Research map / Probabilistic and scientific modeling / Treatment effects
Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data
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 toolbox of model-agnostic Neyman-orthogonal survival meta-learners for heterogeneous treatment effects from censored time-to-event data, with custom weighting functions that target treatment, censoring, and survival overlap.
Problem
Estimating heterogeneous treatment effects (survival probability differences) from censored time-to-event data is biased by censoring and vulnerable to low treatment, censoring, and survival overlap; existing orthogonal survival learners only handle some overlap types.
Approach
Derives an orthogonal (EIF-based) second-stage loss for a weighted target loss with arbitrary positive weighting f of the propensity, survival and censoring functions, yielding a family of learners (DR-, R-, and new C-, S-, and combined weighted learners) that are robust to different overlap violations and model-agnostic.
Key results
- On synthetic data, learners targeted to the overlap type present in the data achieve the lowest PEHE, and targeted weighting reduces estimation variance.
- On the Twins dataset, the T+C and C+S learners show lower estimation variance than others, suggesting censoring overlap is present.
- The C and T+C learners converge fastest on Twins validation loss, whereas the S learner converges slowly.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use overlap-targeted weighting matched to the low-overlap type | Orthogonal survival learners · Synthetic data, Scenarios 1 and 2 | PEHE | Helped |
| Use weighting that does not match the low-overlap type | Orthogonal survival learners · Synthetic data, Scenarios 1 and 2 | PEHE | Hurt |
| Apply censoring- or survival-overlap weighting | Orthogonal survival learners · Synthetic data, low censoring or survival overlap, later time steps | PEHE ratio | Mixed |
| Apply treatment-overlap weighting | Orthogonal survival learners · Synthetic data, low treatment overlap across time steps | PEHE ratio | Helped |
| Use T+C or C+S weighting | Orthogonal survival learners · Twins dataset | estimation variance | Helped |
| Use censoring-overlap weighting | Orthogonal survival learners · Twins dataset, validation during training | convergence speed | Helped |
| Use survival-overlap weighting | Orthogonal survival learners · Twins dataset, validation during training | convergence speed | Hurt |
Surprising findings
- Unsuitable weighting can harm performance relative to unweighted learners.
- Survival-overlap weighting (S-learner) hinders fast convergence on the Twins dataset.
Limitations
- Benefits of targeted weighting diminish over time steps as hazards increase and sample sizes at risk shrink.
- Unobserved ties require approximations (Method 1) or reparametrization (Method 2).
Experimental setup
Datasets: Synthetic Scenario 1 (1-dim confounder, 5 time steps) Synthetic Scenario 2 (10-dim confounders, 30 time steps) Twins dataset
Benchmarks: Synthetic Scenario 1 Synthetic Scenario 2 Twins dataset
Metrics: PEHE PEHE ratio over time validation loss
Baselines: DR-learner (survival, no weighting) R-learner (survival, treatment-overlap weighting)
Builds on: DR-learner R-learner Survival DR-learner Survival R-learner Kennedy orthogonal learning
Closest papers by meaning
- Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring (NeurIPS 2026 · similarity 0.91)
- Model-agnostic meta-learners for estimating heterogeneous treatment effects over time (ICLR 2025 · similarity 0.89)
- Heterogeneous Treatment Effect in Time-to-Event Outcomes: Harnessing Censored Data with Recursively Imputed Trees (ICML 2025 · similarity 0.89)
- Deep Copula-Based Survival Analysis for Dependent Censoring with Identifiability Guarantees (AAAI 2024 · similarity 0.88)
- Quantifying Aleatoric Uncertainty of the Treatment Effect: A Novel Orthogonal Learner (NeurIPS 2024 · similarity 0.88)
- SurvivalPFN: Amortizing Survival Prediction via In-Context Bayesian Inference (NeurIPS 2026 · 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.