A0972
Title: Robust Bayesian change point detection
Authors: Daewon Yang - Chungnam National University (Korea, South) [presenting]
Taeryon Choi - Korea University (Korea, South)
Abstract: A new Bayesian approach for robust change point detection is proposed. The model utilizes the Dirichlet process hidden Markov model for multiple change point detection, which has the advantage of not requiring a predetermined number of change points. Furthermore, for robust estimation, a heavy tail error assumption is introduced based on the Student's t distribution. The model employs a mixture error assumption with Gaussian and Student's t errors to perform robust estimation, which aids in clearer change point detection. The proposed model is applied to an environmental epidemiology application based on a two-stage meta-analysis, analyzing the association between temperature and mortality in Japan from 1974 to 2015. In the first stage, the relationship between temperature and mortality is analyzed across each prefecture of Japan for four-year non-overlapping sub-periods using a distributed lag nonlinear model. In the second stage, the robust Bayesian change point detection methodology is applied to examine how the association between temperature and mortality rates in Japan has changed over time.
