The concept of human brain "connectivity" is an elusive one. On the surface, the idea is straightfoward: a property describing how distal parts of the brain are physically and functionally connected, such that activity in one can drive or influence activity in the other. However, ascertaining this property in the massively connected and largely inaccessible living human brain has proven a daunting task. The best we can usually do is to use techniques which measure the electromagnetic fields produced by brain activity or morphology, and infer connectivity from these. However, while these methods provide much insight into the structure and function of the human brain, they fall short of supporting direct inferences about physical connectivity. As a result, neuroscience has produced numerous metrics bearing the title of "connectivity", but which actually only furnish an indirect estimate of this property. Moreover, due to the complexity of its networks and the activity which propagates through them, it is next to impossible to definitively separate direct from indirect connectivity; a problem which is particularly clear in correlative approaches.
In our recently accepted article in Brain Structure & Function, we compare four different commonly employed correlative estimates of brain connectivity. Two of these approaches use structural covariance to infer connectivity structure, in which covariance in morphometric estimates (cortical thickness or voxel-based morphometry; VBM) across the brain is assumed to capture mutually trophic influences which reflect connectivity patterns. The other two approaches infer functional connectivity from covariance in resting-state BOLD time series, or activation peaks across many task-based fMRI studies (meta-analytic connectivity mapping; MACM). To avoid using arbitrarily defined thresholds in these comparisons, we propose metrics which integrate across a broad range of possible densities, and normalize these to values that would be expected by random chance. We found that the two functional methods had a good agreement, but the two structural ones did not. Indeed, comparisons between structural covariance approaches were no better than those between structural and functional ones. Interhemispheric symmetry and distance from the seed region (or its homotope) were important factors in the degree to which connectivity estimates agreed or disagreed. We hope these findings, and the methodological framework we present, will be useful for future attempts to utilize large multimodal datasets to investigate human brain connectivity from multiple lines of evidence.
