主 题：Survival analysis with longitudinal covariates measured with correlated errors
主讲人：Jianxin Pan教授
主持人：林华珍教授
时 间：2014年7月28日下午3:00-4:00
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：
Jianxin Pan,教授， School of Mathematics,University of Manchester, 他是统计学顶级期刊Biometrics副主编，Fellow of The Royal Statistical Society，Elected Member of The International Statistical Institute 。2002年和2005年分别出版了《Growth Curve Models and Statistical Diagnostics》和《Case-Deletion Diagnostics in Linear Mixed Models》两本书籍，在统计学Journal of the American Statistical Association、Biometrika等期刊发表学术论文80余篇。
内容提要：
When covariates in Cox's proportional hazards model are time-dependent, statistical inferences may be similar to those with time-independent covariates provided that complete knowledge of the true covariates history is available. Time-dependent covariates, however, are usually measured intermittently and very likely to be measured with errors, so that joint modelling of survival and longitudinal data is much preferred. The existing methods such as sufficient statistical methods assume that longitudinal covariates are measured with mutually independent errors, which unfortunately is not always true in practices. In this research, it is evident through simulation studies that violation of the independent errors assumption can lead to very biased estimates of regression coefficients. Generalized least square estimates, rather than ordinary least square estimates, are adopted for time-dependent covariates to account for correlated measurement errors. Furthermore, covariance modelling strategy based on modified Cholesky decomposition is proposed to model the covariance structure of the measurement errors. Simulation studies show that the proposed method performs very well. Real data analysis is provided too.
