acceptodds
Under review as a conference paper at ICLR 2027

Distributionally Robust Survival Models under Subpopulation Shift and Outlier Contamination

Abstract

Learning robust survival models under distribution shift is an important but challenging problem in many applications. In heterogeneous populations, a model that performs well on average may still perform poorly on certain subpopulations, and this issue becomes even more severe when the training data are contaminated by outliers. In this paper, we propose a novel distributionally robust framework for survival analysis that jointly addresses latent subpopulation shift and outlier contamination. The proposed method combines an outer minimization that selects a refined nominal distribution by reducing the influence of contaminated samples and an inner maximization that focuses on the most challenging subpopulation. This formulation directly accommodates non-decomposable survival losses while preserving interactions across samples, including the risk-set structure of the Cox negative partial log-likelihood. We develop an alternating gradient-based algorithm with outer updates derived from the KKT conditions of the inner maximization. Experiments on simulated data and two survival benchmarks demonstrate that the proposed method remains robust when subpopulation shift and outlier contamination occur simultaneously. It stabilizes training in contaminated settings and substantially improves worst-group performance across both linear and nonlinear survival models, while maintaining competitive and sometimes superior overall performance.

Primary area:
optimization
Keywords:
Survival analysis, Distributionally robust optimization, Subpopulation shift, Outlier contamination

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bVSXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a coupled distributionally robust optimization (DRO) framework for survival analysis that jointly addresses latent subpopulation shift and outlier contamination. The method combines outlier downweighting with adversarial reweighting and extends this formulation to non-decomposable Cox partial likelihood through a KKT-based alternating optimization algorithm. Experiments on synthetic data and two survival datasets (METABRIC and FLC) report improvements in worst-group prediction performance.

However, the paper has two major weaknesses: (1) limited methodological novelty and a gap between theoretical guarantees and practical optimization, particularly given the relationship to existing CVaR-DORO methods and the absence of a natural Exact-Cox + CVaR-DORO baseline; and (2) insufficient and potentially confounded experimental evidence for robustness under distribution shift, including the absence of genuine subgroup proportion shifts in real-data experiments, unequal training protocols, and inconsistent improvements in predictive performance.

Critical strengths

The paper addresses a practically relevant challenge in robust survival analysis: standard DRO methods may overemphasize contaminated observations when optimizing worst-group performance. The formulation jointly considers subpopulation robustness and outlier contamination, while explicitly accounting for the non-decomposable structure of survival losses.

Critical weaknesses

1. Limited methodological novelty and a gap between theoretical guarantees and practical optimization.

The central methodological contribution appears closely related to existing distributionally robust optimization approaches. For additive losses, the proposed formulation reduces exactly to CVaR-DORO, as acknowledged by the authors. The main claimed novelty therefore lies in extending this formulation to non-decomposable survival losses.

However, the paper's own Exact-Cox baseline, based on Hu and Chen (2024), reformulates Cox regression using a piecewise-constant baseline hazard, yielding a decomposable full likelihood. This makes it possible to apply existing CVaR-DORO methods directly to the per-observation losses of Exact-Cox. Surprisingly, this natural combination is not evaluated.

Although the resulting full-likelihood formulation is not equivalent to the proposed weighted partial-likelihood objective, the missing comparison is important. Without it, the paper does not convincingly establish whether directly optimizing non-decomposable Cox partial likelihood provides a substantive advantage over existing robust optimization methods applied to an alternative survival likelihood.

Furthermore, the theoretical analysis relies on assumptions that are not satisfied by the practical optimization procedure as implemented. Proposition 1 and the local convergence analysis require an exact global solution to the inner maximization problem under additional regularity conditions. However, the weighted Cox objective is generally non-concave in the sample weights, and the experiments use only one approximate inner gradient update per outer iteration (T=1).

Consequently, the theoretical guarantees do not directly establish the convergence properties of the implemented algorithm, and the accuracy of the approximate inner optimization is not adequately examined.

Overall, the paper does not sufficiently demonstrate methodological advantages over existing DRO formulations, while the theoretical guarantees remain disconnected from the practical optimization procedure.

2. Insufficient and potentially confounded experimental evidence for robustness under distribution shift.

The experimental evaluation does not adequately establish the claimed robustness to simultaneous subpopulation shift and outlier contamination.

First, the real-data experiments on METABRIC and FLC use subgroup-stratified train–test splitting, which approximately preserves subgroup proportions. Therefore, these experiments primarily evaluate worst-group prediction performance rather than robustness under an actual change in subgroup proportions between training and test distributions. Genuine subpopulation shift is explicitly constructed only in the synthetic experiments.

Second, outliers in the real-data experiments are artificially generated by selecting the highest-risk training samples and inverting their fitted Cox risks to generate inconsistent survival outcomes. This construction may favor methods that reduce the influence of high-loss observations. Although an additional contamination setting is considered in the appendix, the robustness of the proposed method to naturally occurring or qualitatively different contamination mechanisms remains insufficiently established.

Third, the training and hyperparameter-tuning protocols are not fully comparable across methods. In the linear setting, f-DRO and Exact-Cox use 100 optimization steps with a learning rate of 0.05 and no regularization, whereas the proposed method uses 200 steps with a learning rate of 0.1 and an elastic-net penalty. Learning rates and regularization settings also differ in the nonlinear experiments. These differences make it difficult to determine whether the reported improvements arise from the proposed objective or more favorable optimization settings.

Finally, the improvements are not consistently substantial and may come at the expense of overall predictive accuracy. For example, on FLC with linear models (Table 2, 10% validation setting), the worst-group C-index improves only from 0.626 (f-DRO) to 0.631, while IBS deteriorates from 0.107 to 0.119 and test log-likelihood decreases from −6.708 to −6.941. The paper reports means and standard deviations over 10 runs but does not provide paired significance analyses to establish the reliability of these differences.

Overall, the experiments do not convincingly establish that the proposed method provides robust and practically meaningful improvements under genuine distribution shift, particularly when accounting for artificial contamination, unequal optimization protocols, and trade-offs in predictive accuracy.

Questions for authors

Q1. Novelty relative to existing robust survival methods.

Since the proposed method reduces to CVaR-DORO for additive losses, and the Exact-Cox formulation provides a decomposable survival likelihood, why is CVaR-DORO applied to the per-observation Exact-Cox losses not included as a baseline?

Can the authors demonstrate that directly optimizing the non-decomposable Cox partial likelihood provides substantial advantages over this alternative under comparable experimental conditions?

Q2. Robustness under genuine subpopulation shift.

The real-data experiments use subgroup-stratified train–test splitting, which does not explicitly introduce changes in subgroup proportions.

How do these experiments demonstrate robustness under subpopulation shift rather than simply improved worst-group prediction within a fixed population?

Can the authors evaluate controlled subgroup proportion shifts on METABRIC and FLC while maintaining a clean, fixed test distribution?

Q3. Fairness of experimental comparisons and statistical reliability.

The proposed method and several baselines use different learning rates, optimization budgets, and regularization settings.

Can the authors provide comparisons under matched computational budgets and appropriately tuned optimization and regularization hyperparameters?

Furthermore, given the relatively small improvements in worst-group C-index on FLC and the deterioration in IBS and test log-likelihood, can the authors provide paired statistical analyses and clarify the practical significance of the reported robustness–accuracy trade-off?

Q4. Applicability of the theoretical guarantees to the implemented algorithm.

The theoretical analysis assumes exact global inner maximization, whereas the implemented algorithm uses only one approximate inner gradient step per outer iteration.

How accurately does the T=1T=1 procedure approximate the inner maximizer, and how sensitive are the experimental results to increasing TT?

Can the authors clarify whether the KKT-based supergradient estimates remain reliable under such approximate inner optimization and whether any convergence guarantees apply to the actual implementation?

Reviewer_zJVROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a distributionally robust optimization (DRO) framework for survival analysis that jointly addresses latent subpopulation shift and outlier contamination. The method formulates a coupled minimax problem in which an outer minimization refines the nominal distribution by downweighting potentially contaminated observations, while an inner maximization emphasizes the most challenging subpopulation. The main technical contribution is the development of an alternating gradient-based algorithm for non-decomposable survival losses, particularly the Cox negative partial log-likelihood, using KKT multipliers, whereas conventional DRO approaches assume additive losses. Synthetic and real-data experiments demonstrated that the proposed method was more stable than conventional DRO approaches under contamination and generally improved worst-group performance.

Critical strengths
  • The simultaneous presence of latent subpopulation shift and outlier contamination is a practically important setting in which conventional DRO approaches can fail in complementary ways. The proposed framework explicitly targets both sources of distributional mismatch without requiring subgroup labels during training.
  • The KKT-based approach to non-decomposable survival losses is a meaningful and technically solid contribution. The proposed approach can directly handle the Cox negative partial log-likelihood, which is one of the most important survival problems, while preserving its risk-set interactions.
  • The paper is well-written and easy to follow. The effectiveness of the proposed method was demonstrated on synthetic and real-world data.
Critical weaknesses
  • The technical novelty of the proposed method is solid but appears somewhat limited. The idea of combining an outer distribution-refinement step for robustness to outliers with an inner worst-case reweighting step is closely related to DORO (distributional and outlier robust optimization) and, more broadly, to nested DRO formulations. The paper itself notes that DORO already addresses subpopulation shift and outlier contamination for additive losses, and that recent work on nested DRO adopts a related optimistic/pessimistic structure. Thus, the main novelty lies more specifically in extending this framework to non-decomposable survival losses and deriving the corresponding optimization procedure.
  • There is a nontrivial gap between the theoretical guarantees and the implemented algorithm. The supergradient characterization and local convergence analysis assume access to an exact, unique, and global maximizer of the inner problem, together with additional smoothness and local regularity conditions. In practice, however, the inner problem is solved only approximately using a finite number of gradient-ascent steps, and the KKT multipliers used in the outer update are estimated from this approximate iterate. The paper would benefit from providing an error bound or a convergence guarantee that connects this practical approximation to the exact theoretical analysis.
  • Although two real datasets are used, the outliers adopted in the main robustness experiments are artificially introduced by modifying the survival times of high-risk training samples through a risk-inversion procedure. Consequently, these experiments demonstrate robustness to a controlled contamination mechanism applied to real data, rather than to naturally occurring contamination. Evaluations under more diverse contamination mechanisms or on datasets containing genuine anomalous observations would substantially strengthen the empirical claims.
  • In the experimental result tables, the best and second-best scores are highlighted, but the more important question of whether these differences are statistically significant is not explicitly addressed. Reporting the results of appropriate hypothesis tests would further strengthen the validity of the empirical findings.
Questions for authors

See Critical Weaknesses.

Reviewer_hFifOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a distributionally robust framework for survival analysis under subpopulation shift and outlier contamination. The method combines an outer minimization over a refined nominal distribution with an inner maximization over adversarial sample weights. For non-decomposable survival losses, the authors introduce a weighted Cox partial likelihood and develop an alternating optimization algorithm, using KKT multipliers to update the nominal weights. Experiments on simulated and real survival datasets report improvements in worst-group C-index and training stability under contamination.

Critical strengths

The paper addresses a relevant problem, as standard DRO methods can overemphasize contaminated observations when targeting worst-group performance. The proposed formulation is reasonably motivated, and the optimization preserves cross-sample interactions in the Cox partial likelihood rather than relying on sample-wise decomposition. The simulations also show meaningful improvements in worst-group discrimination under contamination.

Critical weaknesses
  1. Concerns about the weighted Cox objective. The proposed loss appears to admit degenerate solutions under censoring. When all risk scores are zero, the weighted Cox loss is nonpositive for any probability vector q. If the adversarial weights are concentrated entirely on censored observations, the loss becomes zero regardless of the model parameters. Such a solution is feasible whenever the refined distribution assigns at least α probability mass to censored observations. Consequently, a constant-risk predictor can be a global minimizer of the stated minimax objective. This condition appears to hold for several experimental settings, particularly on FLC. This raises a fundamental concern about whether the proposed objective captures the intended robust survival learning problem.

  2. Limited survival-specific methodological contribution. The alternating optimization and KKT-based updates are largely applicable to general differentiable non-decomposable losses. The main survival-specific component is the construction of the weighted Cox loss, but its statistical interpretation is insufficiently developed. Unlike ordinary sample reweighting, the proposed weights also modify the risk-set denominators. Arbitrary observation-level reweighting can therefore change the risk-set composition and the relative contributions of censored and uncensored observations, rather than simply adjusting subpopulation proportions.

  3. Limited empirical support for the claimed robustness. The real-data experiments use subgroup-stratified train-test splits and therefore do not directly evaluate changes in subpopulation proportions. Moreover, improvements in worst-group C-index are sometimes small, while IBS and test log-likelihood deteriorate substantially. For example, on FLC with linear models, the worst-group C-index improves only from 0.626 to 0.631 relative to f-DRO, whereas IBS is 0.119 compared with 0.100 for Elastic-Cox. The results do not clearly establish an improvement in survival prediction beyond risk discrimination.

Questions for authors
  1. Validity of the objective. Can the authors address the apparent degeneracy of the weighted Cox minimax objective when the adversarial distribution concentrates on censored observations? In particular, what prevents a constant-risk predictor from being globally optimal? Please clarify the normalization and constraints used in the implementation, and whether they differ from the stated formulation.

  2. Statistical interpretation and novelty. What is the population-level interpretation of the proposed weighted Cox objective under latent subpopulation shift, especially when the weights change the event/censoring composition and risk-set denominators? Can the authors examine the learned weights assigned to clean majority, clean minority, and contaminated observations in simulations? A comparison with a simpler trimming-based robust Cox approach would also help isolate the contribution of the coupled optimization.

  3. Experimental validation. Can the authors evaluate explicit changes in subgroup proportions on the real datasets, rather than only reporting worst-group performance under stratified splits? Given the deterioration in IBS, it would also be useful to report worst-group IBS and subgroup calibration, and to examine whether the trade-off depends on the validation criterion used for model selection.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.