This study consists of a comparative analysis of patterns of de-escalation between ages 17-18 and 32, based on data from two well-known prospective longitudinal studies, the Cambridge Study in Delinquent Development (a study of 411 working-class males in London) and the Montreal Two Samples Longitudinal Study (a sample of 470 adjudicated French-Canadian males). Analyses focus on within-individual change, with individuals serving as their own controls. In this regard, the magnitude of measured change is relative to the past degree of involvement in offending. These results are contrasted with predictors of between-individual differences in offending behavior at age 32. We investigate the respective roles of cognitive predispositions and social bonds in the prediction of patterns of de-escalation, and assess whether it is possible to make relatively long-term predictions (over a 15-year period) about offending in adulthood. Findings suggest that traditional measures of social bonds and cognitive predispositions measured at age 17-18 are generally weak predictors of de-escalation up to age 32. However, these measures are stronger predictors of between-individual differences in offending gravity. These findings highlight the difficulties in making accurate long-term predictions about changes in individual offending patterns early in the criminal career.