处理效应异质性分析——机器学习方法带来的机遇与挑战

Analysis of Heterogeneous Treatment Effect: New Opportunities and Challenges with Machine Learning Techniques

处理效应异质性是定量社会科学关注的重点。本文以因果随机森林与贝叶斯叠加回归树为例,指出以算法为导向的新兴分析手段可以克服模型形式和变量选择的限制,并考虑变量间各种交互关系。因果随机森林与贝叶斯叠加回归树分别体现了"匹配"和"模拟"的分析逻辑,以帮助研究者勾勒出异质性处理效应的经验分布并探索该异质性的决定因素。然而,参数设定差异和算法差异都会损害处理效应异质性分析结果的稳健性。

Investigation on heterogeneous treatment effect has been a focus for empirical sociologists.This article uses the causal random forests and the Bayesian additive regression trees to show that new technologies based on algorithm have no model restrictions and can examine all possible interactions between the treatment and the confounding variables.The two methods,respectively,illustrating the ideas of matching and simulation,provide estimates of the individual treatment effect,which helps scholars show the empirical distribution and look into the determinants of the heterogeneity of treatment effect.However,the algorithm-based methods can also bring about new challenges.For instance,arbitrariness in terms of parameter configuration and algorithm can hurt the consistency of empirical results.