关于数据缺失机制的检验方法探讨  

Discussion on Testing the Mechanism of Missing Data

作　　者：孙婕[1] 金勇进[2] 戴明锋[1] 

机构地区：[1]中国人民大学统计学院,北京100872 [2]中国人民大学应用统计科学研究中心,北京100872

出　　处：《数学的实践与认识》2013年第12期166-173,共8页Mathematics in Practice and Theory

基　　金：中国人民大学科学研究基金(中央高校基本科研业务费专项资金资助)项目成果(13XNH192)

摘　　要：在调查研究中,缺失数据是一个非常普遍的问题,各种处理缺失数据的方法都是建立在数据缺失机制的某种假定上.在总结他人研究成果的基础上,分别给出了MCAR、MAR和NMAR机制的检验识别方法,MCAR机制的检验从分布特征入手,通过比较均值和方差是否一致来判定;MAR机制的检验利用Logit模型刻画缺失指示变量R的分布,通过估计参数的显著性来判定,NMAR机制则通过对数据的缺失模式和原因进行分析来识别.In survey research, missing data is a very common problem, and the methods to deal with missing data are almost based on some assumption of data missing mechanism. In this paper, based on the previous research achievement, the authors propose the test methods for MCAR, MAR and NMAR. The test for MCAR starts with the distribution characteristics, and it is determined by whether the comparison of mean and variance is consistent. The test for MAR uses Logit model to describe the distribution of missing indicator variable R, and this test is determined by the significance of estimated parameters. Moreover, the test for NAMR is identified by analyzing the missingness pattern and reason.

关 键 词：数据缺失机制 分布特征 LOGIT模型 缺失模式 

分 类 号：O212.1[理学—概率论与数理统计]