Richard Li aims to design an adaptive survey for verbal autopsy with the support of a Hellman fellowship.
In many low- and middle-income countries throughout the world, when someone dies, a local health official comes to the home and asks the family members of the deceased a long list of questions: whether the person had a cough, a fever, a previous diagnosis, and much more. This process, called verbal autopsy, is the main way health officials determine cause of death amongst their populations after statisticians use algorithms to evaluate survey responses to assign a cause, and it has downstream effects on monitoring population health and designing interventions.
But the lengthy, 200-300 question survey can be emotionally burdensome for both the interviewer and the relatives of the deceased person, and is known to be the biggest ongoing problem in this field. That’s why Zehang “Richard” Li, assistant professor of statistics at UC Santa Cruz’s Baskin School of Engineering, wants to entirely redesign the process of data collection in verbal autopsy. With the support of the UCSC Hellman fellowship behind him, an award to support the research of promising assistant professors, Li aims to create an adaptive survey design model for verbal autopsy that would generate new questions based on previous responses to gather the most accurate information and reduce overall survey time.
“What we are trying to do in this project is to rethink this whole process and have a dynamic survey design that could eventually be used in the field, where potentially if you ask the first question like whether or not the person was hit by a car and the answer is ‘yes,’ you can can stop,” Li said. “Potentially we don’t need to go through all the questions. It would be a brand new type of data collection”
Adaptive survey design is a concept already being used in a variety of fields, from Facebook surveys to Buzzfeed quizzes that guess your celebrity crush. Li believes that now is the right time to bring adaptive surveys into the practice of verbal autopsy, which has been the subject of his research since graduate school, because the infrastructure of internet and tablets needed to carry out an adaptive survey is now in place in many parts of Africa, Latin America, and southeast Asia where verbal autopsy is routinely implemented.
Li and graduate student collaborator Toshiya Yoshida, with the input of collaborators from the openVA team, a multi-university research group focusing on verbal autopsy methods and software, will work to come up with a set of recommendations for these adaptive surveys and guidance for different situations that could arise. They aim to create user-friendly software for adaptive verbal autopsy surveys that interviewers can use in relatively easy and painless fashion in the field.
While a pilot study of this software will be logistically challenging, Li hopes his team can execute one to show that this idea can work not just in theory but in practice, and evaluate the performance of the software his team creates.
Previous attempts to shorten these lengthy surveys meant researchers had to try to determine which questions could be cut with minimal impact on assessing cause of death at the population level. But Li said these decisions introduce inherent bias in the surveys, unlike an adaptive survey, which is tailored and precise to the individual.
By working closely with and learning from practitioners and interviewers in the field, Li aims to connect statisticians back to data collection and the actual people whom their work impacts. In the long term, Li hopes that this work can help integrate the entire verbal autopsy ecosystem by bringing the initial data collection into the fold, and ultimately make the process easier for everyone involved.
“We know that by asking fewer questions we make things easier for the people that are being interviewed, and potentially make this whole process more accepted by people if they know they don’t have to sit there for hours and go through all the details of the death of their relative,” Li said. “It's the same application I've been working on for a long time, but it's taking a different part of the problem that I think is meaningful. We've been looking at how to estimate mortality better and better, but I think we should focus on how to make it easier for people to actually go through the process.”