Bach et al. (2023) DoubleML — An Object-Oriented Implementation of Double Machine Learning in R

Summary

This paper introduces the R package DoubleML, an object-oriented implementation of the double/debiased machine learning (DML) framework of Chernozhukov et al. (2018). DML lets analysts estimate low-dimensional causal parameters (e.g. a treatment effect) while using flexible machine-learning methods to control for high-dimensional nuisance functions, and still obtain valid $\sqrt{n}$ inference. The package builds on three key ingredients — Neyman orthogonality, high-quality ML estimation of nuisances (via the mlr3 ecosystem), and sample splitting / cross-fitting — and supports a range of causal models (partially linear, interactive, and their instrumental-variable extensions) through a flexible R6 object-oriented design.

Key Claims

• Three ingredients of DML. Valid ML-based causal inference needs (i) a Neyman-orthogonal score — a moment condition whose expectation is locally insensitive to errors in the nuisance estimates; (ii) high-quality ML estimation of the nuisance functions (outcome regression $g_0$, propensity/treatment model $m_0$); and (iii) cross-fitting (sample splitting) so the nuisance estimates and the target estimate use different folds.
• Why orthogonality matters. Without an orthogonal score, the regularization bias of ML nuisance estimators (e.g. a random forest's shrinkage of $g_0$ and $m_0$) translates directly into a large bias in the main estimator $\hat\theta_0$ — the "naive" plug-in is not $\sqrt{n}$-consistent.
• Cross-fitting removes overfitting bias. Estimating nuisances on one fold and evaluating the score on the held-out fold eliminates the own-observation overfitting term, restoring asymptotic normality and honest standard errors.
• Supported models. Partially linear regression (PLR), interactive regression (IRM, for ATE/ATTE), and their IV extensions (PLIV, IIVM) — all expressible through orthogonal score functions.
• Software design. Object orientation (R6) with pluggable mlr3 learners makes model specification flexible and the framework easily extendable.

Concepts Introduced or Extended

• Double Machine Learning

Entities Mentioned

• Philipp Bach
• Victor Chernozhukov

Quotes

"The double machine learning framework consist of three key ingredients: Neyman orthogonality, high-quality machine learning estimation and sample splitting."

My Take

The value of this paper is packaging rather than novelty — the intellectual core is Chernozhukov et al. (2018) — but the packaging matters: DML's whole promise is that a practitioner can throw a random forest or gradient booster at the nuisance functions and still get an honest confidence interval, and a clean R6/mlr3 implementation is what makes that promise usable. The paper is also a good on-ramp because it foregrounds the why: it shows concretely that the naive plug-in inherits the ML regularization bias, and that orthogonality plus cross-fitting is exactly what cancels it. For this wiki it anchors the DML concept and connects the tree/forest tools (CART) and the IV/LATE machinery to a modern causal-ML workflow. The honest caveat is that DML's guarantees are asymptotic and lean on the nuisance learners converging fast enough (a rate condition), which is easy to state and hard to verify in finite samples.