变指标系数模型的贝叶斯分位数回归概率论与数理统计专业论文.docxVIP

AbstractWhen Abstract When errors follow a higher peak and fat tailed distribution or a distribution with het- eroscedasticity,the ordinary least squares methods may lead to worse estimation perfor- mance．HoweveL these circumstances are often encountered in practice．Generally speak— ing，the quantile regressions can produce better estimators，since they Call estimate the en- tire distribution of the response conditional on the covariates in theory by selecting different quantiles．We will use Bayesian methods to investigate quantile regressions in the presented Pal；ler． Data in practice may include categorical variables as covariates．Single—index models can not handle such data as the covariates that enter the index should be continuous in general．Although the partially linear single—index models Can be used for analysis，but they assume that the relationship between categorical covariates and the response is linear,and can not model the interactions between the continuous and categorical covariates．To address those problems，we generalize the varying-index coefficient models to quantile regressions SO that new models can deal with more complicated data structure． We develop a fully Bayesian quantile regressions with free—knot splines to analyze the varying—index models．As the existing algorithms to sampling from the full conditional posterior of the index coefficient vectors may no longer converge under the random error with thick tail or asymmetry,we propose a new one that is based on the proposal distribution augmentation algorithm．We verify the effectiveness of our method by simulations，and further analyze a real data set． Key Words：Quantile Regression，Bayesian Analysis，Varying—Index Coefficient Models —II— 万方数据 目 目 录 摘要 ．． I Abstract ． ．． ． ．． ． ．． ． ．．II 目录 ．．HI Contents．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．．1V7 第一章诸论 1 1．1研究背景 1 1．2分位数回归 ． 2 1．3变指标系数模型 ．． 4 1．4贝叶斯分析 ． 6 第二章分位数回归的贝叶斯抽样 ．7 2．1非对称Laplace分布 ．． 7 2．2变指标系数模型分位数回