Shen, J. (STAT) – Bayesian Modeling and Uncertainty Quantification for Verbal Autopsy Data

Verbal autopsy (VA) is a well developed tool to collect information describing deaths outside of hospitals by conducting surveys to the relatives and caregivers of the deceased person. It is routinely-implemented in low and middle income countries, where it often lacks sufficient resources to conduct the autopsy. The main task is to estimate both individual level cause-of-death probabilities and population level cause-specific mortality fractions. In this document, we present three projects dealing with challenges current VA modeling faces. In the first project, we build a shared latent class model for verbal autopsy, which allows causes to share common symptom patterns. A truncated Bayesian nonparametric prior allows the number of latent classes to adapt to data, while retaining a computationally tractable representation. We provide a general framework for few-shot learning of the VA data, where limited labels can be combined in analysis with a potentially much larger collection of unlabeled symptom profile. Two complementary factorizations are considered to account for different types of distribution shift between source and target. In the second project, we develop a conformal prediction procedure for verbal autposy. For each death, we generate a conformal prediction set from existing VA model outputs, which guarantees a marginal coverage of true cause from a fequentist perspective. We also investigate conformal Bayesian procedures that more directly incorporate posterior uncertainty from Bayesian VA models. In the third project, we consider situations where source and target do not share a common cause list. We propose methods using repulsive priors to identify deaths in the target domain whose symptom profiles are not adequately represented by the known causes in the source domain.

Event Host: Jibo Shen, Ph.D. Student, Statistical Science

Advisor: Zehang Richard Li

Zoom: https://ucsc.zoom.us/j/94158273558?pwd=VZORHL8P5O9bfb5JpSMZAL1DOM44uC.1&jst=2

Passcode:  294401