Statistics Seminar: "Total variation smoothed regression for spatial and temporal correlated data"

Speaker: Haochang Shou, University of Pennsylvania

Abstract: The increasing availability of the high-throughput medical data has provided tremendous opportunity for researchers to search for biomarkers that link with pathology and help to improve diagnosis. Appropriate statistical methods that acknowledge the multidimensional structures as well as the correlations induced by temporal and spatial continuity are needed. Motivated by 1D time-varying physical activity tracking data continuously assessed by wearable computing sensors and 3D structural magnetic resonance imaging (MRI) data, we proposed to use a total variation (TV) regularized image-on-scalar regression method that acknowledge the dependency structures while adjusting for confounding by covariates. The estimator is the solution of a penalized regression problem where the objective is the sum of square error plus a TV regularization on the predicted mean across all subjects. We developed a scalable algorithm via alternative direction of methods of multiplier (ADMM). The method has been applied to understanding patterns of activity intensities from NIMH family study of spectrum disorders and the gray matter voxel-based morphometry maps from the attention deficient/hyperactive deficient (ADHD) 200 consortium.

Host: Todd Kuffner