Sampling distribution for single-regression Granger causality estimators

TL;DR

This paper characterizes the sampling distribution of single-regression Granger causality estimators under the null hypothesis, providing explicit distribution approximations and a valid hypothesis test for practical scenarios.

Contribution

It introduces the first distributional analysis of single-regression Granger causality estimators under the null, including explicit formulas and a Neyman-Pearson test.

Findings

Single-regression Granger causality estimator converges to a generalized chi-squared distribution under the null.

The distribution can be well approximated by a Gamma distribution.

The proposed test is asymptotically valid and applicable even with unknown or infinite model order.

Abstract

We show for the first time that, under the null hypothesis of vanishing Granger causality, the single-regression Granger-Geweke estimator converges to a generalised $\chi^2$ distribution, which may be well approximated by a $\Gamma$ distribution. We show that this holds too for Geweke's spectral causality averaged over a given frequency band, and derive explicit expressions for the generalised $\chi^2$ and $\Gamma$-approximation parameters in both cases. We present an asymptotically valid Neyman-Pearson test based on the single-regression estimators, and discuss in detail how it may be usefully employed in realistic scenarios where autoregressive model order is unknown or infinite. We outline how our analysis may be extended to the conditional case, point-frequency spectral Granger causality, state-space Granger causality, and the Granger causality $F$-test statistic. Finally, we discuss approaches to approximating the distribution of the single-regression estimator under the alternative hypothesis.