American Statistical Association
The traditional voxelwise inference scheme entails computation of a test statistic at each voxel and thresholding so that a global error rate is controlled. This scheme implicitly assumes that most voxels behave according to the null distribution and that the test statistics are at most locally dependent. Under these assumptions, a histogram of the observed test statistics should have a large component that follows the null distribution. Data from neuroimaging studies, however, often contradict this, producing shifted and scaled histograms. In this talk, I review some of my own investigations into this phenomenon and describe how high correlation: 1) can affect false discovery rate inference, making it more variable and more conservative; and 2) can sometimes explain most of the histogram behavior even in the presence of no signal. In the latter situation, I show that this can be achieved by a latent factor model, of which Efron's empirical null is a special case.
|Date:||Tuesday, November 18, 2014|
NYU Deptartment of Child & Adolescent Psychiatry
“Central Park” meeting area at the Child Study Center
One Park Avenue (between 32nd and 33rd Streets)
New York, New York