Statistics Senior Thesis Defense by Youngsoo Baek '19, Wednesday, May 8

Bayesian Hierarchical Model for Estimating Particulate Matter (PM10) Levels in South Korea by Youngsoo Baek ’19, Statistics Senor Thesis Defense, Wednesday, May 8, 1:10 – 1:50 pm, Stetson Court Classroom 105

Abstract:  High levels of particulate matter (PM) have raised health concerns in the East Asian region, including South Korea, China, and Japan. The volatile nature of particle transport, and the complexity of dependencies across multiple meteorological factors across space and time, pose various challenges to modeling PM levels and making accurate forecasts. We propose a Bayesian hierarchical regression model for Gaussian process of PM log-levels. The model extends the Multivariate Spatiotemporal Mixed Model (MSTM) framework proposed by Bradley, Holan, and Wikle (2015), through choosing the basis functions applicable to point-referenced dataset. Our major innovation is the extension of Moran’s I basis functions to point-referenced dataset, systematically accounting for historical changes in the spatial coordinates of the stations in the data. We fit the model to a dataset of particulate matter with diameter less than 10 μg/m3 (PM10) collected in South Korea, publicly available from the Korean Meteorological Association (KMA). Due to the computationally taxing nature of Markov Chain Monte Carlo methods, we fit models to three subsamples of the data. Our results show that the model is computationally efficient, produces reliable forecasts of PM levels up to a short lag of 6 to 12 hours ahead, and handles missing values with ease. We benchmark the predictive performance of our model against a similar hierarchical model based on empirical orthogonal functions (EOF), which requires imputation of response data. This thesis marks the first attempt to introduce a systematic Bayesian approach to estimating and forecasting PM level data in South Korean region.