« Back to Results
Estimation of Dynamic Causal Effects in Macro: Promises and Pitfalls
Friday, Jan. 7, 2022
10:00 AM - 12:00 PM (EST)
American Economic Association
University of California-Berkeley
Structural Vector Autoregressions with Imperfect Identifying Information
The problem of identification is often the core challenge of empirical economic research. The traditional approach to identification is to bring in additional information in the form of identifying assumptions, such as restrictions that certain magnitudes have to be zero. In this paper we suggest that what are usually thought of as identifying assumptions should more generally be described as information that the analyst had about the economic structure before seeing the data. Such information is most naturally represented as a Bayesian prior distribution over certain features of the economic structure.
What Can We Learn from Sign-Restricted VARs?
I use a simple business-cycle model to illustrate the workings and limitations of sign restrictions in Structural Vector Autoregressions. Three lessons emerge. First, identification through sign restrictions on impulse responses is vulnerable to “shock masquerading”: linear combinations of other shocks may be mis-identified as the shock of interest. Second, since the popular Haar prior automatically over-weights more volatile shocks, the masquerading problem is particularly severe if the shock of interest does not matter much for business-cycle fluctuations (e.g., monetary policy). Third, adding sign restrictions on structural elasticities — rather than just on impulse responses — can substantially sharpen identification.
Signing Out Confounding Shocks in Variance-Maximizing Identifications
A recent literature has explored the dominant drivers of long-run and business-cycle dynamics of macroeconomic variables using SVARs that rely on variance-maximizing rules for identification. However, identification performance is poor when shocks other than the target of interest also play a large role in driving volatility at the targeted horizon or frequency. The result is that these identifications can capture a hybrid shock rather than a dominant shock (Dieppe et. al. 2021). We suggest a simple enhancement to the identification procedure that reduces the influence of confounding shocks. That fix is to include theoretically-informed sign or exclusion restrictions, if available, in the identification stage of the vector auto-regression.
SVAR Identification from Higher Moments: Has the Simultaneous Causality Problem Been Solved?
Two recent strands of the literature on Structural Vector Autoregressions (SVARs) use higher moments for identification. One of them exploits independence and non-Gaussianity of the shocks; the other, stochastic volatility (heteroskedasticity). These approaches achieve point identification without imposing exclusion or sign restrictions. We review this work critically, and contrast its goals with the separate research program that has pushed for macroeconometrics to rely more heavily on credible economic restrictions and institutional knowledge, as is the standard in microeconometric policy evaluation. Identification based on higher moments imposes substantively stronger assumptions on the shock process than standard second-order SVAR identification methods do. We recommend that these assumptions be tested in applied work. Even when the assumptions are not rejected, inference based on higher moments necessarily demands more from a finite sample than standard approaches do. Thus, in our view, weak identification issues should be given high priority by applied users.
E0 - General
C0 - General