- 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 procedure approximate the inner maximizer, and how sensitive are the experimental results to increasing ?
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?