Improving discrimination in data envelopment analysis: some practical suggestions