[Paper Review] DoubleML -- An Object-Oriented Implementation of Double Machine Learning in Python

Philipp Bach, Victor Chernozhukov+2 more|arXiv (Cornell University)|Apr 7, 2021

Computational Physics and Python Applications19 citations

TL;DR

This paper introduces DoubleML, an open-source Python library that implements the double machine learning (DML) framework for causal inference using Neyman-orthogonal score functions. It enables valid statistical inference on causal parameters—such as average treatment effects—by combining machine learning for high-dimensional nuisance functions with sample splitting and orthogonalization, ensuring robustness to regularization bias.

ABSTRACT

DoubleML is an open-source Python library implementing the double machine learning framework of Chernozhukov et al. (2018) for a variety of causal models. It contains functionalities for valid statistical inference on causal parameters when the estimation of nuisance parameters is based on machine learning methods. The object-oriented implementation of DoubleML provides a high flexibility in terms of model specifications and makes it easily extendable. The package is distributed under the MIT license and relies on core libraries from the scientific Python ecosystem: scikit-learn, numpy, pandas, scipy, statsmodels and joblib. Source code, documentation and an extensive user guide can be found at https://github.com/DoubleML/doubleml-for-py and https://docs.doubleml.org.

Motivation & Objective

• To provide a flexible, extensible, and user-friendly implementation of the double machine learning framework in Python for causal inference.
• To enable valid statistical inference on causal parameters despite the use of high-dimensional, nonparametric machine learning estimators for nuisance functions.
• To support a wide range of machine learning methods and model specifications through an object-oriented architecture.
• To ensure robustness to regularization bias via Neyman orthogonality and sample splitting, particularly in high-dimensional settings.
• To facilitate extensibility for new DML models, including advanced inference methods like clustered standard errors and multiplier bootstrap.

Proposed method

• The library implements Neyman-orthogonal score functions that ensure the causal parameter estimate is robust to first-order bias from machine learning estimators of nuisance functions.
• It uses sample splitting across K-folds with repeated cross-fitting to reduce overfitting bias and improve estimation efficiency.
• The core architecture is object-oriented, with base classes like `DoubleML` and model-specific subclasses such as `DoubleMLIRM` for interactive regression models.
• It integrates with the scientific Python stack (scikit-learn, numpy, pandas, statsmodels, etc.) and supports various ML algorithms including random forests, lasso, and XGBoost via scikit-learn-compatible interfaces.
• The package supports customizable resampling schemes and extensible score functions, enabling implementation of new DML models like difference-in-differences or second-order orthogonal estimators.
• It provides built-in support for inference tools such as cluster-robust standard errors and multiplier bootstrap, enhancing reliability of p-values and confidence intervals.

Experimental results

Research questions

• RQ1 How can double machine learning be effectively implemented in Python to enable valid statistical inference in high-dimensional causal models?
• RQ2 What architectural design enables flexibility and extensibility in implementing diverse DML models while maintaining statistical rigor?
• RQ3 How does sample splitting and Neyman orthogonality jointly reduce regularization bias in machine learning-based nuisance estimation?
• RQ4 To what extent can the library support advanced inference features like clustered standard errors and multiplier bootstrap within a modular, object-oriented framework?
• RQ5 How does the integration of diverse machine learning methods (e.g., random forests, XGBoost) into a unified DML pipeline affect estimation accuracy and inference validity?

Key findings

• DoubleML successfully implements the double machine learning framework in Python, enabling valid inference on causal parameters such as the average treatment effect (ATE) even with high-dimensional, nonparametric nuisance functions.
• The use of Neyman-orthogonal score functions ensures that estimation bias from machine learning methods has no first-order impact on the causal parameter estimate, as demonstrated by reduced bias compared to naive ML approaches.
• Sample splitting with repeated cross-fitting significantly improves estimation efficiency and reduces overfitting bias, as shown in comparative simulations.
• The object-oriented design allows seamless extension to new DML models, such as difference-in-differences or second-order orthogonal estimators, by inheriting from core classes.
• The package achieves 99% test coverage, adheres to PEP8 standards, and is compatible with major ML libraries like scikit-learn, XGBoost, and Keras through a flexible learner interface.
• The library supports advanced inference features such as cluster-robust standard errors and multiplier bootstrap, which are critical for reliable hypothesis testing in complex data structures.