When evaluating decisions founded on data driven procedures, it is crucial to verify that the decisions align with the procedures, to ensure fairness for those affected by the outcomes and to maintain trust. For decisions for which there is a set of ordinal outcomes, fairness considerations extend beyond considering parity in appropriate outcomes (true positives) across different groups. It is also important to consider how inappropriate outcomes (non-true positives) can result in differing consequences depending on if, and by how much, the outcome is an overestimate (falls on a higher position in the ordinal set of outcomes) or an underestimate (falls on a lower position in the ordinal set of outcomes).
In this work, we analyze the fairness of historical decisions through investigating inconsistencies between inappropriate outcomes for different groups in terms of underestimates and overestimates. For a pair of groups, we train a classification model using data from one group, and then test the model with respect to data for a second group. We compare the results to the inverse, a model trained on the second group and tested on the first group. From this, any variations between the models' tendencies to underestimate or overestimate, compared to the actual historical outcomes, can be uncovered. We use a variety of algorithmic models to minimize the potential for any uncovered tendencies to arise simply due to reliance on a single type.
We explore an application case study of the United Kingdom government's Covid-19 tiered restriction decisions, which took a localized strategy to allocate different tiers of restrictions of movement to different geographic areas based on each's data. Here, the different tiers represent a set of ordinal decision outcomes. The decisions sparked public debate regarding their transparency and fairness, and the application of our approach reveals inconsistencies in how the North and the South of England were treated.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)