Abstract
Over the last 25 years, a life-course perspective on criminal behavior has assumed increasing prominence in the literature. This theoretical development has been accompanied by changes in the statistical models used to analyze criminological data. There are two main statistical modeling techniques currently used to model longitudinal data. These are growth curve models and latent class growth models, also known as group-based trajectory models. Using the well known Cambridge data and the Philadelphia cohort study, this article compares the two “classical” models—conventional growth curve model and group-based trajectory models. In addition, two growth mixture models are introduced that bridge the gap between conventional growth models and group-based trajectory models. For the Cambridge data, the different mixture models yield quite consistent inferences regarding the nature of the underlying trajectories of convictions. For the Philadelphia cohort study, the statistical indicators give stronger guidance on relative model fit. The main goals of this article are to contribute to the discussion about different modeling techniques for analyzing data from a life-course perspective and to provide a concrete step-by-step illustration of such an analysis and model checking.
Similar content being viewed by others
Notes
This approach is often referred to under its technical SAS procedure name PROC TRAJ (Jones et al. 2001).
ICPSR study number 8488. http://www.icpsr.umich.edu/ includes the Cambridge data until age 24. David Farrington kindly provided us with the outcome variables until age 40 for the Cambridge sample. A summary of the non-public data is given in Table 1a.
Some of these boys died within the observation period. For the sake of simplicity, data for these boys are not included in our analytical illustration. The variable of interest in this context is the number of convictions per year for each of the 404 males, for whom data are available until age 31.
For an in depth discussion of growth mixture models see Muthén (2001a).
In addition to the BIC, the Akaike Information Criteria (AIC) is sometimes used for model comparison. However, for finite mixture models, the AIC has been shown to overestimate the correct number of components (Soromenho 1994; Celeux and Soromenho 1996). The BIC on the other hand has been reported to perform well (Roeder and Wasserman 1997) and most consistently (Jedidi et al. 1997). For further details and comparisons see McLachlan and Peel (2000) as well as Nylund et al. (in press).
The quadratic growth function is applied to the count part as well as the zero-inflation part of the growth model. We estimated all models with and without growth structure on the zero-inflation part. The results for the models with unstructured zero-inflation part do not differ from those presented here and are therefore not listed in addition. A cubic growth parameter was specified for the conventional growth model with random intercept, but had not significant contribution. The cubic function was then not pursued any further.
Among the seven estimated parameters, three are for the means of the Poisson growth factors, one for the variance for the intercept of the Poisson growth model and three parameters for the quadratic growth model for the inflation part.
The two additional parameters are the variance of the slope and the covariance between slope and intercept.
Note that the three inflation parameters are held equal across classes.
We used an Epanechnikov kernel for the nonparametric smoothing function (Silverman, 1986).
The inflation part was modeled in the same way as it was for the GMM. That is, three parameters are estimated and set equal across the classes.
What again becomes obvious here, is the small number of patterns represented by more than one person in the Cambridge data. It could be that in the case of outliers, the GMM model might have an advantage in as much as allowing for random effects can lower the effects of single influential cases. In a model that allows for variance around the growth factors, a few outliers will increase variance substantially. If the growth factors are not allowed to vary, those cases would be more likely to form a new class. Thus, we were concerned about the effect of single influential cases to the model comparison performed here. We computed the influence statistic for each observation (Cook 1986; Liski 1991). The results of the model comparison did not change after excluding influential cases (patterns).
A similar strategy was employed by Loughran and Nagin (2006) in their analysis of the full data set.
Unlike in the Cambridge analysis, the model here has one inflation parameter per time point, held equal across classes.
For LCGA and the non-parametric GMM those can be computed by hand. For GMM with random effects, numerical integration was used.
Going through the different modeling steps, the non-parametric GMM nicely bridged the results from LCGA and GMM for the Cambridge data. Without having yet looked at covariates or predictive power, the case can be made that there are only three overall patterns in the Cambridge data with high and low variations on the themes.
BIC with random effect for the quadratic slope factor was 2951.3 compared to 2936.5 with just random effects on the intercept and linear slope.
The non-parametric GMM in which the three sub-classes are constrained to have equal probability for later conviction shows a slight decrease in fit. The log-likelihood value for the restricted model is −1467.8 with 19 parameters compared to −1462.0 with 21 parameters in the unrestricted model.
References
Blokland A, Nagin DS, Nieuwbeerta P (2005) Life span offending trajectories of a Dutch conviction cohort. Criminology 43:919–953
Celeux G, Soromenho G (1996) An entropy criterion for assessing the number of clusters in a mixture model. J Class 13:195–212
Clogg CC (1988) Latent class models for measuring. In: Langeheine R, Rost J (eds) Latent trait and latent class models, Plenum, New York
Cook RD (1986) Assessment of local influence (with discussion). J.R. Stat Soc, B 48:133–169
D’Unger AV, Land KC, McCall PL, Nagin DS (1998) How many latent classes of delinquent/criminal careers? Results from mixed Poisson regression analyses. Am J Sociol 103:1593–1630
D’Unger AV, Land KC, McCall PL (2002) Sex differences in age patterns of delinquent/criminal careers: Results from Poisson latent class analyses of the Philadelphia cohort study. J Quant Criminol 18:349–375
Elliot D (1985) National Youth Survey 1976–1980: Wave I-V, Behavioral Research Institute, Inter-University Consortium for Political and Social Research, Ann Arbor, Michigan
Farrington DP, West DJ (1990) The Cambridge study in delinquent development: A long-term follow-up of 411 London males. In: Kerner HJ, Kaiser G (eds) Kriminalität: Persönlichkeit, Lebensgeschichte und Verhalten. Festschrift für Hans Göppinger zum 70. Springer, Geburtstag
Farrington DP, West DJ (1993) Criminal, penal and life histories of chronic offenders: risk and protective factors and early identification. Crim Behav Ment Health 3:492–523
Hall DB (2000) Zero-inflated Poisson and binomial regression with random effects: a case study. Biometrics 56:1030–1039
Heckman J, Singer B (1984) A method for minimizing the impact of distributional assumptions in econometric models for duration data. Econometrica 52:271–320
Hedeker D, Gibbons RD (1994) A random-effects ordinal regression model for multilevel analysis. Biometrics 50:933–944
Hirschi T, Gottfredson M (1983) Age and the explanation of crime. Am J Sociol 89:552–584
Jedidi K, Jagpal HS, DeSarbo WS (1997) Finite-mixture structural equation models for response-based segmentation and unobserved heterogeneity. Mark Sci 16:39–59
Jones B, Nagin D, Roeder K (2001) A SAS procedure based on mixture models for estimating developmental trajectories. Sociol Methods Res 29:374–393
Lambert D (1992) Zero-inflated Poisson regression with application to defects in manufacturing. Technometrics 34:1–14
Laub J, Sampson RJ (2003) Shared beginnings, divergent lives. Delinquent boys to age 70. Harvard University Press, Cambridge
Lawley DN, Maxwell AE (1971) Factor analysis as a statistical method, 2nd edn. American Elsevier, New York
Liski EP (1991) Detecting influential measurements in a growth curve model. Biometrics 47:659–668
Loughran T, Nagin DS (2006) Finite sample effects in group-based trajectory models, Sociol Methods Res 35:250–278
McLachlan G, Peel D (2000) Finite mixture models. John Wiley, New York
Moffitt T (1993) Adolescence-limited and life-course-persistent antisocial behavior: a developmental taxonomy. Psychol Rev 100:674–701
Muthén B (2001a) Latent variable mixture modeling. In: Marcoulides GA, Schumacker RE (eds) New developments and techniques in structural equation modeling. Lawrence Erlbaum Associates, pp 1–33
Muthén B (2001b) Second-generation structural equation modeling with a combination of categorical and continuous latent variables: new opportunities for latent class/latent growth modeling. In: Collins LM, Sayer A (eds) New methods for the analysis of change. APA, Washington, D.C., pp 291–322
Muthén B (2002) Beyond SEM: general latent variable modeling. Behaviormetrika 29:81–117
Muthén B (2003) Statistical and substantive checking in growth mixture modeling: comment on Bauer and Curran. Psychol Methods 8:369–377
Muthén B (2004) Latent variable analysis: growth mixture modeling and related techniques for longitudinal data. In: Kaplan D (ed) Handbook of quantitative methodology for the social sciences. Sage, Newbury Park, pp 345–368
Muthén B, Asparouhov T (in press) Growth mixture modeling: analysis with non-Gaussian random effects. Draft available at: http://www.statmodel.com/download/ChapmanHall06V24.pdf Forthcoming in: Fitzmaurice G, Davidian M, Verbeke G, Molenberghs G (eds) Advances in longitudinal data analysis. Chapman & Hall/CRC Press
Muthén LK, Muthén BO (1998–2007). Mplus user’s guide, 4th edn. Muthén & Muthén, Los Angeles, CA
Nagin DS (1999) Analyzing developmental trajectories: a semi-parametric, group-based approach. Psychol Methods 4:139–157
Nagin DS, Farrington DP, Moffitt TE (1995) Life-course trajectories of different types of offenders. Criminology 33:111–137
Nagin DS, Land KC (1993) Age, criminal careers, and population heterogeneity: specification and estimation of a nonparametric, mixed Poisson model. Criminology 31:327–362
Nieuwbeerta P, Blokland A (2003) Criminal careers in adult Dutch offenders (Codebook and Documentation). NSCR, Leiden
Nylund KL, Asparouhov T, Muthén B (in press) Deciding on the number of classes in latent class analysis and growth mixture modeling: a Monte Carlo simulation study. Draft available at: http://www.statmodel.com/download/LCA_tech11_nylund_v83.pdf To appear in: Structural Equation Modeling: A Multidisciplinary Journal
Piquero AR (2007) Taking stock of developmental trajectories of criminal activity over the life course. In: Liberman A (ed) The long view of crime: a synthesis of longitudinal research. Springer, New York
Piquero AR, Buka SL (2002) Linking juvenile and adult pattern of criminal activity in the providence cohort to the National Collaborative Perinatal Project. J Crim Justice 30:259–272
Piquero AR, Farrington DP, Blumstein A (2003) The criminal career paradigm: background and recent developments. Crime and Justice: A Rev Res 30:359–506
Raudenbush SW, Bryk AS (2002) Hierarchical linear models: applications and data analysis methods, 2nd edn. Sage Publications, Newbury Park
Raudenbush SW (2005) How do we study “what happens next”? Ann Am Acad Pol Soc Sci 602:131–144
Roeder K, Lynch K, Nagin D (1999) Modeling uncertainty in latent class membership: a case study in criminology. J Am Stat Assoc 94:766–776
Roeder K, Wasserman L (1997) Practical Bayesian density estimation using mixtures of normals. J Am Stat Assoc 92:894–902
Sampson RJ, Laub JH (2005a) A life-course view of the development of crime. Ann Am Acad Pol Soc Sci 602:12–45
Sampson RJ, Laub JH (2005b) Seductions of methods: rejoinder to Nagin and Tremblay’s “developmental trajectory groups: fact or fiction?”. Criminology 43:905–913
Schwarz G (1978) Estimating the dimension of a model. Ann Stat 6:461–464
Silverman BW (1986) Density estimation for statistics and data analysis. Chapman and Hall, London
Soromenho G (1994) Comparing approaches for testing the number of components in a finite mixture model. Comput Stat 9:65–78
Tracy P, Wolfgang ME, Figlio RM (1990). Delinquency careers in two birth cohorts. Plenum Press, New York
Acknowledgments
We like to acknowledge everyone who discussed this article with us during the last 5 years at various meetings and conferences. In particular we thank Tihomir Asparouhov for ongoing discussions that shaped our perspective on this article. Shawn Bushway, Booil Jo, John Laub, Katherine Masyn, Daniel Nagin, Paul Nieuwbeerta and three anonymous reviewers provided critical comments to earlier versions of this manuscript that we gladly took into account. Michael Lemay was of great help in data preparation and analysis. The work on this article was partially supported by grant K02 AA 00230 from National Institute on Alcohol Abuse and Alcoholism.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
Zero-inflated Poisson Model
Note that the ZIP model is already a special case of a finite mixture model with two classes. Treating the count outcome variable as zero-inflated at each time point means that a probability is estimated for the observation to be either in the “zero-class” or not. For the zero class a zero count occurs with probability one. For the non-zero class, the probability of a conviction is expressed with a Poisson process.
The interesting feature for the ZIP, or its expression as a two-class model, is that the probability of being in the zero class can be modeled by covariates that are different from those that predict the counts for the Poisson class. The same is true when allowing for a zero class in the growth trajectory modeling.
More formally, for the present application this model can be represented as follows: At each individual time point a count outcome variable U ti (the number of conviction at each time point t for individual i) is distributed as ZIP (Roeder et al. 1999).
The parameters ρ it and λ it can be represented with \( {\text{logit(}}\rho _{{ti}} {\text{) = log[}}\rho _{ti} {\text{/1}} - \rho _{ti} {\text{] = X}}_{ti} \gamma _{{t}} {\text{ }} \) and \( \log (\lambda _{{ti}} ) = {\rm X}_{{ti}} \beta _{i} \).
Notice that the mixture model within the zero-inflated Poisson is a mixture at each time point. The mixture models we discuss in the different growth models are mixtures of different growth trajectories (across all time points).
Rights and permissions
About this article
Cite this article
Kreuter, F., Muthén, B. Analyzing Criminal Trajectory Profiles: Bridging Multilevel and Group-based Approaches Using Growth Mixture Modeling. J Quant Criminol 24, 1–31 (2008). https://doi.org/10.1007/s10940-007-9036-0
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10940-007-9036-0