L_0 Shrinkage for Association Analysis Between Multivariate Imaging and Multivariate Genetics Data

• Activity Number: 121 - Emerging Statistical Methods for Structured and Multimodal Data Analysis
• Type: Topic Contributed
• Date/Time: Monday, August 3, 2020 : 1:00 PM to 2:50 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Shuo Chen*
• Companies: University of Maryland, School of Medicine
• Keywords: l_0 shrinkage; imaging; genetics; graph combinatorics

Abstract

We propose a new statistical framework to extract the associations between multivariate imaging and multivariate genetics data while accounting for the dependency between both imaging and genetics data. In practice, there are one billion possible connections between moderate-sized 10,000 imaging features and 100,000 genetic features. We develop novel algorithms and theories to implement $l_0$ norm regularization to on those potential connections which can lead to highly correlated imaging feature clusters and genetic feature clusters. The $l_0$ norm regularization can accurately extract the organized patterns of associations between two set of multivariate variables. We provide theoretical results for the statistical inference based on graph combinatorics. We apply our method to large sample imaging-genetics studies and perform extensive simulation studies. The results demonstrate the proposed method outperforms existing multivariate statistical methods by simultaneously improving false positive and false negative discovery rates and significantly increase replicability.