Comparison of Adult Offense Prediction Methods based on Juvenile Offense Trajectories using Cross-Validation

Considerable research has found support for the relationship between criminal offending in adolescence and criminal offending in adulthood. Estimating the strength and nature of the relationship has been facilitated by the methodological advances that have been made over the past decade. We add to this literature and describe and apply various prediction methods to examine the extent to which adult (ages 18–33 years) criminal offense trajectories can be predicted by juvenile (ages 9–17 years) offense trajectories. These methods include conventional models based on latent Poisson classes (LPC) and generalised linear models (GLM) and more sophisticated Cox proportional hazards models that predict entire adultoffense timelines. We also present a novel method, based on the exponential distribution, for adjusting the observed offense patterns for time-at-risk using secure custody information and a method for addressing the problem of the offense-conviction date lag. In addition, we discuss how to compare the accuracy of different prediction methods using cross-validation, thus providing a clear, unambiguous measure of prediction accuracy. We apply our methods to a data set comprising 378 male offenders in Toronto, Canada, whose criminal careers were tracked for an average of 12.1 years. Our results show that, for these data, no method can yield very accurate predictions. On the other hand, some prediction methods are able to make better use of pre-18 information to improve the precision in the predictions.