主 题：On Bayesian Quantile Regression with Asymmetric Laplace Likelihood
主讲人：Xuming He 教授
主持人：林华珍教授
时 间：2014年10月24日14：00-14：40
地 点：通博楼B座212学术会议室
主办单位：统计学院 统计研究中心 科研处
主讲人简介：
Xuming He is H.C. Carver Professor of Statistics at University of Michigan. He is an honorary professor of Statistics at the University of Hong Kong, and a Changjiang Visiting Professor at Northeast Normal University. His research interests include theory and methodology in robust statistics and semiparametric regression and applied statistics in biomedical sciences and climate research. He is an elected fellow of the Institute of Mathematical Statistics (IMS), the American Statistical Association (ASA), and the American Association for the Advancement of Science (AAAS). His leadership in the statistical community includes his positions as the 2010 President of the International Chinese Statistical Association (ICSA), Council member of the International Statistical Institute (ISI), Council member of the Institute of Mathematical Statistics (IMS), and editorship of JASA. He has been a keynote speaker at international conferences including the Joint Statistical Meetings in 2007 and the International Conference on Computational Statistics (COMPSTAT) in 2014.
内容提要：
We consider the asymptotic validity of posterior inference of Bayesian quantile regression methods with complete or censored data, when an asymmetric Laplace likelihood is (mis)-specified. The Bayesian asymmetric Laplace quantile regression methods utilize an efficient Markov Chain Monte Carlo algorithm for estimation. However, our asymptotic results suggest that the posterior chain from the Bayesian asymmetric Laplace quantile regression does not lead to valid posterior inference. We propose a simple correction to the covariance matrix of the posterior chain to enable asymptotically valid posterior inference. Our simulation results demonstrate the advantages of the proposed correction in finite samples, especially for cases with heteroscedastic errors. The talk is based on joint work with Yunwen Yang (Drexel University) and Judy Wang (George Washington University).
