New and old methods for thresholding statistic images in functional
neuroimaging

Thomas E Nichols
Assistant Professor of Biostatistics
University of Michigan


Central to any functional neuroimaging experiment is the
identification of significant voxels while controlling the chance of
false positives.  We discuss two approaches to improve sensitivity
while still controlling the risk of false positives.

The standard approach is to control the familywise error (FWE), the
chance of any false positives.  In addition to Bonferroni, random
field methods are used to find thresholds which control FWE.  For low
degrees-of-freedom group studies, we present simulated and real data
that show the random field thresholds to be quite conservative, even
when the data are quite smooth.  We find that this conservativeness is
easily overcome with the permutation test, one of the oldest
statistical methods.

A new statistical method is the use of the false discovery rate (FDR)
to measure false positives.  FDR is the proportion of false positives
among suprathreshold voxels.  FDR is a more lenient measure of false
positives, and hence is more powerful.  Further it is easy to apply
and only requires weak assumptions.  We discuss some of the
limitations of FDR and illustrate the method with real data.