Double Debiased Machine Learning

DDML produces unbiased causal estimates by capturing nonlinear relationships without requiring pre-specified functional forms. DDML was selected to overcome traditional regression limitations, specifically bias from overfitting with high-dimensional covariates and the requirement to pre-specify functional forms. Using only gradient boosting without testing alternative estimators such as random forests or neural networks limits the robustness of the DDML estimates. The nuisance estimation stage uses Gradient Boosting Regressors with 5-fold cross-fitting to prevent overfitting. DDML operates in two stages within a Partially Linear Model framework, first estimating nuisance parameters then estimating the causal treatment effect from residualised data.