We propose Trans-FARM, a transfer learning method for high-dimensional sparse linear models augmented with latent factors, addressing settings where covariates exhibit strong correlations and latent factor structures, which is motivated by the demand for stable and reliable estimations in economic and financial applications. Our framework jointly corrects for latent factor-induced effects and mitigates multicollinearity, reducing potential model misspecification relative to conventional linear sparse regression. When the target dataset is limited by sample size but multiple heterogeneous auxiliary datasets are accessible, we propose transfer-augmented estimators that leverage auxiliary information to improve estimation accuracy. We establish non-asymptotic $\ell_1$- and $\ell_2$-error bounds for the oracle estimator with a known informative source set and derive corresponding minimax lower bounds. This estimator is minimax optimal under suitable regularity and source-target similarity conditions. The minimax lower bounds also reveal a dimension-dependent term arising from observational non-identifiability within the model class, even when the informative domains share the same regression coefficients and the sample sizes are sufficiently large. To guard against negative transfer, we develop a data-driven source detection algorithm with provable consistency in selecting informative auxiliary datasets. In addition, we provide a procedure for constructing simultaneous confidence intervals for the regression coefficients of interest. Simulation studies and a real data analysis illustrate the benefits of the proposed approach in settings with latent factors and heterogeneous sources.
翻译:暂无翻译