Granger Causality

Granger causality defines a cause in terms of predictive information. Variable X is said to 'Granger-cause' Y if past values of X contain statistically significant information for predicting future values of Y, beyond the information contained in past values of Y alone. This is distinct from true physical or mechanistic causality, as it identifies a predictive relationship that may be driven by an unobserved common cause. It answers 'Does X help forecast Y?' rather than 'Does X directly produce Y?'

The test is fundamentally designed for temporal data. It requires observations ordered in time (e.g., stock prices, EEG signals, economic indicators). The core logic relies on temporal precedence: a cause must precede its effect. The analysis involves fitting vector autoregressive (VAR) models to compare:

• A restricted model predicting Y using only its own past.
• An unrestricted model predicting Y using its own past and the past of X. A statistically significant improvement in the unrestricted model's predictive power suggests Granger causality from X to Y.

A major limitation is its vulnerability to unobserved confounding. If a third variable, Z, causes both X and Y, X may appear to Granger-cause Y even without a direct link. The test also assumes all relevant variables are included in the model; omitted variable bias can lead to spurious causal inferences. For reliable results, the analyst must include all potential common causes in the VAR model, which is often impossible with observational data. This is why Granger causality is best seen as evidence of a predictive relationship, not proof of a causal one.

The simplest test is bivariate, considering only the relationship between X and Y. This is highly prone to spurious results due to confounding. The robust approach is multivariate Granger causality, where the VAR model includes other potentially relevant variables (W, Z, etc.). Conditioning on these other variables helps block some backdoor paths. Modern implementations, like conditional Granger causality, explicitly test if X provides unique information about Y's future, conditional on the past of a set of other observed variables.

Granger causality is implemented as a hypothesis test (e.g., using an F-test on VAR model residuals or a likelihood ratio test). The output is typically a p-value indicating whether the null hypothesis ("X does not Granger-cause Y") can be rejected. It is not a continuous measure of causal strength. However, derived metrics like the Granger causality index or the magnitude of the F-statistic are sometimes used to quantify the degree of predictive influence. The core result remains binary: rejection or failure to reject the null hypothesis.

Granger causality is a primary tool in causal discovery algorithms for time series data. It forms the basis for methods like:

• PC algorithm adapted for temporal data.
• Granger causal graph construction.
• Transfer entropy, an information-theoretic analogue. These methods use pairwise or conditional Granger tests to infer the structure of a causal graph among multiple time series variables. It provides a computationally tractable and widely understood starting point for modeling dynamic causal systems, bridging statistical time series analysis and causal inference.

Causal inference is the overarching field of drawing conclusions about cause-and-effect relationships from data. It moves beyond detecting mere statistical associations to answer questions about the impact of interventions.

• Goal: To estimate what would happen if a variable were manipulated.
• Contrast with Granger Causality: While Granger causality tests for predictive precedence in time series, causal inference seeks to establish true causal mechanisms, often requiring stronger assumptions about confounding.
• Key Methods: Include randomized controlled trials (the gold standard), instrumental variables, regression discontinuity, and methods based on structural causal models.

A Structural Causal Model (SCM) is a formal mathematical framework that represents causal relationships using a system of structural equations, often visualized as a causal graph.

• Components: Consists of endogenous variables, exogenous variables (noise), and functions that assign each variable a value based on its direct causes.
• Contrast with Granger Causality: Granger causality is a statistical test within time series. An SCM provides a full, manipulable model of the system, enabling reasoning about interventions (do-calculus) and counterfactuals.
• Use Case: The 'source code' of a causal system, allowing simulation of changes and estimation of effects beyond prediction.

A causal graph is a directed acyclic graph (DAG) where nodes represent variables and edges represent direct causal relationships. It encodes conditional independence assumptions via d-separation.

• Function: Serves as a map of assumed causal structure, separating direct causes from confounders and mediators.
• Link to Granger Causality: A discovered Granger-causal relationship (X → Y) might correspond to a directed edge in a causal graph for time-series variables, but spurious temporal prediction can occur due to unobserved confounding not depicted in the graph.
• Key Tools: Used to apply the backdoor criterion and frontdoor criterion for identifying causal effects from observational data.

An intervention, denoted by the do-operator (e.g., do(X=x)), is the act of externally forcing a variable to a value, simulating an experiment. Do-calculus is a set of rules to compute interventional probabilities from observational data given a causal graph.

• Purpose: To answer 'what if' questions by cutting a variable's equation from its natural causes.
• Contrast with Granger Causality: Granger causality analyzes observational temporal data without intervention. Do-calculus provides the mathematical machinery to predict the outcome of an intervention, which is the essence of causal effect estimation.
• Example: P(Y | do(X=x)) is the distribution of Y if we set X to x, distinct from the conditional probability P(Y | X=x).

Causal discovery refers to algorithms that automatically infer the causal structure (e.g., a causal graph) from data. Methods include constraint-based (testing conditional independencies), score-based, and those using functional causal models.

• Relationship to Granger Causality: Granger causality tests are a form of bivariate causal discovery for time series. Modern multivariate causal discovery algorithms (e.g., PC, FCI, LiNGAM) can handle more complex scenarios with latent confounders and contemporaneous effects.
• Challenges: Requires assumptions like the Causal Markov Condition and Causal Faithfulness. Results are often a set of possible graphs (Markov equivalence class) unless strong assumptions are made.

Invariant Risk Minimization (IRM) is a machine learning paradigm that aims to find data representations whose optimal predictor remains invariant across multiple training environments. It promotes the learning of causal features over spurious correlations.

• Objective: Achieve out-of-distribution generalization by discovering features with stable causal relationships to the target.
• Philosophical Link: While Granger causality identifies temporally stable predictive relationships, IRM seeks representations that are invariant across environments, a related concept for robustness. Both approaches aim to move beyond brittle statistical patterns.
• Application: Used to build models that perform well under distribution shift, a key concern in deploying causal models in non-stationary real-world settings.