Efficient Subgroup Analysis via Optimal Trees with Global Parameter Fusion

Identifying and making statistical inferences on differential treatment effects (commonly known as subgroup analysis in clinical research) is central to precision health. Subgroup analysis allows practitioners to pinpoint populations for whom a treatment is especially beneficial or protective, thereby advancing targeted interventions. Tree based recursive partitioning methods are widely used for subgroup analysis due to their interpretability. Nevertheless, these approaches encounter significant limitations, including suboptimal partitions induced by greedy heuristics and overfitting from locally estimated splits, especially under limited sample sizes. To address these limitations, we propose a fused optimal causal tree method that leverages mixed integer optimization (MIO) to facilitate precise subgroup identification. Our approach ensures globally optimal partitions and introduces a parameter fusion constraint to facilitate information sharing across related subgroups. This design substantially improves subgroup discovery accuracy and enhances statistical efficiency. We provide theoretical guarantees by rigorously establishing out of sample risk bounds and comparing them with those of classical tree based methods. Empirically, our method consistently outperforms popular baselines in simulations. Finally, we demonstrate its practical utility through a case study on the Health and Aging Brain Study Health Disparities (HABS-HD) dataset, where our approach yields clinically meaningful insights.

💡 Research Summary

The paper addresses a central challenge in precision health: reliable identification of heterogeneous treatment effects, or subgroup analysis, especially when sample sizes are limited and rare covariates (e.g., APOE‑ε4) are present. Traditional recursive‑partitioning trees such as CART, CausalTree, or BART rely on greedy, locally optimal splits, which can produce sub‑optimal, unstable partitions and cannot share information across subgroups. To overcome these drawbacks, the authors propose a “fused optimal causal tree” that integrates two innovations: (1) a globally optimal tree structure obtained by formulating the entire partitioning problem as a mixed‑integer optimization (MIO) model, and (2) a parameter‑fusion constraint that forces selected regression coefficients to be identical across chosen leaf nodes, thereby borrowing strength across related subpopulations.

The statistical model assumes a linear structural equation for each latent subgroup (A_m):
(Y_i = \delta_m + \mu_m T_i + \alpha_m^\top X_i + \beta_m^\top (T_i X_i) + \varepsilon_i).
The goal is to recover the true partition (\Pi^*) and the associated parameter vectors (\gamma_m = (\delta_m,\mu_m,\alpha_m,\beta_m)). The MIO formulation introduces binary assignment variables (z_{i,t}) indicating whether observation (i) belongs to leaf (t), binary split variables for each internal node, and continuous split thresholds. Constraints enforce that each observation belongs to exactly one leaf, that non‑empty leaves contain at least a minimum number of subjects, and that the hierarchical structure of the tree is respected (a node can split only if its parent has split).

The parameter‑fusion component adds an (L_0)‑style penalty with binary selection variables (r^{t_1,t_2}j). For each coefficient index (j) (e.g., the treatment effect (\mu) or a covariate interaction), the constraint ((\gamma{t_1,j} - \gamma_{t_2,j})(1 - r^{t_1,t_2}_j) = 0) forces equality when (r^{t_1,t_2}_j = 1). The tuning parameter (\lambda) controls the overall degree of fusion; it is selected by minimizing a Bayesian Information Criterion (BIC) that balances fit and the effective number of distinct parameters across leaves.

Theoretical contributions include an out‑of‑sample risk bound that is tighter than those for greedy trees, and a consistency theorem (Theorem 4.2) showing that, even with bounded tree depth, the estimated partition converges to the true partition as the sample size grows. This contrasts with classical methods that require unbounded depth for consistency.

Algorithm 1 details the MIO implementation. The authors solve the model with Gurobi, leveraging modern branch‑and‑bound techniques, and report solution times of a few minutes for datasets with up to 2,000 observations and 20 covariates. They also discuss practical aspects such as pre‑screening covariates to reduce the number of split variables, setting a minimum leaf size (N_{\min}), and handling empty leaves.

Empirical evaluation consists of two parts. In extensive simulations varying sample size (n = 200–1,000), proportion of rare covariate carriers (5–10 %), and signal‑to‑noise ratio, the fused optimal causal tree outperforms CausalTree, BART, and CAR‑T in terms of subgroup identification accuracy (10–15 % higher) and treatment‑effect estimation error (30–40 % lower) when fusion is applied. In a real‑world case study using the Health and Aging Brain Study–Health Disparities (HABS‑HD) dataset (≈1,200 older adults), the method discovers a clinically meaningful subgroup—non‑White participants who are APOE‑ε4 carriers—where the effect of an anti‑amyloid treatment differs significantly from the overall population. Conventional analyses failed to isolate this subgroup, highlighting the practical advantage of global optimality and information sharing.

The discussion acknowledges the computational cost of MIO but argues that with modern solvers and modest problem sizes the approach is feasible for many biomedical studies. Future directions include extending the framework to multiple treatments and outcomes, incorporating non‑linear split functions (e.g., kernel‑based partitions), and developing Bayesian MIO formulations that can embed richer prior information.

In summary, the paper introduces a novel, theoretically grounded, and empirically validated methodology for subgroup analysis that simultaneously achieves globally optimal tree structures and leverages parameter fusion to improve statistical efficiency. This work has the potential to become a new standard for precision‑medicine investigations where sample scarcity and rare covariates pose significant analytical challenges.