No label defined (Q4040694)

This monograph deals with the statistical estimation and inference related to a system of equations that are linked statistically through the jointness of the error terms' distribution. Such systems are called ``seemingly unrelated regression equations'' (SURE), since the equations are unrelated structurally. A common situation which suggests a SURE specification is where regression equations explaining a certain economic activity in different geographical locations are to be estimated. Consider the linear model \[ y=x\beta +u\quad with\quad E(u)=0,\quad E(uu')=(\Sigma \otimes I_ T)=\Psi. \] On replacing the variance- covariance matrix \(\Sigma\) by a sample estimate S we obtain a feasible generalized least squares (FGLS) estimator of \(\beta\) as \[ b_ F=[x'(S^{-1}\otimes I_ T)x]^{-1}(S^{-1}\otimes I_ T)y. \] Several properties of this FGLS estimator are discussed, viz., asymptotic approximation of bias, some exact finite sample properties and the comparison of its efficiency properties with other estimators like ordinary least squares and Stein-rule estimators. Problems of inference are discussed in relation to vector autoregression models where disturbances follow SURE specification. Different cases of prior information are also analyzed. By providing suitable exercises at the end of each chapter and also simple illustrations, the monograph serves well as an advanced text.