B1013
Title: Interpretable scalar-on-covariance regression with applications to functional connectivity
Authors: Xiaomeng Ju - New York University (United States) [presenting]
Hyung Park - New York University (United States)
Thaddeus Tarpey - New York University (United States)
Abstract: A novel regression methodology is introduced utilizing covariance matrices to predict scalar outcomes. The motivation is the need to reduce dimensionality in order to improve the predictive performance when modeling scalar outcomes using high-dimensional subject-specific covariance matrices obtained from functional magnetic resonance imaging (fMRI) signals. Dimension reduction is achieved by projecting signals onto a data-driven orthonormal basis. The proposal yields a parsimonious regression model that offers meaningful interpretations of the projection matrix and regression coefficients, which are jointly estimated within a Bayesian framework. To enable variable selection, the incorporation of sparsity is investigated into the orthonormal projection matrix. The performance of the method is evaluated by comparisons to existing alternatives in various simulation settings and is illustrated through a case study.
