Network Analysis Reveals Hidden Risk Patterns Behind Identical Scores

Published in Applied Network Science, the study introduces a network-based framework to examine configuration-level heterogeneity among score-equivalent cases. By analyzing 17,272 administrative assessments from the Chilean prison system, the authors show that additive scoring obscures how risk factors are organized. Identical total scores map onto distinct and recurring network architectures, indicating that relational structure contains substantively relevant information beyond simple severity totals. This finding challenges the assumption that a single sum score fully captures the complexity of multidimensional risk profiles in applied settings.

The study addresses a specific analytic cost associated with composite measurement instruments. These tools are widely used to summarize complex, multidimensional information into operational scores, often through additive rules that combine multiple indicators into a single measure of severity. While this approach is operationally efficient and easy to communicate, it abstracts from how underlying components are distributed. Observations with the same total may differ significantly in the configuration of indicators that produced that total. The authors argue that this limitation becomes consequential when equivalent aggregate totals imply different intervention needs or decision-making requirements in case management contexts.

To investigate this issue, the researchers utilized data from the Instrumento de Gestión de Intervención (IGI), the standard assessment tool used in the Chilean prison system. The IGI operationalizes criminogenic risk through a structured inventory of 43 items organized into eight domains, including criminal history, education, employment, family, leisure, peers, substance use, pro-criminal attitudes, and antisocial behavior. The dataset consists of assessments conducted by Gendarmería de Chile between 2015 and 2022 as part of routine case management. The records were provided in de-identified form, ensuring that the analysis reflects conditions of applied risk assessment in a real institutional setting rather than a protocol designed specifically for research purposes.

The methodological approach shifts attention from additive severity to relational structure. The authors recoded item responses into binary present/absent indicators to focus on the structure of co-occurrence rather than within-item severity gradations. They conducted the analysis within narrow, overlapping score windows to hold overall severity approximately constant. This design reduces the risk of conflating structural differences with simple differences in the number of activated items. By conditioning on small ranges of the sum score, the study isolates the internal organization of indicators among individuals who are otherwise comparable in overall severity.

Network psychometrics provided the tools to describe this relational structure. Items were treated as nodes and conditional associations as edges, allowing the construct to be analyzed as a system of interacting components. The researchers used Ising models for binary data to estimate conditional dependence patterns. This statistical approach separates consistent co-activation or decoupling from links that arise merely because some indicators are common. The analysis involved three stages: identifying score-equivalent subsets, clustering individuals by similarity in their activation patterns using Jaccard distance, and estimating separate networks for each configuration to describe specific conditional dependence structures.

Results indicated that the mid-range of IGI scores contains multiple stable configurations of needs. The raw item-prevalence map showed that item activation rates vary substantially across the score range, meaning two people can land on the same sum score while carrying very different mixes of active needs. Some items remained highly prevalent across almost the entire score range, contributing little to differentiating configurations. In contrast, other items displayed graded and informative patterns that a single severity label would obscure. Clustering via Jaccard distance identified three distinct configurations within each score window, providing a parsimonious description of the main patterns without fragmenting the sample into thin slices.

Stability checks confirmed that many estimated edges were robust under resampling. Bootstrap edge-weight estimates showed that the strongest edges had relatively narrow intervals with a consistent sign, while others were imprecisely estimated or overlapped zero. This variability highlights the importance of robustness diagnostics in network analysis. The researchers also conducted focal robustness tests to rule out selection artifacts driven by fixing the sum score. When controlling for a restricted score that freed the focal pair from the sum constraint, 92.5% of the edge significance patterns remained stable, confirming that the reported configurations capture robust central tendencies in the data rather than methodological byproducts.

The study further examined whether these configuration-specific patterns predict recidivism. Recidivism was defined as any recorded re-entry within two years of the index assessment. Logistic models estimated separately for each configuration and score window showed that a subset of configuration-specific relational indicators is associated with recidivism even among score-equivalent individuals. Some joint presence indicators were associated with re-entry in specific windows and configurations, and several of these associations persisted after adjustment for standard covariates such as age, sex, relationship status, and mental health. This suggests that relational features carry prognostic information not captured by the sum score alone.

The findings illustrate how the same severity can encode distinct organizational patterns. In the consumption domain, relational structures separated substance-use problems into profiles that differed in whether consequences appeared fragmented across life areas, broadly co-activated, or partly anchored in preserved routines such as employment. These forms imply different intervention problems, with some profiles requiring better integration across needs and others reflecting more entrenched impairment. The antisocial attitude–behavior domain showed that the meaning of coherence can shift with severity, where alignment among items may make lower-score profiles easier to identify but reflect a more consolidated pattern in mid-range windows.

The authors emphasize that sum scores remain practical summaries for screening and monitoring when the relevant question concerns aggregate severity. However, their limitations become consequential when the configuration of needs may inform intervention priorities. A configuration-aware reading complements domain-based assessment by highlighting when a profile is tightly integrated versus dispersed. It flags relational patterns that may be hard to see case-by-case, offering a lens for interpreting why some less immediately salient profiles can still show elevated recidivism within the same score band. This shift from severity-only assessment to configurational assessment within severity represents a significant methodological contribution to applied measurement instruments.

Limitations of the study include the specific institutional context of the Chilean prison system, where configurations and their relationship to re-entry may shift across systems with different programming and release regimes. The binarization of items is a modeling decision driven by the research question, compressing intensity information that could matter for prediction. Additionally, conditioning on score windows can induce selection artifacts, so the safest reading is conditional: these are structures and associations within score bands, not population-level estimates that necessarily generalize across the full score range. Future research should replicate these findings across jurisdictions and examine whether alternative aggregation choices preserve or transform configuration-level heterogeneity.

The framework developed in this study has broader implications for other administrative score-based systems where equifinality may obscure substantively meaningful variation. Network psychometrics can uncover latent structural heterogeneity in applied measurement instruments, providing tools to make visible forms of within-score heterogeneity that aggregation leaves implicit. By combining network structure with score-conditioned designs, the study offers a workable bridge between relational patterns and outcomes, extending the composite-indicator literature on weighting, aggregation, and robustness. This approach allows researchers and practitioners to move beyond additive scoring to understand the complex architectures of risk that underlie operational decisions.

Subject of Research: Network-based analysis of configuration-level heterogeneity in composite criminological risk measures

Article Title: Same score, different network: configuration-level heterogeneity in composite risk measures

Article References: Fábrega, J., & Olivera, A. (2026). Same score, different network: configuration-level heterogeneity in composite risk measures. Applied Network Science, 11(1), Article 64. https://doi.org/10.1007/s41109-026-00815-1

Image Credits: AI Generated

DOI: 10.1007/s41109-026-00815-1

Keywords: network psychometrics, composite risk measures, criminogenic risk, configuration-level heterogeneity, Ising models, score-equivalence, recidivism prediction, administrative data, relational structure, additive scoring, Same, score