Bayesian Toda-Yamamoto Causality Test
Also known as: Bayesian TY causality, Bayesian modified Wald causality, Bayesian Granger non-causality in VAR, BTY causality
The Bayesian Toda-Yamamoto causality procedure combines the Toda-Yamamoto VAR augmentation strategy — which sidesteps the need for pre-testing integration and cointegration — with Bayesian prior-posterior updating. It tests Granger non-causality between time series that may be integrated or cointegrated without requiring differencing or error-correction modeling, while incorporating prior information and producing full posterior distributions over the causal parameters.
Key highlights
- Handles integrated and cointegrated series without requiring pre-testing for the order of integration or imposing a cointegrating rank.
- Produces posterior probabilities and credible intervals rather than a binary p-value, giving richer inference about causal strength.
- Prior specification (e.g., Minnesota prior) regularises the augmented VAR, which can be large, reducing overfitting in moderate samples.
- Avoids distortions from sequential pre-testing (unit-root test then causality test) that inflate size in frequentist workflows.
- Bayesian model comparison via Bayes factors allows formal comparison of causal and non-causal models.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Bayesian Toda-Yamamoto causality when you have time series that may be integrated (I(1) or I(2)) or cointegrated and you want to test Granger causality without committing to a specific cointegration specification. The Bayesian framing is especially valuable with small or moderate samples where asymptotic frequentist tests are unreliable, when prior information from theory or earlier studies is available, or when a full posterior distribution over causal parameters is more informative than a binary reject/fail-to-reject decision. Avoid it when the series are clearly stationary (standard Granger causality in a VAR is simpler) or when very large samples make prior choice inconsequential and frequentist tests reliable.
Strengths & limitations
- Handles integrated and cointegrated series without requiring pre-testing for the order of integration or imposing a cointegrating rank.
- Produces posterior probabilities and credible intervals rather than a binary p-value, giving richer inference about causal strength.
- Prior specification (e.g., Minnesota prior) regularises the augmented VAR, which can be large, reducing overfitting in moderate samples.
- Avoids distortions from sequential pre-testing (unit-root test then causality test) that inflate size in frequentist workflows.
- Bayesian model comparison via Bayes factors allows formal comparison of causal and non-causal models.
- Prior choice is consequential: a poorly calibrated prior can dominate the posterior in small samples, introducing subjectivity.
- The augmented VAR grows quickly with lag order and number of series, making posterior computation expensive for large systems.
- Requires familiarity with Bayesian computation (MCMC, conjugate updating); less accessible than the simple frequentist Toda-Yamamoto Wald test.
- If d_max is mis-specified (e.g., the series are actually I(2) but treated as I(1)), the asymptotic validity guarantee breaks down.
- Software implementations are not yet standardised; analysts often code their own MCMC routines.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
Why add extra lags equal to d_max?
Adding d_max extra lags to the estimated VAR order k ensures the joint chi-squared limit of the Wald statistic holds even if some series have unit roots or are cointegrated, because the extra lags absorb the non-standard asymptotic behaviour. The causality test is then applied only to the first k lag coefficients.
What prior should I use for the Bayesian version?
The Minnesota (Litterman) prior is the standard choice: it shrinks own-lag coefficients toward a random-walk value of 1 and cross-lag coefficients toward 0, with tighter shrinkage at longer lags. This is well-suited to macroeconomic and financial time series and stabilises the augmented VAR when the system is large.
How do I report the causality result in a Bayesian framework?
Report the posterior probability that all k relevant lag coefficients are jointly zero (non-causality), a 95% highest posterior density interval for the causal coefficients, and — if feasible — a Bayes factor comparing the causal to the non-causal model. State the prior used and include a sensitivity analysis.
Can I use this test when I am unsure whether my series is I(1) or I(2)?
Yes — set d_max to the highest plausible integration order (e.g., 2 if you suspect some series may be I(2)). The test remains valid, though adding more extra lags reduces power slightly. This is one of the key advantages of the Toda-Yamamoto strategy over standard Granger tests.
How does this differ from a Bayesian VECM approach?
A Bayesian VECM explicitly models the cointegrating relationships and imposes them as restrictions, which is more efficient when the cointegrating rank is known. The Bayesian Toda-Yamamoto approach makes no assumption about cointegrating rank — it is more robust when rank is uncertain but pays a cost in efficiency.
Sources
- 1.Toda, H. Y., & Yamamoto, T. (1995). Statistical inference in vector autoregressions with possibly integrated processes. Journal of Econometrics, 66(1-2), 225-250.
- 2.Zellner, A. (1971). An Introduction to Bayesian Inference in Econometrics. Wiley.ISBN 978-0471982326
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Bayesian Toda-Yamamoto Causality. ScholarGate. https://scholargate.app/econometrics/bayesian-toda-yamamoto-causality