Brain imaging datasets are difficult to analyze because they are very large and rectangular (i.e. the number of voxels is much larger than the number of images). The traditional approach to this problem computes one parametric statistic per voxel. Because it assumes voxel independence, this approach requires a drastic correction for multiple comparisons. An alternative to the voxel approach is pattern-based analysis, which minimizes the number of comparisons and takes advantage of the dependence between voxels. In one example of this approach, OToole et. al. (2005) re-analyzing data from Haxby et. al. (2001) used pattern-based classification to determine the functional distance between visual categories of objects. The data were fMRI scans from participants who viewed pictures from eight categories. The scans were processed by a classifier that predicted the category of the viewed pictures. The performance of the classifier was evaluated with a between category distance matrix (d matrix). Unfortunately, this d matrix has no associated measure of reliability, and therefore could not be used to test statistical hypotheses. Here, we use bootstrap resampling to create a non-parametric estimate of reliability for d matrices. To analyze a d matrix, we create many estimates of the d matrix using a bootstrap resampling protocol and then compare these estimates using a variant of multidimensional scaling (called DISTATIS). We then transform these d matrices into a map that displays the categories as confidence ellipses, with the best estimate of each category as the center of its ellipse. |