Essays on Identification and Estimation of Dynamic Stochastic General Equilibrium Models.

Abstract

The estimation of dynamic stochastic general equilibrium (DSGE) models is the subject of a rapidly growing literature. This dissertation contributes to the existing body of work by focusing on issues related to parameter identification.

In the first essay I show that DSGE models are characterized by a set of cross-equation and covariance restrictions, which can be used to determine the identifiability, and estimate the parameters of such models. I derive conditions for identification, and propose a two-step minimum distance method for estimating the parameters of DSGE models. I show that the estimator is asymptotically efficient, and provide simulation evidence that it has good small sample properties.

In the second essay I show how the Information matrix of DSGE models can be evaluated analytically. This is achieved by a factorization of the matrix as a product of two terms: the Jacobian matrix of the mapping from deep to reduced-form parameters, and the Information matrix of the reduced-form model. I show that both terms can be derived analytically. This result is useful for the estimation of DSGE models, as well as for detecting identification problems.

In the third essay I develop a methodology for analyzing identification in linearized DSGE models. Specifically, I show how to address the following questions: are the parameters of the model identifiable; how strong is identification; if there are identification problems, do they originate in the model or in the data; which parameters are not well-identified and why. I apply this methodology to study identification of a model estimated in Smets and Wouters (2007). I find that identification is generally very weak, and the problems are largely in the structure of the model, and thereby cannot be resolved by using more informative data. I estimate the model with maximum likelihood, and find substantial differences with the estimates obtained with Bayesian methods. I conclude that the use of DSGE models for policy analysis should be done with caution since the results are likely to be strongly influenced by the prior distribution.