Regularized Bayesian transfer learning for population-level etiological distributions

• DOI: 10.1093/biostatistics/kxaa001
• Publication Date: 2020-01-14T20:10:50Z
• Authors: Abhirup Datta; Jacob Fiksel; Agbessi Amouzou; Scott L Zeger
• Keywords: Transfer of learning; Baseline (sea)

Computer-coded verbal autopsy (CCVA) algorithms predict cause of death from high-dimensional family questionnaire data (verbal autopsy) a deceased individual, which are then aggregated to generate national and regional estimates cause-specific mortality fractions. These may be inaccurate if CCVA is trained on non-local training different the local population interest. This problem special case transfer learning, i.e., improving classification within target domain (e.g., particular population) with classifier in source-domain. Most learning approaches concern individual-level person's) classification. Social health scientists such as epidemiologists often more interested understanding etiological distributions at population-level. The sample sizes their sets typically orders magnitude smaller than those used for common applications like image classification, document identification, etc. We present parsimonious hierarchical Bayesian framework directly estimate population-level class probabilities domain, using any baseline source-domain, small labeled target-domain dataset. To address sizes, we introduce novel shrinkage prior error rates guaranteeing that, absence or when perfectly accurate, our agrees direct aggregation predictions classifier, thereby subsuming default practice case. extend approach use an ensemble classifiers producing unified estimate. Theoretical empirical results demonstrate how model favors most accurate classifier. analyses demonstrating utility approach.