Granger-Causality Analysis

• Granger-causality analysis is a statistical framework that uses VAR model comparisons to determine if one time series improves the prediction of another.
• Modern techniques extend classical methods by incorporating local, dynamic, and time-resolved analyses to capture evolving causal relationships in non-stationary systems.
• Advanced methods address high-dimensional and networked data through sparse estimation and state-space modeling, enhancing reliable causal inference.

Granger-causality analysis is a suite of statistical methodologies for quantifying directed predictive relationships between components of multivariate time series. Formally, in the context of a vector autoregressive (VAR) model, $X_j$ is said to "Granger-cause" $X_i$ if knowledge of the past of $X_j$ significantly improves the prediction of $X_i$'s future, over and above information in the histories of all other components. This predictive notion, first introduced by Clive Granger in the 1960s, underpins much of modern functional network inference in econometrics, neuroscience, climatology, and beyond. Recent theoretical advances extend Granger-causality to dynamic, high-dimensional, non-stationary, nonlinear, frequency-resolved, and local settings, enabling fine-grained analysis of complex, time-heterogeneous systems.

1. Formalism: VAR Models and Classical Granger-Causality

Consider an $n$-dimensional, zero-mean, discrete-time process $X_t = [X_1(t),...,X_n(t)]^\top$ modeled as a stable VAR($p$): $X_t = \sum_{k=1}^p A_k\, X_{t-k} + \epsilon_t,$ where $A_k$ are $n \times n$ coefficient matrices and $X_i$0 is zero-mean white noise with covariance $X_i$1 of full rank. Granger-causality from component $X_i$2 to $X_i$3 is quantified by comparing the predictive accuracy (specifically, the residual one-step-ahead variance) of two models for $X_i$4:

• Full model: includes the past of all variables,
• Restricted model: omits the past of $X_i$5.

The standard (averaged) Granger-causality statistic is

$X_i$6

where $X_i$7 and $X_i$8 denote the innovations under the full and restricted models, respectively. In multivariate (group-to-group or ensemble) settings, the generalized variance—i.e., the determinant of the residual covariance matrices—is compared, leading to the multivariate Granger-causality (MVGC) measure: $X_i$9 This form, originally due to Geweke, admits further refinement for assessing "partial" causality and causal density in complex systems (Barrett et al., 2010).

2. Local and Time-Resolved Granger-Causality

The classical measure $X_j$0 summarizes predictive influence over the entire sample, obscuring dynamic changes in causal strength. Local (time-resolved) Granger-causality, as developed by Stramaglia et al., defines a local increment $X_j$1 at each $X_j$2 such that

$X_j$3

where $X_j$4, $X_j$5 collects the past of $X_j$6 up to lag $X_j$7, $X_j$8 collects all other lags, and the result exploits the joint Gaussianity of the VAR process (Stramaglia et al., 2020). Algebraically, $X_j$9 is proportional to the log-ratio of conditional densities and, crucially, under Gaussianity, is exactly twice the local transfer entropy: $X_i$0 The mean over $X_i$1 recovers the standard $X_i$2. This machinery enables the tracking of instantaneous, context-dependent causal interactions, making it suitable for analyzing non-stationary and adaptive systems.

Computationally, local GC is extracted post-hoc from a fitted global VAR via matrix operations involving the covariance and concentration (precision) matrices of the stacked past state variables. No model re-estimation is required per time point (Stramaglia et al., 2020).

3. Dynamic, Local, and Adaptive Extensions

Beyond local GC, a suite of methodologies targets dynamic and locally-stationary contexts:

• Dynamic Window-level Granger Causality (DWGC): Causal inference via F-statistics on sliding windows of length $X_i$3, supplemented by learned "causality indexing" (re-weighting series by per-time weights $X_i$4) to suppress autocorrelation-driven artifacts and enhance detection of cross-channel effects. Theoretical results show DWGC with indexing strictly dominates naive windowed-GC with respect to power for any fixed $X_i$5, and DWGC degenerates to classical GC as $X_i$6 (Zhang et al., 2020).
• Spatio-Temporal Granger Causality (stGC): The framework introduced by Luo et al. partitions both time and space (e.g., windows and voxels in fMRI), fitting local VARs in each segment. Two monotonicity theorems establish that fine-scale windowing increases the cumulative and average Granger indices ($X_i$7 and $X_i$8), formalizing a resolution-dependent sensitivity principle. Optimal window division and scale trade-offs are solved by minimizing a penalized cost (prediction error plus reciprocal of GC information) subject to minimal window size, with BIC for model selection (Luo et al., 2012).
• Statistical Inference for Local Granger Causality (locally stationary VAR): For non-stationary multivariate time series, local Whittle likelihood-based estimation of time-varying spectra yields time- and frequency-specific causality measures, with asymptotic normality derived for local estimators under kernel smoothing. This enables statistically valid, fine-grained inference of evolving causal structure (Liu et al., 2021).

4. Variable-lag Causality and Generalizations

Classical Granger-causality tests restrict the predictive influence of $X_i$9 on $n$0 to a fixed lag structure. Recent advances introduce:

• Variable-lag Granger Causality (VLGC): The lag from $n$1 to $n$2 is allowed to vary arbitrarily by time, identified via dynamic time warping (DTW). The full regression is then performed with the DTW-aligned series, and model fit compared to standard fixed-lag approaches via BIC or F-statistics. VLGC can recover relations missed by fixed-lag models, especially in phenomena with leader-follower dynamics of variable delay (Amornbunchornvej et al., 2019, Amornbunchornvej et al., 2020).

This class of methods expands the scope of Granger-causal inference, maintaining formal model selection criteria and residual-based statistical decision rules.

5. High-dimensional, Network, and Model-based Settings

Scaling Granger-causality to high-dimensional time series and complex network inference necessitates robust estimation and regularization:

• Sparse and Structured VAR Estimation: Use of LASSO, group-LASSO, and sparse-group LASSO allow for controlled model selection in regimes where the number of parameters exceeds the data length. Recent theoretical work provides debiased inference and false positive control for LASSO-based GC statistics under restricted eigenvalue and deviation conditions (Das et al., 2021, Babii et al., 2019).
• Dimension Reduction: Techniques such as time-ordered variable selection, non-uniform mixed embedding, and penalized partial mutual information (PMIME, RCGCI, RGPDC) significantly improve inference and specificity in high-$n$3, moderate-$n$4 settings, especially for network or graph recovery in EEG/fMRI and financial applications (Siggiridou et al., 2019).
• State-Space Granger Causality: For processes with moving-average components, down-sampling, or observational noise, state-space models and EM algorithms with sparsity priors provide a statistically principled framework. Recovery