MerQuR+Statistics Glossary

Statistics Glossary → Granger causality

Granger causalityconcept

The idea, and test, of whether the past values of one time series help predict the future values of another beyond what that series' own past already provides.

Saying that x Granger-causes y means that, with y's own lags already in the model, the lags of x add significantly to the prediction of y. In a VAR model the test checks the hypothesis that all lag coefficients of x in the y equation are jointly zero, using an F or Wald test. The relationship can run both ways: x can Granger-cause y and y can Granger-cause x.

Despite its name, it captures predictive precedence, not true causation: a third variable that affects both x and y, but affects x earlier, can also produce Granger causality. The results are sensitive to the number of lags, which should be chosen with information criteria. If the series are not stationary the standard test can mislead, and differencing, cointegration analysis or the Toda-Yamamoto approach may be needed. The sampling frequency also matters: effects that play out within a month look simultaneous in monthly data and cannot be detected.

An everyday example

A supermarket chain studies the link between ice-cream and sunscreen sales. Last week's ice-cream sales seem to help predict this week's sunscreen sales. That does not mean ice cream drives sunscreen sales; most likely warmer weather affects both, and ice-cream sales simply respond to the heat faster.

In MerQur

As the Granger causality tests in the VAR analysis; whether each variable can be predicted from the lags of the others is reported together with lag selection, impulse-response functions and variance decomposition.

Analyses where this concept appears

VAR (Vector Autoregression)

Related concepts

Impulse response function (IRF) and Cholesky decomposition