Granger Causality Testing

Description

Tests whether past values of one time series improve prediction of another beyond its own past, a disciplined but limited form of temporal causal evidence.

X Granger-causes Y if lagged values of X significantly improve a model predicting Y from its own lags. Estimated within vector autoregression (VAR) models with stationarity checks and lag selection by information criteria, the test provides evidence about temporal precedence and predictive information. It is not proof of causation: common causes and unmeasured variables can produce Granger causality, and instantaneous effects are missed. Used carefully, it helps untangle feedback between metrics, such as whether support contacts predict churn or churn predicts support contacts, and whether engagement drives content creation or vice versa.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Days

Steps

  1. Choose the time series pairs or sets and an appropriate aggregation interval.
  2. Test for stationarity (ADF, KPSS) and difference or detrend as needed.
  3. Fit a VAR model selecting lag length by AIC or BIC.
  4. Run Granger tests in both directions and check residual diagnostics.
  5. Report results as predictive precedence, and check robustness to lag choices and added control series.
  6. Corroborate with experiments or other causal designs before acting.

Sources

  • Granger, Investigating causal relations by econometric models and cross-spectral methods (1969), Econometrica
  • Lütkepohl, New Introduction to Multiple Time Series Analysis (2005)

Also known as

  • Granger causality
  • Vector autoregression causality
  • Predictive causality

Categories

  • Analytics & Behaviour
  • Statistics & Modelling