Bayesian Latent Class Models for Verbal Autopsy Data from Multiple Domains

• Activity Number: 23 - Bayesian Methods and Approaches in Big Data Analysis
• Type: Topic-Contributed
• Date/Time: Sunday, August 8, 2021 : 1:30 PM to 3:20 PM
• Sponsor: International Society for Bayesian Analysis (ISBA)
• Author(s): Zehang Richard Li* and Zhenke Wu and Irena Chen
• Companies: University of California, Santa Cruz and University of Michigan, Ann Arbor and University of Michigan
• Keywords: latent class models; verbal autopsy; cause of death; global health

Abstract

Verbal autopsy (VA) is a survey-based tool for assigning a cause to deaths when traditional autopsy and cause certification are not available. It has been routinely used for mortality surveillance in low-resource settings. In the last decade, several statistical and machine learning methods for inferring cause-of-death using VA data have been developed. Generalizability has been a common challenge with most of the probabilistic VA algorithms, as data collected from different domains, e.g., locations or time periods, often exhibit different relationships between causes and symptoms. As a result, the choice of training data has strong implications on the performance of VA algorithms. In this talk, I will present statistical approaches to characterize the joint distribution of symptoms while accounting for the heterogeneity of data from different domains. We propose a novel latent class model that classifies causes-of-death by learning the similarities between the new domain and the existing domains. I will demonstrate the performance and interpretability of the method using a gold-standard VA dataset collected from multiple study sites.