主 题：Evaluating alternation in regression coefficients directed by low-effect variables
主讲人：Peter Song
主持人：林华珍教授
时 间：2013年12月10日下午3：30-4：30
地 点：通博楼B座212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：
Dr. Peter Song is currently Full Professor of Biostatistics in the School of Public Health, University of Michigan, Ann Arbor. He received his PhD degree in Statistics from University of British Columbia, Vancouver in 1996.Prior to the appointment of professorship at University of Michigan in Ann Arbor,he served as faculty in University of Waterloo and York University in Canada. Dr. Song is an elected member of International Statistical Institute, and the recipient of John von Neumann award from Technical University of Munich, Germany in 2013.He is now Associate Editor of Statistica Sinica, Canadian Journal of Statistics, Sankhya, and a Co-Guest Editor of Statistics and Its Interface. His current research is funded by multiple grants from National Science Foundation and National Institute of Health. Dr. Song has published over 80 articles in top-tier statistical and biomedical journals, including a single-authored monograph "Correlated Data Analysis: Modeling, Analytics and Applications" by Springer. Some of his published works have been reported by CNN, the Wall Street Journal and CBS.
内容提要：
Regression analysis routinely assumes constant covariate effects. This assumption does not address dynamics of interest in the study of growth mechanism or characterization of disease progression, because the rate of growth or disease progression is intervened by some variables such as dietary intake, environmental exposure, medication or genetic mutation. Interestingly, most of such interveners are of small size in their effects, and the traditional statistical method fails to detect their statistical significance. In this talk I will introduce a new modeling strategy that incorporates a type of principle component in the formation of regression coefficients,termed as index coefficients that allow us to combine weak covariates into possibly strong variable-groups. Statistical estimation and inference in such model is challenging because it contains nonlinear interactions between groups of low-effect variables and covariates of interest (e.g. age or time). I will present a conceptually simple and numerical stable estimation procedure by profiling least squares method with B-splines, while to estimate nonparametric functions by a spline backfitted local linear procedure. I will briefly discuss estimation consistency and asymptotic normality for the proposed estimators of index coefficients as well as the oracle property of the nonparametric function estimator. The proposed models and methods are illustrated by both simulation studies and an analysis of body fat data. This is a joint work with Dr. Shujie Ma from University of California at Riverside.
