Addressing Observational Biases in Algorithmic Fairness Assessments