics including specificity. Results The estimated proportion of adult deaths due to HIV/AIDS varied between methods, from 30–53% in Kisesa, and 58–73% in Manicaland. It was not possible to conclude with certainty which estimate was most accurate, nor was there any relationship between the estimated proportions and the performance measured by validity metrics. The methods had variable performance, with physician review having the highest specificity, followed by InterVA-4 and the Lopman algorithm. Findings were broadly consistent with the published literature. Analysis of the Lopman algorithm provided a clear illustration of the problems of using data-derived methods, even where reference-standard data are available to train them. Conclusion: Using validity to assess the quality of real-world VA findings is flawed. Cause-specific mortality estimation should move from seeking single best estimates based on assessment of validity to seeking plausible estimates using synthesis of multiple sources of data – including VA.