Identification and estimation of heterogeneous dynamic causal effects using local projection instrumental variable

Abstract

The magnitude of the government spending multiplier has been at the center of debates among economists and policymakers, yet there is no consensus. Such disagreement is partly due to the endogeneity of government spending. That is, not only government spending affects GDP, but also GDP affects government spending. To solve this endogeneity problem, the instrumental variables (IV) method has been used and it has gained considerable popularity recently in the literature. The IV method uses an external source of variation such as news on wars to identify the causal link from government spending to GDP. A potential drawback of the existing approach is that it treats government spending as a single random variable, ignoring the fact that it is actually the sum of sectoral spending which may have very different multipliers. In this paper, we show that the government spending multiplier identified by using the IV method is a weighted average of sector-specific multipliers and the weights depend on the instrumental variables (external source of variation). We apply our method to real US data to decompose the government spending multiplier into defense and non-defense spending multipliers. Our result shows that the non-defense spending multiplier can be larger than one (so non-defense spending increases net GDP) and persistent over time.

Short Bio

Dr Seojeong (Jay) Lee is a senior lecturer in Economics at UNSW. He received his PhD in Economics at the University of Wisconsin-Madison in 2012 after completing his MA in Economics and BAs in Economics and Political Science at Seoul National University. His research agenda is statistical inference under model misspecification. Specifically, he is interested in analysing the behavior of popular econometric methods such as the instrumental variables (IV) and the generalised method of moments (GMM) when the underlying economic model is misspecified. Dr Lee also works on cluster sampling and computationally intensive methods such as the bootstrap and the complete subset regressions.