海量数据下多指标含大量缺失的因果推断  

Causal Inference of Multiple Indicators With Missing Data Under Big Data

作　　者：韩锋 Han Feng(Teachers' College,Beijing Union University,Beijing 100011,China)

机构地区：[1]北京联合大学师范学院

出　　处：《统计与决策》2019年第11期9-12,共4页Statistics & Decision

基　　金：国家自然科学基金后期资助项目(18FTJ003)

摘　　要：因果推断中,常存在大量的缺失数据,特别是当协变量和结局变量都存在着缺失数据问题,如果处理不好,获得的估计可能会存在着偏误。文章在基于倾向评分逆概加权方法估计处理效应的基础上,调整权重为不只是倾向评分加权,还有协变量的缺失机制和结局变量缺失机制的加权,给出处理效应估计方法。应用delta方法给出估计量的渐近方差,借助模拟研究验证了因果效应估计量及其渐近方差估计的正确性和可行性,并与传统方法做比较,本文得到的估计量的Bias和MSE都较优于传统方法。In causal inference,there is often a large amount of missing data,and especially when both covariates and outcome variables have the problem of missing data,if not handled properly,the obtained estimation may have bias. This paper relies on the practice that the processing effect is estimated by propensity-score-based inverse generalized weighting method to adjust the weights,not only for the propensity score weighting,but also for the missing mechanism of covariates and the missing mechanism of outcome variables,with the processing effect estimation method offered. And then the paper uses the delta method to give the asymptotic variance of the estimator,verifies the correctness and feasibility of the causal effect estimator and its asymptotic variance estimation with the help of simulation studies,and also makes a comparison with the traditional method. Both the bias and MSE of the estimators obtained in the paper are better than traditional methods.

关 键 词：平均处理效应(ATE) 倾向评分 渐近方差 随机缺失机制(MAR) 

分 类 号：O212[理学—概率论与数理统计]