Social connection features predicting loneliness: A longitudinal, interpretable machine-learning analysis

Background: Loneliness is increasingly recognized as a major global health concern. As a marker of poor social health, loneliness may emerge when certain aspects of social connection are missing. Yet, it remains unclear to what extent and which features of social connection contribute to a vulnerability of loneliness. In this study, we examined how 24 features of individuals’ social connections spanning structural, functional and quality dimensions predict loneliness both at the same time point and over a two-year follow-up. Methods: Using fine-grained, tie-specific data from a population-based Dutch cohort aged 18–93 years at baseline (N = 6,852) and at follow-up (N = 3,175), we evaluated which social connection features are predictive of loneliness using random-forest models at both timepoints and across age groups. Findings: Across both analyses, network quality was the strongest predictor of loneliness. Good-quality and strained relationships, friends-know-family, and network size consistently predicted loneliness. Being in a steady relationship was a top predictor at baseline but less important at follow-up. Functional aspects were generally less predictive. Models explained 24% of baseline and 44% of follow-up variance.Interpretation: These findings suggest that loneliness is most strongly linked to the quality of one’s social ties rather than their function, and that individuals embedded in low-quality or strained networks are particularly vulnerable. The results also underscore that loneliness is shaped by a broader constellation of factors beyond social connections.