Accuracy and Fairness for Juvenile Justice Risk Assessments

Berk, Richard A.

doi:10.1111/jels.12206

Cited by 37 publications

(24 citation statements)

References 45 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Furthermore, removing relevance/increases accuracy as a possible answer for property assignment questions does not substantially increase the predictive power of any other properties. This result is especially important because fairnessrelated machine learning research often references an "unavoidable" accuracy-fairness trade-off [e.g., 4,18,19,29,31,41]. In contrast, our results point to an important correlation between accuracy and fairness (specifically perceived process fairness).…”

Section: Conclusion and Discussionmentioning

confidence: 47%

The Role of Accuracy in Algorithmic Process Fairness Across Multiple Domains

Albach

Wright

2021

Proceedings of the 22nd ACM Conference on Economics and Computation

View full text Add to dashboard Cite

Machine learning is often used to aid in human decision-making, sometimes for life-altering decisions like when determining whether or not to grant bail to a defendant or a loan to an applicant. Because of their importance, it is critical to ensure that the processes used to reach these decisions are considered fair. A common approach is to enforce some fairness constraint over the outcomes of a decision maker, but there is no single, generally-accepted definition of fairness. With notable exceptions, most of the literature on algorithmic fairness takes for granted that there will be an inherent trade-off between accuracy and algorithmic fairness. Additionally, most work focuses only on one or two domains, whereas machine learning techniques are used in an increasing number of distinct decision-making contexts with differing pertinent features.In this work, we consider six different decision-making domains: bail, child protective services, hospital resources, insurance rates, loans, and unemployment aid. We focus on the fairness of the process directly, rather than the outcomes. We also take a descriptive approach, using survey data to elicit the factors that lead a decision-making process to be perceived as fair. Specifically, we ask 2157 Amazon Mechanical Turk workers to rate the features used for algorithmic decision-making in one of the six domains as either fair or unfair, as well as to rate how much they agree or disagree with the assignments of eight previously (and one newly) proposed properties to the features. For example, a worker could be asked to rate the feature of "criminal history" as fair or unfair to use in bail decisions, and then rate how much they agree or disagree that "criminal history" is a reliable feature. We show that, in every domain, disagreements in fairness judgements can be largely explained by the assignments of properties (like reliability) to features (like criminal history). We also show that fairness judgements can be well predicted across domains by training the predictor using the property assignments from one domain's data and predicting in another. These findings imply that the properties act as moral determinants for fairness judgements, and that respondents reason similarly about the implications of the properties in all the decision-making domains that we consider. Although our results are mostly consistent across domains, we find some important differences within specific demographic groups in the hospital and insurance domains, indicating that at least some differences in fairness judgements are introduced by demographic differences. However, a single property usually holds the majority of the predictive power. With some exceptions, predictors learning from only the "increases accuracy" property perform better (in all domains) than predictors learning from any combination of the other seven properties, implying that the primary factor affecting respondents' perceptions of the fairness of using a feature for prediction is whether or not a feature increases the accuracy...

show abstract

Section: Conclusion and Discussionmentioning

confidence: 47%

The Role of Accuracy in Algorithmic Process Fairness Across Multiple Domains

Albach

Wright

2021

Proceedings of the 22nd ACM Conference on Economics and Computation

View full text Add to dashboard Cite

show abstract

“…One merely weights the training data so that post‐release arrests of Black offenders for violent crimes are effectively reduced relative to post‐release arrests of White offenders for violent crimes. We have undertaken such exercises elsewhere (Berk, 2019; Elzarka, 2019) showing that changing bases rates can dramatically alter the results and impact fairness.…”

Section: Altering the Base Ratesmentioning

confidence: 99%

Almost politically acceptable criminal justice risk assessment

Berk

Elzarka

2020

Criminology & Public Policy

Self Cite

View full text Add to dashboard Cite

Modern, algorithmic risk tools are trained in the sense that they inductively seek associations between predictors (e.g., prior record) and an outcome of interest (e.g., a rearrest). There is no model. 3 The associations are used to construct a measure of risk. Risk can be represented as a numerical score, a probability of an outcome, or a particular outcome class. Because criminal justice decisions are typically categorical (e.g., release on parole or not), often built into the algorithmic is machinery to translate a numerical score or probability into an appropriate categorical outcome. If not, some less formal means are required to generate an outcome class to serve as the forecast. A probability of a rearrest, for instance, is not actionable until it is high enough to warrant assignment to the outcome class of, say, "recidivist." Once training is completed, the risk algorithm can be used to forecast outcomes for new cases when their outcomes are unknown. Data for such forecasting must be collected and properly managed, as is done for the training data. The process can be demanding because the data used for forecasting must have the same predictors as the data used for training and be realized in the same fashion. Instructive forecasts might well be obtained, for example, from training data and forecasting data properly realized from the same jurisdiction, in roughly the same time period, and for the same criminal justice setting (e.g., parole decisions). And just as for the training data, there will usually be important data quality concerns. The forecasting data are provided to a trained algorithm that, in turn, produces forecasts for each case. Ideally, these are forecasted outcome classes such as a post-release arrest for a violent crime, a post-release arrest for a nonviolent crime, or no post-release arrest whatsoever. Often, the reliability of those forecasts also can be determined (Berk, 2018b). With the forecasts and (preferably) reliabilities in hand, decisions can be made and actions can be taken. The recent risk literature makes clear that machine learning, algorithmic methods can have demonstrable, superior performance compared with subjective or model-based methods (Berk, 2020; Berk & Bleich, 2014; Berk, Sorenson, & Barnes, 2016). It has been long known that even simple statistical methods produce better accuracy relative to subjective approaches (Dawes, Faust, & Meehl, 1989; Meehl, 1954), and as a mathematical matter, machine learning algorithms can adaptively find far more complex relationships in the data, should they exist, compared with conventional models such as logistic regression (Hastie, Tibshirani, & Friedman, 2009). When the relationships between predictors and an outcome are not complex, algorithmic methods properly applied will perform no worse than conventional models.

show abstract

“…They will also likely need to define how algorithms should make trade-offs between forecasting accuracy and other values, such as fairness. 25 Toward this end, environmental officials will need to continue to engage with elected officials, members of the public, environmental groups, and industry representatives to forge clarity and consistency over how various risk and regulatory objectives should be specified. At the same time that government officials will need to strengthen their analytic and technological skills, they will continue to need to strive for excellence in social engagement.…”

Section: Building Capacity For Algorithmic Governancementioning

confidence: 99%

Deploying Machine Learning for a Sustainable Future

Coglianese¹

2019

A Better Planet

View full text Add to dashboard Cite

To meet the environmental challenges of a warming planet and an increasingly complex, high techeconomy, government must become smarter about how it makes policies and deploys its limited resources. It specifically needs to build a robust capacity to analyze large volumes of environmental and economic data by using machine-learning algorithms to improve regulatory oversight, monitoring, and decision-making. Three challenges can be expected to drive the need for algorithmic environmental governance: more problems, less funding, and growing public demands. This paper explains why algorithmic governance will prove pivotal in meeting these challenges, but it also presents four likely obstacles that environmental agencies will need to surmount if they are to take full advantage of big data and predictive analytics. First, agencies must invest in upgrading their information technology infrastructure to take advantage of computational advances. Relatively modest technology investments, if made wisely, could support the use of algorithmic tools that could yield substantial savings in other administrative costs. Second, agencies will need to confront emerging concerns about privacy, fairness, and transparency associated with its reliance on Big Data and algorithmic analyses. Third, government agencies will need to strengthen their human capital so that they have the personnel who understand how to use machine learning responsibly. Finally, to work well, algorithms will need clearly defined objectives. Environmental officials will need to continue to engage with elected officials, members of the public, environmental groups, and industry representatives to forge clarity and consistency over how various risk and regulatory objectives should be specified in machine learning tools. Overall, with thoughtful planning, adequate resources, and responsible management, governments should be able to overcome the obstacles that stand in the way of the use of artificial intelligence to improve environmental sustainability. If policy makers and the public will recognize the need for smarter governance, they can then start to tackle obstacles that stand in its way and better position society for a more sustainable future.

show abstract

Accuracy and Fairness for Juvenile Justice Risk Assessments

Cited by 37 publications

References 45 publications

The Role of Accuracy in Algorithmic Process Fairness Across Multiple Domains

The Role of Accuracy in Algorithmic Process Fairness Across Multiple Domains

Almost politically acceptable criminal justice risk assessment

Deploying Machine Learning for a Sustainable Future

Contact Info

Product

Resources

About