Abstract
Understanding how neural populations interact is crucial to understanding brain function. Most common approaches to infer neural interactions are based on Granger causality (GC) analyses and effective connectivity (EC) models of neural time series. However, an in-depth investigation of the similarity and complementarity of these approaches is currently lacking. Here, we show that in the context of functional magnetic resonance imaging (fMRI) data analysis and first-order autoregressive models, GC and EC share common assumptions and are mathematically related. By defining a "corrected" version of GC accounting for unequal noise variances, we demonstrate that the two measures are linked by an approximately quadratic relation. While GC quantifies the amount of predictability (information flow), EC quantifies the amount and sign of the interaction. Our study provides a common mathematical framework to make grounded methodological choices in the reconstruction of directed brain networks.
Introduction
A fundamental goal in neuroscience is to map the "directed" interactions between brain regions—understanding not just that two areas are connected, but which one is driving the other. In the field of functional MRI (fMRI), two dominant frameworks have emerged for this purpose: Granger Causality (GC) and Effective Connectivity (EC).
Granger Causality is a statistical approach based on predictability: if the past values of Region A significantly improve the prediction of Region B's future, A is said to "Granger-cause" B. Effective Connectivity, conversely, is a model-based approach that seeks to estimate the physical parameters (the "coupling strengths") of the underlying dynamical system that generated the observed data. Historically, these have been viewed as complementary but distinct. This research provides a mathematical bridge between them, demonstrating that in the context of fMRI, they are essentially two perspectives on the same underlying phenomenon.
The Mathematical Bridge: Corrected GC and EC
The researchers analyzed these methods using first-order Multivariate Autoregressive (MAR) models—the foundation of GC—and Multivariate Ornstein-Uhlenbeck (MOU) processes, which are standard generative models for EC.
A significant hurdle in comparing these methods is that standard GC is highly sensitive to noise imbalances. If Region A has a higher noise variance than Region B, the standard GC estimate will be biased. To resolve this, the authors introduced "corrected GC" ($cG$). By accounting for local noise variances at the source and target nodes, the relationship between the statistical measure (GC) and the physical model (EC) becomes clear.
The central discovery is that $GC$ is approximately proportional to the square of $EC$ ($GC \approx EC^2$). This quadratic relationship implies that while both measures capture the strength of an interaction, GC loses the "sign" of the connection. Whether an interaction is excitatory (positive EC) or inhibitory (negative EC), it results in a positive GC value.
Empirical Validation and Group Stability
The theoretical relationship was tested using both synthetic simulations and real-world data from the Human Connectome Project (HCP).
- Simulations: In idealized scenarios with infinite data, the quadratic relationship between corrected GC and EC achieved a near-perfect fit ($R^2 \approx 1.0$).
- Human Connectome Project: Using resting-state fMRI data from 100 subjects, the authors found a robust quadratic relationship ($R^2 = 0.72$) when looking at group-averaged data.
A critical finding for practitioners is the issue of reliability. At the single-subject or single-connection level, directed connectivity estimates are often noisy and unreliable. However, the study demonstrates that these measures become remarkably stable at the group level. Specifically, when averaging over 20 or more subjects, the predicted GC-EC relationship emerges clearly, and more than 80% of significant connections become replicable across different sessions.
Sampling Rates and Instantaneous Causality
The study also highlights how the sampling rate (the "TR" in fMRI) affects these measures. In fMRI, where sampling is relatively slow compared to neural activity, much of the interaction between regions appears as "instantaneous causality" (IC)—correlation that happens within the same time bin—rather than "lagged" causality. The authors show that the balance between GC and IC shifts predictably based on how fast the data is collected.
Conclusion and Future Directions
This work provides a unified framework for understanding directed interactions in the brain. It shows that GC and EC are not competing ideologies but are mathematically linked, with EC providing the additional benefit of distinguishing between excitation and inhibition.
However, several challenges remain. The current model assumes a linear relationship and does not fully account for the complexities of the Hemodynamic Response Function (HRF), which translates neural activity into the BOLD signal measured by fMRI. Because the HRF varies across brain regions, it can introduce timing delays that confound causal inferences. Future research integrating these hemodynamic factors with the GC-EC mathematical bridge will be essential for refining our map of the brain's functional architecture.
Summary updated