大数据背景下面板数据政策评估的估计和推断  

Estimation and Inference of Treatment Effects with Panel Data in the Big Data Era

作　　者：沈艳[1] 李星宇  Shen Yan;Li Xingyu;Zhou Qiankun

机构地区：[1]北京大学国家发展研究院 [2]路易斯安那州立大学经济学系

出　　处：《复印报刊资料（统计与精算）》2022年第5期14-29,共16页STATISTICS AND ACTUARIAL SCIENCE

基　　金：国家社会科学基金重大项目“数字普惠金融的创新、风险与监管”(18ZDA091)的资助。

摘　　要：研究目标:介绍大数据背景下基于面板数据模型的政策评估方法的最新进展与相关应用。研究方法:回顾双重差分法、合成控制法、面板数据方法、因子估计方法和机器学习方法这几类方法在估计面板数据因果效应方面的最新进展后,介绍现有研究中基于上述估计量的推断方法,最后报告已有文献对于不同方法的对比,并提供实证应用建议。研究发现:当实证应用问题中随时间变化的因子个数超过一个时,特别要关注基于双向固定效应的双重差分法的适用性。运用双向固定效应设定模型不恰当时,可考虑使用基于交互固定效应模型的因子模型类估计和推断方法。研究创新:从大数据时代的政策评估需求出发,梳理基于面板数据的因果效应估计和推断方法并给出应用建议。研究价值:为实证研究者提供了选择政策评估方法的参考指南。Research Objectives:A literature review of programme evaluation methods based on panel data models.Research Methods:We first review the recent advances in treatment effect estimation,including difference in differences,synthetic control methods,panel data approaches,factor estimation methods,and machine learning methods.Then we introduce the existing inference methods based on the above estimators.Finally,we report the comparison of different methods in the existing literature and provide guidelines for their empirical application.Research Findings:This paper suggests that empirical researchers should pay particular attention to the applicability of the two-way fixed effects-based difference in differences method when the number of time-varying factors is more than one.When models with two-way fixed effects are not appropriate,one can consider estimation and inference methods based on factor models(panel data models with interactive fixed effects).Research Innovations:We introduce the estimation and inference methods for causal effects based on panel data from the needs of policy evaluation in the big data era and give suggestions on their applications.Research Value:Providing aguidance for empirical researchers on how to choose programme evaluation methods.

关 键 词：面板数据 处理效应 因子模型 双重差分 合成控制法 机器学习 

分 类 号：O212.1[理学—概率论与数理统计]