A comparison study of least squares and ridge estimators in the presence of heteroscedasticity and multicollinearity under normal and nonnormal disturbances

In this chapter the sample properties of the least squares (LS) and of some ridge estimators (and predictors) are studied for alternative models of heteroscedasticity at various levels of multicollinearity, under normal and non-normal disturbances with small and large variances. The present simulation study shows that when the regression coefficient vector β is aligned to the normalized eigenvector corresponding to thelargest eigenvalue of the X’X matrix the Generalized Cross-Validation estimator in almost all cases is superior to the other estimators examined in the study. It is also “confirmed” that the LS estimator exhibits a poor performance independent of the level of multicollinearity for all the examined criteria with the notable exception of the multiple correlation coefficient.