A1545
Title: High-dimensional multiple testing under latent confounding
Authors: Shota Katayama - Faculty of Economics, Keio University (Japan) [presenting]
Abstract: Detecting group differences based on treatments or disease status is a critical challenge in high-dimensional genetic data analysis. In observational studies where random assignment is infeasible, confounding variables can affect both the variables of interest (such as treatment variables) and the outcome variables. While observable confounders can be addressed by incorporating them into the model, latent factors such as ancestry, analytical environment, and equipment may also act as confounders. Consequently, statistical inference performed without accounting for such latent confounding variables will lead to biased conclusions. A multiple testing procedure tailored for high-dimensional settings with latent confounding is proposed. Specifically, high-dimensional treatment parameters and latent variables are first estimated using L1-norm and nuclear norm regularizations, respectively, and then test statistics are constructed based on these estimators. Finally, a multiple testing procedure that can asymptotically control the false discovery rate (FDR) is proposed.
