Low-burden data-driven work addiction screening: Comparing demographic and psychosocial predictors using machine learning

Woropay-Hordziejewicz, Natalia A. and Charzyńska, Edyta and Buźniak, Aleksandra and Czerwiński, Stanisław K. and Schneider, Zuzanna and Atroszko, Paweł A. (2026) Low-burden data-driven work addiction screening: Comparing demographic and psychosocial predictors using machine learning. JOURNAL OF BEHAVIORAL ADDICTIONS, 15 (2). pp. 900-914. ISSN 2062-5871

Preview

Text 2006-article-p900.pdf - Published Version Available under License Creative Commons Attribution Non-commercial. Download (2MB) | Preview

Abstract

Background and Aims Identifying employees at risk for work addiction is critical for early intervention, yet systematic screening approaches face practical challenges for organizations. This study compared two screening strategies: a demographic-only model that requires no additional data collection versus a psychosocial model that incorporates job satisfaction and job stress assessments. Methods We analyzed data from employees across 85 countries or territories. Dataset 1 (psychosocial model, n = 22,136) included demographic, occupational, and psychosocial variables. Dataset 2 (demographic model, n = 23,219) used only readily available information. Work addiction risk was assessed using the International Work Addiction Scale. Five machine learning algorithms (Logistic Regression, Random Forest, XGBoost, Neural Networks, Ensemble Voting) were evaluated on both full and Elastic Net-selected feature sets using cross-validation. Results The psychosocial model demonstrated superior performance. The best-balanced model (Logistic Regression with Elastic Net selection) achieved balanced accuracy of 0.72 ± 0.01, AUC-ROC of 0.80 ± 0.01, and Youden's J of 0.45 ± 0.02, correctly identifying 72.7% of at-risk individuals while maintaining 72.2% specificity. The demographic model demonstrated moderate performance (Ensemble Voting: balanced accuracy 0.61 ± 0.01, AUC-ROC 0.65 ± 0.01), yet retained practical utility, detecting 57.0% of at-risk cases. High-sensitivity configurations achieved detection rates of 71%, but with elevated false-positive rates, presenting important trade-offs for screening design. Discussion and Conclusions While psychosocial variables provide superior prediction, demographic-only models offer a practical screening alternative that can identify over half of at-risk individuals without an additional data-collection burden. This trade-off between predictive performance and implementation feasibility has important implications for organizational screening programs.

Actions (login required)

Edit Item