The KI-KRIM project combines fundamental criminological research with Social Data Science and addresses the following question: What can be learned from decades of criminological research when data from different studies are brought together and examined using new methodological approaches? To this end, survey data collected by the Criminological Research Institute of Lower Saxony (KFN) over more than two decades will be consolidated into a common database and analysed using machine learning (ML) methods to identify hidden patterns and potential mechanisms underlying juvenile delinquency.
Background
Datasets documenting past findings can provide insights into what may happen in the future (Gerritsen, 2020). In particular, combining data from numerous individual studies opens up new possibilities in this regard: patterns that remain hidden in individual surveys, change over time, or occur only in specific contexts may become apparent only when examined within a large, integrated database. Such a data foundation can also serve to advance social science theories through AI-supported methods – an approach that has so far received little attention in criminology (Campedelli, 2022; Shrestha et al., 2021).
Objectives
From a criminological perspective, the project examines how well established theories explaining juvenile delinquency hold up under different social, regional, and historical conditions. Where appropriate, it will also generate new hypotheses concerning previously overlooked relationships.
Methodologically, the project aims to demonstrate how ML methods can be adapted to the specific characteristics of social science data (e.g. measurement error, rare events, and complex structures), validated, and improved in terms of their interpretability.
Data
The project draws primarily on self-report surveys of unrecorded crime conducted by the KFN, particularly school-based surveys in Lower Saxony, supplemented by studies from cities and rural districts across Germany. This will result in a database comprising an estimated 200,000 respondents.
The surveys include information on self-reported offending and victimisation experiences, covering offences ranging from violent and property crimes to substance use, bullying, and cyberbullying. They also capture a wide range of explanatory factors, including parenting practices, school attachment, leisure activities, peer groups, media use, self-control, attitudes, and normative orientations. This makes it possible to operationalise key criminological theories.
Methods
Data integration: Since the studies often measure similar constructs using different questions and response formats, the data will be harmonised using model-based approaches (Integrative Data Analysis; Curran & Hussong, 2009). Psychometric methods such as SEM Trees (Brandmaier et al., 2013) will help assess the comparability of measurements across time points, regions, and groups.
Analysis: The project focuses on hybrid ML approaches that combine data-driven flexibility with theory-guided modelling (Molina & Garip, 2019). Model-based tree methods, for example, will examine whether theoretical models hold equally well across different contexts (e.g. social groups, regions, and time points).
Prediction-oriented methods will identify new explanatory factors, whose contributions can then be examined using interpretability methods such as Shapley values (Lundberg & Lee, 2017). Methods at the intersection of ML and causal inference (Künzel et al., 2019) will investigate for whom and under what conditions particular factors are especially influential. The limits of predictability will also provide valuable insights for theory development (Garip, 2020).
Data infrastructure
The integrated database will be established at the KFN as a permanent, extensible, and well-documented research data infrastructure.
Selected References
Brandmaier, A. M., von Oertzen, T., McArdle, J. J. & Lindenberger, U. (2013). Structural equation model trees. Psychological Methods, 18(1), 71–86.
Campedelli, G. M. (2022). Machine learning for criminology and crime research: At the crossroads. Routledge.
Curran, P. J. & Hussong, A. M. (2009). Integrative data analysis: The simultaneous analysis of multiple data sets. Psychological Methods, 14(2), 81–100.
Garip, F. (2020). What failure to predict life outcomes can teach us. PNAS, 117(15), 8234–8235.
Gerritsen, C. (2020). Big data and criminology from an AI perspective. In B. Leclerc & J. Cale (Hrsg.), Big data (S. 29–39). Routledge.
Künzel, S. R., Sekhon, J. S., Bickel, P. J., & Yu, B. (2019). Metalearners for estimating heterogeneous treatment effects using machine learning. PNAS, 116(10), 4156–4165.
Lundberg, S. M. & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30.
Molina, M. & Garip, F. (2019). Machine learning for sociology. Annual Review of Sociology, 45, 27–45.
Shrestha, Y. R., He, V. F., Puranam, P., & von Krogh, G. (2021). Algorithm supported induction for building theory: How can we use prediction models to theorize? Organization Science, 32(3), 856–880.
