Background: InterVA is a probabilistic method for interpreting verbal autopsy (VA) data. It uses a priori approximations of probabilities relating to diseases and symptoms to calculate the probability of specific causes of death given reported symptoms recorded in a VA interview. The extent to which InterVA's ability to characterise a population's mortality composition might be sensitive to variations in these a priori probabilities was investigated.
Methods: A priori InterVA probabilities were changed by 1, 2 or 3 steps on the logarithmic scale on which the original probabilities were based. These changes were made to a random selection of 25% and 50% of the original probabilities, giving six model variants. A random sample of 1,000 VAs from South Africa, were used as a basis for experimentation and were processed using the original InterVA model and 20 random instances of each of the six InterVA model variants. Rank order of cause of death and cause-specific mortality fractions (CSMFs) from the original InterVA model and the mean, maximum and minimum results from the 20 randomly modified InterVA models for each of the six variants were compared.
Results: CSMFs were functionally similar between the original InterVA model and the models with modified a priori probabilities such that even the CSMFs based on the InterVA model with the greatest degree of variation in the a priori probabilities would not lead to substantially different public health conclusions. The rank order of causes were also similar between all versions of InterVA.
Conclusion: InterVA is a robust model for interpreting VA data and even relatively large variations in a priori probabilities do not affect InterVA-derived results to a great degree. The original physician-derived a priori probabilities are likely to be sufficient for the global application of InterVA in settings without routine death certification.