Robust Granger Causality Test
Also known as: bootstrap Granger causality, heteroscedasticity-robust Granger causality, non-asymptotic Granger causality test, RGC
Robust Granger causality extends the classic Granger causality framework by using bootstrap-based or heteroscedasticity-robust critical values rather than asymptotic chi-squared tables. This makes the test reliable in finite samples and when the data exhibit non-normality, heteroscedasticity, or near-integration, settings where the standard F- or Wald-based test is known to over-reject.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use robust Granger causality when you suspect that standard asymptotic critical values are unreliable: small or moderate samples (T < 100), residuals that show ARCH effects or non-normality, or series that are near-integrated (close to a unit root without clear cointegration). It is also the preferred choice after a Toda-Yamamoto augmentation when the augmented system still shows non-normal or heteroscedastic errors. Do not use it as a causal identification tool in the structural or policy sense — Granger causality is strictly a predictability concept, not a proof of mechanism; for structural causality use instrumental variables or structural VAR.
Strengths & limitations
- Maintains correct test size in finite samples where the standard chi-squared approximation breaks down.
- Robust to heteroscedasticity and non-normality of VAR residuals without requiring parametric error-distribution assumptions.
- Compatible with integrated or near-integrated data when combined with the Toda-Yamamoto augmentation strategy.
- Easy to implement in standard software (R packages 'vars', 'tsDyn'; Stata user commands) once the VAR is estimated.
- Provides reliable inference without needing to pre-test for unit roots or cointegration.
- Bootstrap inference is computationally heavier than a simple F-test, though the added cost is modest for typical macro/finance sample sizes.
- The test still requires correct lag-length selection; an under-fitted VAR can distort both the point estimates and the bootstrap distribution.
- Granger causality detects predictive precedence, not structural or counterfactual causation — a finding of Granger causality can reflect a common driver or an anticipation effect rather than a true mechanism.
- With many variables the VAR becomes high-dimensional, consuming degrees of freedom and making the bootstrap less reliable; consider bivariate or factor-augmented specifications in such cases.
Frequently asked
How does robust Granger causality differ from standard Granger causality?
Both tests share the same null hypothesis — that past values of X carry no additional predictive power for Y — and both compute a Wald statistic from a VAR. The difference is that the standard test compares the statistic to an asymptotic chi-squared critical value, which can be inaccurate in small or heteroscedastic samples. The robust version replaces those critical values with ones generated by bootstrapping the residuals, preserving the correct test size under weaker assumptions.
How many bootstrap replications are sufficient?
A common recommendation is 1,000 replications for a 5% significance level and 5,000 for a 1% level. More replications reduce Monte Carlo noise in the p-value estimate but increase computation time; 1,000 is usually adequate for applied work.
Should I use wild bootstrap or iid bootstrap?
Wild bootstrap (e.g., the Rademacher or Mammen weights) is preferred when the VAR residuals are heteroscedastic, because it preserves the heteroscedasticity structure in each pseudo-sample. The simpler iid bootstrap is acceptable only if residuals show no ARCH effects.
Can I apply this test when variables are I(1) or cointegrated?
Yes. The Toda-Yamamoto (1995) strategy — augmenting the VAR with extra lags equal to the maximum integration order — allows standard Wald tests on the original-lag block even with integrated variables, and adding bootstrap critical values on top of that augmentation gives a fully robust procedure.
Does a Granger causality result imply economic causation?
No. Granger causality is a statement about predictive precedence: X Granger-causes Y if X's past improves forecasts of Y. A common third factor, or market anticipation, can produce the same statistical result without any direct mechanism linking X to Y. Use economic theory and structural identification to move from predictive causality to economic causation.
Sources
- Hacker, R. S., & Hatemi-J, A. (2006). Tests for causality between integrated variables using asymptotic and bootstrap distributions: Theory and application. Applied Economics, 38(13), 1489–1500. DOI: 10.1080/00036840500405763 ↗
- Granger, C. W. J. (1969). Investigating causal relations by econometric models and cross-spectral methods. Econometrica, 37(3), 424–438. DOI: 10.2307/1912791 ↗
How to cite this page
ScholarGate. (2026, June 3). Robust Granger Causality Test. ScholarGate. https://scholargate.app/en/econometrics/robust-granger-causality
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Cointegration TestEconometrics↔ compare
- Granger CausalityEconometrics↔ compare
- Toda-Yamamoto CausalityEconometrics↔ compare
- VAR ModelEconometrics↔ compare