异质性政策效应评估与机器学习方法:研究进展与未来方向

Policy Effect Heterogeneity Evaluation and Machine Learning Methods:Research Progress and Future Orientation

对公共政策的准确评估,是制定科学公共政策的重要前提,科学的公共政策将有助于促进国家治理体系和治理能力现代化。异质性政策效应评估作为一种新兴的研究范式,其重要性在政策效应评估文献中已经获得广泛认可。本文总结了异质性政策效应评估的重要价值,以及代表性传统方法的逻辑和局限性。在此基础上,文章重点梳理了机器学习方法在异质性政策效应评估中的重要价值和具体应用:更好地筛选和切分异质性变量、更好地评估多重异质性政策效应、更好地估计个体政策效应等。本文也指出机器学习在异质性政策效应评估的算法可接受性、过程可检验性以及结论稳健性中存在局限性。进一步,文章提出了异质性政策评估和机器学习的重点发展方向:引入和发展机器学习方法,重视异质性政策评估的政策价值及提升机器学习的可接受性;结合传统分析范式,拓展机器学习在异质性政策评估中的新模式;规范研究数据的采集和处理,推动数据和代码的公开透明等。

Accurate assessment of public policies is a prerequisite for the formulation of scientific public policies,which can contribute to the modernisation of the country's governance system and capacity.As a new research paradigm,evaluating the effects of policies based on the treatment effect heterogeneity has been widely recognized to be of great importance in the literature on policy evaluation.This article states the significance of treatment-effect heterogeneity of policies,then introduces its representative traditional methods and further makes a comment on the limitation of these methods.On this basis,the key point of the article is elaborating the value of machine learning methods and its applications in policy effect heterogeneity evaluation.The specific applications are as follows:better screening and segmentation of heterogeneous variables;identify the multiple heterogeneous treatment effect in policy effect evaluation;estimate the policy effects on individuals and so on.Also,this paper points out that machine learning have doubts about the operability of recommendations,the testability of processes and the robustness of conclusions in the evaluation of heterogeneous policy effects.This paper proposes that under the background of big data,the key development directions of heterogeneous policy evaluation and machine learning include:(1)Introduce and develop machine learning methods,pay attention to the policy value of policy effect heterogeneity evaluation and improve the acceptability of machine learning;(2)Combining with the traditional analysis paradigm,expand the new model of machine learning in policy effect heterogeneity evaluation;(3)Standardize the collection and processing of research data,and promote the openness and transparency of data and code.