回归模型中有影响变量的检测标准及其灵敏度分析

本文从修正的似然入手，得到了回归模型中自变量影响的一种有效度量，它的表达式简洁易在常见的统计软件上实现．这种影响度量可帮助我们建立一个合适的回归模型，作为特例，它可用来检测回归模型中的异常点或强影响点．利用局部影响分析的思想，我们还研究了这种影响度量的灵敏度，据此可成功地发现对选模有较大影响的观察数据．最后我们给出了一个应用实例．

The Criterion for Detecting lnfluential Variables in Linear Regression Modelsand lts Sensitivity Analysis

Based on the modified profile log-likelihood,we derive an efficient influence measure of explanatory variables. The computation of the measure is simple and can be realized in the usualstatistical software. Applying our measure,we can establish an appropriate model in model building. Using the idea of local influence,we also study the sensitivity of the measure and then can detect successfully the innuential observations for model selection.