Research map / Probabilistic and scientific modeling / Treatment effects
Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring
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 partial-identification CATE bounds on survival time under informative (dependent) censoring and a doubly robust, quasi-oracle-efficient two-stage meta-learner (SurvB-learner) to estimate them.
Problem
Estimating conditional average treatment effects on survival time is biased when dropout/censoring is informative (depends on survival time), and point identification is impossible without strong assumptions.
Approach
Replaces non-informative censoring with sensitivity functions (censoring strength plus post-dropout survival bound) to derive Manski-style CATE bounds, and estimates them with a model-agnostic doubly robust two-stage pseudo-outcome meta-learner (SurvB-learner) that corrects plug-in bias.
Key results
- On synthetic data, SurvB-learner achieves lower RMSE to oracle bounds than the plug-in learner across all settings (e.g., exponential, censoring 0.2 domain knowledge: 2.541 vs 3.244), about 1.43x smaller on average.
- Conservative bound widths shrink as censoring strength decreases toward zero, consistent with the theory that width equals gamma times censoring strength.
- The paper claims SurvB-learner identifies benefiting subgroups on the ADJUVANT NSCLC trial in a small-sample setting.
Limitations
- Bounds can be wide when censoring is high and domain knowledge for post-dropout survival is absent.
- Sensitivity function gamma is treated as known and user-specified rather than learned from data.
- Relies on unconfoundedness and overlap; hidden confounding handled only in appendix.
Experimental setup
Datasets: ADJUVANT synthetic censored survival datasets (over 60 scenarios)
Benchmarks: synthetic censored survival datasets ADJUVANT
Metrics: RMSE vs oracle bounds bound width
Baselines: plug-in learner oracle bounds
Builds on: Manski bounds worst-case survival bounds DR-learner cross-fitting
Closest papers by meaning
- Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.91)
- Heterogeneous Treatment Effect in Time-to-Event Outcomes: Harnessing Censored Data with Recursively Imputed Trees (ICML 2025 · similarity 0.88)
- Unveiling the Potential of Robustness in Selecting Conditional Average Treatment Effect Estimators (NeurIPS 2024 · similarity 0.88)
- Learning Representations of Instruments for Partial Identification of Treatment Effects (ICML 2025 · similarity 0.88)
- Quantifying Aleatoric Uncertainty of the Treatment Effect: A Novel Orthogonal Learner (NeurIPS 2024 · similarity 0.88)
- Meta-Learners for Partially-Identified Treatment Effects Across Multiple Environments (ICML 2024 · similarity 0.87)
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.