Long-term Prediction of Youth Crime

• In Event: Crime and its Consequences: Studies Using Nordic Register Data

Abstract

The efficacy of early prevention is heavily dependent upon the extent to which one can predict later outcomes from childhood (Caspi et al 2016). Data rich cohort studies have been used to identify a number of potential risk factors, but generally the wealth of data far exceeds the limitations of traditional analytic methods. We investigate the extent to which administrative records can provide a basis for predicting involvement in crime during childhood and adolescence. Administrative records provide extensive information, but also introduce complicated and unknown correlation structures between potential predictors and offending. In this setting, machine learning should be the favoured approach precisely because it is a data-driven approach. We use a sample of the total population of persons born in Norway in the years 1992-1996. For these cohorts, we predict whether the individual was ever charged with a crime by age 18. All covariates are on the parents’ characteristics and their neighbourhood at the time of birth, as available in Norwegian administrative data.

Authors

• Torbjorn Skardhamar, University of Oslo

• Torkild Lyngstad, University of Oslo

• Richard Berk, University of Pennsylvania