Bij welke sociaal-demografische groepen komt suïcide het meest voor?

Guus Berkelmans, Renske Gilissen, Sandjai Bhulai, Rob van der Mei, Lizanne SchwerenTSG 2023 · DOI

Abstract

Identifying high-risk groups is essential for effective suicide prevention. While individual risk factors are well-documented, the complex interactions between multiple socio-demographic variables remain under-explored. This study utilizes machine learning methods to analyze Dutch national microdata, offering a flexible, data-driven approach to identifying high-risk clusters. The results reveal disproportionately high suicide rates among specific groups, including individuals on disability benefits who have never married, isolated seniors with low incomes, and widowed men. These findings provide a granular foundation for targeted public health interventions.

Introduction: Beyond Linear Risk Factors

Suicide prevention strategies traditionally rely on identifying broad risk categories, such as gender or age. However, human behavior and mental health are influenced by a complex interplay of life circumstances. Traditional statistical models often struggle to capture these "non-linear" interactions—where the combined effect of two or more factors is significantly greater than the sum of their parts. For instance, while being male or being unemployed are known risks, the specific risk profile of a widowed man who is also unfit for work may be far more extreme.

This research addresses this complexity by using machine learning to sift through millions of data points, identifying specific socio-demographic "pockets" where risk is concentrated. By moving away from pre-defined hypotheses, the study allows the data to reveal which combinations of traits—such as marital status, income, and employment—create the highest levels of vulnerability.

Methodology: A Data-Driven Pipeline

The researchers developed a four-step machine learning pipeline designed to balance discovery with interpretability:

  1. Data Partitioning: The dataset was split into training (50%), validation (40%), and test (10%) sets to ensure the findings were robust and not the result of overfitting to specific noise in the data.
  2. Interaction Detection: An algorithm scanned the training data to identify multi-way interactions. This allowed the model to find complex patterns, such as the intersection of age, household role, and income level.
  3. Model Estimation: The discovered interactions were incorporated into a logistic regression framework. This approach combines the pattern-finding power of machine learning with the interpretability of traditional statistics, providing clear Odds Ratios (OR) for each identified group.
  4. Validation: The model's predictive accuracy was confirmed using Area Under the Curve (AUC) and log-likelihood metrics.

Data Sources: The Power of National Microdata

The study utilized the Statistics Netherlands (CBS) microdata, a high-quality longitudinal database containing socio-demographic information for every inhabitant of the Netherlands.

  • Scale: The analysis covered approximately 137.6 million person-years, representing all Dutch residents aged 10 and older between 2011 and 2019.
  • Scope: The dataset included 16,417 recorded instances of suicide.
  • Variables: Features analyzed included gender, age, migration background, personal and household income, education level, healthcare costs, and various social security benefits (e.g., disability or unemployment).

Key Findings: Identifying Ultra-High-Risk Clusters

The national average suicide rate in the Netherlands is 11.8 per 100,000. The machine learning model identified 17 significant interactions that defined groups with risks far exceeding this baseline.

The Disability and Isolation Nexus

The highest risk was found among individuals receiving disability benefits (unfit for work) who had never been married. This group faced a suicide rate of 88.48 per 100,000—nearly 7.5 times the national average. Similarly, men receiving disability benefits showed a rate of 58.56 per 100,000. These findings suggest that the loss of professional identity combined with a lack of a primary domestic partner creates a severe vulnerability.

Socio-Economic "Perfect Storms"

The model highlighted a specific cluster of seniors (aged 55–69) who live alone, have never been married, and possess a low household income. This group had a suicide rate of 57.22 per 100,000. This "perfect storm" of social isolation and financial strain in late middle age represents a critical area for policy intervention.

Gendered Vulnerability in Bereavement

A striking insight from the data was the role of widowhood. While losing a spouse is a significant risk factor for men, the same effect was not observed for women. This disparity may stem from differences in social support networks or the traditional division of domestic labor, where men may be less equipped to manage social and household life following the loss of a partner.

Education and Youth

Among younger adults (ages 25–39), a low education level was identified as a significant risk multiplier. While individual rates for these groups are lower than for older cohorts, the interaction between youth and limited educational attainment significantly increased the odds of suicide (OR 1.63).

Conclusion and Policy Implications

This study demonstrates that suicide risk is not evenly distributed but concentrated in specific socio-demographic intersections. While the model is not intended as a clinical screening tool for individuals, it provides a roadmap for "bottom-up" public health strategies.

By identifying these high-risk pockets, policymakers can design more effective, targeted interventions. For example, outreach programs could be integrated into disability benefit processes or social services for isolated seniors. Future research should focus on qualitative studies to understand the lived experiences of these groups, bridging the gap between data-driven identification and empathetic, effective care.

Summary updated

Abstract

Samenvatting Inleiding Voor suïcidepreventie is het van belang om groepen met een verhoogd risico op suïcide zo goed mogelijk te identificeren. Tot nog toe is er weinig bekend over interacties van meerdere risicofactoren. Machine learning-methoden bieden nieuwe mogelijkheden voor flexibel, datagedreven, hypothesevrij en robuust onderzoek naar de interacties van risicofactoren voor suïcide. Methode In dit onderzoek is gebruikgemaakt van microdata van het Centraal Bureau voor de Statistiek met sociaal-demografische gegevens van alle Nederlandse inwoners. Er is een machine learning-model ontwikkeld en uitgevoerd om interacties van risicofactoren voor suïcide te vinden. Resultaten De resultaten laten disproportioneel hoge suïcide-incidenties zien onder 1) mensen met een arbeidsongeschiktheidsuitkering die nooit getrouwd zijn geweest, 2) arbeidsongeschikte mannen, 3) mensen tussen de 55 en 69 jaar oud, die alleen wonen, nooit getrouwd zijn en wiens huishoudinkomen laag is. Daarnaast hebben we hoge suïcidecijfers gevonden onder mensen tussen de 25 en 39 jaar met een laag opleidingsniveau en onder verweduwde mannen. Conclusie De bevindingen helpen om gericht interventies in te kunnen zetten om het aantal suïcides te verminderen.
BibTeX
@inproceedings{berkelmans2023bij,
  author = {Guus Berkelmans and Renske Gilissen and Sandjai Bhulai and Rob van der Mei and Lizanne Schweren},
  title = {{Bij welke sociaal-demografische groepen komt suïcide het meest voor?}},
  booktitle = {TSG},
  year = {2023},
  doi = {10.1007/s12508-023-00398-7},
}
Figure