A1348
Title: Bayesian sparse-group SLOPE for controlling the false discovery rate
Authors: Marina Evangelou - Imperial College London (United Kingdom) [presenting]
Fabio Feser - Imperial College London (United Kingdom)
Abstract: Selecting the right variables for controlling the false discovery rate (FDR) is a challenge that remains an open problem. Recent work has proposed a new penalised regression approach that incorporates available group structure, named sparse-group SLOPE (SGS). SGS incorporates both the SLOPE and group-SLOPE penalties for selecting both variables and groups, and with appropriate penalty sequences SGS controls the FDR at both levels. Selecting the tuning parameter for SGS and other SLOPE-based models remains an outstanding challenge for non-orthogonal data where the aim is FDR control. A Bayesian sparse-group SLOPE (BSGS) model is presented that is built by assuming a continuous spike-and-slab prior to the model coefficients. The BSGS model inherits many of the advantages of Bayesian formulations, including uncertainty quantification and adaptivity to the underlying sparsity structure. Through extensive analysis of both synthetic and real datasets, the proposed model is shown to consistently control FDR while achieving high power. In comparison with other major model selection strategies, it is found to outperform them.
