A Generalized View on Continuum Regression (Q4255153)

This article focuses on investigation of interrelationships between such important regularized regression methods as continuum regression (CR) and ridge regression (RR) used for analysis of data with near collinearity in explanatory variables. The CR approach by \textit{M. Stone} and \textit{R. J. Brooks} [J. R. Stat. Soc., Ser. B 52, No.2, 237-269 (1990; Zbl 0708.62054)] is wide enough to include ordinary and partial least squares regression as well as principal component regression as special cases. The authors demonstrated that RR is in a certain sense an even more fundamental method than CR by showing that any reasonable maximization criterion which depends only on regression sample correlation and sample variation will yield a regressor that is proportional to a ridge regressor. Based on this result, a generalized formulation of CR is proposed. It is proved that for this formulation the coefficient of determination and length of estimators are decreasing functions of the continuum parameter. Applications to real data sets are presented.