Bayesian Granger Causality
Bayesian Granger Causality Analysis · Also known as: Bayesian Granger test, Bayesian predictive causality, BGC, Bayesian causality in mean
Bayesian Granger causality tests whether past values of one time series carry predictive information about another, framing the hypothesis through Bayesian inference rather than frequentist p-values. It combines a vector autoregressive (VAR) structure with prior distributions over coefficients and evaluates causal claims via posterior probabilities or Bayes factors, providing a probabilistic and nuanced alternative to the classical Granger test.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Bayesian Granger causality when you work with short or moderate time series where classical p-value tests over-reject under small samples, when you want to incorporate prior economic theory about causal direction, or when you need a quantitative measure of evidence strength rather than a binary significant/not-significant verdict. It is especially valuable when comparing multiple causal hypotheses simultaneously across a VAR system or when avoiding the pretesting bias of classical unit-root-then-causality pipelines. Do not use it as a substitute for genuine structural identification — Granger causality measures predictive precedence, not structural economic causation. Avoid when the computational cost of MCMC is prohibitive and a classical test on a large, well-behaved sample would suffice.
Strengths & limitations
- Provides a continuous evidence measure (Bayes factor or posterior probability) rather than an arbitrary significance threshold.
- Naturally regularises coefficient estimation through priors, reducing spurious causality findings in short samples.
- Handles uncertainty about lag length by averaging across models rather than committing to a single lag selection.
- Allows explicit incorporation of prior theoretical knowledge about causal directions.
- Robust to misspecification of the lag order compared to frequentist AIC/BIC selection followed by testing.
- Results depend on the choice of prior; poorly specified priors can strongly influence posterior causality probabilities, especially in small samples.
- Computationally intensive when using MCMC for non-conjugate priors or high-dimensional VAR systems.
- Like all Granger tests, it captures predictive precedence, not structural causality — confounders and omitted variables can generate spurious findings.
- The Bayes factor scale (e.g., Jeffreys thresholds) is a convention, not a formal decision rule, and different thresholds can lead to different practical conclusions.
Frequently asked
What is the difference between Bayesian and classical Granger causality?
Classical Granger causality uses an F-test on VAR coefficient restrictions and returns a p-value; rejection depends on an arbitrary alpha level. Bayesian Granger causality computes a posterior probability of causality and a Bayes factor, providing a continuous, interpretable evidence measure that incorporates prior information and naturally handles small-sample uncertainty.
Does Bayesian Granger causality require stationary series?
The standard VAR-based setup works best with stationary series. For integrated or cointegrated series, the Bayesian VECM (vector error correction model) extension or the Bayesian Toda-Yamamoto approach (which pads the VAR with extra lags) is recommended to avoid spurious causality findings.
How do I choose the prior for Bayesian Granger causality?
The Minnesota/Litterman prior is the most widely used default: it shrinks own-lag coefficients toward 1 (random walk) and cross-lag coefficients toward 0. The hyperparameters governing tightness can be set by cross-validation or marginal likelihood maximisation. The key is that the prior should encode reasonable economic beliefs, not merely replicate the flat-prior frequentist test.
What Bayes factor threshold indicates strong Granger causality?
Using Jeffreys's scale: BF10 > 3 is moderate evidence, BF10 > 10 is strong evidence, and BF10 > 100 is decisive. However, these are conventions — report the full Bayes factor and posterior probability rather than a binary threshold decision.
Can Bayesian Granger causality handle many variables?
Yes, but large VAR systems require either conjugate priors with analytical posteriors or computationally efficient MCMC. Bayesian shrinkage priors such as the Minnesota prior are especially important in high-dimensional systems to avoid overfitting and spurious causality claims.
Sources
How to cite this page
ScholarGate. (2026, June 3). Bayesian Granger Causality Analysis. ScholarGate. https://scholargate.app/en/econometrics/bayesian-granger-causality
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian VAR modelEconometrics↔ compare
- Bayesian VECMEconometrics↔ compare
- Granger Causality TestEconometrics↔ compare
- Panel Granger CausalityEconometrics↔ compare
- Toda-Yamamoto causality testEconometrics↔ compare
- Vector AutoregressionEconometrics↔ compare