Bayesian additive regression trees and the next generation of juvenile justice risk assessments

This dissertation involves two studies that center on estimation of recidivism risk in juvenile justice. Criminal and juvenile justice systems throughout the U.S. are experiencing reform focused on aligning practice with research evidence. With this, standardized risk assessment tools are brought to the forefront of legislatively required processes in the field. Criminal and juvenile justice literature has involved empirically based risk assessment instruments for decades. However, the field may be shifting toward the adoption of more flexible methods that have only recently become accessible through advancements in mainstream statistical software packages. The studies in this dissertation help to fill important research gaps by applying state-of-the-art tree based methods alongside a conventional traditional approach to recidivism risk estimation for multiple samples gathered from Utah's juvenile justice system database. Each study compares the predictive validity of Bayesian additive regression trees (BART), random forests, and logistic regression with the purpose of informing researchers and practitioners about which of these methods present the best option for assessing risk to reoffend.