Koop (1992b) Aggregate Shocks and Macroeconomic Fluctuations: A Bayesian Approach

Summary

This paper develops a Bayesian treatment of structural VAR analysis of aggregate shocks. Instead of reporting point estimates of a series' response to an aggregate shock, Koop computes the entire posterior density of any feature of interest — impulse responses, variance decompositions — using Monte Carlo integration and Gibbs sampling. Crucially, the identifying restrictions that give the shocks a structural interpretation are imposed in a way that propagates the uncertainty about those restrictions into the posterior, unlike classical least-squares-plus-bootstrap procedures that treat the restrictions as exactly known. Applied to the Blanchard-Quah (1989) and Lee-Pesaran-Pierse (1992) models, the headline finding is that Bayesian measures of uncertainty — posterior standard deviations — are substantially larger than their classical counterparts, and often markedly skewed.

Key Claims

• Posterior densities, not point estimates. For structural objects such as impulse responses and variance decompositions, the method delivers exact finite-sample posterior densities of virtually any feature, evaluated by Monte Carlo or Gibbs sampling — a range of tasks classical bootstrap methods cannot match.
• Structural VAR identification. From the Wold moving-average form $X_t=\sum_j C_j u_{t-j}$ (reduced-form innovations $u_t$, covariance $\Omega$) to the structural form $X_t=\sum_j A_j e_{t-j}$ with $A_j=C_j A_0$ and $\Omega=A_0A_0'$: because $\Omega$ is symmetric, its estimate is not enough to recover $A_0$, so economic identifying restrictions are required.
• Uncertainty in the restrictions is carried through. The Bayesian scheme imposes the identifying restrictions while treating them as uncertain, so the posterior for impulse responses reflects doubt about the identification — an improvement over classical methods that condition on the restrictions being exactly true.
• Two applications. The Blanchard-Quah model (long-run restriction: demand shocks have no permanent effect on output) is analysed by Monte Carlo integration; the Lee-Pesaran-Pierse model by Gibbs sampling from the posterior.
• Bayesian uncertainty is larger. Posterior standard deviations exceed classical standard errors and the posteriors show substantial skewness — classical inference understates the true uncertainty about aggregate-shock effects.

Concepts Introduced or Extended

• Bayesian SVAR Identification
• Structural Identification
• Impulse Response Function

Entities Mentioned

• Gary Koop

Quotes

"Rather than calculate point estimates of the response of a time-series to an aggregate shock, we calculate the whole probability density function of the response."

"The proposed techniques impose identification restrictions in a way that includes the uncertainty in these restrictions, and thus are an improvement over traditional approaches that typically use least-squares techniques supplemented by bootstrapping."

My Take

This is a notable early statement of the idea that later matures in Baumeister-Hamilton (2019): identifying restrictions are not facts to condition on but objects carrying their own uncertainty, and a Bayesian posterior is the natural place to register that. Koop (1992b) does not yet put rich informative priors on structural elasticities — it works within the Blanchard-Quah long-run-restriction tradition — but the methodological point, that honest error bands for impulse responses must widen once identification uncertainty is admitted, is exactly the same, and the empirical payoff (posterior SDs larger and skewed relative to classical bootstrap intervals) prefigures the modern literature. For the wiki it is the Bayesian-SVAR bridge between the BVAR forecasting tradition (Doan-Litterman-Sims) and structural impulse-response analysis.