Technological and data sharing advances have led to a proliferation of high-resolution structural and functional maps of the brain. Modern neuroimaging research increasingly depends on identifying correspondences between the topographies of these maps; however, most standard methods for statistical inference fail to account for their spatial properties. Recently, multiple methods have been developed to generate null distributions that preserve the spatial autocorrelation of brain maps and yield more accurate statistical estimates. Here, we comprehensively assess the performance of ten such published null frameworks in controlling the family-wise error rate in statistical analyses of parcellated neuroimaging data. We apply each framework on two prototypical analyses: (1) testing the correspondence between brain maps (e.g., correlating two activation maps) and (2) testing the spatial distribution of a feature within a partition (e.g., quantifying the specificity of an activation map within an intrinsic functional network). In agreement with previous reports, we find that naive null models that do not preserve spatial autocorrelation consistently yield unrealistically liberal statistical estimates. Performance of spatially-constrained null models depended on research context; model performance was generally consistent when testing correspondences between brain maps, but considerably more variable when testing partition specificity. Throughout these analyses, we observe minimal impact of parcellation and parcel resolution on null model performance. Altogether, our results highlight the need for continued development and standardization of statistically-rigorous methods for comparing brain maps.
bioRxiv Subject Collection: Neuroscience