What Impulse Response Do Instrumental Variables Identify?

with Bonsoo Koo (Monash University) and Myung Hwan Seo (Seoul National University)

Macro shocks are often composites, yet their implications are overlooked in impulse response analysis. When an instrumental variable (IV) is used to identify a composite shock, it violates the common IV exclusion restriction. We show that the Local Projection-IV estimand is represented as a weighted average of component-wise impulse responses, but with possibly negative weights, which occur when the IV and shock components have opposite signs of correlation. An LP-IV estimand with negative weights does not have a causal interpretation. However, when appropriately combined with other LP-IV estimands, such a non-causal LP-IV estimand can provide identification of the component-wise impulse responses. We develop identification strategies based on additional granular information or sign restrictions. In contrast, conventional approaches combining multiple instruments, such as two-stage least squares, do not provide identification of structural parameters. Our applications confirm the composite nature of monetary policy shocks and reveal a non-defense spending multiplier exceeding one.

• pdf (Apr 2024)

• arXiv (Aug 2023 version)

• The following figure is Figure 6 from the paper. On the left panel, we present the IV estimates of the cumulative government spending multiplier using the narrative military news shock of Ramey and Zubairy (2018, JPE) and the current defense spending following Blanchard and Perotti (2002, QJE) as an IV one at a time. The IV-estimated multipliers are consistent with those reported in Ramey and Zubairy (2018). On the right panel, we decompose these IV estimates into the cumulative sectoral (non-defense and defense) spending multipliers using the method proposed in the paper. Interestingly, the non-defense spending multiplier is larger than one after four years, even though the IV estimates on the left panel are well below one.