Debiased Machine Learning U-statistics

We propose a method to debias estimators based on U-statistics with Machine Learning (ML) first-steps. Standard plug-in estimators often suffer from regularization and model-selection biases, producing invalid inferences. We show that Debiased Machine Learning (DML) estimators can be constructed within a U-statistics framework to correct these biases while preserving desirable statistical properties. The approach delivers simple, robust estimators with provable asymptotic normality and good finite-sample performance. We apply our method to three problems: inference on Inequality of Opportunity (IOp) using the Gini coefficient of ML-predicted incomes given circumstances, inference on predictive accuracy via the Area Under the Curve (AUC), and inference on linear models with ML-based sample-selection corrections. Using European survey data, we present the first debiased estimates of income IOp. In our empirical application, commonly employed ML-based plug-in estimators systematically underestimate IOp, while our debiased estimators are robust across ML methods.