Description:

Amilcar Velez (Northwestern University): Non-Standard Asymptotic Approximations for Debiased Machine Learning Estimators

Abstract: Beyond prediction tasks, machine learning methods have become popular tools for estimating parameters of interest in economics through debiased machine learning (DML). This approach is suited for economic models where the parameter of interest depends on unknown nuisance components that require estimation. However, existing asymptotic theory does not distinguish between the alternatives practitioners face, such as the two available estimators, DML1 and DML2, introduced by Chernozhukov et al. (2018). This paper addresses these limitations by proposing non-standard asymptotic approximations. It shows that DML2 outperforms DML1 in terms of bias and mean-square error, formalizing a previous conjecture supported by simulation evidence. Additional results of this paper provide practical guidance for enhancing the implementation of DML