Data-driven controlled subgroup selection in clinical trials

Subgroup selection in clinical trials is essential for identifying patient groups that react differently to a treatment, thereby enabling personalised medicine. In particular, subgroup selection can identify patient groups that respond particularly well to a treatment or that encounter adverse events more often. However, this is a post-selection inference problem, which may pose challenges for traditional techniques used for subgroup analysis, such as increased Type I error rates and potential biases from data-driven subgroup identification. In this paper, we present two methods for subgroup selection in regression problems: one based on generalised linear modelling and another on isotonic regression. We demonstrate how these methods can be used for data-driven subgroup identification in the analysis of clinical trials, focusing on two distinct tasks: identifying patient groups that are safe from manifesting adverse events and identifying patient groups with high treatment effect, while controlling for Type I error in both cases. A thorough simulation study is conducted to evaluate the strengths and weaknesses of each method, providing detailed insight into the sensitivity of the Type I error rate control to modelling assumptions.

💡 Research Summary

The paper introduces a rigorous statistical framework for data‑driven subgroup selection in clinical trials that simultaneously controls the Type I error rate. Traditional subgroup analyses either require pre‑specified groups, suffer from post‑hoc selection bias, or rely on two‑stage validation that reduces power. Here the authors define the target subgroup as the super‑level set Sτ = {x : η(x) ≥ τ}, where η(x) is the conditional mean response and τ a clinically relevant threshold (e.g., a safety probability bound or a minimum treatment effect). The goal is to select a region that is entirely contained in Sτ with probability at least 1 − α. Two concrete methods are proposed. The first builds on generalized linear models (GLM): η(x) is modeled via a link function and linear predictor, parameters are estimated by maximum likelihood, and confidence bands for the regression function are obtained through bootstrap or multiple‑testing adjustments. The resulting band yields a data‑driven cutoff that guarantees the desired error control when the model is correctly specified. The second method, Isotonic Subgroup Selection (ISS), imposes only a monotonicity constraint on η(x). Using order‑statistics‑based isotonic regression, the method estimates a non‑decreasing regression curve and identifies the smallest covariate values for which the estimated response exceeds τ. ISS is robust to model misspecification but typically has lower power than GLM. Extensive simulations explore a range of data‑generating mechanisms (linear, nonlinear, varying noise, high‑dimensional covariates) and compare the two approaches in terms of Type I error, power, and cutoff accuracy. Results show that GLM achieves high power when its assumptions hold, while ISS maintains reliable error control across diverse settings. The authors also contrast their methods with existing tools such as STEPP, SIDES, interaction trees, and causal trees, highlighting that their approach avoids a separate confirmation stage yet still provides strong inferential guarantees. Two motivating clinical questions—identifying a low‑risk safety subgroup and a high‑effectiveness efficacy subgroup—are used to illustrate practical implementation. The paper concludes with recommendations for hybrid strategies (e.g., using ISS for exploratory screening followed by GLM for confirmatory analysis) and discusses regulatory relevance, emphasizing that the framework aligns with recent FDA guidance on AI‑assisted decision‑making while offering a principled path toward personalized medicine.