Generalized multivariate ridge regression estimate and criteria Q(c) for choosing matrix \(K^*\) (Q1261506)

This paper considers the standard multivariate linear regression model \(Y= XB+\varepsilon\) where \(Y\) is an \(n\times q\) matrix, \(X\) an \(n\times p\) full rank matrix of known constants, and \(B\) a \(p\times q\) matrix of unknown coefficients to be estimated. When multicollinearity is present, a generalized ridge estimate \(\beta(K)\) of the regression coefficient \(\beta=\text{vec} (B)\) is considered. The MSE of the above estimate is less than the MSE of the least squares estimate by choosing an appropriate ridge parameter matrix \(K\). It is pointed out that the MSE criterion for choosing \(K\) has several weaknesses. In order to overcome these weaknesses a new family of criteria \(Q(c)\) is proposed. Some good properties of the criteria \(Q(c)\) are proved and discussed from the theoretical point of view. The statistical meaning of the scalar \(c\) is explained and methods of determining \(c\) are also given.