Cause-specific mortality data is critical information for public health decision-making, but in many areas of the world it is not routinely recorded. In the past several decades, Verbal Autopsy (VA) has emerged as an interview-based approach to fill this gap. Until recently, however, the methods for implementation and analysis of VA were costly, time-consuming, and not validated. I will present an overview of a recently completed multisite validation study where researchers meticulously collected over 10,000 VA interviews for which the underlying cause of death was known with near certainty.
It was when the preliminary data from this validation study arrived at IHME that I learned about this quantitative challenge, and I knew that I had to get involved. Many, many work hours later, we have now wrapped up an analysis, which was recently published in a special issue of Population Health Metrics (http://www.pophealthmetrics.com/series/verbal_autopsy). In this work, we realized that even measuring how well a VA analysis method performs is non-trivial. I will tell you about our new metrics that quantify the quality of VA analysis methods and I will also tell you about the results of the "horserace" of analytic methods that we ran. We reached the conclusion that a variant of the Random Forest classifier yields the highest quality results, better even than human experts.
