Zha (2008) Vector Autoregressions

Summary

An encyclopedia entry for the New Palgrave Dictionary of Economics by Tao Zha. Provides a compact and rigorous survey of Vector Autoregression (VAR) methodology: model structure, structural identification, Bayesian estimation via the Sims-Zha prior, model comparison via marginal data density, error band construction, and Markov-switching extensions. Not an original research contribution but an authoritative methodological synthesis reflecting the state of the field in 2005.

Key Claims

• VARs emerged as a principled reaction to the "incredible" exclusion restrictions in Cowles Commission structural models; the key philosophical move is treating all variables as endogenous.
• Structural shocks must be mutually independent — this is what gives them economic interpretability (e.g., separating technology from policy shocks).
• The Sims-Zha (1997) Bayesian prior is the standard baseline; it prevents overfitting in large VARs without reducing model dimension.
• Normalization of the contemporaneous matrix $A_0$ is not cosmetic — incorrect normalization generates spuriously multi-modal posteriors and artificially wide error bands.
• Pointwise error bands mischaracterize impulse-response uncertainty; the Sims-Zha (1999) eigendecomposition method captures joint shape uncertainty.
• US business cycle fluctuations are dominated by variance changes in private-sector shocks, not shifts in monetary policy coefficients (Sims-Zha 2005a).
• The outstanding open challenge is VARs with cross-equation restrictions implied by Dynamic Stochastic General Equilibrium (DSGE) models.

Concepts Introduced or Extended

• Vector Autoregression
• Structural Identification
• Impulse Response Function
• Sims-Zha Prior
• Marginal Data Density
• Error Bands
• Markov-Switching VAR

Entities Mentioned

• Tao Zha
• Christopher Sims
• Daniel Waggoner

Quotes

"The philosophy of VAR modeling begins with a multivariate time series model that has minimal restrictions and gradually introduces identifying information, with emphasis always placed on the model's fit to data."

"It is sometimes argued that identified VARs are unreliable because certain conclusions are sensitive to the specific identifying assumptions. This argument is a sophism."

My Take

The paper is most valuable as a methods reference, not a literature survey — the treatment of the Sims-Zha prior, the Chib marginal data density (MDD) algorithm for restricted VARs, and the Waggoner-Zha normalization fix are the durable contributions. The framing of VARs vs. DSGE models is dated (the DSGE-VAR literature has moved significantly since 2005), but the core identification and estimation machinery remains standard.