Using a small simulation experiment, this study examined whether semiparametric group-based modeling (SPGM) is suitable for testing the existence of subpopulations not directly observable; and it considers whether similar trajectories might equally well result from mechanisms suggested by general theories.
The author concludes that as generally applied, SPGM cannot provide evidence either for or against a taxonomy and that the usual findings can be explained by competing theories. This paper further argues that this result is not only because of the methodology's characteristics but also because of the modeling strategy applied. In the experiment reported in the current article, the author generated longitudinal data without any meaningful groups. He then analyzed these data using SPGM and checked the diagnostics and estimated group-specific effects of time-varying covariates. If seemingly distinct groups can be recovered from such data, then the use of SPGM to distinguish between general and taxonomic theories of crime will be called into question, and the empirical support for the taxonomic approach must be reconsidered. Moffitt (1993, 2006) proposed what is perhaps the most influential contemporary taxonomic theory of crimes. She has suggested that adolescent-limited and life-course persistent antisocial behaviors are committed by two "qualitatively distinct categories of individuals . . . each in need of its own distinct theoretical explanation." These antisocial types are not directly observable, which leads to empirical challenges. Moffitt's approach was spurred by the development of SPGM, a technique that offers a methodological basis for distinguishing between groups of individuals with different patterns of offending across time (Nagin and Land, 1993). This novel technique promised "the capability to identify rather than assume distinctive groups of trajectories" (Nagin, 1999: 139), making it possible to test for the existence of offender categories on a more objective basis. The analysis reported in this article challenges this methodology. 3 tables, 2 figures, and 53 references