Offender risk assessment

Risk assessment tools that predict the likelihood of an individual committing an offence (or a particular type of offence) based on Big Data include Correctional Offender Management Profiling for Alternative Sanctions (COMPAS). Greater use of these tools in bail, parole or sentencing has been advocated by the Conference of Chief Justices Conference  of  State  Court  Administrators (National Center for State Courts, 2007) and the Arnold Foundation (2013).

Reasons for promoting these tools include public safety, reduction in recidivism and cost savings through the release of ‘low risk’ offenders. The data used in making predictions may be based on information from a person’s police record as well as on interviews.

The use of risk assessment tools in particular jurisdictions will depend on the laws of that jurisdiction. In some jurisdictions, these tools are expressly permitted or man- dated by legislation. For example, in Ohio, the Department of Rehabilitation and Correction is required to select a single validated risk assessment tool for adult offenders to be used by various entities including the parole board and sentencing courts (Ohio Rev Code Ann § 5120. 114 (2016)). In State v Loomis (2016), the Supreme Court of Wisconsin held that risk assessments tools could be used in sen- tencing provided that they were not the only factor taken into account, that they were not used in decisions about the severity of a sentence, and that reports were accompanied by appropriate warnings.

In many cases, offender risk assessment tools are designed to optimise predictive accuracy rather than assist with explanation or focus on the kinds of variables that might generally be considered appropriate in making decisions about bail and parole. As Berk and Bleich (2013: 517) argue, any variable that can operate as a good predic- tor should be used even if, like shoe size, it is not obvious why this variable might be important or whether it is something that would have traditionally been taken into account. Offender risk assessment tools are not just concerned with more accurate weighting of traditionally relevant factors (such as number of prior offences and type of offences committed in the past), but also with identifying new variables that cor- relate with recidivism.

As in the case of predictive policing, there are diverse tools that can be used. One example is the random forest approach used in Berk (2012: 65–6). This involves machine learning from historic data concerning recidivism in order to predict whether a particular person should be flagged as a recidivism risk (the output is thus binary, in that individuals are classified as either high risk or low risk). A random forest is an ensemble of classification trees. A classification tree is essentially a deci- sion tree that, at each branch, requires the user to decide which is the relevant branch based on the value of a particular variable. An example of a (hypothetical) classifica- tion tree is shown in Figure 11. 3.

In a random forest algorithm, multiple classification trees are constructed based on different data sets drawn from the training set. In particular, random samples of size N are drawn with replacement from a training data set comprising N obser- vations. Each sample is used to construct a classification tree. First, a small random sample of predictors is drawn. The tree is then constructed through a series of partitions of the data, each time selecting the predictor (from the random sample) that most reduces the heterogeneity of outcomes at each node. At each terminal node, the Bayes classifier is used to assign a class to that node, so that the class assigned (e. g. high or low risk) is the most probable class for that node. The Bayes classifier can be determined mathematically by choosing the class that

FIgure 11. 3 Example of a hypothetical classification tree

minimizes the expected prediction error (for a formal explanation, see Berk (2012: 43–7)). The training data that was not used to construct that particular tree is then dropped down the tree (according to the partition criteria) and is then assigned the class associated with the relevant terminal node (in this case, either high risk or low risk). For each observation in the training data set, classification (in this case, as high risk or low risk) is by vote over those classification trees that it was dropped down (that is, where it was not used to build the tree). Once one has a number of classification trees, one can record actual against predicted clas- sifications, because for training data it is already known whether the person did or did not re-offend within the relevant period of time. Given a new offender, one then works out the terminal node for each classification tree, notes the various classifications as high risk or low risk and votes over the trees to determine the class into which the individual falls.

What is interesting about this technique, and many similar techniques, is that they do not provide an explanation for why an individual is high or low risk. Each new classification will be the result of multiple classifications according to many variables, some of which may have no obvious relevance. At most, one can determine the con- tribution that a particular predictor makes to forecasting accuracy or the extent to which a particular predictor influenced the classification given to a particular indi- vidual (Berk, 2012: 66–9). There is a random element to the construction of the

random forest which means that, even with access to the same data, one might con- struct a random forest that makes different predictions for the same person. Nevertheless, it has been claimed that these techniques are often more effective at predicting risk than those that might be easier to explain (Berk and Bleich, 2013).