In survival analysis, we rarely get to see the full story. We want to know when an event—like a heart attack or a machine failure—will happen, but often the "clock" stops early. A patient might move away, or a study might end before the event occurs. This is censoring.
For decades, the standard toolkit of survival analysis has relied on a convenient but fragile assumption: independent censoring. This assumes that the reason a subject leaves the study has nothing to do with their underlying risk of the event. If a patient drops out because they moved to a different city, that is independent. But if they drop out because they are too sick to continue the trial, the censoring is informative.
When censoring is informative or dependent, traditional models like Kaplan-Meier or the Cox Proportional Hazards model break down. They produce biased estimates, typically overestimating survival times because the "high-risk" individuals are the ones disappearing from the data.
The Identifiability Trap
The fundamental challenge of dependent censoring is that it is a counterfactual problem. As Ali Hossein Gharari Foomani and Michael Cooper point out in Copula-Based Deep Survival Models for Dependent Censoring, we only ever observe the minimum of the event time ($T$) and the censoring time ($C$). We never see both.
Because we cannot observe the joint distribution of $T$ and $C$ directly, the model is not "identifiable" without making specific assumptions about how they are linked. If we ignore this link, our metrics lie to us. Christian Marius Lillelund and colleagues demonstrate in Overcoming Dependent Censoring in the Evaluation of Survival Models that when censoring rates exceed 20%, standard metrics like the C-index can show improved scores even as the model's actual bias increases.
Linking Distributions with Copulas
To solve this, researchers are turning to copulas—mathematical functions that join marginal distribution functions to form a multivariate distribution. By using a copula, we can explicitly model the dependency between the time-to-event and the time-to-censoring.
Weijia Zhang and Chun Kai Ling introduced DCSURVIVAL, a framework that uses neural networks to parameterize Archimedean copulas. Instead of researchers guessing the dependency structure, the model learns it from the data. This approach provides "identifiability guarantees," meaning that under certain conditions, the model can actually recover the true underlying survival curve even when dropouts are correlated with the outcome.
Disentangling Latent Factors
Another way to think about this dependency is through shared underlying causes. A patient’s "overall health" might influence both their risk of a relapse (the event) and their likelihood of dying from an unrelated cause (the censoring).
In Explicitly Modeling Censoring Produces Superior Survival Predictors, Shi-ang Qi and Russell Greiner propose a model called SALaD. It works by disentangling features into three latent spaces:
- Event-specific factors (only affect the event).
- Censoring-specific factors (only affect dropout).
- Shared factors (affect both).
By explicitly modeling these shared factors, SALaD accounts for the "path of bias" that usually degrades the accuracy of survival predictors.
Proxies and Double Robustness
Sometimes, we know the dependency exists but we haven't measured the variables causing it. Andrew Ying addresses this in Proximal Survival Analysis to Handle Dependent Right Censoring. This framework uses "proxy" variables—imperfect indicators of unmeasured health status—to adjust for informative censoring.
This line of research often aims for double robustness. For example, in Doubly Robust Conformalized Survival Analysis with Right-Censored Data, Matteo Sesia and Vladimir Svetnik developed a method where the final prediction remains valid as long as either the survival model or the censoring model is correctly specified. This provides a safety net for researchers working with complex, real-world clinical data.
Toward Reliable Evaluation
If standard metrics like the C-index are unreliable under dependent censoring, how should we evaluate these models?
Shi-ang Qi and others have proposed MAE-PO, a metric based on pseudo-observations. This technique essentially "fills in" the missing data for censored individuals using the surrounding data distribution, allowing for a more accurate calculation of Mean Absolute Error. Similarly, David Rindt and Dino Sejdinovic introduced optHSIC, which uses optimal transport to create a synthetic, uncensored version of the dataset for more rigorous independence testing.
By moving away from the "nuisance" view of censoring and treating it as a process to be modeled, this research direction is making survival analysis more robust for high-stakes applications in medicine and engineering.