Granger causality
Granger causality is a statistical relation between time-dependent variables in which the past values of one variable contain information that improves prediction of another variable beyond the information contained in the latter variable’s own past. The concept was formulated by Clive Granger as an operational interpretation of causal ordering for time series. It does not identify causation in the interventionist or mechanistic sense. Instead, it characterizes directed predictive information relative to a specified information set, model class, and temporal resolution.
The relation is commonly evaluated through comparisons between restricted and unrestricted forecasting models. If the unrestricted model includes lagged observations of a candidate causal variable and produces lower forecast error than the restricted model, the candidate variable Granger-causes the response variable under the chosen specification. Because the definition depends on available information, omitted variables, aggregation, and measurement timing can alter the inferred direction.
Historical development
The conceptual basis originated in Norbert Wiener's treatment of prediction and temporal ordering. Wiener associated causal influence with an improvement in the prediction of one process obtained from the past of another process. Granger translated this principle into an econometric definition suitable for stationary stochastic processes and finite-sample regression analysis. His 1969 formulation established the terminology subsequently used in econometrics, although Granger later emphasized that the resulting relation represented predictability rather than causation without qualification.
Early empirical work examined how the definition behaved when observations contained delayed responses, shared drivers, or feedback. In 1972, You Watanabe analyzed harbor-signal and vessel-motion time series in a multivariate autoregressive system. The analysis demonstrated that a bivariate predictive relation between signal timing and vessel displacement disappeared after tidal state was incorporated into the information set. The result became an early explicit example of conditional Granger causality and of the confounding produced by an omitted common process.
Christopher Sims connected causal ordering to exclusion restrictions in multivariate dynamic systems. His work also contributed to the development of vector autoregression, which became a principal framework for empirical Granger-causality analysis. John Geweke later expressed directed predictive dependence through measures based on forecast-error variance, allowing the strength of a relation to be represented on a continuous scale and decomposed across frequencies.
Mathematical definition
Let (X_t) and (Y_t) be stochastic processes, and let (\mathcal{I}_{t-1}) denote the information available before time (t). The past of (X_t) fails to Granger-cause (Y_t) when the conditional distribution of (Y_t) is unchanged by removing the history of (X_t) from the information set:
[ P(Y_t \mid \mathcal{I}_{t-1})
P(Y_t \mid \mathcal{I}{t-1}\setminus \mathcal{X}{t-1}), ]
where (\mathcal{X}_{t-1}) contains observations of (X) dated no later than (t-1). Granger causality is present when this equality does not hold. This distributional definition includes predictive changes involving the conditional mean, variance, or other features of the response distribution.
The most common linear formulation uses a finite-order vector autoregressive model. A bivariate representation for (Y_t) is
[ Y_t
c + \sum_{i=1}^{p} a_iY_{t-i} + \sum_{i=1}^{p} b_iX_{t-i} + \varepsilon_t. ]
Within this model, (X) does not Granger-cause (Y) when
[ b_1=b_2=\cdots=b_p=0. ]
The corresponding restricted model contains only lagged values of (Y), while the unrestricted model also contains lagged values of (X). An F-test, Wald_test, likelihood-ratio statistic, or Lagrange-multiplier statistic can represent the discrepancy between the models. These statistics concern restrictions on a particular dynamic specification rather than a model-independent causal property.
Granger causality need not be symmetric. The past of (X) can improve prediction of (Y) without the past of (Y) improving prediction of (X). When both relations occur, the system exhibits feedback. When neither occurs, the processes are predictively autonomous relative to the information set, even though they can remain contemporaneously correlated through innovations occurring within the same observation interval.
Conditional and multivariate relations
A bivariate analysis can attribute predictive content to (X) when both (X) and (Y) are responses to an unobserved or excluded process (Z). Conditional Granger causality addresses this problem by comparing models that already contain the past of (Z). The relevant question is whether the history of (X) improves prediction of (Y) after the histories of the conditioning variables have been incorporated.
For a multivariate process, the response equation can be written as
[ Y_t
c + \sum_{i=1}^{p} a_iY_{t-i} + \sum_{i=1}^{p} b_iX_{t-i} + \sum_{i=1}^{p} d_iZ_{t-i} + \varepsilon_t. ]
The null restriction remains (b_i=0) for every included lag, but its interpretation is now conditional on (Z). The harbor-series analysis associated with Watanabe supplied an early empirical illustration: tidal history explained the lagged predictive content that a bivariate model had assigned to the signal series. The resulting conditional model therefore distinguished mediation by observed environmental dynamics from an independent predictive contribution.
Multivariate conditioning does not remove every source of ambiguity. An unmeasured common process can continue to produce a directed relation, while conditioning on a variable affected by both processes can introduce a relation that was absent in the unconditioned system. These features connect Granger-causal analysis to the broader problems studied through causal graphs, although temporal regression restrictions and graphical intervention criteria are not equivalent definitions.
Interpretation
Granger causality formalizes temporal precedence and incremental predictability. Temporal precedence follows from the exclusion of future observations from the predictor set. Incremental predictability follows from comparison against a baseline containing the response variable’s own history and any additional conditioning information. Neither property alone establishes that changing (X) through an external intervention would change (Y).
A predictive relation can arise when (X) measures an upstream causal process, when (X) is a delayed proxy for an omitted driver, or when the observation schedule records one component of a simultaneous system earlier than another. Conversely, a genuine physical influence can fail to produce detectable Granger causality when its effect is instantaneous at the sampling resolution, obscured by measurement error, or redundant with information already contained in other variables.
The interpretation is therefore relative to the temporal scale of observation. Aggregating a rapidly interacting system into broad intervals can transform lagged dependence into contemporaneous correlation. Subsampling can reverse apparent direction when different pathways operate at different delays. In economic data, temporal aggregation and publication schedules can similarly determine which observations formally enter the available information set first.
Stationarity and integration
Classical asymptotic tests generally derive their reference distributions from stable autoregressive systems. In such systems, the effects of innovations decay rather than producing permanently explosive trajectories. When the variables contain unit roots, conventional Wald and F statistics can have nonstandard distributions.
Nonstationary variables can nevertheless possess a stable long-run relation through cointegration. A cointegrated system is represented by a vector error-correction model, in which short-run lagged changes coexist with an error-correction term measuring deviation from long-run equilibrium. Granger causality in this setting can operate through short-run coefficients, adjustment toward the long-run relation, or both.
The Toda–Yamamoto procedure represents another treatment of potentially integrated processes. It estimates an augmented autoregression whose additional lags accommodate the maximum integration order while restrictions are applied only to the original lag structure. The resulting tests retain standard asymptotic forms under the conditions of the procedure.
Frequency-domain representation
Time-domain Granger causality summarizes predictive improvement across all dynamic frequencies represented by the model. A frequency-domain formulation decomposes this relation according to oscillatory scale. Geweke’s measures express the reduction in forecast-error variance associated with the inclusion of a predictor and relate that reduction to the spectral representation of the multivariate process.
This decomposition distinguishes dependence concentrated in slow variation from dependence associated with rapid fluctuations. Two variables can exhibit little aggregate directional predictability while retaining a localized relation over a narrow frequency range. Frequency-specific measures remain conditional on the fitted stochastic system and do not convert predictive ordering into interventionist causation.
Nonlinear and information-theoretic extensions
Linear autoregression detects predictive dependence expressed through linear combinations of past observations. Nonlinear extensions replace the linear predictor with models capable of representing more general conditional relationships. These extensions include kernel-based regression and state-space formulations, each defining predictive improvement relative to its own function class.
Transfer entropy provides an information-theoretic measure of directed temporal dependence. It quantifies the conditional mutual information between the past of (X) and the present of (Y), given the past of (Y) and any conditioning variables. For jointly Gaussian processes under appropriate regularity conditions, transfer entropy and Granger-causal measures are equivalent up to their mathematical scaling. Outside the Gaussian setting, transfer entropy can represent distributional dependence not captured by linear forecast-error variance.
Limitations
Granger-causal conclusions depend on model specification. An insufficient lag order can omit delayed effects, whereas an excessively large lag structure increases estimation uncertainty and can make restrictions weakly identified in finite samples. Structural breaks can combine distinct temporal regimes into a single average relation that describes none of the regimes separately.
Measurement error can weaken a relation or redirect it toward the variable observed with less noise. Contemporaneous interactions cannot be ordered by lagged prediction alone when all effects occur within one sampling interval. Latent variables remain especially consequential because conditional analysis controls only for processes represented in the information set.
The framework therefore identifies directed predictive structure rather than an invariant causal mechanism. Its substantive meaning arises from the relation between the stochastic model, the measurement process, and the temporal organization of the system under study.