Estimation of the constant term when using ridge regression

Abstract

In most applications of ordinary ridge regression the usual practice is to estimate the constant term in the model by least squares. This contrasts with the theoretical development of ridge regression in which all regression coefficients are treated in the same way. A variety of results are given which support the use of the response sample mean to estimate the constant term when the predictor data matrix is in correlation form. This conclusion also applies to generalized ridge regression.