Estimation and Prediction in Interdependent Systems in the Presence of Specification Errors

Abstract: Mult i relät ional models involving simultaneously dependent linear equations are often employed in prediction and regulation of economic systems. The prediction purposes focus their attention on the reduced form of the model, relating each endogenous variable to exogenous variables, predicted outside the system. Different principles for estimating the reduced form parameters have been suggested. These concentrate to a different extent upon sample information and structural a priori information. Knowledge of the properties <bf predictions based on different principles of reduced form estimation is essential when choosing between competing estimation techniques. Reduced form estimation is often subject to many potential sources of specification error. In this work different types of specification errors, essentially due to incorrect omission of exogenous and endogenous variables, are studied in the context of reduced form estimation. A strategy of estimating reduced form parameters in the case of structural forms characterized by heterogenous specification quality is proposed and evaluated analytically as well as by a Monte Carlo study. The Monte Carlo study supports the conjecture that this strategy is better (when considering prediction quality measured by Ball's Q ) than purely direct (unrestricted) or indirect (restricted) reduced form estimation.