Technical Summary: Nonlinear Permuted Granger Causality (NPGC)

Problem Statement

Granger causal inference is a widely used method for determining if one time series predicts another, traditionally defined by the reduction in prediction error variance when including a covariate set. However, adapting this framework to nonlinear data presents significant challenges. Existing nonlinear adaptations often rely on deep learning and representation learning but suffer from two primary issues:

1. Overfitting and In-Sample Bias: Many methods utilize in-sample tests (comparing restricted and unrestricted models) which, when paired with universal approximators like Artificial Neural Networks (ANNs), can lead to spurious correlations. ANNs can approximate arbitrary, data-specific dependencies even when no true predictive relationship exists, making in-sample error reduction an unreliable metric for causality.
2. Individual vs. Collective Inference: Many approaches focus on isolating individual covariates through penalized variable selection (e.g., Lasso). This ignores the collective behavior of the covariate set, introduces multiplicity problems, and often fails to provide a global decision on whether a set of variables is Granger causal. Furthermore, the reliance on specific penalty parameters (e.g., $\lambda$) makes results sensitive to hyperparameter tuning.

The paper argues that the definition of Granger causality should shift from comparing model residuals to explicitly testing out-of-sample predictability while maintaining the collective nature of the covariate set.

Methodology: Nonlinear Permuted Granger Causality (NPGC)

The authors propose Nonlinear Permuted Granger Causality (NPGC), a framework that redefines Granger causality using a permutation structure and out-of-sample testing.

Core Definitions

The method replaces the traditional comparison between restricted and unrestricted models with a comparison between a null model (using permuted covariates) and the original model.

• Permutation: The rows of the covariate matrix $X_{<t}$ are randomly permuted to create $\tilde{X}_{<t}$. This breaks the temporal dependence between the covariates and the response $Y$ while preserving the internal dependence structure (intradependence) of the covariates.
• Hypothesis: $X_t$ is conditionally Granger causal for $Y_t$ if the variance of the prediction error using the original data is significantly lower than the variance using the permuted data:
  $$\text{Var}[Y_t - P(Y_t | Y_{<t}, Z_{<t}, X_{<t})] < \text{Var}[Y_t - P(Y_t | Y_{<t}, Z_{<t}, \tilde{X}_{<t})]$$

Implementation Steps

1. Featurization: The data is transformed into a high-dimensional feature space using randomly generated Feed-Forward Neural Networks (FNNs). Unlike standard training, the network weights are fixed and randomly generated (acting as a random feature map) to approximate nonlinear relationships without overfitting to specific weights.
2. Out-of-Sample Estimation: The data is split into $K$ folds. For each permutation $m$ (where $m=1$ is the original data and $m>1$ are permuted versions) and each random featurization $r$, the model is trained on $K-1$ folds and tested on the remaining fold.
3. Variance Estimation: The out-of-sample prediction residuals are aggregated to estimate the variation parameter $\hat{\vartheta}_m$ for each permutation.
4. Decision Rule: A null distribution is constructed from the $M$ permutations. The p-value (quantile estimate $\hat{Q}_M$) is calculated based on where the original data's error variance ($\hat{\vartheta}_1$) falls relative to the permuted errors. If $\hat{Q}_M \leq \alpha$, the null hypothesis of no causality is rejected.

Theoretical Guarantees

The paper establishes asymptotic properties under three mild conditions:

1. Data Properties: The data is continuous, stationary, and sampled on a sufficiently fine grid.
2. Featurization Regularity: The random FNNs produce consistent estimates of the underlying covariance matrix with finite variance.
3. Error Independence: Model errors are independent across time points given the relevant history.

Under these conditions, the estimator $\hat{\vartheta}_m$ is consistent, and the test statistic converges to a uniform distribution under the null hypothesis.

Simulation Results

The authors evaluated NPGC against several competitors, including penalized cMLP methods (Group Lasso, Sparse Group Lasso, Hierarchical Lasso), a naive Restricted/Unrestricted (R/U) model comparison, and a Gaussian Noise Substitution (GNS) method.

• Performance: NPGC demonstrated high power in detecting Granger causal relationships in both Lorenz-96 (chaotic) and Threshold Autoregressive (TAR) models across varying time series lengths ($T=250, 500, 1000$).
• Comparison:
  • Penalized Methods (cMLP): Results were highly sensitive to the regularization parameter $\lambda$. Without a clear global decision rule, these methods produced inconsistent results.
  • Naive R/U: Failed to identify causal pairs in the Lorenz-96 trials, likely due to overfitting.
  • Gaussian Noise Substitution: Correctly identified causal pairs but suffered from uncontrolled Type I error rates (false positives).
  • NPGC: Successfully controlled Type I error rates close to the nominal level while maintaining high detection power, effectively distinguishing true functional connectivity from noise.

Application Study

The method was applied to neuronal responses in the primary auditory cortex of an anesthetized rat in response to acoustic stimuli.

• Setup: The response variable was a whole-cell recording. The covariate set included a "jaguar mating call" sound fragment (known to be causal based on experimental design) and two non-causal fragments (humpback whale and Knudsen's frog).
• Outcome: NPGC correctly identified the jaguar call as Granger causal ($\hat{Q}_M = 0.0025$) and correctly classified the whale and frog sounds as non-causal. In contrast, the penalized cMLP methods failed to isolate the correct causal stimulus, often flagging non-causal sounds or failing to detect the causal one depending on the penalty parameter.

Key Contributions and Significance

The paper claims the following contributions:

1. Redefinition of Granger Causality: It explicitly redefines Granger causality for nonlinear data using a permutation framework that prioritizes out-of-sample predictability over in-sample fit.
2. Collective Inference: The framework addresses the "collective inference" principle, testing the covariate set as a whole rather than relying on individual variable selection, thereby avoiding multiplicity issues and hidden assumptions about specific model coefficients.
3. Robustness to Overfitting: By using random feature maps and permutation testing, the method mitigates the overfitting risks inherent in deep learning models, providing a safeguard against identifying data-specific noise as functional dependence.
4. Elimination of Hyperparameter Tuning: The approach removes the burden of selecting regularization parameters (like $\lambda$ in Lasso), which often dictates the inferential outcome in penalized methods.
5. Theoretical Foundation: The paper provides a theoretical analysis of the variance estimates and decision rules, showing consistency and asymptotic uniformity under the null hypothesis.

The authors conclude that while Granger causality cannot establish true mechanistic causality, the NPGC framework offers a targeted, robust tool for identifying predictive relationships and functional connectivity in complex, nonlinear systems where prior knowledge of causal mechanisms is limited.