Bootstrap Inference for Impulse Response Functions in
Factor-Augmented Vector Autoregressions
Abstract
This paper investigates structural identification and residual-based bootstrap inference schemes for impulse response functions (IRFs) in factor-augmented vector autoregressions (FAVARs). I first discuss general conditions for structural identification, which also resolve the random rotation of the principal components estimates. I also provide empirically popular three such identification schemes: short-run, long-run and contemporaneous restrictions with sign restrictions. Second, two bootstrap procedures for the identified structural IRFs are compared: A) bootstrap with factor estimation and B) bootstrap without factor estimation. Although both procedures are asymptotically valid in the first-order under √T/N→0 (T and N are the time and the cross sectional dimensions), the errors in the factor estimation produce higher-order discrepancies. The asymptotic normal intervals also tend to provide smaller coverage ratios and are quite erratic. Monte Carlo simulations and an empirical example confirm the theoretical findings.
