In a ridge regression for an univariate linear regression model, it is common that an optimal ridge parameter is determined by minimizing an information criterion, e.g., Mallows'
Cp criterion (Mallows (1973, 1995)). Since the solution to the minimization problem of the information criterion is not expressed by a closed form, an additional computational task is required. On the other hand, a generalized ridge regression proposed by Hoerl and Kennard (1970) has multiple ridge parameters, but optimal ridge parameters are obtained by closed forms. In this paper, we extend the generalized ridge regression to a multivariate linear regression case. Then,
Cp criterion for optimizing ridge parameters in the multivariate generalized ridge regression is considered as an estimator of a risk function based on the mean square error of prediction. By correcting a bias of the
Cp criterion completely, a bias-corrected
Cp criterion named by modified
Cp (
MCp) criterion is proposed. It is analytically proved that the proposed
MCp has not only smaller bias but also smaller variance than an existing
Cp criterion and is the uniformly minimum variance unbiased estimator of the risk function. We show that the criterion has useful properties by means of numerical experiments.
View full abstract