Improved prediction in the presence of multicollinearity

In this paper Monte Carlo techniques are used to examine the performance of several estimators for the linear statistical model under a squared error of prediction loss measure when the data are multicollinear. Under this measure of performance the Stein-like rules that shrink toward the principal components estimator perform very well relative to other minimax estimators for alternative specifications of the characteristics root spectrum. The sampling performance of a non-minimax pretest rule is also considered.