On the interpretability and computational reliability of frequency-domain Granger causality

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Granger causality (GC) analysis faces interpretation challenges in neuroscience, partly due to its name.
• A recent study highlighted issues with GC, using an outdated formulation.
• This formulation's limitations may lead to unwarranted dismissal of GC's utility.

Purpose of the Study:

• To comment on and address limitations in Granger causality (GC) analysis as presented in Stokes and Purdon (2017).
• To demonstrate how updated GC methods, particularly frequency-domain approaches, can resolve identified pitfalls.
• To provide a balanced perspective on the applicability of GC in scientific research, including neuroscience.

Main Methods:

• Replication of simulations from Stokes and Purdon (2017).
• Implementation of an updated Granger causality (GC) methodology.
• Utilization of spectral and causal information in the updated GC analysis.

Main Results:

• The updated GC implementation, incorporating spectral and causal information, mitigates or resolves the pitfalls identified in the original study.
• Simulations demonstrate that advanced frequency-domain GC methods are less susceptible to the issues highlighted by Stokes and Purdon.
• The findings suggest that previous criticisms of GC may stem from the use of suboptimal or outdated methodologies.

Conclusions:

• The interpretation issues associated with Granger causality (GC) can be addressed through modern, frequency-domain implementations.
• Dismissing GC based on outdated formulations is premature, given recent advancements in methodology and interpretation.
• Updated GC approaches offer a more robust and reliable tool for causal inference in complex systems, including neuroscience.