The impact of heterogeneous spatial autocorrelation on comparisons of brain maps

Abstract

It is increasingly common to statistically compare brain maps to assess how spatially similar they are. However, statistical inference can be challenging due to the presence of spatial autocorrelation. Therefore, random permutation approaches based on null models are widely used to address this concern. Here, we show how that the presence of heterogeneity in the spatial autocorrelation across brain maps impacts the validity of statistical inference in common approaches for spatially correlated maps. Furthermore, we illustrate how a Bayesian spatial regression approach can be applied to compare functional and structural cortical brain maps, yielding valid statistical inferences even in the presence of heterogeneity. Explicitly modelling spatial properties provides more valid inferences about whole brain spatial maps allowing a wider and more sophisticated range of neurobiological questions to be answered about the relationship between brain maps than are possible with current approaches.