Cluster thresholding in multi-voxel pattern analysis (MVPA) is an open problem, as many figures of merit are possible, few (if any) of which have been sufficiently analyzed to permit a parametric solution or a guarantee of compatibility with pre-computed simulations. Stelzer, et al. 2012 represents probably the most conservative approach, constructing voxel-wise and then cluster-wise null distributions at the group level, based on permuting the training labels at the individual level.
In Mapping the cortical representation of speech sounds in a syllable repetition task, we adapted this approach to skip the voxel-wise null distribution,