Double/Debiased Machine Learning for Treatment and Structural Parameters
Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, James Robins
Summary
The paper develops a general framework for estimating low-dimensional treatment or structural parameters when high-dimensional nuisance components are estimated with machine learning methods. By combining Neyman-orthogonal (debiased) moment conditions with sample-splitting/cross-fitting, the approach removes regularization and overfitting biases. The resulting estimators are root-N consistent, asymptotically normal, and valid for inference despite slowly converging nuisance estimates.
Key findings
- Neyman orthogonality makes moment conditions insensitive to first-order errors in nuisance estimates.
- Cross-fitting (sample splitting) eliminates bias from overfitting and weakens entropy/complexity requirements.
- Delivers root-N consistent, asymptotically normal estimators usable with a wide range of ML learners (lasso, random forests, neural nets, etc.).
Subjects & keywords
Cite this paper
Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, & James Robins (2018). Double/Debiased Machine Learning for Treatment and Structural Parameters. The Econometrics Journal. https://doi.org/10.1111/ectj.12097
@article{chernozhukov2018doubledebiased,
author = {Victor Chernozhukov and Denis Chetverikov and Mert Demirer and Esther Duflo and Christian Hansen and Whitney Newey and James Robins},
title = {Double/Debiased Machine Learning for Treatment and Structural Parameters},
journal = {The Econometrics Journal},
year = {2018},
doi = {10.1111/ectj.12097},
url = {https://doi.org/10.1111/ectj.12097}
}