Limited information estimation and evaluation of dynamic macroeconomic models

Dynamic models provide powerful tools for modern macroeconomics. Consequently, economists desire to obtain accurate and stable model parameter estimates so that any subsequent analysis is reliable. Since most dynamic macroeconomic models are highly nonlinear and linear approximation is a common strategy, the performance of these methods in this situation should be an important issue to both researchers and practitioners. On the other hand, many recent studies show that heterogeneity across sectors, firms or individuals plays an important role in explaining macroeconomic facts. Since most previous empirical studies were based on aggregate series, it would be important to incorporate disaggregate information into macroeconomic estimation so that potential heterogeneity can be addressed. In addition, predictive ability of a model draws more attention in recent years since either academic researchers or policy makers desire to build or choose a model that can provide forecasts in good quality. Therefore, my dissertation attempts to shed some light on these issues.;The first chapter employs a stylized dynamic stochastic general equilibrium (DSGE) model used in An and Schorfheide (2007) to compare the most widely used estimators, including MLE, GMM and Bayesian. The model is solved by first-order, second-order and third-order approximations, then artificial data are generated based on these solutions. Monte Carlo simulation evidence suggests that MLE and Bayesian generally perform better than GMM, but good guess of initial values in the numerical routines and very tight priors are crucial to success of full information methods. When the model is dynamically misspecified, the bias of an estimator is determined by the difference between the first order solution dynamics and the higher order ones. The evidence also suggests that the performance of GMM could be improved by fixing some troublesome parameters, but this improvement depends on the structure of the model. On the other hand, this strategy does not work for Bayesian method when priors are not very informative.;The second chapter focuses on the evaluation of a structural equation, the so-called New Keynesian Phillips curve (NKPC), using disaggregate data. This approach helps to address possible heterogeneity across sectors in macroeconomic analysis. Panel data evidence based on productivity data for 462 U.S. manufacturing industries strongly supports the relative importance of the forward-looking term in NKPC as well as a limited role of intrinsic inflation persistence. Therefore, it may not be appropriate to treat intrinsic persistence as the deep structure of the economy. The evidence also suggests that only a small fraction of the firms adopt a backward-looking rule to set prices. This confirms that forward-looking price setting is prevalent among U.S. industries. Unlike previous studies based on aggregate series, unit labor cost is not the appropriate proxy of real marginal cost since it only accounts for a small fraction of total cost and is countercyclical in the industry-level data. The analysis suggests that unit material cost works better in the NKPC framework. In addition, the main conclusions are robust to different measures of the driving variable and additional instruments used in the estimation procedures.;The final chapter further explores which source of inflation persistence, intrinsic or extrinsic, is more important by forecast-based evaluation. To identify the source of persistence in each model, the driving force in the purely forward-looking NKPC follows an AR(1) process while the driving variable in the hybrid model does not have any persistence. Thus the source of inflation persistence in the hybrid model is only intrinsic, and the source in the purely forward-looking one is solely extrinsic. Two models are then used to generate one-period-ahead and path forecasts. Both root mean squared error (RMSE) statistics and predictive ability test based on Giacomini and White (2006) generally indicate that extrinsic persistence is more helpful in prediction. Thus the importance of intrinsic persistence may not be emphasized in macroeconomic analysis.