Multicollinearity’s Effect on Regression Prediction Accuracy with Real Data Structures

Abstract

Recommendations from popular statistics texts regarding avoidance of predictor variable multicollinearity in the use of multiple regression are considered from the perspective of the alternate purposes of explanation and prediction. As opposed to prior studies that consider the effect of multicollinearity on prediction accuracy by varying a constant proportion eigenvalue decrement, a method for manipulating multicollinearity while maintaining a real data set’s eigenvalue structure is used. For 21 data sets examined, it is shown that multicollinearity has no effect in respect to either relative or absolute prediction accuracy.