Bayesian Inference on Fully and Partially Identified Structural Vector Autoregressions

We show that under standard assumptions the elements of the impact matrix of the structural vector autoregression (SVAR) are always at least set identified and bounded by the standard deviations of the corresponding reduced form errors. This result facilitates valid Bayesian inference without additional restrictions when only some (or none) of the columns of the impact matrix are point identified due to non-Gaussianity or heteroskedasticity. Moreover, the properties of the shocks can be assessed to find out which of them (if any) are point identified. We expand identification results put forth in the previous literature to models where all or part of the structural shocks are orthogonal but mutually dependent. To exploit deviations from Gaussianity, we propose using versatile error distributions and discuss their implementation in Bayesian analysis. Simulation results and an empirical application to U.S. fiscal policy lend support to the usefulness of efficiently accounting for non-Gaussianity