Addressing Fairness in Prediction Models by Improving Subpopulation Calibration