From Unsupervised Subgroups to Hypothetical State-Intervention Policies: An Evaluation of Selected Subgrouping Methods in Observational Health Data

This study investigates whether patient subgroups constructed solely from pretreatment characteristics—without using information about treatment, outcomes, or estimated treatment effects—can function as effective, interpretable units for prioritizing health interventions under budget constraints. The authors aim to provide a transparent alternative to complex, supervised heterogeneous treatment effect (HTE) methods, which can be unstable and difficult to interpret in observational biomedical data.

The authors propose an end-to-end framework that includes causal-discovery-informed covariate selection, sample splitting to prevent overfitting, and inductive unsupervised clustering (e.g., K-means, Fuzzy C-means, Bayesian Gaussian mixture models). These methods are compared against a supervised causal-forest-derived CATE-tree comparator. The framework evaluates policies for hypothetical state shifts—such as moving from an obese to a non-obese state or reducing glucose levels—using the PIMA Indians Diabetes dataset and NHANES data. Policy performance is assessed using held-out doubly robust estimators, with uncertainty-aware gating mechanisms (Empirical Bernstein gating and hierarchical Bayesian pooling) used to prioritize subgroups.

The study found that phenotype-first subgrouping is a feasible approach for policy prioritization, with several methods achieving high estimated utilities (e.g., 0.799 for BMI policy using Bayesian GMM). However, the results show that no single clustering algorithm consistently dominates across different policy experiments. While supervised CATE-tree comparators produced greater within-group effect homogeneity, they did not consistently yield higher held-out policy utility. Additionally, after adjusting for multiple comparisons, no statistically significant differences in policy-risk were found between the different subgrouping methods, suggesting that the choice of algorithm is less critical than the underlying causal assumptions.