Testing Granger Non-Causality in Panels with Cross-Sectional Dependencies

Lenon Minorics, Ali Caner Türkmen, David Kernert, Patrick Blöbaum, Laurent Callot, Dominik JanzingAISTATS 2022 · OpenReview

Introduction

In time-series analysis, Granger causality is a fundamental tool for determining if one variable's history can predict another's future. However, applying this to panel data—datasets containing multiple entities over time—presents a significant hurdle: cross-sectional dependency. Traditional methods, such as the Dumitrescu-Hurlin (DH) test, assume that entities within a panel are independent. In reality, global systems like financial markets or pandemic spreads are deeply interconnected. This paper introduces Quantile p-value Panel Adjustment (QPPA), a robust framework designed to maintain statistical integrity even when panel members are correlated.

The Challenge of Cross-Sectional Dependency

Standard panel causality tests often aggregate individual test statistics. This approach relies heavily on the assumption of cross-sectional independence. When entities are dependent—due to shared external shocks or direct interactions—these aggregated statistics can lead to a high False Discovery Rate (FDR), mistaking correlation for causation.

Visualizing different dependency structures in panel data, from independent members (left) to complex cross-sectional dependencies (right).

As shown in the figure above, real-world data often exhibits complex inter-member dependencies. If a test assumes the left-most structure but the data follows the right-most, the resulting "causal" findings are often spurious.

The QPPA Methodology

QPPA shifts the focus from aggregating statistics to aggregating p-values. The core innovation lies in using a quantile-based aggregation technique borrowed from high-dimensional regression. The process follows three steps:

  1. Individual Testing: Conduct a standard Granger causality test for every member $i$ in the panel to generate a set of p-values ($p_1, p_2, \dots, p_N$).
  2. Quantile Aggregation: For a chosen threshold $\gamma$, the aggregated p-value is calculated as a scaled version of the $\gamma$-quantile of the individual p-values.
  3. Theoretical Robustness: The authors provide a mathematical proof that this method approximately bounds the Type I error by the chosen significance level, regardless of the underlying dependency structure between panel members.

Experimental Evaluation

The researchers evaluated QPPA against the DH test and its block-bootstrap variant using synthetic AR-1 processes. The results highlight a critical failure in traditional methods: under cross-sectional dependency, the DH test's FDR surged to over 40%, while QPPA maintained an FDR near zero.

Performance comparison of QPPA vs. DH tests under cross-sectional dependencies. Note the high FDR for DH-test compared to QPPA's stability.

As the table demonstrates, while QPPA can be more conservative in small samples ($T=10$), it achieves perfect power ($1.0$) with zero FDR as the time horizon ($T$) and panel size ($N$) increase, outperforming state-of-the-art alternatives in reliability.

Real-World Application: COVID-19

The utility of QPPA was demonstrated on COVID-19 data (confirmed cases vs. deaths) across 225 regions. While the DH test erroneously suggested that deaths "cause" confirmed cases—a biological impossibility in this context—QPPA correctly identified only the unidirectional causal link from cases to deaths. This "COVID Paradox" highlights how cross-sectional dependencies (e.g., global travel patterns) can mislead traditional tests into finding "wrong-way" causality. To ensure stationarity, the researchers analyzed the "differenced" data (daily changes) rather than cumulative totals.

Limitations and Future Work

While QPPA is a significant step forward, it is not without limitations. Like most Granger-based tests, it requires data stationarity and assumes "causal sufficiency"—the absence of hidden confounders. Furthermore, the current implementation assumes linear relationships. Future research may extend this quantile-based aggregation to non-linear causal discovery methods, such as those utilizing neural networks.

Conclusion

QPPA provides a necessary safeguard for causal inference in interconnected systems. By leveraging p-value aggregation rather than test statistics, it offers a distribution-free solution to the problem of cross-sectional dependency, ensuring that discovered causal relationships in panel data are statistically sound and practically meaningful.

Summary updated

Abstract

This paper proposes a new approach for testing Granger non-causality on panel data. Instead of aggregating panel member statistics, we aggregate their corresponding p-values and show that the resulting p-value approximately bounds the type I error by the chosen significance level even if the panel members are dependent. We compare our approach against the most widely used Granger causality algorithm on panel data and show that our approach yields lower FDR at the same power for large sample sizes and panels with cross sectional dependencies. Finally, we examine COVID-19 data about confirmed cases and deaths measured in countries/regions worldwide and show that our approach is able to discover the true causal relation between confirmed cases and deaths while state-of-the-art approaches fail.
BibTeX
@inproceedings{minorics2022testing,
  author = {Lenon Minorics and Ali Caner Türkmen and David Kernert and Patrick Blöbaum and Laurent Callot and Dominik Janzing},
  title = {{Testing Granger Non-Causality in Panels with Cross-Sectional Dependencies}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2022},
  url = {https://openreview.net/forum?id=yE3xd6AX9I6},
}
Figure