Data‐driven identification in SVARs—When and how can statistical characteristics be used to unravel causal relationships?

Abstract

Structural vector autoregressive analysis aims to trace the contemporaneous linkages among multiple economic time series back to underlying orthogonal structural shocks. Traditionally, researchers rely on economically motivated restrictions to identify these shocks. However, in the presence of heteroskedasticity or non‐Gaussian independent components, only these statistical properties allow a locally unique identification. In this paper, we compare alternative statistical identification procedures under distinct covariance changes and distributional frameworks. We find that statistical identification schemes are robust under distinct data structures to some extent and support researchers in detecting shocks that feature an economic underpinning. The detection of independent components appears most flexible.

How certain are we about the role of uncertainty in the economy?

“…In sum, we consider three distinct identification strategies, that is, (i) economically‐motivated narrative event and correlation constraints from the benchmark study of LMN, (ii) identification based on changes in the covariance structure (Rigobon, 2003) in the vein of ABCF, and (iii) ICA‐based non‐Gaussian shocks as suggested by Matteson and Tsay (2017). Approaching the identification problem by means of ICA is motivated by (i) a minimal number of assumptions required for uniqueness of the structural model, (ii) strong deviations of uncertainty shocks from Gaussianity as documented, for example, in LMN, and (iii) the robust performance of ICA‐based identification within a rich variety of SVARs generated from heteroskedastic or non‐Gaussian distributed model residuals (Herwartz, Lange, & Maxand, 2022). 7 Overall, these circumstances make principles of ICA particularly well suited to identify uncertainty shocks, and to critically reassess core findings of both benchmark studies.…”

“…Meanwhile the econometric literature has developed a variety of specific ICA‐based identification schemes (see, e.g., Lanne et al., 2017; Moneta et al., 2013). In their large scale simulation study Herwartz, Lange, and Maxand (2022) provide guidance for the choice of a most promising identification method in the realistic case that actual data features are heterogeneous and unknown, and economic theory is not sufficiently conclusive to deliver fully convincing external information. With this background and noticing the conflicting theoretical considerations delivered in the benchmark studies of LMN and ABCF, we opt for the flexible and agnostic identification approach of Matteson and Tsay (2017).…”

“…Kilian and Lütkepohl (2017) provide a textbook treatment of alternative branches of identification in SVARs with an embedding of data‐based identification schemes such as ICA‐based approaches. Herwartz, Lange, and Maxand (2022) provide a large scale simulation study of alternative identification techniques applying under heterogeneous data generating processes. These authors conclude that the identification approach suggested by Matteson and Tsay (2017) performs remarkably robust and is particularly helpful if an analyst lacks trustworthy information on the actual data generating process.…”

Do Energy Efficiency Improvements Reduce Energy Use? Empirical Evidence on the Economy-Wide Rebound Effect in Europe and the United States

“…Journal Pre-proof studied in the econometric literature in the context of SVAR models (Herwartz, 2018). Herwartz et al (2021) show that distance covariance is the most robust of various approaches to data-based identification of SVARs, though nGML performs better if the shocks are actually t-distributed and homoskedastic. To test the robustness of our results, we also compute the Choleski decomposition of the residual variance matrix, which gives similar results (see Table E.7 for a comparison of the different rebound estimates).…”

Identification of Structural VAR Models Via Independent Component Analysis: A Performance Evaluation Study

“…Herwartz (2018) also undertakes a performance evaluation analysis, with a focus, however, on the discriminatory power of several identification schemes in detecting structural shocks embedded in a simple DSGE model for the Euro Area. Herwartz et al (2022) compare different ICA methods with identification by heteroskedasticity. Our paper shares some features with these simulation studies, but it introduces the following novelties.…”