基于监督机器学习的Rubin因果范式研究：模型构建与设定检验

Rubin因果范式与监督机器学习方法在经济学研究中广泛应用，但实践中也存在一些问题：①因果推断领域中多种研究范式并存，之间存在激烈争论。Rubin范式中非混淆性假设的内涵、混淆变量的选取准则、模型设定检验等问题缺乏系统性解决方案；②监督机器学习方法是预测性统计模型，其算法是自适应性的“黑箱”方法，不适合直接用于因果推断。.本课题结合Rubin因果范式与监督机器学习，从几个方面解决上述问题：①系统比较Rubin范式与Pearl范式，厘清Rubin范式的理论基础与核心假设；②引入高维LASSO与弹性网，构建高维混淆变量选取方法；③结合递归分割与回归树、系统方程等方法，构建Rubin因果范式的模型设定检验；④研究因果机制分析的中介分析方法，分析序贯非混淆性假设与有向无环图的内在联系。⑤将上述理论成果应用于中国教育生产函数、教育收益率的实证分析，为政策制定者提供审慎的政策建议。

The Rubin Causal Model (RCM) and the supervised machine learning are widely applied in empirical researches. However, several issues remain unresolved, as following. First, there are several paradigms in causal inference, among which there are fierce debates. Several important issues in RCM, including connotation of the Unconfoundedness Assumption, principle of confounding variable selection, and model specification test lack systematic approaches. Consequently, researchers just include all confounding variables linearly, without much systematic effort to find more compelling specifications. Second, supervised machine learning techniques are predictive statistic models, which are adaptive "black box" methods, could not directly serve as tools for causal inference. Some scholars have warned about the danger of "naively" applying machine learning methods..In our research project, we want to combine machine learning methods and the RCM, to explore several important issues in treatment effect study. First, we will compare the RCM and the Pearl Causal Model, therefore to clarify the foundation of RCM. Second, we will combine the LASSO and the Elastic Net method with the RCM, to build a confounding variable selection method for high dimensional data. Third, we will adopt the Recursive Partition and Regression Tree, and the System of Equations Regression, to build model specification tests for the RCM. Fourth, we will study the Causal Mediation Analysis method, analyze the relation between the Sequential Unconfoundedness Assumption and the corresponding Directed Acyclic Graph. Last, we will adopt these techniques in the study of the production function of education, and return of education in China, to offer prudent advice for policy makers.