|
Activity Number:
|
274
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Tuesday, August 8, 2006 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #306824 |
|
Title:
|
An Algorithm for Regression of Scalars on Images
|
|
Author(s):
|
Philip Reiss*+ and R. Todd Ogden
|
|
Companies:
|
Columbia University and Columbia University
|
|
Address:
|
Department of Biostatistics, New York, NY, 10032,
|
|
Keywords:
|
brain imaging ; b-splines ; dimension reduction ; functional data analysis ; image regression ; partial least squares
|
|
Abstract:
|
When neuroimaging techniques are used to measure a quantity at each voxel (volume unit) of the brain for each of a set of subjects, it is common practice to relate this quantity to a scalar variable of interest through a "mass-univariate" approach: Regress the imaged quantity on the scalar variable separately at each voxel and identify the voxels at which a significant effect is observed. This paper adopts the opposite approach; we treat the scalar as the outcome and regress it on the imaged quantity at all voxels simultaneously to produce a coefficient image. As this single model typically has many more predictors than cases, one must reduce the dimension of the predictors drastically. We present an algorithm combining three dimension-reduction techniques: projection onto a spline basis, selection of the most relevant predictor components, and thresholding.
|
