Technical Summary: Overcoming Dependent Censoring in the Evaluation of Survival Models

1. Problem Statement

Survival analysis differs from standard regression due to the presence of censored outcomes, where the exact event time is unknown but known to exceed a certain value. A fundamental assumption underlying most survival evaluation metrics, such as the Integrated Brier Score (IBS), is independent censoring. This assumption posits that the censoring time $C$ and the event time $E$ are conditionally independent given observed covariates $X$ ($E \perp\!\!\!\perp C \mid X$).

However, in many real-world scenarios, dependent censoring occurs ($E \not\perp\!\!\!\perp C \mid X$), often due to unobserved confounding variables (e.g., tumor grade affecting both relapse and patient dropout). When this assumption is violated, standard evaluation metrics relying on Inverse Probability of Censoring Weighting (IPCW) become invalid. Specifically, IPCW weights are derived from the marginal censoring distribution (typically estimated via the Kaplan-Meier estimator), which assumes independence. Under dependent censoring, these weights are misspecified, leading to biased performance estimates and distorted comparisons between survival models. While dependent censoring has been addressed in the context of model estimation, its impact on model evaluation remains largely unaddressed.

2. Methodology

The authors propose a Dependent Brier Score and its integrated version, IBS-Dep, designed to evaluate survival models under dependent censoring. The methodology consists of two primary components:

A. Copula-Graphic (CG) Estimator for Margin-Time Imputation

Instead of relying on the Kaplan-Meier estimator, the method utilizes the Copula-Graphic (CG) estimator to model the joint distribution of event and censoring times.

• Copula Modeling: The dependence between $E$ and $C$ is modeled using an Archimedean copula (e.g., Clayton or Frank), defined by a generator function $\phi_\theta$ and a dependence parameter $\theta$.
• Margin-Time Imputation: For censored instances, the method imputes a "margin time," defined as the conditional expectation of the event time given that the event time exceeds the censoring time ($E > t_i$).
  $$\tilde{e}_i = E[E \mid E > t_i] = t_i + \frac{\int_{t_i}^{\infty} S_E(u) du}{S_E(t_i)}$$
  Since the true survival function $S_E$ is unknown, the CG estimator $\hat{S}^{CG}_E$ is used to compute this imputed time. This imputation accounts for the dependence structure induced by the copula, unlike standard methods that assume independence.

B. Uncertainty Weighting

To account for the uncertainty inherent in imputing event times for censored instances, the authors introduce an adaptive weighting mechanism.

• Weights: Uncensored instances receive a weight of $w_i = 1$. Censored instances receive a weight $w_i = 1 - \hat{S}^{CG}_E(t_i)$, which corresponds to the estimated probability of the event having occurred by time $t_i$.
• Rationale: Early censoring (low survival probability) results in a smaller weight, reflecting higher uncertainty, while later censoring yields a larger weight.

C. The Dependent Brier Score

The proposed metric, $BS_{Dep}$, replaces the binary outcome indicator in the standard Brier score with the imputed margin time status and applies the uncertainty weights:
$$BS_{Dep} = \frac{1}{\sum w_i} \sum_{i=1}^N w_i \left( \mathbb{1}[\tilde{e}_i > t^*] - \hat{S}(t^* \mid x_i) \right)^2$$
The Integrated Dependent Brier Score (IBS-Dep) is the time-average of this score over a specified interval.

3. Theoretical Contributions

The paper establishes the theoretical properties of the proposed estimator:

• Consistency: The authors prove that the CG-based margin-time estimator is consistent for its target (the margin time induced by the assumed Archimedean copula model).
• Asymptotic Normality: Using a delta-method expansion and the functional central limit theorem, they demonstrate that the estimator converges at a rate of $N^{-1/2}$ and follows an asymptotic normal distribution.
• Validity: These properties imply that IBS-Dep is a consistent estimator for the corresponding margin-time surrogate Brier risk, even when the true censoring mechanism is dependent.

4. Experimental Evaluation

The authors evaluated the metric using two frameworks:

A. Synthetic Experiments

Using fully synthetic data with known ground-truth event times, the authors simulated dependent censoring via Archimedean copulas (Clayton and Frank) with varying Kendall's $\tau$.

• Findings: As dependence ($\tau$) increased, the standard IPCW-based IBS became increasingly unstable and biased. In contrast, IBS-Dep remained stable and achieved lower estimation error (closer to the oracle IBS).
• Robustness: Even under copula misspecification (e.g., using a Frank copula when data was generated by a Clayton copula), IBS-Dep remained less sensitive to dependent censoring than IPCW.

B. Semi-Synthetic Experiments

To evaluate performance on real data where ground truth is unavailable, the authors constructed semi-synthetic datasets. They learned event and censoring mechanisms from 12 real-world datasets (e.g., WHAS, METABRIC, MIMIC-IV, SEER) and resampled outcomes while preserving covariate structures. Dependent censoring was induced by omitting informative features (feature selection strategies).

• Results: Across 12 datasets and 4 feature selection strategies, the proposed metric reduced estimation error relative to the IPCW baseline.
  • The CG-based estimator alone reduced error by 4–8%.
  • The uncertainty-weighted variant (IBS-Dep with UW) achieved the largest reductions, averaging 12–16% lower error.
• Ranking: While IBS-Dep significantly reduced absolute estimation error, it did not consistently improve the ranking of models compared to the oracle IBS, suggesting that bias reduction does not always translate to better model separation for closely performing models.

5. Key Contributions and Significance

• First Metric for Dependent Censoring: To the authors' knowledge, this is the first metric specifically designed for survival analysis evaluation that supports dependent censoring.
• Bridging the Evaluation Gap: The work addresses a critical mismatch where models are increasingly designed to handle dependent censoring (e.g., via copula-based estimation), yet are evaluated using metrics that assume independence.
• Practical Impact: The proposed method reduces estimation bias in model evaluation, providing a more accurate assessment of model performance in real-world settings where unobserved confounders likely induce dependent censoring.
• Implementation: The authors provide a semi-synthetic framework for generating realistic dependent censoring scenarios and release source code to facilitate adoption.

The paper concludes that while dependent censoring is often treated as a rare edge case, it is a relevant trait of many real-world datasets. By shifting the focus from model estimation to model evaluation, this work offers a flexible, copula-based extension of the Brier score that yields more reliable performance estimates when the independence assumption is violated.