Quantifying biases in reconstructed brain networks

Abstract Brain network reconstruction from neuroimaging data is subject to systematic biases that arise from the statistical approaches used to handle variability across subjects and from the arbitrariness introduced by thresholding procedures that generate connectivity matrices with varying density, affecting the evaluation of graph metrics. However, how these biases influence the estimation of network properties remains unclear. Here, we quantify how different types of bias impact network measures by analyzing their dependence on network density. For analytically tractable cases, we estimate the dependence of topological metrics on network density and then extend the analysis to real and synthetic networks preserving brain-like features. Our analytical results show that global clustering coefficient estimates depend on network density, regardless of the statistical approach. Consistent trends are observed across multiple scales in multimodal neuroimaging data and synthetic ensembles. Our findings set limits on threshold-based connectome analyses and call for caution in interpreting results.