Title: A Bayesian credible subgroups approach to statistical inference in fMRI
Authors: Mark Fiecas - University of Minnesota (United States) [presenting]
Abstract: The multiple testing problem in fMRI will be discussed. Multiple comparisons due to the hundreds of thousands of voxels in the data inflate the family-wise error rate, and the strong correlation between voxels make exacerbate the problem by making naive multiplicity corrections too conservative. We extend the idea of subgroup identification developed for clinical trials to neuroimaging studies, where we translate the problem to finding the voxels whose time course covaries with the stimulus presentation in the experiment. We develop a Bayesian credible subgroups approach using hierarchical linear models that constructs the subgroups of voxels using information about model parameters from the hierarchical model. These subgroups correspond to one of the following possibilities: i) the voxels with strong evidence of association with the stimulus, ii) the voxels with no evidence of association with the stimulus, and iii) the voxels for which there is insufficient evidence of association with the stimulus. These subgroups allow for one to draw conclusions on the effect of the stimulus on each voxel, and when to defer classification to further collection of data. Finally, our approach fully accounts for the multiplicity of the problem.