Research map / Probabilistic and scientific modeling / Molecules and proteins
Deep Copula-Based Survival Analysis for Dependent Censoring with Identifiability Guarantees
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 DCSurvival, a deep copula-based survival model that learns an Archimedean copula (via a neural completely-monotone generator) jointly with event and censoring marginals, avoiding the need to specify the true copula under dependent censoring.
Problem
Survival analysis under dependent censoring, where the event and censoring times are correlated; existing copula-based methods require the practitioner to specify the true copula, and misspecification causes substantial bias.
Approach
Parameterize the Archimedean copula generator with a neural network built from convex combinations of negative exponentials (guaranteeing complete monotonicity), learn it end-to-end with neural density estimator marginals by maximizing the dependent-censoring likelihood, and prove identifiability under Weibull/log-normal-type marginals and this generator class.
Key results
- On the STEEL semi-synthetic dataset, DCSurvival achieves Survival-l1 of 0.092±0.015 versus 0.105±0.013 for KnownCopula and 0.157 for the best non-copula baseline (DeepHit).
- On Airfoil, DCSurvival reaches 0.176±0.047 Survival-l1, better than KnownCopula (0.215±0.060) and CoxPH (0.235±0.054).
- On synthetic data, DCSurvival recovers Clayton, Frank, Gumbel and Independence copula contours that closely match ground truth under Weibull marginals.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Learn the copula and survival marginals jointly instead of using a specified copula | DCSurvival · STEEL test | Survival-l1: 0.105\pm0.013 -> 0.092\pm0.015 | Helped |
| Learn the copula and survival marginals jointly instead of using a specified copula | DCSurvival · Airfoil test | Survival-l1: 0.215\pm0.060 -> 0.176\pm0.047 | Helped |
| Learn the copula from censored observations | DCSurvival with parametric survival marginals · Linear-Risk synthetic data | copula recovery | Helped |
Surprising findings
- KnownCopula's performance deteriorates as dependency increases, attributed to numerical difficulties in computing inverses of exponentials of theta.
- Under independent censoring, algorithms beating DCSurvival are restricted to proportional-hazard models that match the data generating process.
Limitations
- Identifiability holds only for Archimedean copulas representable by the neural generator and for specific marginal families (Weibull, log-normal, log-logistic, log-Student); the NDE instantiation lacks guaranteed necessary identifiability.
- Restricted to single-risk right-censored settings; competing risks and higher-dimensional copulas are left for future work.
- Real-world datasets lack ground truth, so evaluation relies on calibration plots only.
Experimental setup
Datasets: Linear-Risk (synthetic) Nonlinear-Risk (synthetic) STEEL Airfoil SEER GBSG2
Benchmarks: Linear-Risk Nonlinear-Risk STEEL Airfoil SEER GBSG2
Metrics: Survival-l1 calibration plots copula contour recovery
Baselines: CoxPH DeepSurv CoxCC RSF DeepHit PCHazard KnownCopula
Builds on: Archimedean copula neural density estimator (monotonic) Bernstein-Widder theorem Cox proportional hazards KnownCopula
Closest papers by meaning
- Survival Analysis via Density Estimation (ICML 2025 · similarity 0.89)
- Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.88)
- Censor Dependent Variational Inference (ICML 2025 · similarity 0.88)
- Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring (NeurIPS 2026 · similarity 0.87)
- SurvivalPFN: Amortizing Survival Prediction via In-Context Bayesian Inference (NeurIPS 2026 · similarity 0.87)
- Toward Conditional Distribution Calibration in Survival Prediction (NeurIPS 2024 · 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.