Designing Against Bias: Identifying and Mitigating Bias in Machine Learning and AI

Bias in machine learning algorithms is one of the most important ethical and operational issues in statistical practice today. This paper describes common sources of bias and how to develop study designs to measure and minimize it. Analysis of disparate impact is used to quantify bias in existing and new applications. New open-source packages such as Fairlearn and AI Fairness 360 Toolkit quantify bias by automating the measurement of disparate impact on marginalized groups, offering great promise to advance the mitigation of bias. These design strategies are described in detail with examples. Also, a comparison algorithm can be developed that is designed to be fully transparent and without features subject to bias. Comparison to this bias-minimized model can identify areas as bias in other algorithms.