Assessing Algorithmic Fairness without Sensitive Information