基于交叉验证的分位数处理效应估计方法  

Quantile Treatment Effect Estimation Method Based on Cross Validation

作　　者：燕明琪 石洪波[2] Yan Mingqi;Shi Hongbo(School of Statistics,Shanxi University of Finance and Economics,Taiyuan 030006,China;School of Information,Shanxi University of Finance and Economics,Taiyuan 030006,China)

机构地区：[1]山西财经大学统计学院,太原030006 [2]山西财经大学信息学院,太原030006

出　　处：《统计与决策》2024年第20期49-54,共6页Statistics & Decision

基　　金：教育部人文社会科学研究规划项目(22YJAZH092)。

摘　　要：分位数处理效应(QTE)可以用于研究干预处理对整个结果分布的影响,评估干预处理对不同分位点的异质性影响。双机器学习方法(DML)在因果推断分析中可以克服传统方法的不足,有效解决高维协变量带来的问题,消除正则化偏差和过拟合偏差,但是DML主要用于平均处理效应的估计,在估计分位数处理效应时存在计算过程复杂、难以实现的问题。文章提出一种基于交叉验证的分位数处理效应估计方法,该方法以交叉验证思想构造估计框架,采用机器学习算法估计潜在结果的预测值,并直接利用潜在结果的序列信息得到分位数值;利用线性时间选择算法,无需对所有的样本进行排序,即可在线性时间内得到分位数;交叉验证估计处理效应减小了机器学习方法估计带来的偏差。理论分析和实证研究表明,基于交叉验证的分位数处理效应估计方法的估计结果均较好,得到的估计值接近真实值,估计的标准误也较小。Quantile treatment effect(QTE)can be used to study the effect of intervention on the overall outcome distribution and evaluate the heterogeneous effect of intervention on different quantiles.In causal inference analysis,double machine learning(DML)can overcome the shortcomings of traditional methods,effectively solve the problems caused by high-dimensional covariables,and eliminate regularization bias and overfitting bias.However,DML is mainly used to estimate the average treatment effect,and there exist problems of huge calculation and low efficiency when used to estimate the quantile treatment effect.This paper proposes a quantile treatment effect estimation method based on cross validation.The method constructs an estimation framework based on the idea of cross validation,uses machine learning algorithm to estimate the predicted value of the potential result,and directly uses the sequence information of the potential result to get the quantile value.By using linear time selection algorithm,quantile can be obtained in linear time without sorting all samples.Cross validation of the estimated processing effect reduces the bias caused by machine learning method estimation.Theoretical analysis and empirical study show that the quantile effect estimation method based on cross validation has good estimation results,and that the estimated value is close to the true value,with small estimated standard error.the estimation results of the quantile treatment effect estimation method based on cross validation are better,the estimated value is close to the true value,and the estimated standard error is smaller.

关 键 词：因果推断 分位数处理效应 DML 交叉验证 

分 类 号：C81[社会学—统计学]