Responses to Reviewers' Comments and Suggestions

R1 suggests that anyone could have "anticipated the various factors that define the clusters ... [would lead] to certain offending trajectories." The reviewer further adds that anyone who has studied offending trajectories using the CSDD would be familiar with our results.

Certainly, we anticipated that the adolescent-basedbehavioral, psychological and social characteristics would yield clusters of individuals with different offending trajectories. However, we submit that our offending trajectories are distinct in important ways from the offending trajectories identified retrospectively using the same dataset (CSDD).

Below are the trajectories identified in Piquero, et al (2007):

The findings based on retrospective analyses result in offending trajectories that differ in both frequency and shape. This supports the idea that different causal processes give rise to qualitatively distinct kinds of offenders.

By contrast, our prospectively identified trajectories are notably less distinct from one another. They vary in frequency but not in shape. As we set up on p. 4 and conclude on p. 25, this is more supportive of a general causal process driving the age-crime relationship across offenders.

R1 writes that it is not surprising that "(k)ids with low IQs, who are impulsive and do naughty things are likely to offend more than other kids."R1 urged us to make a stranger argument for the advantages of a person-centered approach over a variable-centered approach.

While such a finding may not be a surprising variable-centered conclusion, as we know from previous literature that both low IQ and impulsivity independently predict offending, our person-centered analyses suggest that it is an inaccurate conclusion. Our most impulsive cluster ("impulsive rowdies") has the lowest percentage of kids with low IQs (p. 18). We believe this difference is precisely what demonstrates the utility of the person-centered approach as a means of more accurately summarizing predictors of offending, and developing programs that are targeted at whole persons rather than ad-hoc predictors. We appreciate the opportunity to clarify this on page 23.

R1 recommends that we make a stronger argument for the advantages of a person-centered approach over a variable-centered approach.

Our central purpose was to identify offending trajectories prospectively using adolescence-based social and behavioral variables. This prospective analysis is better accomplished through a person-centered approach rather than a variable-centered approach because it allows us to capture interactions between the predictor variables. While interaction terms can be added to variable-centered analyses, interpretation quickly becomes complex. Person-centered approaches ease this interpretation.

Further, as stated noted above, a person-centered approach yields a more accurate picture of how key predictors cluster together. With the goals of adolescent intervention and treatment in mind, it is important to consider the typical clusters of setbacks and psychological problems adolescents face so that multidimensional interventions address the person as a whole rather than one variable at a time. For this reason, LCA offers a clear advantage over variable-centered approaches in allowing one to focus on prototypical patterns in the data (Bauer and Shanahan 2007). We have strengthened this argument on pp. 7-8.

R1 asks us to provide more detail aboutour measures and how we defined the offending trajectories.

We have revised the Data and Measures section of the manuscript to provide further clarification of our measures (pp. 9-13). We also edited the section to ensure that the names of the variables match the names used in Table 3. On p. 14 we provide additional detail about hierarchical linear modeling procedures we used to identify the offending trajectories. We suspect that R1 may have shared R2s concern (below) about being unable to discern how the trajectories relate to one another, and anticipate that our newly added ANOVA tests also described on p. 14 allow the reader greater insight into how the trajectories relate to one another (p. 19 and Figure 1).

R1 further suggeststhat we rethink our presentation of the “robustness” or “sensitivity”results.

We have modified the introduction to this section tomake it clear that this is simply a sensitivity analysis. We removed the word "robust" from the manuscript. We agree that the discussion of the sensitivity results was redundant to the central results. Given that scholars (Laub and Sampson 2003; Moffitt et al. 1996; Sampson and Laub 2005) typically to use the median to assign risk, as we have in our central analysis, we present those finding in full. We continue to think it important to discuss findings from the sensitivity analyses in order to demonstrate that our findings were not the result of measurement choices. (We discuss this intent in the first footnote on p. 14 and on p. 21). That said, we eliminated all tables and figure associated with the sensitivity analysis and now only describe the analysis very briefly on pp. 21-22. We appreciate the suggestion as these changes greatly improved the flow of the manuscript.

R1 suggests that we broaden our literature review of the use of AICs and BICs

On p. 16we review the relative strengths and weaknesses of BIC and AIC and include Brame and colleagues'(2006)work. Our paper also presents a number of statistical criteria, which guided our decisions about the appropriate number of clusters to include in our models, so that it is clear we are not overly reliant one criterion (pp. 15-17).

R1 suggests that we include literature on cumulative disadvantage.

We appreciate the suggestion and have woven literature on cumulative disadvantage throughout the manuscript and especially on pp. 4, 8, 9, 22 and 23. The newly added literature on cumulative disadvantage also works to strengthen the case for taking a person-centered approach. The idea that problems accumulate and cluster within individuals, is consistent with a person centered approach, as we note on pp. 7-8.

We do not frame the paper centrally around the cumulative disadvantage literature because we are concentrating on measures at a single time point, adolescence, rather than how these psychological and social disadvantages accumulative over the life course.

R1 notes that by inferring prediction, we "tread on risky ground."

To the contrary, we ultimately argue that the different clusters result in different rates of crime rather than patterns of crime. Regardless of early problems, we expect desistence in adulthood. By prospectively identifying the criminal trajectories of groups identified on the basis of adolescent personal and social problems, we actually fail to predict the crime types commonly seen in the literature. We regret this was not more clear in the original manuscript and have clarified these points on pp. 25 and 25.

R2 points out that no formal tests were applied to demonstrate whether the differences in criminal trajectories are statistically significant. R2 suggests that we conduct tests of differences across groups and by age grouping.

We agree with R2 that the trajectories do not stand on their own and require formal testing. We have added one-way analysis of variance (ANOVA) tests that allow us to assessthe differences between our clusters. Rather than conducing these tests across age groupings, we conducted the tests at each age. We felt this would offer the most transparency. Moreover, this allowed us to incorporate the results of these tests in Figures 1. We have added superscripts to the ages along the x-axis to indicate which comparisons are statistically significant. Substantively, the figures show that rates of offending become more similar to one another in the later years. The ANOVA results add weight to this claim.As we write on p. 19 the pair-wise comparisons are more consistently significant in early adulthood than at later ages.

R2 points out that group-based trajectory models could also be estimated on the post-18 data.

We agree that this could be an interesting piece of analysis. It would also provide another approach to substantiating any differences in the crime trajectories. We found R2's first suggestion to test across groups by age a clearer way to communicate significant differences. Further, wished to remain consistent in our emphasis on prospective identification of the trajectories.

R2 recommends that we soften our emphasis on the Moffitt model and more broadly discuss our effort to predict adult offending from adolescence.

To this end, we have significantly revised the manuscript, shifting the focus to predicting offender groups and offending more generally. We further agree that some of our points are already familiar to the reader, and we have substantially trimmed them down on pp. 3-6.

Thank you once again for your careful review and thoughtful comments.

References

Bauer, D. J., & Shanahan, M. J. (2007). Modeling complex interactions: Person-centered and variable-centered approaches. Modeling contextual effects in longitudinal studies, 255-283.

Brame, R., Nagin, D. S., & Wasserman, L. (2006). Exploring some analytical characteristics of finite mixture models. Journal of Quantitative Criminology, 22(1), 31-59.

Laub, J. H., & Sampson, R. J. (2003). Shared beginnings, divergent lives. Cambridge, Massachusetts and London, England: Harvard University Press.

Moffitt, T. E., Caspi, A., Dickson, N., Silva, P. A., & Stanton, W. (1996). Childhood-onset versus adolescent-onset antisocial conduct problems in males: Natural history from ages 3 to 18 years. Development and psychopathology, 8, 399-424.

Piquero, A. R., Daigle, L. E., Gibson, C., Piquero, N. L., & Tibbetts, S. G. (2007). Research note: Are life-course-persistent offenders at risk for adverse health outcomes? Journal of Research in Crime and Delinquency, 44(2), 185-207.

Sampson, R. J., & Laub, J. H. (2005). A life-course view of the development of crime. The Annals of the American Academy of Political and Social Science 602, 12-45.