Longitudinal modeling of population heterogeneity: Methodological challenges to the analysis of empirically derived criminal trajectory profiles

Frauke Kreuter, Bengt MuthénMADOC 2008 · OpenAlex

Introduction

In the field of criminology, researchers frequently attempt to identify "trajectory profiles"—distinct groups of individuals who follow similar patterns of criminal behavior over the life course. These profiles, such as "adolescence-limited" or "life-course persistent" offenders, have become foundational to developmental criminology. However, a significant methodological debate persists: is population heterogeneity truly discrete (consisting of distinct types of people) or is it continuous (a spectrum of variation)?

This paper evaluates three powerful statistical frameworks—Growth Mixture Modeling (GMM), Latent Class Growth Analysis (LCGA), and Non-parametric GMM (NP-GMM)—to determine how these modeling choices influence the interpretation of criminal careers. The authors argue that while different models may yield similar statistical fits, they often produce divergent narratives about offender groups, highlighting the risk of "reifying" statistical clusters as real-world categories without external validation.

Methodological Framework

The study evaluates three primary longitudinal modeling strategies, all utilizing a Zero-Inflated Poisson (ZIP) framework. This framework is essential for handling the "excess zeros" typical of crime data, where the majority of a population has zero arrests in any given year.

  • Growth Mixture Models (GMM): These models assume the population is a mixture of several groups but allow for individual variation (random effects) within those groups. This acknowledges that even individuals in the same "class" will not behave identically.
  • Latent Class Growth Analysis (LCGA): A more restrictive version of GMM where all variation is assumed to be between groups. Individuals within a specific group are assumed to follow the exact same trajectory, with no internal variance.
  • Non-parametric GMM (NP-GMM): This approach does not assume a normal distribution for individual variation. Instead, it utilizes "mass points" (additional sub-classes) to approximate the actual shape of the distribution, providing a flexible alternative when data is highly skewed.

Data and Analysis

The authors applied these models to two classic longitudinal datasets:

  1. The Philadelphia Cohort Study: Data on 13,160 males born in 1958, tracking police contacts from ages 4 to 26. A random subset of 1,000 individuals was used for this comparative analysis.
  2. The Cambridge Study in Delinquent Development: A smaller, long-term dataset following 403 males from London from ages 10 to 40, focusing on official convictions.

Key Insights: The "Reification" Trap

A central finding of the paper is the danger of treating statistical artifacts as social realities. Because LCGA does not allow for within-group variation, it often requires a higher number of classes to achieve a good statistical fit. For example, a researcher might identify five distinct "types" of criminals using LCGA, whereas a GMM might suggest only three types with significant internal "wiggle room."

Furthermore, the authors note that high "Entropy" scores—a measure of how cleanly a model classifies individuals—can be misleading. High entropy is often a mathematical byproduct of a model's constraints rather than a reflection of distinct human categories. If the only difference between two identified groups is the "intercept" (the starting level of offending), they may simply be variations of the same developmental theme rather than truly different types of offenders.

Comparative Results

  • Model Parsimony: In the Philadelphia data, the GMM provided a better (lower) Bayesian Information Criterion (BIC) than the LCGA, despite using fewer parameters (15 vs. 22). This suggests that allowing for within-class variation is a more efficient way to model the data than creating numerous rigid classes.
  • Trajectory Convergence: In the smaller Cambridge dataset, the NP-GMM and LCGA results were remarkably similar. This suggests that in smaller samples, the choice of model may have a less dramatic impact on the visual representation of trajectories.
  • Temporal Shifts: Model choice also affected the identified "peak" of criminal activity. In the Philadelphia data, LCGA identified late-peaking classes at age 19, while GMM and NP-GMM placed those peaks at age 17. Such shifts are critical, as they can change how criminologists interpret the impact of late-adolescent life events on desistance.

Conclusion and Future Directions

The authors conclude that statistical fit indices like Log-Likelihood and BIC are insufficient for determining which model is "correct." Instead, the validity of a trajectory profile should be tested against "distal outcomes" (e.g., does group membership predict future recidivism or employment?) and "antecedents" (e.g., does it correlate with childhood trauma?).

Ultimately, successful modeling requires a synthesis of technical proficiency and substantive expertise. Statistics can identify clusters, but only criminological theory can determine if those clusters represent meaningful human behavior. Future research should expand these models to allow for random variation not just in the starting point (intercept) but also in the rate of change over time (slopes).

Summary updated

BibTeX
@inproceedings{kreuter2008longitudinal,
  author = {Frauke Kreuter and Bengt Muthén},
  title = {{Longitudinal modeling of population heterogeneity: Methodological challenges to the analysis of empirically derived criminal trajectory profiles}},
  booktitle = {MADOC},
  year = {2008},
}

Loading related content...

Figure