Finite-Sample Simulation-Based Inference in VAR Models with Application to Granger Causality Testing

Summary

Dufour and Jouini (2006) document that standard asymptotic and bootstrap tests in vector autoregression (VAR) models can have catastrophic finite-sample size — rejection rates up to 97% at a nominal 5% level — and propose the Maximized Monte Carlo (MMC) test as a provably exact solution. The MMC procedure maximizes a simulated p-value over the nuisance parameter space, yielding valid level-α inference regardless of integration order, cointegration structure, or dimensionality. Applied to a 4-variable quarterly U.S. VAR (non-borrowed reserves, federal funds rate, GDP, GDP deflator), MMC identifies a monetarist causal chain: money → interest rate → output (M → r → y).

Key Claims

• Standard chi-square asymptotic tests for Granger non-causality have empirical rejection rates as high as 97% at a nominal 5% level ($k=4$, $p=5$, $T=50$, $\rho=0.99$). Even with $T=300$ the problem persists for larger systems.
• Local Monte Carlo (LMC) tests — equivalent to parametric bootstrap — reduce but do not eliminate overrejection: rates up to 59% at 5% nominal level in the same configurations.
• The lag-augmented Wald test (Dolado-Lütkepohl / Toda-Yamamoto) also severely over-rejects in many configurations, even with bootstrap critical values.
• The MMC test controls size exactly throughout: rejection rate ≤ α for all configurations tested. It is the only procedure that achieves this.
• MMC requires only the ability to simulate the model under the null hypothesis. No assumption about asymptotic distribution, integration order, or cointegration rank is needed.
• Key identity: if $\alpha(N+1)$ is an integer, $P[\hat{p}_N(S_0) \leq \alpha] = \alpha$ exactly — the MC p-value has a uniform $[0,1]$ distribution under the null.

Concepts Introduced or Extended

• Granger Causality

Entities Mentioned

• Jean-Marie Dufour
• Tarek Jouini

Quotes

"Standard tests based on asymptotic critical values can have catastrophic size properties, with rejection frequencies as high as 0.97 (instead of 0.05)."

"The MMC procedure offers protection against failures of the bootstrap. A non-significant bootstrap p-value entails a non-significant MMC p-value."

My Take

The key insight is that the nuisance parameter problem in VAR causality testing — which makes both asymptotic and bootstrap tests unreliable — is solved not by eliminating the nuisance parameters (as cointegration rank pre-testing attempts) but by maximizing over them. The simulation evidence is striking: the problem is not confined to near-integrated or small-sample edge cases but is severe even at $T=300$ in higher-dimensional VARs. The main practical cost is computational intensity — the MMC maximization over the nuisance parameter space requires multiple rounds of VAR simulation — but the authors show it is feasible. The monetarist M → r → y finding contrasts with the prior literature on money-income causality, which was sensitive to test method.