Misspecification and the small sample properties of econometric estimators.

The purpose of this paper is to examine the effect of a variety of misspecifications broadly defined to include both over- and underspecifications upon three econometric estimators (i.e., ordinary least squares (DLS) on the structural equations, two stage least squares (2SLS) and limited information single equation (LISE) maximum likelihood). The analysis which follows in this paper differs in several important respects from that of Cragg. First, our Monte Carlo experiments have been conducted with an economically relevant model following the suggestions of SMITH (1971) in earlier work. Second, our application of 2SLS and LISE assumes knowledge of all the pre-determined variables which are arguments in the 'true' unrestricted reduced form. In CRAGG'S (1968) paper he suggests that with such single equation estimators a misspecification involving only one equation will not alter these techniques' performance patterns for the other equations in the model (CRAGG, 1968, p. 65). However, if the misspecification calls for the omission of a pre-determined variable and the variable appears only in the equation in question, then the first stage instruments for the right hand side endogenous variables constructed with 2SLS and LISE will (a) not satisfy the properties for consistent estimators with this equation,[3] and (b) not provide efficient estimates for all equations of the model (see MCCARTHY, 1971). [ABSTRACT FROM AUTHOR]