On the determinants of intensity and duration in institutional long-term care in Switzerland: New insights from random forest modeling

This study examines the intensity of care and duration of stay among elderly residents in long-term care institutions in the canton of Geneva, Switzerland. We use a dataset of 26 060 individuals spanning 1998–2024 and described by 55 covariates, including demographic information, medical records, and quality-of-life indicators. Building on prior work based on classical regression methods, we propose a three-step machine learning approach: a random forest to identify the key determinants of each outcome and estimate their relative contributions, density-based clustering using the forest’s distance measure, and a classification tree to identify the drivers of cluster membership. We identify ten distinct clusters for intensity of care and five for duration of stay. The level of dependence is the dominant predictor across both outcomes, with effects most pronounced at high dependence levels. Quality-of-life indicators prove stronger predictors than medical diagnoses; in particular, gender outperforms primary diagnosis in predicting duration of stay. The random forest outperforms classical models in predictive accuracy, though it yields a lower C-index, consistent with the homogeneity of the institutionalized population. Cluster-level estimates of total care volume per resident profile provide actionable guidance for planning infrastructure, healthcare personnel, and financial resources, relevant to policymakers, insurers, care institutions, and individuals alike.