When Relative and absolute information matter: compositional predictor with a total in generalized linear models