Abstract

There is an increasing trend towards combining machine learning methods with traditional econometric methodologies. Starting from comparing features and internal relations of two mainstream causal inference frameworks, this paper proposes that causal inference can be significantly improved with the introducing of machine learning methods in two ways, one is sample matching and one is counterfactual prediction. Firstly, machine learning techniques can enhance matching qualities by pairing samples directly or improving the accuracies of propensity score predictions. This can make the matched samples more similar to samples collected from randomized controlled trials. Secondly, machine learning techniques can improve the accuracies of counterfactual predictions by modeling complex relations, using cross-validation, and adopting regularization. This paper then introduces the theoretical foundations of combining machine learning techniques and causal inferences by reviewing four specific methods:Matching, regression discontinuity, difference-in-difference, and synthetic control method. At the meantime, several application cases are provided in each method section for researchers in applied econometrics as references.

Citation

Haoqi Qian, Yanran Gong, Libo Wu (2021). More Accurate Causal Inference: A Perspective of Machine Learning. China Journal of Econometrics, https://doi.org/10.12012/CJoE2021-0027