Optimization hierarchy for fair statistical decision problems

Data-driven decision making has drawn scrutiny from policy makers due to fears of potential discrimination, and a growing literature begun develop fair statistical techniques. However, these techniques are often specialized one model context based on ad hoc arguments, which makes it difficult perform theoretical analysis. This paper develops an optimization hierarchy, is sequence problems with increasing number constraints, for problems. Because our hierarchy the framework problems, this means provides systematic approach developing studying versions hypothesis testing, making, estimation, regression, classification. We use insight that qualitative definitions fairness equivalent independence between output technique random variable measures attributes desired. construct lends itself numerical computation, we tools variational analysis set theory prove higher levels lead consistency in sense asymptotically imposes as constraint corresponding demonstrate effectiveness using several data sets, fairly automated dosing morphine.