Causal Modelling and Brain Connectivity in Functional Magnetic Resonance Imaging

Neuroimaging studies that investigate the involvement of brain regions in various cognitive and perceptual tasks have become increasingly prevalent. Functional magnetic resonance imaging (fMRI) studies are especially popular, due to their non-invasive nature and high spatial resolution. With recent advances in data analysis and modelling, it is now possible to use fMRI data to ask not only which brain regions are involved in these tasks, but also how they communicate with one another; for example, one can ask, “Is attentional modulation of visually evoked responses mediated by top-down (task-driven) or bottom-up (stimulus-driven) connections in the brain?” There are two state-of-the-art approaches for understanding the communication among distributed brain systems using neuroimaging. They reflect two distinct approaches to understanding connectivity. One approach—dynamic causal modelling (DCM)—tries to model how activity in one brain area is affected by activity in another (using models of effective connectivity), while the other—Granger causal modelling (GCM)—tests for the signature of these influences by looking for correlations in the activity of two or more regions (using models of functional connectivity). Previously, the relative accuracies of these methods, in disclosing patterns of communication among brain regions, were unknown. In a recent issue of PLoS Biology, Olivier David et al. compared them directly and provided evidence that may have a profound influence on their application [1]. Here, we consider the motivation behind the two techniques, their underlying assumptions, and the implications of David et al. [1] for their continued use.