•  
  •  
 

Abstract

Recidivism prediction instruments (RPI) are increasingly employed as an algorithmic arm of the criminal justice system to assist in bond, sentencing, and probation determinations. Due process and ethics guidelines require that RPIs, such as the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS), apply a fair assessment in making recidivism predictions. To establish fair RPIs, we must first define standards of fairness and methods to measure bias in RPI outputs. However, defining algorithmic fairness can be cryptic and context-dependent. Scholars have proposed different criteria for what defines a fair RPI. Unfortunately, these criteria are conflicting and mutually exclusive; it is mathematically impossible for a system to satisfy all of them simultaneously. Thus, to ensure fairness in RPIs, we need to separate the wheat from the chaff and identify the most suitable fairness criterion based on philosophical, normative, and practical considerations.

In this article, I propose that false positive parity is the best statistical measure of parity for use in assessing the fairness of RPIs. I reach this conclusion by examining the three-pronged debate over the appropriate measure for assessing bias in COMPAS and evaluating the validity and justifications of the arguments.

I attempt to answer the following two important questions. (1) What are the most essential factors for substantiating fairness in the criminal justice system and are thus necessary conditions in any algorithmic bias measurement? (2) Which statistical measure of parity best meets these necessary conditions in assessing fairness in RPIs? Furthermore, I aim to present these statistical concepts and their practical meaning in a manner which is easily digestible without a statistics background.

Download

Share

COinS