On the Within-Group Fairness of Screening Classifiers