因果推断中基于能源距离的协变量分布平衡

Covariate Distribution Balance via Energy Distance for Causal Inference

本文基于倾向得分构造处理组和对照组协变量的经验加权分布,采用能源距离度量加权经验分布与总体协变量经验分布的差异,通过最小化分布差异最大化协变量平衡,进而估计倾向得分和平均处理效应。本文最优化不带任何约束并且保证解的唯一性。同时,平均处理效应估计量是√n相合,估计量精度较高。将本文方法应用于考察养老保险对生育意愿的影响,相比回归分析本文结果可靠性更佳。

This paper constructs the empirical weighted distributions of covariates with respect to the treatment and control groups based on propensity score,and utilizes energy distance to quantify the distribution differences between the above empirical weighted distributions and the empirical distribution with respect to the overall covariates.The optimal covariate balance can be achieved by minimizing the distribution differences,and the estimators of propensity score and average treatment effect can be obtained accordingly.Our optimization is without only constraint,and exists unique solution.Moreover,the average treatment effect estimator is root n consistent and obtains good performance in accuracy.This paper applies the method to investigate the effect of pension insurance on fertility intention,the result is more reliable compared with regression analysis.