Research map / Probabilistic and scientific modeling / Molecules and proteins
Doubly Robust Conformalized Survival Analysis with Right-Censored 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 DR-COSARC, a doubly robust conformal method that imputes censoring times and calibrates lower prediction bounds for survival times under right-censoring.
Problem
Constructing valid lower prediction bounds for survival times from right-censored data, where censoring times are unobserved for individuals who experience the event, which breaks existing type-I censoring conformal methods.
Approach
Impute latent censoring times for uncensored-event individuals by sampling from the estimated censoring density conditioned on exceeding the observed time, convert the data to semi-synthetic type-I censored data, apply weighted conformal calibration (fixed or adaptive cutoffs), and clip with an uncalibrated survival-model quantile to obtain asymptotic double robustness.
Key results
- In synthetic setting 1 (hardest), DR-COSARC achieves coverage close to the 90% target while Uncalibrated and KM Decensoring under-cover.
- On seven real datasets with four survival models, DR-COSARC consistently achieves average estimated coverage closer to the 90% nominal level than Uncalibrated and KM Decensoring, while Naive CQR is overly conservative.
- The adaptive-cutoff version consistently outperforms the fixed-cutoff version in the synthetic experiments and is recommended.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DR-COSARC instead of Uncalibrated or KM Decensoring | DR-COSARC · Synthetic setting 1 | coverage | Helped |
| Use DR-COSARC instead of Uncalibrated or KM Decensoring | DR-COSARC · Synthetic setting 2 | coverage | Helped |
| Use DR-COSARC instead of other methods | DR-COSARC · Synthetic setting 3 | average LPB | No clear effect |
| Use DR-COSARC instead of Uncalibrated or KM Decensoring | DR-COSARC · Seven real datasets | estimated coverage | Helped |
| Use adaptive cutoffs instead of fixed cutoffs | DR-COSARC · Synthetic experiments | average LPB | Helped |
| Reduce the censoring-model training sample size, lowering imputation quality (ablation) | DR-COSARC · Synthetic setting 1 | coverage | Hurt |
| Increase the censoring-model training sample size (ablation) | DR-COSARC · Synthetic setting 1 | coverage | Helped |
| Use more covariates to fit the censoring model (ablation) | DR-COSARC · Synthetic data | coverage | Hurt |
| Vary the censoring-model training sample size (ablation) | DR-COSARC · Synthetic settings 2 and 3 | coverage | No clear effect |
| Add the double-robustness adjustment that takes the minimum of the calibrated bound and the uncalibrated quantile (ablation) | DR-COSARC · Synthetic experiments | average LPB | No clear effect |
| Use adaptive cutoffs instead of fixed cutoffs | DR-COSARC · Imputation randomness stability experiments | stability | Helped |
Surprising findings
- Uncalibrated survival-model LPBs are often reasonably well calibrated on the real datasets, and the double-robustness clipping step often has little practical effect.
Limitations
- Focuses only on lower prediction bounds rather than two-sided or upper bounds.
- Asymptotic double robustness only; finite-sample bounds are too loose to be practically useful.
- Fixed-cutoff version is sensitive to the choice of cutoff c0 and results depend on random data splits and imputation randomness.
Experimental setup
Datasets: VALCT PBC GBSG METABRIC COLON HEART RETINOPATHY synthetic log-normal/exponential settings
Benchmarks: synthetic settings 1-3 VALCT PBC GBSG METABRIC COLON HEART RETINOPATHY
Metrics: empirical coverage average lower bound value (normalized by oracle) estimated coverage bounds (beta_low, beta_upp, midpoint)
Baselines: Uncalibrated Naive CQR KM Decensoring Oracle
Builds on: conformalized survival analysis under type-I censoring (Gui et al.) adaptive cutoff conformal survival (Davidov & Ravid-type / Gui et al. adaptive) conformalized quantile regression weighted conformal inference KM decensoring conformal (Candès et al.)
Closest papers by meaning
- Conformalized Survival Distributions: A Generic Post-Process to Increase Calibration (ICML 2024 · similarity 0.89)
- Toward Conditional Distribution Calibration in Survival Prediction (NeurIPS 2024 · similarity 0.89)
- Censor Dependent Variational Inference (ICML 2025 · similarity 0.87)
- History-Aware Conformal Prediction Sets for Censored Time-to-Event Outcomes (NeurIPS 2026 · similarity 0.86)
- SurvivalPFN: Amortizing Survival Prediction via In-Context Bayesian Inference (NeurIPS 2026 · similarity 0.86)
- Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring (NeurIPS 2026 · similarity 0.86)
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.