Abstract:
Some cohort studies have claimed crime can be predicted at early childhood with high accuracy, but typically using far more detailed data than would be available in an applied setting. Routinely available administrative records would be more realistic although far less detailed. We examine to what extent future recorded crime can be predicted at the time of birth, based on family-level and neighborhood data available in administrative data. We assess the realism of early prediction, with emphasis on the tradeoffs between accuracy and the costs of errors as well as biases across sub-populations. The results suggests that while forecasting might be better than one could expect given the limitations of the data, fairness is much harder.