Granger Causality

The core axiom of Granger causality is that the cause must occur before the effect. The test evaluates whether past values of time series X contain statistically significant information that helps predict future values of time series Y, beyond the information contained in the past values of Y alone.

• Strict ordering: X(t-1), X(t-2), ... must precede Y(t)
• Lag selection: The number of past time points (lags) included is critical and is often determined by information criteria like AIC or BIC
• Practical example: In biomarker studies, testing whether a protein expression trajectory Granger-causes a downstream metabolite concentration change

A fundamental requirement for valid Granger causality testing is that the time series are covariance stationary. This means the mean, variance, and autocorrelation structure do not change over time. Non-stationary data can produce spurious regression results, indicating a causal relationship where none exists.

• Augmented Dickey-Fuller (ADF) test: Commonly used to check for unit roots
• Differencing: If non-stationary, series are often differenced (e.g., using returns instead of prices) to achieve stationarity
• Cointegration caveat: If two non-stationary series are cointegrated, a Vector Error Correction Model (VECM) must be used instead of the standard VAR framework

Granger causality is typically implemented within a Vector Autoregression (VAR) model. In a bivariate system, two equations are estimated: one predicting Y using lags of Y and X, and another predicting X using lags of both. An F-test or Wald test then assesses the joint significance of the lagged coefficients of the causal variable.

• Restricted vs. Unrestricted model: The test compares the predictive error of a model using only past Y to a model using past Y and past X
• Bidirectional testing: The test is run in both directions (X→Y and Y→X) to establish the direction of predictive flow
• Instantaneous causality: A separate test for whether current X helps predict current Y, requiring different theoretical assumptions

Granger causality is a test of conditional independence within a specific information set. The result is entirely dependent on the variables included in the model. Omitting a confounding variable Z that drives both X and Y can lead to a false positive where X appears to Granger-cause Y.

• Omitted variable bias: The most common pitfall in practical applications
• Exogeneity requirement: All relevant predictors must be included; there should be no serial correlation in the residuals
• Biomedical context: In biomarker time series, failing to account for a circadian rhythm driver could falsely suggest one biomarker causes another

A critical distinction is that Granger causality does not imply true structural or mechanistic causality. It is strictly a measure of incremental predictive utility. A rooster's crow may Granger-cause the sunrise in a purely predictive model, but this is not a mechanistic relationship.

• Predictive vs. structural: Granger causality is necessary but not sufficient for true causality
• No counterfactual: The test does not answer what would happen to Y if we intervened on X
• Complementary methods: In biomedicine, Granger findings are often validated against Mendelian Randomization or causal discovery algorithms to establish mechanistic plausibility

Spectral Granger causality decomposes the predictive relationship across different frequency components. This reveals whether the causal influence is concentrated in slow oscillations (long-term trends) or fast oscillations (short-term fluctuations).

• Fourier transform of VAR coefficients: The time-domain coefficients are transformed to the frequency domain
• Band-specific causality: Useful in fMRI and EEG analysis to isolate causal connectivity in specific neural frequency bands (alpha, beta, gamma)
• Biomedical application: Identifying whether a neural signal in one brain region drives another specifically in the gamma band during a cognitive task

A graphical representation of causal assumptions where nodes represent variables and directed edges represent direct causal effects, containing no feedback loops. Unlike Granger causality, which is purely predictive, DAGs encode structural causal knowledge. In biomarker time series, a DAG can distinguish between a variable that Granger-causes another due to a common driver versus a true mechanistic link.

An instrumental variable method using genetic variants as proxies for exposures to estimate causal effects on outcomes. While Granger causality relies on temporal precedence, MR leverages the random assortment of genes at conception to minimize confounding. In drug target validation, MR is often the gold standard for confirming that a biomarker identified via Granger analysis has a lifelong causal impact on disease.

A class of algorithms, such as the PC algorithm, that infers causal structures directly from observational data by testing conditional independencies. These methods extend Granger causality by handling instantaneous effects and non-temporal data. In multi-omics integration, causal discovery can map regulatory networks from static gene expression data where time-series ordering is unavailable.

A framework that estimates what would have happened to an outcome if a different exposure had occurred. Granger causality answers 'does the past of X help predict Y?', while counterfactuals answer 'what would Y be if we intervened on X?'. In clinical biomarker studies, counterfactual models estimate the expected change in disease progression if a specific protein target is modulated.

A multivariate technique that models complex relationships between observed and latent variables, combining factor analysis and path analysis. SEM provides a parametric alternative to Granger causality, allowing for the modeling of instantaneous feedback loops and measurement error. In systems biology, SEM is used to test hypothesized causal pathways between metabolites and clinical endpoints.

An information-theoretic measure of directed information flow between time series, often considered the non-linear generalization of Granger causality. It quantifies the reduction in uncertainty about the future of Y given the past of X. In neural biomarker analysis, transfer entropy is preferred for detecting causal interactions in EEG and fMRI data where relationships are highly non-linear.