高华川,白仲林.一种基于机器学习的时变面板数据政策评估方法[J].数量经济技术经济研究,2019,(8):111-128 |
一种基于机器学习的时变面板数据政策评估方法 |
A Time Varying Panel Data Approach for Program Evaluation Based on Machine Learning |
|
DOI: |
中文关键词: 面板数据 政策评估 机器学习 时变LASSO 脱欧公投 |
英文关键词: Panel Data Program Evaluation Machine Learning Time Varying LASSO Brexit |
基金项目:本文得到天津市哲学社会科学研究规划项目“大数据背景下的宏观经济实时预测”(TJTJ16-001Q)、教育部人文社会科学研究规划基金项目“宏观经济政策因果效应分析方法及其应用研究”(8YJA790005)、国家自然科学基金青年科学基金项目“ACD模型中的密度估计与模型检验”(11801399)资助。 |
|
中文摘要: |
研究目标:基于个体间关系时变的面板数据,提出一种估计政策因果效应的反事实机器学习算法——时变LASSO方法。研究方法:基于真实数据生成过程未知的经典文献中的实际数据和时变的模拟数据,分别将时变LASSO方法与两种常系数方法(HCW方法和常系数LASSO方法)进行对比研究;以英国“脱欧公投”对英镑汇率的因果效应分析为例,研究了政策信息提前披露时政策因果效应的动态行为。研究发现:相对于两种常系数方法,时变LASSO方法能更准确地估计反事实结果,更适用于样本时期数较长、政策干预时间点相对较晚的政策评估;实证研究表明,常系数方法对脱欧公投的影响存在一定的高估。研究创新:提出了一种基于机器学习的时变LASSO面板数据政策评估方法,以及基于此的两种反事实估计方法。研究价值:对于政策信息提前披露的政策,提供了评价政策因果效应动态行为的工具,完善了面板数据政策评估理论方法。 |
英文摘要: |
Research Objectives:Based on the panel data of time-varying relationships among individuals,a counterfactual machine learning algorithm,time-varying LASSO method,is proposed to estimate the policy causal effect. Research Methods:Using the real data in classical literatures whose data generation process is unknown and time-varying simulated data,the time-varying LASSO method is compared with two constant coefficient methods (HCW method and LASSO method),and by taking analysis of impact of “Brexit” on sterling exchange rate as an example,the dynamic behavior of policy causal effect is studied when policy information is disclosed in advance. Research Findings: Compared with the two constant coefficient methods,the time-varying LASSO method can estimate the counterfactuals more accurately,and is more suitable for program evaluation with long sample periods and with relatively late intervention time,which improves the timeliness of policy evaluation. Empirical studies show that the constant coefficient method could overestimate the impact of Brexit. Research Innovations: Propose a time-varying LASSO panel data policy evaluation method based on machine learning,and two counterfactual prediction methods. Research Value: For the policy of which the information is disclosed in advance,it provides a tool to evaluate the dynamic behavior of causal effect,and improves the theoretical methods of panel data program evaluation. |
查看全文 |