Spatial modeling of brain connectivity data via latent distance models with nodes clustering

Brain network data—measuring structural interconnections among brain regions of interest—are increasingly collected for multiple individuals. Moreover, recent analyses provide additional information on the brain regions under study. These predictors typically include the three-dimensional anatomical coordinates of the regions, and their membership to hemispheres and lobes. Although recent studies have explored the spatial effects underlying brain networks, there is still a lack of statistical analyses on the net connectivity structure which is not explained by the physical proximity of the brain regions. We answer this question via a predictor-dependent latent space model for replicated brain network data which provides a meaningful representation for the net connectivity architecture via a set of latent positions having a mixture of Gaussians prior. This model allows for flexible inference on brain network patterns which are not explained by the anatomical structure, and facilitates clustering among brain regions according to local similarities in the latent space. Our findings offer novel insights on wiring mechanisms among subsets of brain regions which interestingly departs from the anatomical proximity structure.