Title: Connectivity regression
Authors: Jeffrey Morris - University of Pennsylvania (United States) [presenting]
Veerabhadran Baladandayuthapani - University of Michigan (United States)
Neel Desai - Rice University (United States)
Abstract: An important problem posed by modern big data is the regression of multivariate associations on predictors. One example in neuroimaging involves functional connectivity, discerning associations among brain regions and assessing how these vary according to discrete and continuous factors. We introduce a general connectivity regression framework that can determine which factors impact connectivity and characterize which graph edges vary by each significant factor. Our approach involves projecting the subject-specific connectivity estimates into an alternative space for which Gaussian assumptions are justified and positive definiteness in the original space in ensured, and in which we perform multivariate regression using a multivariate-spike-and-slab lasso to simultaneously perform variable selection on covariate effects on edges and detect edge-to-edge associations. This penalty increases efficiency in estimation and covariate selection through the principles of seemingly unrelated regressions, and we demonstrate small sample properties by simulation. We apply this method to data from the Human Connectome Project and discuss the generality and extendibility of the framework.