Whence Is A Model Fair? Fixing Fairness Bugs via Propensity Score Matching