Double/Debiased Machine Learning for Economists: Practical Guidelines, Best Practices, and Common Pitfalls

Abstract In an era of increasing data complexity and volume, integrating machine learning (ML) into economics has become essential. Double/debiased machine learning (DML) provides a framework for causal inference using ML algorithms. However, the transition from theoretical concepts to practical implementation requires clear, accessible guidance. This paper bridges this gap by first providing an intuitive understanding of the DML framework, focusing on the partially linear model. It then establishes generalizability, measured through out-of-sample performance, as the primary criterion for model selection. The paper presents best practices for achieving superior generalizability, including appropriate evaluation metrics, cross-validation procedures, and effective hyperparameter tuning. A key contribution is clarifying the distinction between cross-validation for model selection and cross-fitting for unbiased estimation, two complementary procedures often confused in practice. These elements are synthesized into a systematic two-stage implementation framework with guidance on learner selection. By providing step-by-step procedures while emphasizing proper evaluation, this paper makes ML-based causal inference more accessible and enables more accurate and robust causal inferences in economic research.