The Constrained Network-Based Statistic: A New Level of Inference for Neuroimaging.

Abstract

Neuroimaging research aimed at dissecting the network organization of the brain is poised to flourish under major initiatives, but converging evidence suggests more accurate inferential procedures are needed to promote discovery. Inference is typically performed at the cluster level with a network-based statistic (NBS) that boosts power by leveraging known dependence within the local neighborhood. However, existing NBS methods overlook another important form of dependence-shared membership in large-scale brain networks. Here, we propose a new level of inference that pools information within predefined large-scale networks: the Constrained Network-Based Statistic (cNBS). We evaluated sensitivity and specificity of cNBS against existing standard NBS and threshold-free NBS by resampling task data from the largest openly available fMRI database: the Human Connectome Project. cNBS was most sensitive to effect sizes below medium, which accounts for the majority of ground truth effects. In contrast, threshold-free NBS was most sensitive to higher effect sizes. Ground truth maps showed grouping of effects within large-scale networks, supporting the relevance of cNBS. All methods controlled FWER as intended. In summary, cNBS is a promising new level of inference for promoting more valid inference, a critical step towards more reproducible discovery in neuroscience.