ROC-AUC (discriminatory power) for models built only with the 3 correct predictors versus adjusted deviance explained (i.e. model fit). Model fit and discriminatory power are not always inversely correlated. The model built with "true" absences achieved high values for both. Thus ROC-AUC and adjusted deviance measure very different aspects of a model's performance and one should never be used as a surrogate for the other.