Prediction using regression models with multicollinear predictor variables

Linear regression models are widely used for forecasting and prediction of new observations from the underlying modeled process. This article explores the use of regression models in this context when the regressor or predictor variables exhibit multi-collinearity, or near-linear dependence. It is shown that multicollinearity can severely impact the predictive performance of a regression model and that biased estimation methods can be an effective countermeasure when multicollinearity is present. Several biased estimation methods are described and evaluated, including a new method for selecting the biasing parameter in ordinary ridge regression. A simulation study is performed to provide some guidelines for the choice of an estimation method.