When relative and absolute information matter: Compositional predictor with a total in generalized linear models