Research map / Probabilistic and scientific modeling / Treatment effects
Dynamic Subgroup Identification in Covariate-adjusted Response-adaptive Randomization Experiments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a dynamic subgroup identification CARA design that adaptively identifies and merges tied best subgroups, reallocating treatment to maximize correct best-set identification, with asymptotic inference for the best-subgroup treatment effect.
Problem
Identifying the set of best-performing subgroups (those with the largest treatment effects, possibly tied) in randomized trials, where traditional designs pre-specify subgroups and fixed allocation, and existing adaptive methods target only a single best subgroup or cannot handle ties.
Approach
A CARA design that dynamically identifies the tie set of best subgroups via bootstrap-based near-tie tests with data-driven hyperparameters, merges tied subgroups, and reallocates treatment probabilities by solving a large-deviation-motivated max-min optimization under a resource constraint.
Key results
- In the synthetic PBC study with T=15 stages, the proposed design's correct selection probability approaches 1 and exceeds complete randomization, epsilon-greedy and UCB1.
- Table 1: the proposed design has SD 40.94 versus about 80-81 for the baselines, with sqrt(N)-scaled bias 29.98 versus 28.84-34.82.
- Theoretical results: consistency of best-tie-set identification (Theorem 1), convergence of allocation to oracle optimum (Theorem 2), and asymptotic normality with consistent variance estimator (Theorem 3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed adaptive design instead of complete randomization | Synthetic PBC experiment · Synthetic PBC | correct selection probability | Helped |
| use the proposed adaptive design instead of ϵ-greedy | Synthetic PBC experiment · Synthetic PBC | correct selection probability | Helped |
| use the proposed adaptive design instead of UCB 1 | Synthetic PBC experiment · Synthetic PBC | standard deviation: 80.85 -> 40.94 | Helped |
| use the proposed adaptive design instead of complete randomization | Synthetic PBC experiment · Synthetic PBC | standard deviation: 80.29 -> 40.94 | Helped |
| use the proposed adaptive design instead of ϵ-greedy | Synthetic PBC experiment · Synthetic PBC | standard deviation: 81.19 -> 40.94 | Helped |
| use the proposed adaptive design instead of ϵ-greedy | Synthetic PBC experiment · Synthetic PBC | √N-scaled bias: 28.84 -> 29.98 | Hurt |
| use single rather than double bootstrap for hyperparameter selection (ablation) | Proposed dynamic subgroup identification design · Synthetic PBC | √N-scaled bias: 41.69 -> 34.44 | Helped |
| use separate rather than naïve bootstrap for merging (ablation) | Proposed dynamic subgroup identification design · Synthetic PBC | √N-scaled bias: 34.20 -> 29.98 | Helped |
| vary the bootstrap variants for hyperparameter selection and merging (ablation) | Proposed dynamic subgroup identification design · Synthetic PBC | correct selection probability | No clear effect |
Surprising findings
- The epsilon-greedy and UCB1 bandit baselines performed no better than complete randomization in the PBC study, with near-identical estimates and SDs.
Limitations
- Assumes outcomes are observed immediately without delay at the end of each stage.
- Only addresses constrained treatment budgets; does not incorporate early stopping for adverse subgroup effects.
- Evaluation relies on synthetic data generated from a Gaussian model calibrated to a real trial rather than real adaptive trial outcomes.
Experimental setup
Datasets: Mayo Clinic primary biliary cirrhosis (PBC) trial data National Supported Work (NSW) job training dataset
Benchmarks: synthetic PBC cirrhosis trial simulation NSW job dataset simulation
Metrics: correct selection probability normalized mutual information sqrt(N)-scaled bias standard deviation 95% confidence interval coverage/width
Baselines: Complete randomization design epsilon-greedy algorithm upper confidence bound 1 algorithm causal tree model
Builds on: covariate-adjusted response-adaptive randomization (CARA) large deviation theory Thompson sampling upper confidence bound
Closest papers by meaning
- Using Surrogates in Covariate-adjusted Response-adaptive Randomization Experiments with Delayed Outcomes (NeurIPS 2024 · similarity 0.88)
- Stochastic Grouping Conformal Prediction for Effective Subgroup Reliability (NeurIPS 2026 · similarity 0.85)
- Active Adaptive Experimental Design for Treatment Effect Estimation with Covariate Choice (ICML 2024 · similarity 0.84)
- On Efficient Estimation of Distributional Treatment Effects under Covariate-Adaptive Randomization (ICML 2025 · similarity 0.84)
- Safe Exploration in Dose Finding Clinical Trials with Heterogeneous Participants (ICML 2024 · similarity 0.84)
- Benchmarking Estimators for Natural Experiments: A Novel Dataset and a Doubly Robust Algorithm (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.