Complete and Partial Identification of the A- and B-Models in the Context of Heteroskedastic SVARs
15 Pages Posted: 21 Aug 2014
Date Written: August 20, 2014
Abstract
We investigate the identification problem in the setting of heteroskedastic structural vector autoregressions (SVARs) under two alternative normalizations: i) the A-model where a unit diagonal is imposed on the matrix of contemporaneous parameters, and ii) the B-model when the unconditional structural variance matrix is the identity matrix. We prove that both models are identified under the same condition of linear independence of conditional variance equations, which may take any form. The case of partial identification is also considered. If, out of the total of K structural innovations, there are 0 < k < K - 1 shocks which exhibit linearly independent heteroskedasticity while the remaining K - k innovations are homoskedastic, then at least k rows (columns) of the contemporaneous parameter matrix in the A-model (B-model) are identified. A simulation study demonstrates the results for a B-model with GARCH innovations.
Keywords: Identification via heteroskedasticity, structural vector autoregression
JEL Classification: C30, C13
Suggested Citation: Suggested Citation