Using canonical correlation for hydrological predictions

Abstract

Hydrologists often encounter analyses involving multiple dependent variables. Previously it had been suggested that canonical correlation might be used as a suitable multivariate analogue to multiple regression in such situations (Rice, 1967). However, in order to use canonical correlation [or prediction, all roots must be retained. This study of the use of canonical correlation for prediction purposes led to the discovery that, when all roots are retair.ed, a canonical correla1ion solution is mathematically equivalent to a set of multiple regression solutions. Consequently, multiple regressions are preferable because they give the hydrologist more latitude in the choice of variables in his prediction equations. Results obtained in this study suggest that the D2 statistic is not a suitable descriptor of the differences between hydrographs described by the parameters of a fitted gamma distribution. © 1972 Taylor & Francis Group, LLC.