Evaluating Subgrouping Methods for Health Intervention Policy Prioritization

Analyzing observational health data to derive actionable intervention policies is challenging, especially when aiming for stable and interpretable conclusions about patient subgroups. Traditional methods often struggle with the inherent limitations of observational data, such as the absence of true individual treatment effects and uncertain causal structures. This paper introduces a comprehensive framework to investigate whether unsupervised subgroups, formed solely from pretreatment characteristics, can serve as reliable units for prioritizing budget-constrained health policies. The framework integrates causal-discovery-informed covariate selection, sample splitting for evaluation, inductive unsupervised clustering (comparing methods like K-means, Fuzzy C-means, and Bayesian Gaussian mixture models), uncertainty-aware subgroup selection, and doubly robust policy evaluation. While the study found similar estimated utilities across different clustering methods for hypothetical obesity, glucose, and smoking-history interventions, it highlighted that these methods often prioritize distinct individuals, emphasizing the need for careful consideration in policy design.