Data Bias, Algorithmic Fairness and Evaluating Discriminatory Impacts in Predictive Policing

• Activity Number: 50 - Statistical Methods Applied to Discrimination: Recent Contributions from the Journal
• Type: Invited
• Date/Time: Sunday, July 28, 2019 : 4:00 PM to 5:50 PM
• Sponsor: Statistics and Public Policy
• Author(s): P. Jeffrey Brantingham* and George Mohler
• Companies: UCLA Department of Anthropology and Indiana University – Perdue University Indianapolis

Abstract

The move towards data-driven policing has been underway for several decades, but recently algorithmic methods have become more complex and automated in day-to-day police decision making. The use of predictive policing requires added scientific and policy scrutiny because of the special powers that police hold in society and the potential for discriminatory impacts. We discuss a comprehensive model based on the life-cycle of crime event data from crime reporting, through analysis, police decision-making, police actions in the field and evaluation of outcomes. The comprehensive model considers the origins and extent of bias in police data, how to build fairness directly into algorithmic methods and approaches to evaluating the outcomes of predictive policing in the field. We use real-world data drawn from observational and experimental studies in Los Angeles and Indianapolis.