Structural Break Granger Causality
Granger Causality Testing with Structural Breaks · Also known as: break-robust Granger causality, Granger causality under regime change, time-varying Granger causality, structural change Granger test
Structural break Granger causality extends the classic Granger causality framework to accommodate regime shifts and parameter instability in time series. By detecting break points and testing causality within sub-samples or via rolling/recursive windows, it reveals whether a predictive relationship between variables switches on, switches off, or changes direction over time.
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 structural break Granger causality when you have a moderately long time series (at least 60-80 observations recommended for rolling analysis) and economic or institutional reasons to believe the causal relationship may not be stable over the full sample — for example, before and after a financial crisis, a policy reform, or a structural change in energy markets. Avoid it when the series are very short (fewer than 40 observations), when no theoretical reason for instability exists, or when you need a simple, easily communicated baseline result; in those cases plain Granger causality or Toda-Yamamoto suffices.
Strengths & limitations
- Detects time-varying causal relationships that a full-sample test would obscure or miss entirely.
- Toda-Yamamoto augmentation preserves asymptotic validity even when series are I(1) or cointegrated, without requiring pre-testing for unit roots.
- Rolling-window variant produces a visual timeline of causality that is immediately interpretable for policy analysis.
- Bootstrap critical values maintain correct size in small windows and with non-normal errors.
- Accommodates multiple break dates via Bai-Perron, avoiding the need to choose a single ad-hoc split point.
- Requires a long enough series: very short samples yield unstable rolling-window estimates and unreliable break-date detection.
- Multiple testing inflates the probability of false positives across many sub-windows; some correction (e.g., sequential testing, FDR) is advisable.
- Bootstrap procedures are computationally intensive, especially with multivariate VAR and many rolling windows.
- Break dates estimated from the data carry uncertainty; treating them as known fixed points understates total parameter uncertainty.
- Results depend on the chosen window size and VAR lag order, so sensitivity analysis across those choices is essential.
Frequently asked
How is this different from standard Granger causality?
Standard Granger causality assumes a single stable causal structure over the whole sample. Structural break Granger causality allows that structure to shift at one or more break dates, or to vary continuously across rolling windows, so it reveals when a causal relationship exists rather than just whether it exists on average.
Do I still need to test for unit roots before applying this method?
When using the Toda-Yamamoto augmented VAR, you do not need to pre-test for unit roots or cointegration to preserve asymptotic validity — the d_max augmentation handles it. You should still detect break dates first, because integrated series with structural breaks can produce spurious unit-root and cointegration results.
What window size should I use for rolling-window analysis?
A common starting point is 15-20% of the total sample length, large enough to estimate the VAR reliably but small enough to detect regime changes. Report results for at least two window sizes as a robustness check, since conclusions can be sensitive to this choice.
How do I correct for multiple testing across rolling windows?
One approach is to use the sequentially rejective Bonferroni procedure or the Benjamini-Hochberg false discovery rate correction across the sequence of p-values. At minimum, report the proportion of windows showing rejection and flag whether results cluster around economically meaningful episodes.
Can I apply this in a multivariate setting with more than two variables?
Yes. A multivariate VAR can accommodate several variables simultaneously, and you can test block-Granger causality from a set of predictors to a target variable. Computational cost rises with system size, and small rolling windows become more problematic as the number of estimated parameters grows.
Sources
- Toda, H. Y., & Yamamoto, T. (1995). Statistical inference in vector autoregressions with possibly integrated processes. Journal of Econometrics, 66(1-2), 225-250. DOI: 10.1016/0304-4076(94)01616-8 ↗
- Balcilar, M., Ozdemir, Z. A., & Arslanturk, Y. (2010). Economic growth and energy consumption causal nexus viewed through a bootstrap rolling window. Energy Economics, 32(6), 1398-1410. DOI: 10.1016/j.eneco.2010.05.015 ↗
How to cite this page
ScholarGate. (2026, June 3). Granger Causality Testing with Structural Breaks. ScholarGate. https://scholargate.app/en/econometrics/structural-break-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.
- Granger CausalityEconometrics↔ compare
- Toda-Yamamoto CausalityEconometrics↔ compare
- Vector AutoregressionEconometrics↔ compare