Abstract
In the last few years, fMRI studies of visual object representations have begun to probe within-category exemplar distinctions. Most of them defined an exemplar discriminability index (EDI) as the difference between the average of the pattern distances between different exemplars (i.e. the effect) and the average of the pattern distances for repetitions of identical exemplars (i.e. the noise). (Note that for the popular correlation distance (1-r), the difference of distances is equal to the difference of correlations: (1-r1)-(1-r2)=r2-r1.) The EDI’s significance is then commonly assessed using a t-test (H0: group-average EDI = 0) across subjects. This approach is valid under the assumption that the EDI is 0-mean normal under H0. However, the EDI is a difference of two non-normal variables with different variance, and thus not in general 0-mean or normal. Here we address whether t-testing EDIs still provides acceptable protection against false positives in practice, and explore alternative tests. In simulations, we explore a wide parameter space for data and analysis (number of voxels, number of conditions, regional-mean activation level, pattern modulation across voxels, distance measure, pattern normalisation), and show that the EDI is usually nearly 0-mean normal under H0 (although the theoretically expected violations do occur). This means that the previous results are probably trustworthy. We describe and validate alternative tests of the EDI that use randomisation and do not require EDI normality. Moreover, these tests allow single-subject and group analysis (with subject as a fixed or random effect). Application to simulated and real fMRI data suggests that these non-parametric tests have similar power as the t-test approach. These methods may enable vision scientists to safely and sensitively detect subtle pattern differences including subordinate distinctions.
Meeting abstract presented at VSS 2012