Selecting a Regression Estimator with Integrated Mean Squared Error as a Criterion.

摘要

Integrated mean squared error is employed as a criterion for choosing an estimator of a multiple linear regression model when the regressor variables are multicollinear. Three estimators of the regression coefficients are examined: ordinary least squares, principal components regression, and ridge regression. The integrated variance and integrated squared bias of the corresponding prediction equations are evaluated for a general class of weight functions. Comparisons of the predictors are made on the basis of integrated variance and squared bias separately and combined as integrated mean squared error.