回归模型的经验似然拟合优度检验.docVIP

回归模型的经验似然拟合优度检验 吴鑑洪1，许王莉2 1 浙江工商大学统计与数学学院，杭州　 310018 2 中国人民大学统计学院，北京 100872 摘要：经验似然是一个强大的统计工具, 可用于经典的参数回归模型和时间序列自回归模型的 检验统计量的的构造. 但是, 当直接用残差进行统计量的构造时, 常用的经验似然方法不是分布 自由的. 本文提出一个纠偏的方法, 用于调整残差, 使得导出的检验是渐近卡方的. 该检验是尺 度不变的, 即不同于已有的一些检验, 不需要对于作为常数因子的渐近方差进行估计. 这一点能 提高检验的功效, 因为渐近方差依赖于模型, 在备择模型下的估计要大于在零假设模型下的估 计. 该检验能检测到以参数的速度收敛到零假设的备择, 当有多个可能的备择时, 可构造渐近分 布自由的极大极小检验. 最后通过模拟研究和实际数据分析进行了说明. 关键词：回归模型; 自回归模型; 经验似然; 拟合优度; 极大极小检验 中图分类号： O212 Empirical Likelihood Goodness-of-Fit Tests for Regression Models WU Jianhong1, XU Wangli 2 1 Zhejiang Gongshang University, Hangzhou 310018 2 Renmin University of China, Beijing 100872 Abstract: Empirical likelihood (EL) is a powerful tool for constructing tests for regression models including classical parametric regression and time series autoregressive models. However, when the residuals are used to construct tests, the classical EL based tests are not asymptotically distribution-free. In this paper, a bias correction is proposed to adjust the residuals so that the tests are asymptotically chi-squared. The tests also have the advantage that they are self-scale invariant,and then, unlike the existing tests, an estimator of the limit variance, as a normalizing factor, is not needed. This is useful for enhancing the power of the tests because the limiting variance is model dependent and gets larger under alternatives than it does under the null. The tests can detect the directional alternatives that are district from the null at a rate n-1/2. When there are several candidates of alternative models,asymptotically distribution-free Maximin tests are constructed. We also discuss how to use Maximin test procedure to partly deal with saturated alternatives. A simulation study is carried out and the application to a real data set is illustrated. 基金项目： National Nature Science Foundation of China , Specialized Research Fund for the Doctoral Program of Higher Education (20103326120002) 作者简介： Wu Jianhong（1976-），male，Associate profe