Federated double machine learning for high-dimensional semiparametric models

Federated learning enables the training of a global model while keeping data localized; however, current methods face challenges with high-dimensional semiparametric models that involve complex nuisance parameters. This paper proposes a federated double machine learning framework designed to address high-dimensional nuisance parameters of semiparametric models in multicenter studies. Our approach leverages double machine learning (Chernozhukov et al., 2018a) to estimate center-specific parameters, extends the surrogate efficient score method within a Neyman-orthogonal framework, and applies density ratio tilting to create a federated estimator that combines local individual-level data with summary statistics from other centers. This methodology mitigates regularization bias and overfitting in high-dimensional nuisance parameter estimation. We establish the estimator's limiting distribution under minimal assumptions, validate its performance through extensive simulations, and demonstrate its effectiveness in analyzing multiphase data from the Alzheimer's Disease Neuroimaging Initiative study.