Predicting Future Criminal Justice Involvement in the Face of Social Change

• In Event: New Advances in the Prediction of Crime and Recidivism

Abstract

Risk assessment instruments (RAIs), including those that utilize machine learning algorithms, are increasingly prevalent in criminal justice, healthcare, and social services. In this paper, we argue that the risk assessment enterprise has proceeded with little to no incorporation of a crucial point about the future: as time goes on it is not just individuals who age and develop, but societies themselves which contemporaneously change. Using data from the Project of Human Development in Chicago Neighborhoods we test the implications of this argument and show the consequences of this common assumption’s failure on the calibration, precision, and accuracy of models trained to predict a key indicator of future criminal justice involvement as measured by arrest.

Authors

• Erika Montana, Carnegie Mellon University

• Alexandra Chouldechova, Carnegie Mellon University

• Dan Nagin, Carnegie Mellon University

• Roland Neil, University of Pennsylvania

• Robert J. Sampson, Harvard University