By Min-Taec Kim, Fan Cheng
The Bureau of Crime Statistics and Research (BOCSAR) has released a new study that considered whether linked data from health, education, child protection (and more) can be used to predict which children are likely to offend before turning 18. The study found: Whole-of-population level: Machine learning methods applied to large-linked data can identify future young offenders with reasonable accuracy. Aboriginal young people: The models were far less accurate for Aboriginal young people and could not be reliably used for this group. Offending predictions for Aboriginal young people were more likely to be wrong than for non-Aboriginal young people – a higher rate of false positives. This raises significant ethical concerns about the practical use of such models for Aboriginal people and other sub-populations. Predictive factors over time: The study also identified key predictors of offending and how these change with age. At birth, gender and parental characteristics are most influential. Between ages 5 and 10, child protection reports and disability become important. By age 15, school suspension is the strongest predictor of youth offending.