Evaluating Fairness of Algorithmic Risk Assessment Instruments: The Problem With Forcing Dichotomies

• Author: Evan M. Lowder,Sarah L. Desmarais,Samantha A. Zottola,Sarah E. Duhart Clarke
• DOI: 10.1177/00938548211040544
• Published date: 01 March 2022
• Date: 01 March 2022
• Subject Matter: Articles

CRIMINAL JUSTICE AND BEHAVIOR, 2022, Vol. 49, No. 3, March 2022, 389 –410.

DOI: https://doi.org/10.1177/00938548211040544

Article reuse guidelines: sagepub.com/journals-permissions

© 2021 International Association for Correctional and Forensic Psychology

389

EVALUATING FAIRNESS OF ALGORITHMIC

RISK ASSESSMENT INSTRUMENTS

The Problem With Forcing Dichotomies

SAMANTHA A. ZOTTOLA

SARAH L. DESMARAIS

Policy Research Associates

EVAN M. LOWDER

George Mason University

SARAH E. DUHART CLARKE

RTI International

Researchers and stakeholders have developed many definitions to evaluate whether algorithmic pretrial risk assessment

instruments are fair in terms of their error and accuracy. Error and accuracy are often operationalized using three sets of

indicators: false-positive and false-negative percentages, false-positive and false-negative rates, and positive and negative

predictive value. To calculate these indicators, a threshold must be set, and continuous risk scores must be dichotomized. We

provide a data-driven examination of these three sets of indicators using data from three studies on the most widely used

algorithmic pretrial risk assessment instruments: the Public Safety Assessment, the Virginia Pretrial Risk Assessment

Instrument, and the Federal Pretrial Risk Assessment. Overall, our findings highlight how conclusions regarding fairness are

affected by the limitations of these indicators. Future work should move toward examining whether there are biases in how

the risk assessment scores are used to inform decision-making.

Keywords: Risk assessment; pretrial; bias; algorithmic fairness; Public Safety Assessment; Virginia Pretrial Risk

Assessment Instrument; Federal Pretrial Risk Assessment

INTRODUCTION

The fairness of algorithmic risk assessment instruments has received significant public

and legal attention in recent years (Angwin et al., 2016; Hamilton, 2018; Hellman, 2020;

Mayson, 2019; Minow et al., 2019; Pretrial Justice Institute, 2020; Slobogin, 2018;

Southerland & Woods, 2018; U.S. Department of Justice, 2014). This attention has included

discussion regarding whether algorithms developed from criminal justice data can only

AUTHORS’ NOTE: Correspondence concerning this article should be addressed to Samantha A. Zottola,

Policy Research Associates, 345 Delaware Ave, Delmar, NY 12054; e-mail: szottola@prainc.com

1040544CJBXXX10.1177/00938548211040544Criminal Justice and BehaviorZottola et al. / Forcing Dichotomies

research-article2022

390 CRIMINAL JUSTICE AND BEHAVIOR

exacerbate racial disparities by carrying forward the racism present in existing criminal

justice practices (e.g., disparate arrest rates experienced by black people; Kochel et al.,

2011) and whether algorithms will convey a false, and therefore harmful, sense of objectiv-

ity (Eckhouse et al., 2019; Robinson & Koepke, 2019). As part of these discussions, con-

cerns have been raised regarding whether the instruments produce risk estimates that

demonstrate evidence of bias in general and as a function of race, specifically. While these

concerns are relevant to algorithmic risk assessment instruments in general, discussion has

centered on instruments used to inform pretrial decision-making (Vincent & Viljoen, 2020).

Briefly, algorithmic pretrial risk assessment instruments are designed to evaluate the likeli-

hood that a person will fail to appear in court or be arrested for a new crime on pretrial

release. The likelihood of these outcomes (collectively referred to as pretrial misconduct) is

calculated via an algorithm that weights and sums ratings on items such as criminal history,

current charges, or personal characteristics to produce a scaled risk score representing the

likelihood of pretrial misconduct.

Currently, there is a great deal of debate around whether algorithmic risk assessments

should be part of pretrial reform efforts sweeping the United States. Those who support the

use of algorithmic risk assessment instruments assert that there is one strategy that may help

promote transparency, accuracy, and consistency in pretrial decisions (Desmarais & Lowder,

2019). They also point out that there is relatively limited empirical evidence demonstrating

differential predictive validity across groups (Desmarais et al., 2020; Vincent & Viljoen,

2020), but that when identified, differential prediction may highlight biases in the criminal

justice system that are otherwise difficult to quantify (Mayson, 2019; Viljoen & Vincent,

2020). Critics of these instruments, in contrast, assert that algorithmic pretrial risk assess-

ment instruments are racially biased and exacerbate racial disparities in pretrial decisions

(Angwin et al., 2016; Pretrial Justice Institute, 2020). These differing viewpoints on whether

there is racial bias in algorithmic pretrial risk assessments may reflect, at least in part, the

various definitions—and consequently, indicators—used for determining whether an algo-

rithmic pretrial risk assessment instrument produces “fair” or “biased” results (see for

example, Berk et al., 2021; Chouldechova, 2017; Kleinberg et al., 2018; Mayson, 2019).

Indeed, stakeholders with differing views on the issue tend to rely on evidence derived

using different indicators of predictive accuracy (Hamilton, 2018). Our goal is to examine

three sets of indicators that are being emphasized by advocates, legal scholars, and other

stakeholders as measures of fairness: (a) true-positive, false-positive, false-negative, and

true-negative percentages; (b) false-positive (FPRs) and false-negative rates (FNRs); and

(c) positive (PPVs) and negative predictive values (NPVs). In the sections that follow, we

describe how these indicators are being used to evaluate fairness of algorithmic risk assess-

ment instruments in the pretrial context.

INDICATORS USED TO EVALUATE FAIRNESS

Much of the discussion about fairness in algorithmic risk assessment instruments is con-

cerned with whether the error and accuracy of assessment results are the same across sub-

groups of people. There are dozens of indicators that can be employed to compare error and

accuracy in instruments, both overall and across subgroups. A full discussion of all indica-

tors is beyond the scope of this paper (for information on indicators not covered here, see:

American Educational Research Association [AERA], American Psychological Association,

& National Council on Measurement in Education, 2014; Berk et al., 2021; Chouldechova,