Panel Granger Causality Test
Panel Data Granger Causality Test · Also known as: panel causality test, Dumitrescu-Hurlin test, heterogeneous panel causality, panel Granger test
The Panel Granger Causality test examines whether past values of one variable help predict another variable across multiple cross-sectional units observed over time. It extends the classical Granger causality framework to panel data, accounting for cross-sectional heterogeneity and enabling more powerful inference by pooling information across units.
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 Panel Granger Causality when you have panel data (multiple units observed over multiple periods) and want to assess whether one variable helps predict another after controlling for the variable's own past. It is most appropriate when T is moderate to large (T ≥ 10), units are heterogeneous, and the series are stationary or have been transformed to stationarity. Prefer Dumitrescu-Hurlin over pooled tests whenever slope heterogeneity across units is plausible — which is almost always in cross-country or cross-firm panels. Do NOT apply the test to non-stationary series without pre-filtering; do not interpret Granger causality as structural or economic causality; and avoid when N is very small (N < 5) since the asymptotic approximation degrades.
Strengths & limitations
- Allows heterogeneous causal relationships across panel units, avoiding the homogeneity bias of pooled tests.
- Gains statistical power relative to individual time-series Granger tests by pooling information across units.
- The Dumitrescu-Hurlin Z-bar statistic has a known asymptotic distribution, enabling straightforward inference.
- Compatible with unbalanced panels and robust to moderate cross-sectional dependence when bootstrap critical values are used.
- Easily extended to the full system by testing causality in both directions simultaneously.
- Assumes stationarity; applying the test to integrated series without differencing or a VECM framework yields spurious results.
- The standard asymptotic distribution requires both N and T to be sufficiently large; results are unreliable in short or very narrow panels.
- Granger causality is predictive, not structural — rejection does not establish economic or policy causation.
- Cross-sectional dependence (common shocks) inflates test size; bootstrap or CD-robust versions are needed when cross-sectional dependence is detected.
- Choosing the lag length K can meaningfully affect results, and there is no universally optimal criterion.
Frequently asked
What is the difference between Dumitrescu-Hurlin and pooled panel Granger causality?
Pooled (homogeneous) panel Granger tests constrain the causal coefficients to be identical across all units. The Dumitrescu-Hurlin test allows each unit to have its own causal coefficients and averages the resulting Wald statistics, making it robust to slope heterogeneity — the more realistic assumption in most empirical panels.
Do I need to difference the data before running the panel Granger test?
Yes, if the series are integrated (I(1) or higher). Run a panel unit root test first. If variables are cointegrated, a panel VECM framework is preferable because it captures both short-run Granger causality and long-run adjustment; applying the standard Granger test to non-stationary levels produces spurious inference.
How do I handle cross-sectional dependence in panel Granger causality?
First test for cross-sectional dependence using the Pesaran CD test. If dependence is detected, use bootstrap critical values for the Dumitrescu-Hurlin statistic, or employ a cross-sectionally augmented version of the panel VAR to filter common factors before testing.
What does a rejection of the null tell me practically?
Rejection means that past values of X contain statistically significant information about future values of Y beyond Y's own history, for at least some panel units. It does not tell you how many units exhibit causality, nor does it imply a structural causal mechanism — additional analysis or theory is needed to interpret the direction and magnitude of the relationship.
How many lags should I use?
Select lags using AIC or BIC on individual unit regressions or on the pooled system. In practice, one or two lags are common when T is moderate. Be cautious with long lags in short panels because degrees of freedom diminish rapidly, reducing power and reliability.
Sources
- Dumitrescu, E.-I., & Hurlin, C. (2012). Testing for Granger non-causality in heterogeneous panels. Economic Modelling, 29(4), 1450–1460. DOI: 10.1016/j.econmod.2012.02.014 ↗
- Holtz-Eakin, D., Newey, W., & Rosen, H. S. (1988). Estimating vector autoregressions with panel data. Econometrica, 56(6), 1371–1395. DOI: 10.2307/1913103 ↗
How to cite this page
ScholarGate. (2026, June 3). Panel Data Granger Causality Test. ScholarGate. https://scholargate.app/en/econometrics/panel-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 Causality TestEconometrics↔ compare
- Panel ARDL Bounds TestEconometrics↔ compare
- Panel Johansen CointegrationEconometrics↔ compare
- Panel VECMEconometrics↔ compare
- Toda-Yamamoto causality testEconometrics↔ compare