Source-linked AI summary
In reply to Faes et al. and Barnett et al. regarding "A study of problems encountered in Granger causality analysis from a neuroscience perspective"
Patrick A. Stokes, Patrick L. Purdon
TL;DR
The reply examines how Granger-Geweke causality is computed and how system dynamics appear in the measure, with emphasis on neuroscience interpretation. It argues that interpretation depends on model details and that directed information flow may be read as a physical mechanism in practice.
Problem
The paper addresses how Granger-Geweke causality's statistical properties and representation of system dynamics affect its interpretation in neuroscience.
Method
The authors characterize traditional Granger-Geweke causality computation and analyze how system dynamics are represented in the resulting measure.
Results
The analysis finds that effect-node dynamics are absent from Granger-Geweke causality and that interpreting it requires the selected model and its component dynamics.
Takeaways & Limitations
Meaningful neuroscience inferences require models and derived quantities to correspond appropriately to the scientific questions of interest.
Takeaways & Limitations
Granger-Geweke causality may change with a different model, and inference is worthless when the model is inadequate for the data or question.
Abstract
from arXiv · showhide
This reply is in response to commentaries by Barnett, Barrett, and Seth (arXiv:1708.08001) and Faes, Stramaglia, and Marinazzo (arXiv:1708.06990) on our paper entitled "A study of problems encountered in Granger causality analysis from a neuroscience perspective." (PNAS 114(34):7063-7072. 2017). In our paper, we analyzed several properties of Granger-Geweke causality (GGC) and discussed potential problems in neuroscience applications. We demonstrated: (i) that GGC, estimated using separate model fits, is either severely biased, particularly when the true model is known, or a high variance is introduced to overcome the bias; and (ii) that GGC does not reflect some component dynamics of the system. The commentaries by both Faes et al. and Barnett et al. point out that the computational problems of (i) are resolved by using recent computational methods. We acknowledge that these problems are indeed resolved by these methods. However, the traditional computation using separate model fits continues to be presented and applied. More fundamentally, the interpretational problems stemming from (ii) are not in anyway addressed by the improved methods because they are inherent to the definition of GGC. These properties are indeed acknowledged by both commentaries. We have no misconception of the GGC measure and do not claim that these properties are facially wrong. But we do discuss at length how these properties make it inappropriate and misleading for common types of scientific questions, how presentation of GGC results without model estimates are not decipherable, and how the absence of clear statements of questions of interest present further opportunities for misinterpretation.