Source-linked AI summary
Model-based Recursive Partitioning for Subgroup Analyses
Heidi Seibold, Achim Zeileis, Torsten Hothorn
TL;DR
The paper tackles the challenge of finding patient subgroups with different treatment effects using statistically appropriate, data-driven procedures. It applies model-based recursive partitioning to segment the protocol-specified overall treatment-effect model by predictive factors. The framework is illustrated for functional and survival outcomes in ALS patients receiving Riluzole and retains model-specific treatment parameters for resulting subgroups.
Problem
Subgroup analyses aim to identify patients with different treatment effects, but post-hoc analyses can be biased and difficult to perform with full statistical error control.
Method
Model-based recursive partitioning detects parameter instabilities in an overall treatment-effect model and links resulting subgroup-specific treatment parameters to predictive factors.
Results
The method is applied to ALS patients to obtain subgroups with differential Riluzole treatment effects across functional and survival models.
Takeaways & Limitations
The framework directly segments the model used for the overall treatment effect and can focus subgroup searches on predictive factors while holding other model terms fixed.
Takeaways & Limitations
Confidence intervals after selecting both the partitioning variable and split point have unclear validity, so subgroup findings require confirmation in follow-up trials.
Abstract
from arXiv · showhide
The identification of patient subgroups with differential treatment effects is the first step towards individualised treatments. A current draft guideline by the EMA discusses potentials and problems in subgroup analyses and formulated challenges to the development of appropriate statistical procedures for the data-driven identification of patient subgroups. We introduce model-based recursive partitioning as a procedure for the automated detection of patient subgroups that are identifiable by predictive factors. The method starts with a model for the overall treatment effect as defined for the primary analysis in the study protocol and uses measures for detecting parameter instabilities in this treatment effect. The procedure produces a segmented model with differential treatment parameters corresponding to each patient subgroup. The subgroups are linked to predictive factors by means of a decision tree. The method is applied to the search for subgroups of patients suffering from amyotrophic lateral sclerosis that differ with respect to their Riluzole treatment effect, the only currently approved drug for this disease.
1. Introduction
The paper addresses the difficulty of identifying treatment-effect subgroups with appropriate error control, especially when covariate interactions are complex. It introduces model-based recursive partitioning to connect predictive factors with differential treatment effects across ALS endpoints.
- Motivation: Post-hoc subgroup analyses seek patients who benefit, do not benefit, or fare worse under a new therapy, but can be biased and difficult to conduct with full error control.
- Motivation: In parallel-group trials, subgroup analysis can be framed as searching for treatment × covariate interactions because individual treatment effects are generally unavailable.
- Methodological challenge: Classical trees identify covariate interactions, but restricting them to treatment × covariate interactions and blending them with treatment-effect models remains technically challenging.
- Methodological challenge: The analysis distinguishes prognostic factors, which affect outcomes independently of treatment, from predictive factors, which explain differential treatment effects.
- Application: The paper applies the procedure to functional and survival outcomes in ALS, using normal GLM, proportional-odds, Weibull, and Cox models according to endpoint scale.
- Proposed approach: Model-based recursive partitioning segments functional and survival treatment-effect models to identify ALS subgroups with differential Riluzole effects while retaining the overall analysis model's measurement scale.
2. Model-based recursive partitioning for subgroup identification
Model-based recursive partitioning extends an overall-treatment-effect model into subgroups by detecting parameter instability through partial-score tests and recursive splits. The resulting tree identifies subgroup-specific intercepts and treatment effects while retaining the base model framework.
- Model formulation: The method partitions a protocol-specified model when intercept or treatment-effect parameters vary with patient characteristics.Subgroup-specific parameters are estimated within a partition of the patients, while selected covariate and nuisance effects may remain constant.
- Model formulation: The segmented model permits subgroup-dependent intercepts and treatment effects while expressing these parameters as functions of the partitioning variables.For the illustrated linear model, the conditional endpoint distribution uses α(z) and β(z) alongside treatment, stratum, and variance terms.
- Detection algorithm: Model-based recursive partitioning detects instability by testing independence between partial scores for the intercept or treatment effect and candidate partitioning variables.The procedure tests both score types across variables and selects the variable with the strongest detected association.
- Detection algorithm: The algorithm repeatedly estimates cut-points and refits partition-specific models until score-variable deviations from independence are no longer detected.Root, inner, and leaf nodes represent the full sample, splits, and final subgroups, respectively.
- Content interpretation: Partial-score patterns can reveal treatment-effect instability even when least-squares residuals do not reveal the corresponding cut-point.In the example, changing treatment scores led to a split near z1 < 0 while intercept scores fluctuated around zero.
- Content interpretation: Whether a partitioning variable is predictive or prognostic must be determined from the fitted model parameters, not from the score responsible for splitting.Variation in β indicates predictive involvement, whereas constant β with varying intercepts indicates prognostic involvement.
- Relation to established procedures: Compared with the Gs procedure, recursive partitioning tests partial scores and can avoid dependence on a cut-point near the mean of an ordinal variable.The supplied comparison states that Gs may have lower power or fail to split when its mean-based dichotomisation is poorly located.
3. Partitioning effects of Riluzole on ALS patients
The analysis applies model-based recursive partitioning to PRO-ACT data to identify ALS subgroups with differing Riluzole effects on functional and survival outcomes. Splits based on disease timing and patient characteristics reveal heterogeneous treatment patterns across endpoints.
- Method: Model-based recursive partitioning was applied to functional and survival models, allowing instability in both baseline and Riluzole treatment-effect parameters.Bonferroni-adjusted permutation tests assessed independence between partial score functions and partitioning variables, with tree depth restricted to two levels.
- 3.1. ALSFRS: The functional endpoint was ALSFRS six months after treatment start, modeled with a log-link Gaussian GLM adjusted for baseline ALSFRS.The model describes expected relative ALSFRS change over six months under Riluzole or no Riluzole.
- 3.1. ALSFRS: Time from disease onset to treatment, FVC, and phosphorus balance formed the ALSFRS tree’s splits, with negative Riluzole effects in one early-treatment, higher-FVC subgroup and little effect after longer delays.The early-treatment subgroup was defined by fewer than 468 days between onset and treatment start; longer delays generally indicated slower disease progression.
- 3.2. ALSFRS items: The item-level analysis used proportional-odds models with baseline-specific treatment effects and assessed instability across 250 parameters for ten ALSFRS items.The resulting tree split on time from onset to treatment, FVC, and lymphocyte percentage, while coefficient colors indicated positive, negative, or absent effects.
- 3.3. Survival time: Survival analyses used Weibull and Cox models, with age and time from onset to treatment forming similar partitions and little benefit among very young patients.For other groups, Riluzole showed prolonged life expectancy in the Weibull analysis and a slight tendency toward lower death risk in the Cox analysis.
- Cox model: In the Cox model, martingale residuals served as a surrogate score for the baseline hazard, while score residuals represented treatment parameters.Permutation tests were used because martingale residuals are not normally distributed; proportional hazards were assumed within each partition.
4. Discussion
The discussion emphasizes that model-based recursive partitioning directly segments the protocol-specified overall treatment model while controlling variable selection, but inference after selection remains unresolved and some data-model extensions are needed.
- Methodological advantages: Direct segmentation preserves the study protocol’s model for the overall treatment effect while producing subgroup-specific treatment effects.This avoids estimating overall and partitioned effects with different procedures and can avoid suboptimal models.
- Methodological advantages: Allowing splits in both intercept and treatment scores helps retain predictive factors, although it can also detect prognostic factors.The analysis determines whether partitioning variables are prognostic or predictive when interpreting the segmented model.
- Methodological advantages: Variable selection is error controlled, but many uninformative partitioning variables can reduce the chance of selecting an informative one.The procedure controls the probability of selecting a variable when all candidates are independent of the scores.
- Limitations: Confidence-interval validity after variable and break-point selection is unclear because the relevant inferential problem lacks established literature.The authors recommend treating subgroup confidence intervals conservatively as ranges of possible values rather than significance measures.
- Limitations: The ALS framework could be extended to longitudinal mixed models and joint models that combine functional and survival information.The current functional-endpoint modeling excludes patients who died within six months after treatment start.
- Conclusion: The authors judge the procedure broadly consistent with EMA requirements because it offers statistical error control and unbiased variable selection.An open-source implementation is available for applying the method elsewhere.
Computational details
The paper provides open-source code and data resources, including a Weibull-model implementation that computes score matrices and supplies them to a recursive-partitioning tree.
- Resources: The partykit package contains open-source implementations of the paper’s methods, while PRO-ACT data and analysis code are also made available.Database-reading and cleaning code is provided in the TH.data package, and analysis code appears in the supplementary material.
- Weibull implementation: The code listing is specifically a ctree()-based Weibull model implementation in model-based recursive partitioning.It connects the survival-model score computation to the tree-fitting procedure.
- Weibull implementation: The Weibull example defines a function that fits a weighted survival model and returns its score matrix for recursive partitioning.The function fits a Weibull survival regression with Riluzole as the treatment term and places score contributions into the returned matrix.
- Weibull implementation: The tree is computed with ctree() using the custom Weibull transformation and a maximum depth of two.The code snippet shows the recursive-partitioning call and its depth restriction.
Affiliation:
The paper lists affiliations for Heidi Seibold at the University of Zurich and Achim Zeileis at the University of Innsbruck.
- Affiliations: Heidi Seibold is affiliated with the Department of Biostatistics, Epidemiology, Biostatistics and Prevention Institute at the University of Zurich.The listed address is Hirschengraben 84, CH-8001 Zurich, Switzerland.
- Affiliations: Achim Zeileis is affiliated with the Department of Statistics, Faculty of Economics and Statistics, at the University of Innsbruck.The listed address is Universitätsstr. 15, A-6020 Innsbruck, Austria.