\pretitle\posttitle\preauthor\postauthor\predate\postdate

Conclusion and Next Steps

Criminal justice institutions are in constant search of policy alternatives to increase efficiency in maximizing public safety while curbing incarceration rates. Data-driven policy is the latest trend in achieving this goal by shaping decision making processes in different stages of the criminal justice process. Two technological developments from the last couple of decades have placed machine learning methods as the cutting-edge tool to explore critical problems of criminology and shape policy. I have provided an overview of machine learning methods used to visualize and analyze criminal justice data. Clustering and principal component analysis; linear methods for regression,including penalized regression, variable and model selection; tree methods and, bootstrapping, boosting, bagging are approaches to unveil relationships among variables and predict events of interest for the institutions in the criminal justice system.
I have provided a summary of the critiques and empirical examples supporting and challenging the scope and reach of such computational statistical methods. My review has shown that there are two main focus of critique to machine learning methods. First, the claims of mathematical complexity and opacity which has gained them the name of black box methods. Second, complexity and opacity in policy design and implementation. I have proposed that contrary to a trending belief reproduced by scholars and media, black box methods, as applied to criminal justice, have a policy dimension that needs to be addressed in order to tackle the issue of mathematical opacity. I claim this dimension of policy making is more challenging and dangerous that the issue presented by mathematical complexity alone.
Based on the analysis of front-end statistical risk assessment, fours steps are suggested to ensure that the selection of an computational statistical tool for policy has gone through ethical and methodological considerations: objective, design, implementation plan and validation. Objective refers to a definition of policy purpose. Design relates to identifying, defining and justifying relevant variables associated with the goal of the instrument. Implementation plan refers to identifying the institution, public official and stage in the process at which the instrument will be applied, as well as the role of the tool in the decision process. Finally, empirical validation of the instrument will make it time and location specific to serve its purpose. These steps will expand the extent to which machine learning can help to maximize the objective function of the criminal justice system and provide public safety at the same time that curbs the incarceration trend.
As for the existing policies that rely on machine learning methods, an approach to assess their efficacy is by conducting empirical analyses comparing geographical units -such states or counties- where a given policy has been implemented with one where it has not. One potential case of study is at the early stages in the criminal justice process, pre-trial detention, when a judge assess if the defendant has a flight risk or more likely to appear in court during the process. An analysis of rates of failure to appear by group in both scenarios, one where the judge decision was supported by a statistical tool and one where the judge decided with clinical assessments would provide additional information about the effect of introducing a machine learning methods to approach criminal justice policy.