Fairness tradeoffs may appear rigid when the data used to measure them are treated as fixed. Yet, in many settings, today’s decisions impact the data on which tomorrow’s decisions rely. I study what that feedback implies for the tradeoff highlighted by the algorithmic fairness impossibility theorem. Focusing on policing, I develop a model in which stops lead to arrests, justifying more stops in certain neighborhoods. Under diminishing returns, this selective over-enforcement inefficiently inflates arrest incidence, lowers arrest yield (the fraction of stops resulting in arrest) and increases false-positive (FP) burden independently of latent crime levels. A policy disruption can then improve policing efficiency while expanding the full administrative fairness frontier. I study this mechanism using the sequence of political shifts that ended New York City’s stop-and-frisk program. In historically over-policed neighborhoods, arrest and weapons yield increased sharply, while stop-generated arrest incidence, weapons incidence, and FP stop incidence all fell sharply, with no increase in reported crime. Under common true positive rates, the implied racial FP gap narrowed as well. When institutions learn from data impacted by their own past actions, feedback-disruptive policy can enlarge the scope for improving equity and efficiency simultaneously.

Prof. McMillon, an economist and UM math alum, will give a talk about his research and his career path. His research interests include systemic discrimination, educational inequality, criminal dynamics, algorithmic fairness, and the political economy of reparative reforms.

The talk will be followed at 4pm by a reception in the atrium.