On the Origins of Sampling Bias: Implications on Fairness Measurement and Mitigation