社会科学背景下因果推论的统计方法

Statistical Methods of Causal Effects in Social Sciences

如何根据非随机数据估计变量间的因果关系是社会科学研究中一个迫切的方法论问题。上世纪70年代,Rubin等人指出因果问题本质上是一个反事实的问题,认为某些统计方法可保证混淆变量和分组安排独立,并将这种方法推广到观察数据的分析中。倾向分数、工具变量和回归间断点是三种常用的方法,其中倾向分数居核心地位。以实际数据为例建立计算倾向分数的logistic模型,报告了模型的整体检验、预测变量的显著性检验和多重共线性检验、建立匹配组和分析结果报告。

People before stuck to the apparent relation between objects to infer the causal effects. After applying the manipulation and control means to research, they have made great progress but also faced more difficulties. In 1970s, Rubin pointed out that the nature of the causal effects is the issue of counter fact and that some statistical methods used to guarantee the confounding variables is independent from subject assignment. Now, these methods are extensively used to analyze the observable data. The present paper introduces three statistical methods for causal effects inference and explains how to apply the propensity score to data analysis.