Bayesian Inference on Fully and Partially Identified Potentially Non-Gaussian Structural Vector Autoregressions *

Abstract

We introduce a new approach to Bayesian inference in potentially non-Gaussian structural vector autoregressive models. It relies on the result that the elements of the impact matrix of the model are always at least set identified with relatively narrow bounds under standard assumptions, and, therefore, an efficient simulation algorithm should be able to explore the parameter space of the model even when identification of some (or all) of its parameters fails. Hence, it can be checked which of the shocks (if any) are point identified. To exploit deviations from Gaussianity, we recommend employing versatile error distributions and discuss their implementation in Bayesian analysis. Simulation results and an empirical application to U.S. fiscal policy highlight the usefulness of the proposed methods and lend support to efficiently accounting for non-Gaussianity.

Keywords: Non-Gaussian structural vector autoregression, identification, Bayesian methods, fiscal policy

JEL Classification: C11, C32, C51, C54, E62