Interpretable subgroup identification across survival and generalized-linear outcomes
Abstract
A subgroup analysis searching a trial's covariates or many prespecified subgroups often singles out a subgroup for its extremity, interpreted as if prespecified. Selection inflates the estimate; for a given selection rule, a correction removes the inflation. Two questions belong to the rule: which candidate to return, and whether to declare none. A third is whether inference extends to the patients left outside. We treat all three for interpretable subgroups, defined by conjunctions of covariate cuts, generated by forest search, a causal forest or a cross-fit surface estimator under one set of rules, each a modelling choice trading size against effect. Forest search screens candidates on splitting consistency, the rate at which effects replicate across half-splits; the screen is complementary-pairs stability selection in closed form, and the law of the maximum over the family gives the family-wise false-declaration rate when every candidate's effect is at least a clinically specified benefit. Refit-free multiplier resampling gives adjusted one-sided bounds (subgroup and complement), joined in a two-subgroup claim, with coverage asymptotically exact under separated selection, characterized under local competition, and conditional on a family proposed from the outcomes. Two trial applications, and simulations built from them, cover survival, binary and continuous endpoints.
Related
- Co-Supervised Tree Synthesis for Interpretable Subgroup Identification and Honest Treatment Effect Inference in Randomized Trials
- Sharp Partial Identification for Survival Model Comparison Without Target Outcomes
- On subgroup perfect codes and regular sets in generalized Cayley graphs
- Generalized survival energy models with exogenous covariates for mortality prediction
- Max-Stable Survival Copulas under Dependent Censoring: Sharp Identification and Efficient Inference
BibTeX
@misc{leon2026interpretable,
title = {{Interpretable subgroup identification across survival and generalized-linear outcomes}},
author = {Larry F. Leon and Keaven M. Anderson},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.28379},
eprint = {2610.28379},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.28379},
}
RIS
TY - UNPB
TI - Interpretable subgroup identification across survival and generalized-linear outcomes
AU - Larry F. Leon
AU - Keaven M. Anderson
PY - 2026
DA - 2026/10/20
PB - arXiv
DO - 10.48550/arxiv.2610.28379
UR - https://arxiv.org/abs/2610.28379
AB - A subgroup analysis searching a trial's covariates or many prespecified subgroups often singles out a subgroup for its extremity, interpreted as if prespecified. Selection inflates the estimate; for a given selection rule, a correction removes the inflation. Two questions belong to the rule: which candidate to return, and whether to declare none. A third is whether inference extends to the patients left outside. We treat all three for interpretable subgroups, defined by conjunctions of covariate cuts, generated by forest search, a causal forest or a cross-fit surface estimator under one set of rules, each a modelling choice trading size against effect. Forest search screens candidates on splitting consistency, the rate at which effects replicate across half-splits; the screen is complementary-pairs stability selection in closed form, and the law of the maximum over the family gives the family-wise false-declaration rate when every candidate's effect is at least a clinically specified benefit. Refit-free multiplier resampling gives adjusted one-sided bounds (subgroup and complement), joined in a two-subgroup claim, with coverage asymptotically exact under separated selection, characterized under local competition, and conditional on a family proposed from the outcomes. Two trial applications, and simulations built from them, cover survival, binary and continuous endpoints.
ER -