Essays on non-Gaussian structural vector autoregressions

This dissertation develops new identification and estimation methods for Non-Gaussian Structural Vector Autoregressions to address two key challenges in empirical macroeconomics: the identification of structural shocks and the detection of structural breaks in macroeconomic relationships. Traditional SVAR models rely on economic theory-based restrictions and impose constant contemporaneous causal effects over time. These assumptions may not hold in practice, leading to potential biases in empirical applications.

The first essay introduces a novel identification strategy that exploits the non-Gaussian properties of structural shocks. Specifically, it demonstrates that when at least \( n - 1 \) structural shocks exhibit excess kurtosis, the impact matrix is globally identified with a small number of suitably selected moment conditions. This approach provides an alternative to traditional exclusion and sign restrictions, enabling identification through statistical properties alone. An empirical application to an SVAR model of the global oil market confirms the effectiveness of this method in distinguishing between supply and demand shocks.

The second essay advances partial global identification in non-Gaussian Structural Vector Autoregression (SVAR) models, showing that structural shocks with skewness can be identified when co-skewness is absent, while those with excess kurtosis are identified when co-kurtosis is zero. The former case is particularly useful as it accommodates dependent conditional heteroskedasticity, a common feature in macroeconomic data. In both scenarios, the remaining shocks are only set-identified, and these results can be combined to identify both skewed and non-mesokurtic shocks. To account for non-Gaussian characteristics, the framework employs flexible error distributions. A Bayesian SVAR model with skewed t-distributed shocks and dependent stochastic volatility is developed, enabling an evaluation of identification conditions and the validity of external instruments. The methodology is applied to U.S. monetary policy, demonstrating its ability to identify structural shocks and enhance inference in macroeconomic research.

The third essay focuses on detecting and estimating structural breaks in SVAR models. Macroeconomic relationships evolve over time due to policy changes, financial crises, and technological advancements, yet standard SVAR models assume constant contemporaneous causal effects. This dissertation introduces a Partial Sample Generalized Method of Moments (PSGMM) estimator, which endogenously detects structural breaks without requiring pre-specified break dates. Applying this method to U.S. monetary policy data from 1954 to 2024 identifies two major structural shifts: the Volcker disinflation period in April 1981 and the transition to unconventional monetary policies in November 2008. These findings highlight the importance of accounting for structural instability in empirical macroeconomic modeling.

Contact Keyan Liu

Email: keyanricky@gmail.com
Home page: https://keyanliu1.github.io/