Zubin Jelveh and Michael Luca
Prediction algorithms are increasingly used to forecast outcomes of processes that are societally sensitive. In response, algorithms have been developed to produce fair classifications but at the potential cost of accuracy. In this work, we present a framework for modeling the pathways by which sensitive variables influence—and are influenced by—nonsensitive variables. These pathways allow us to discern between two types of accuracy loss: justified reduction due to underlying discrimination in the data, and overadjustment due to the removal of nonsensitive predictive information. We also present a framework for adjusting input data to remove the association between sensitive and nonsensitive predictors and assess its ability to produce fair classifications. Finally, we apply our methodology to a new dataset in the criminal justice domain.