Designing verbal autopsy studies

Abstract

Background: Verbal autopsy analyses are widely used for estimating cause-specific mortality rates (CSMR) in the vast majority of the world without high-quality medical death registration. Verbal autopsies -- survey interviews with the caretakers of imminent decedents -- stand in for medical examinations or physical autopsies, which are infeasible or culturally prohibited.

Methods and findings: We introduce methods, simulations, and interpretations that can improve the design of automated, data-derived estimates of CSMRs, building on a new approach by King and Lu (2008). Our results generate advice for choosing symptom questions and sample sizes that is easier to satisfy than existing practices. For example, most prior effort has been devoted to searching for symptoms with high sensitivity and specificity, which has rarely if ever succeeded with multiple causes of death. In contrast, our approach makes this search irrelevant because it can produce unbiased estimates even with symptoms that have very low sensitivity and specificity. In addition, the new method is optimized for survey questions caretakers can easily answer rather than questions physicians would ask themselves. We also offer an automated method of weeding out biased symptom questions and advice on how to choose the number of causes of death, symptom questions to ask, and observations to collect, among others.

Conclusions: With the advice offered here, researchers should be able to design verbal autopsy surveys and conduct analyses with greatly reduced statistical biases and research costs.

Figure 1

For sample sizes of 3,000 (left graph), 1,000 (middle), and 500 (right), the figure gives the mean square error as symptoms are removed using our detection diagnostic. The mean square error first declines, as bad symptoms are removed and bias drops, and then increases, as unbiased symptoms are dropped and variance increases. The procedure selects all biased symptoms (open circles) before unbiased symptoms (solid disks). The vertical line indicates where our automated procedure would indicate that we should stop.

Figure 2

Validation of Mean Square Error in Tanzania, where the true cause of death is known in both samples.

Figure 3

Simulation Results for 5 (left graph), 10 (middle), and 15 (right) causes of death.