Abstract
Collinearity is viewed as an expression of the shape of the domain defined by the regressor variables. By applying the singular value decomposition to the matrix of regressors, a geometrical figure is derived, called the “effective prediction domain” (EPD), within which prediction from the regression value is valid. The further away the prediction point is from the EPD, the less confidence one can have in the predicted value.
Additional information
Notes on contributors
John Mandel
Dr. Mandel is a Senior Statistical Consultant at the National Measurement Laboratory. He is a Fellow of ASQC.