The Added Value of Machine Learning to Economics: Evidence from Revisited Studies

Recently, promising new methods based on Machine Learning (ML) have been introduced to conduct causal inference, see for an overview Section 4 of Athey (2017). These methods can pick up complex and high-dimensional nuisance relationships such that they improve plausibility of the widely used unconfoundedness and IV assumptions that identify causal effects. Accordingly, causal inference might become more credible than with established methods like difference-in-difference, OLS, fixed effects or two stage least squares estimation. However, the merits of these new methods in empirical applications have not been studied yet. The purpose of this paper is therefore to employ ML-based methods to conduct causal inference on the Average Treatment Effect (ATE) and on Heterogeneous Treatment Effects (HTEs) by revisiting two well-known and well-cited applied papers. We compare to the original results the new results from different ML-based causal inference methods for the ATE (Double Machine Learning and Approximate Residual Balancing). We find that our ML-based methods for the ATE give estimates that deviate from original ones. This implies that these methods might improve causal inference from established methods to a great extent. Then, we extend the original papers by applying different ML-based methods for HTEs (heterogeneous DML and Causal Forests). This gives us additional relevant findings which could not have been obtained with established causal inference techniques.