Selecting Items to Compose a Brief Version of the Youth Level of Service/Case Management Inventory Using Machine Learning, Sparse Modeling

This study aims to briefly compose a well-known tool, Youth Level of Service/Case Management Inventory, and to obtain findings on the efficient use of tools. We hypothesize that sparse modeling can identify item combinations with high predictive validity. Using follow-up data on 516 justice-impacted youths (age range = 13-19 years) in Japanese juvenile classification homes, this study obtains area under the curves (AUCs) of 0.71 and 0.72 for seven- and eight-variable models itemized using LASSO and Elastic Net, respectively, as machine learning methods. Analysis demonstrates that the brief version using sparse modeling insignificantly differed from original version, although its AUC values were slightly lower, supporting the hypothesis. Through the effort to incorporate many types of risk factors into the tool, the use of sparse modeling to extract combinations of items with high predictive power for recidivism could lead to the creation of a user-friendly, item-count scale tool for forensic settings.