This paper proposes a framework for undertaking bias monitoring with respect to binary classification algorithms. We present reproducible methods using a reproducible dataset rich in sensitive attribute information that can help identify both problematic bias and problematic proxies. Our demonstration uses the IPUMS International 10% random sample of the 2016 Irish Census survey courtesy of the Central Statistics Office in Ireland and centres on whether sex or ethnicity or both and their interaction significantly contribute to the prediction of our owner or renter target controlling for our social indicator of occupation and age. We develop our bias monitoring framework around this demonstration.

Our focus is global as we are interested in monitoring the overall model performance as well as the contribution of each feature rather than locally picking out individual row instances. We deploy explainable AI functions (ELI5, RFECV and SHAP) on both clear/transparent (logistic regression) and opaque/black-box (random forest classifier) models in order to assess the level of inferential agreement on the underlying dataset despite the algorithms having different predictive capabilities. Our proposed framework can be extended to any classification or regression task as it is designed to be model agnostic so long as there is access to a structured, tabular dataset. The framework is designed to be as fair as possible to practitioners whilst also providing robust bias detection that citizens can have confidence in.