UC San Diego

This dissertation contains three essays exploring how macroeconomists can identify and estimate dynamic causal effects in models where researchers have doubts about identifying assumptions.

Chapter 1 proposes a new Markov Chain Monte Carlo algorithm to estimate a sign-restricted structural vector autoregression on time series that are subject to regime shifts. My approach can incorporate useful prior information about both model parameters and hidden states while transparently imposing sign restrictions. I illustrate my method by revisiting the literature on asymmetric effects of conventional monetary policy during recessions and expansions. My evidence suggests that previous empirical research found asymmetric effects by questionable identification schemes and neglecting changes in the variances of structural shocks. I find little difference in the structural parameters, and thus I do not find evidence of asymmetry.

Chapter 2 studies the method of instrumental variables in set-identified models. I develop a proxy structural vector autoregression in which prior information from both theory and the empirical literature is incorporated about signs and magnitudes of certain parameters and equilibrium impacts. I use my method to investigate the relevance and validity of three popular instruments for monetary policy shocks, developed by Romer and Romer (2004), Sims and Zha (2006), and Smets and Wouters (2007). I find that all of them are strongly relevant but only that of Smets and Wouters is valid. Furthermore, the empirical analysis demonstrates that my framework can combine information from a relevant and valid instrument with prior information about sign restrictions to improve inference about structural impulse-response functions.

Chapter 3 develops new methods to study dynamic causal effects in a data-rich environment. Current development in high-dimensional statistics fails to address the main interest of economists: causal inference with credible assumptions. I first review the literature on high-dimensional linear regression models and dynamic factor models. Then, I develop several new Bayesian numerical algorithms that combine the techniques in high-dimensional statistics with recent advances in dynamic causal inference. In particular, I discuss how to make causal statements from a high-dimensional structural model when researchers have doubts about identifying assumptions. Finally, I extend those algorithms to the case of Markov-switching models to accommodate nonlinearities in economic time series.