Multivariate Granger Causality and Generalized Variance

Abstract: Granger causality analysis is a popular method for inference on directed interactions in complex systems of many variables. A shortcoming of the standard framework for Granger causality is that it only allows for examination of interactions between single (univariate) variables within a system, perhaps conditioned on other variables. However, interactions do not necessarily take place between single variables, but may occur among groups, or "ensembles", of variables. In this study we establish a principled framework for Granger causality in the context of causal interactions among two or more multivariate sets of variables. Building on Geweke's seminal 1982 work, we offer new justifications for one particular form of multivariate Granger causality based on the generalized variances of residual errors. Taken together, our results support a comprehensive and theoretically consistent extension of Granger causality to the multivariate case. Treated individually, they highlight several specific advantages of the generalized variance measure, which we illustrate using applications in neuroscience as an example. We further show how the measure can be used to define "partial" Granger causality in the multivariate context and we also motivate reformulations of "causal density" and "Granger autonomy". Our results are directly applicable to experimental data and promise to reveal new types of functional relations in complex systems, neural and otherwise.

• The paper extends univariate Granger causality to analyze multivariate interactions using generalized variance as a principled residual error measure.
• It demonstrates numerical stability, invariance under nonsingular transformations, and equivalence to transfer entropy under Gaussian assumptions.
• The framework supports spectral decomposition and offers actionable insights for analyzing complex systems in neuroscience, economics, and biological networks.

Multivariate Granger Causality and Generalized Variance: A Summary and Analysis

The paper, "Multivariate Granger Causality and Generalized Variance," presents a comprehensive framework for extending Granger causality (G-causality) to multivariate sets of variables. Granger causality, a cornerstone method in econometrics, is instrumental in inferring causal interactions from time series data. The traditional approach, however, focuses on univariate time series and does not directly handle interactions among groups of variables—a limitation that this paper addresses.

The Framework and Its Justifications

The study establishes a principled framework for multivariate G-causality by utilizing generalized variance as a measure of residual error. This development builds on the foundational work of Geweke (1982), emphasizing the determinant of the covariance matrix of residuals. The use of the generalized variance provides several advantages:

1. Invariance Under Transformation: Multivariate G-causality maintains invariance under any nonsingular linear transformation, a property not shared by alternative methods such as those based on total variance (trace of the covariance matrix).
2. Equivalence with Transfer Entropy: Under Gaussian assumptions, the proposed multivariate measure equates to transfer entropy, an information-theoretic measure of causality. This equivalence suggests that linear models suffice in capturing all relevant causal information in Gaussian settings.
3. Numerical Stability: Contrary to concerns about numerical instability due to high dimensionality, the determinant-based measure is shown to be numerically stable, akin to methods focusing on total variance.
4. Spectral Decomposition: The measure supports a spectral decomposition, facilitating analysis in the frequency domain. This feature is crucial for understanding interactions within systems exhibiting oscillatory or frequency-specific dynamics.

Practical and Theoretical Implications

The implications of this study are far-reaching. Practically, multivariate G-causality allows for more nuanced insights into the complex systems prevalent in neuroscience, economics, and biological networks, where interactions often involve groups rather than individual variables. In neuroscience, for instance, regions of interest (ROIs) in fMRI data can be more appropriately represented as multivariate ensembles, capturing interactions that are lost in univariate analyses.

Theoretically, the extension to multivariate frameworks aids in identifying macroscopic variables and in decomposing complex networks into functionally relevant ensembles. The generalized measure's robustness to measurement error and invariance to linear transformations make it particularly suitable for empirical data often plagued by such challenges.

Future Directions

Future research could explore multivariate G-causality's application to non-Gaussian data, potentially augmenting the linear framework with nonlinear dynamics to capture subtler interaction patterns. Moreover, integrating this approach with machine learning could enhance causal inference in high-dimensional datasets.

In sum, this paper extends G-causality to incorporate multivariate interactions using a theoretically sound and empirically robust measure. By advocating for the use of generalized variance, it addresses the limitations of previous methods and opens new avenues for analyzing complex systems. This study represents a step forward in the methodological toolkit available for causal inference, particularly in fields where understanding multivariate interactions is crucial.