A key challenge in meta-analysis is that populations across studies differ from target populations in unpredictable ways. We introduce MetaHunt, which leverages shared low-rank structures to predict function-valued quantities using only study-level information rather than individual data. The methodology extends the Successive Projection Algorithm to functional settings with a denoised basis-hunting component, achieving consistency under mild conditions and enabling flexible modeling of relationships between study covariates and mixing weights. A key advantage is its privacy-preserving nature—analysts need only study-level estimates, not raw data. The approach includes conformal prediction intervals for uncertainty quantification, demonstrated to achieve asymptotically valid coverage under exchangeability assumptions. Effectiveness is validated through simulations and real-world applications.

MetaHunt: Privacy-preserving Meta-analysis through Low-Rank Basis Hunting — GitHub