Machine Learning Risk Assessments in Criminal Justice Settings

This chapter provides a general introduction to forecasting criminal behavior in criminal justice settings. A common application is to predict at a parole hearing whether the inmate being considered for release is a significant risk to public safety. It may surprise some that criminal justice forecasts of risk have been used by decision-makers in the United States since at least the 1920s. Over time, statistical methods have replaced clinical methods, leading to improvements in forecasting accuracy. The gains were at best gradual until recently, when the increasing availability of very large datasets, powerful computers, and new statistical procedures began to produce dramatic improvements. But, criminal justice forecasts of risk are inextricably linked to criminal justice decision-making and to both the legitimate and illegitimate interests of various stakeholders. Sometimes, criticisms of risk assessment become convenient vehicles to raise broader issues around social inequality. There are, in short, always political considerations, ethical complexities, and judgement calls for which there can be no technical fix. The recent controversy about “racial bias” in risk instruments is a salient example. 1.1 Some Introductory Caveats There are claims, increasingly common, that one can sensibly apply machine learning without understanding the underlying fundamentals. Not surprisingly, one hears this a lot from business firms trying to sell “data analytics” software. At the other extreme, are claims from academics in computer science and statistics that unless a data analyst is well-versed in those underlying fundamentals, that analyst will know just enough to be dangerous. Complicating these debates is that there is a great deal about machine learning that is not well understood by any academic discipline. In many instances, practices has evolved much faster than the requisite theory. An effort is made in the pages ahead to strike a balance between the two extremes. The danger is that both camps will be dissatisfied. And that is OK. There is an old saying that you know when you are doing something right if everyone is mad at you. © Springer Nature Switzerland AG 2019 R. Berk, Machine Learning Risk Assessments in Criminal Justice Settings, https://doi.org/10.1007/978-3-030-02272-3_1 1