Social and Environmental Determinants Show Greater Predictive Attribution than Health Behaviours for Multimorbidity: A Population-Based Machine Learning Study

Background/Objectives: Multimorbidity is usually attributed to ageing and health behaviours; social and environmental exposures are rarely quantified in the same model. We estimated how much of the predictive attribution for chronic condition burden, beyond age, falls to social, environmental, and behavioural domains. Methods: We analysed 16,590 respondents aged ≥15 years in the 2023–2024 Chilean National Quality of Life and Health Survey, counting 19 physician-diagnosed chronic conditions. After removing age with design-weighted splines, we modelled the residual count on 151 upstream determinants with 17 supervised algorithms, using survey weights and disjoint sampling units for training and testing. SHapley Additive exPlanations (SHAP) values were aggregated into pre-specified domains, summed, and averaged per variable. Results: Among 12,070 complete cases, weighted multimorbidity prevalence was 40.1%; age alone explained 23.2% of variance. The random forest used for attribution reached test R2 = 0.117 (95% confidence interval 0.091–0.138), leaving 88.3% of residual variation unexplained. Predictive attribution was highest for social isolation and support (23.1%), neighbourhood environment (20.9%), and socio-economic position (20.3%), whereas health behaviours received 12.0% and the lowest attribution per variable of any substantive domain. Each of these domains exceeded health behaviours with age modelled directly, with size-normalised or grouped attribution, with reweighting for exclusion, and without mental health diagnoses in the count; among adults aged 15–44 years, however, health behaviours received the largest summed attribution (20.9%). Conclusions: After accounting for age, social isolation, neighbourhood conditions, and material deprivation received greater predictive attribution within the final model than health behaviours did, whereas behaviours led among younger adults, suggesting age-specific pathways into multimorbidity. Attribution is not a causal estimate; most variation remained unexplained, and reverse causation cannot be excluded. Nevertheless, the ordering questions behaviour-centred framings of prevention, and these domains warrant evaluation as prevention targets in longitudinal and intervention studies.