1: \begin{abstract}

2: \noindent

3: Motivated by recent work on studying massive  imaging data in  various

4: neuroimaging  studies,  we propose  a novel spatially varying coefficient

5: model (SVCM)  to capture the  varying association between

6:   imaging measures  in a three-dimensional  (3D)

7: volume (or 2D surface) with a set of covariates.

8: Two stylized features of  neuorimaging data are the presence of

9: multiple piecewise  smooth  regions   with unknown edges and jumps and substantial spatial correlations.

10:  To specifically account for these two features, SVCM includes

11:   a measurement model with multiple varying coefficient functions,

12:     a jumping surface  model    for each varying coefficient function, and  a functional principal component model.

13: We develop a three-stage estimation procedure to simultaneously estimate the varying coefficient functions and  the spatial correlations.

14: The estimation procedure includes a  fast multiscale adaptive

15: estimation and testing procedure

16:  to  independently estimate each   varying  coefficient function, while preserving its edges among different piecewise-smooth regions.

17:   We   systematically  investigate the asymptotic properties (e.g.,   consistency and asymptotic normality) of the  multiscale adaptive parameter

18: estimates.    We also establish the uniform

19: convergence rate of the estimated spatial covariance function

20: and its associated eigenvalues and eigenfunctions.     Our Monte Carlo simulation and real data analysis have confirmed

21:   the excellent performance of SVCM.

22: \end{abstract}

23: