Estimated Generalized Least Squares for Random Coefficient Regression Models (Q4255154)

The class of models under consideration are random coefficient regression (RCR) models which allow differences in the behaviour over cross-sectional units, but show constant behaviour for a given cross-sectional unit. For the \(i\)-th cross-sectional unit, \(i=1,\dots,N\), such a model can be written as \(Y_i=X_i\beta_i+\epsilon_i\), where the response vector \(\beta_i\) is regarded as a random vector \(\beta_i=\beta+\nu_i\) with the mean \(\beta\) and the unknown covariance matrix \(E(\nu_i {\nu_i}')=\Delta\), \(E(\nu_i {\nu_j}')=0\), \(i \not=j\), \(E(\epsilon_i)=0\), \(E(\epsilon_i {\epsilon_j}')=0\), \(i \not=j\), \(E(\epsilon_i {\epsilon_i}')=\sigma_{ii}I\). The generalized least squares (GLS) estimator of \(\beta\), i.e., \(\beta^* = (X^{'}\Omega^{-1}X)^{-1}X^{'}\Omega^{-1}y\) involves the unknown covariance matrix of the error components depending on the unknown \(\Delta\) and \(\sigma_{ii}\). A new method to estimate \(\Delta\) is suggested and asymptotic properties of the GLS estimator based on the proposed new estimate \(\Delta^*\) are established. Procedures for testing the linear restrictions on the mean vector are derived. Finally, a test for non-randomness in RCR models is discussed and its asymptotic distribution is obtained.