Impulse Response Priors for Discriminating Structural Vector Autoregressions

The structural vector autoregression (SVAR) has become a central tool for research in empirical macroeconomics. Because the vast majority of these models are exactly identified, researchers have traditionally relied upon the informal use of prior information to compare alternative specifications. This paper surveys some of the structural dynamic restrictions used to evaluate SVARs. I provide a method for constructing prior distributions that incorporates this information on impulse responses. Based upon these Impulse Response Priors (IRPs) I employ a formal Bayesian model selection procedure for comparing SVAR specifications. I use this procedure to compare several alternative, six variable SVAR models of the interaction of real and monetary sectors of the U.S. economy. I make these comparisons under a variety of assumptions regarding the nature of the money supply rule, and lag length. Emprically, I find strong evidence in favor of interpreting shocks to the federal funds rate as monetary policy shocks, as opposed to shocks to nonborrowed reserves. The most favored identification is one in which monetary policy reacts to contemporaneous movements in real variables and the price level. There is less evidence that monetary policy reacts as quickly to fluctuations in money demand.