We propose a variational Bayesian framework for variable selection with multiple responses across heterogeneous populations. The prior dependence among the inclusion indicators allows variable-selection information to be shared at different levels across populations and responses, while regression coefficients remain population- and response-specific. When the appropriate sharing structure is unknown, a data-driven rule adaptively determines the scope of information sharing from marginal association evidence. The resulting variational algorithm is computationally scalable and can be implemented using either individual-level or summary data. Under regularity conditions, we establish selection consistency and convergence properties of the variational parameters at fixed iterations of the algorithm for fixed dimension case. In high dimensions, we establish model-selection consistency and coefficient contraction in both $\ell_1$ and $\ell_2$ norms. Simulation studies show competitive variable selection performance and demonstrate the potential gains from joint analysis. We further illustrate the summary-data implementation using genome-wide association summary statistics from the 2021 Global Lipids Genetics Consortium study for four lipid traits across five populations.

翻译：暂无翻译