Bayesian Transfer Learning for Enhanced Estimation and Inference

Music City Center 

Transfer learning enhances model performance in target populations with limited samples by leveraging related studies. While much work focuses on predictive performance, statistical inference remains challenging. Bayesian methods offer uncertainty quantification but often require single-source or individual-level data. We propose TRAnsfer leArning via guideD horseshoE prioR (TRADER), enabling multi-source transfer via pre-trained models in high-dimensional linear regression. TRADER shrinks target parameters toward a weighted average of source estimates, accounting for differing source scales. Theoretical results show TRADER achieves faster posterior contraction rates than standard priors when sources align well with the target while mitigating negative transfer from heterogeneous sources. Finite-sample analysis shows TRADER maintains frequentist coverage probabilities even for moderate signals, where standard priors falter. Numerical studies and a real-data application estimating blood glucose–insulin use associations in a Hispanic diabetic population demonstrate TRADER's superior estimation and inference accuracy over standard priors and state-of-the-art methods.

Data heterogeneity

global-local shrinkage prior

high-dimensional inference

transfer learning

sparsity 

Biopharmaceutical Section