The goal is to combine the information about the custody arrival context (such as, arrested offence, restraining equipment used, detained person demographics, injury, mental health problems, alcohol drugs, etc.) and the behaviour of the offender for each hour after his arrival to develop a model that would help to predict accurately whether the detained person would be fully compliant or generally compliant, would show any sign of antisocial behaviour, or would attempt to harm himself or others.
The project will use of data-driven insights to define a model with concrete decision rules for the assessment of custody risk decisions. Researchers will estimate a diverse set of models, which will later be tested in a set of RCTs. Among these models, the team will formulate ordinal logistic regression (using regularisation), decision trees and fast and frugal trees. Fast and frugal decision trees have the major advantage of being easy for human decision makers to implement.
Although Fast and frugal decision trees are very simple, the trees typically perform as well as more complicated models whilst maintaining sufficient simplicity to be implemented in the field. The construction methodology for the fast-and-frugal trees is already established. Scaling up the trees with the new, larger data set is the first stage of the deployment of the project and the basis for the randomized controlled trial.
This project is funded by the WBS Impact and Dissemination Fund.