acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Experimentation for Censored Survival Outcomes

Abstract

Adaptive experimentation enables efficient estimation of causal effects, but existing methods are not designed for survival data with censoring, where event times are only partially observed (e.g., overall survival in cancer trials but with dropout). In this paper, we develop a novel framework for adaptive experimentation to estimate causal effects under right censoring. For this, we derive the semiparametric efficiency bound for the average survival effect curve as a function of the treatment allocation policy and thereby obtain a closed-form efficiency-optimal allocation policy. The policy generalizes classical Neyman allocation to survival settings by prioritizing patient strata where both event and censoring dynamics induce high uncertainty. Building on this, we propose the *Adaptive Survival Estimator (ASE)*, an adaptive framework that learns the allocation policy and estimates the average survival effect curve sequentially. Our framework has three main benefits: (i) it accommodates arbitrary machine learning models for nuisance estimation; (ii) it is guided by a closed-form efficiency-optimal allocation policy; and (iii) it admits strong theoretical guarantees, including asymptotic normality via a martingale central limit theorem. We demonstrate our framework across various numerical experiments to show consistent efficiency gains over uniform randomization and censoring-agnostic baselines.

Primary area:
causal reasoning
Keywords:
Adaptive experimentation, Causal inference, Survival analysis.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6ySGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the design of efficient adaptive experiments with the existence of right censored outcomes, which is an important topic in causal inference and survival analysis. It first derives the semiparametric efficiency bound for the average survival effect curve, then obtain an efficiency-optimal allocation policy by minimizing the variance bound. Based on this, it proposes the adaptive survival estimator (ASE) that adaptively learns the optimal allocation policy from the streaming data, and estimates the average survival effect curve sequentially. Theoretical analysis shows its consistency and asymptotic normality, using martingale CLT to handle the adaptively collected data. Finally, they demonstrate the empirical advantages of their methods over baselines, using several synthetic and semi-synthetic experiments.

Critical strengths
  1. This paper proposes a principled framework for optimizing treatment allocation with right censored outcomes. The motivation and flow of the methodology are very easy to follow.

  2. This paper provides a consistent estimator (ASE) for the average survival effect curve. The asymptotic normality of the estimator is proved using martingale CLT to handle the adaptive data structure.

  3. The assumptions of the theorems are clearly states and well motivated. The authors also provide enough explanations for why these assumptions are needed and reasonable.

Critical weaknesses
  1. The novelty of this paper is limited, and the proposed method should be considered as a methodological extension rather than a substantially new contribution that reaches the ICLR publication standard. There is already substantial related work (Zhang et al., 2023; Wei et al., 2023; Cook et al,. 2024) that proposes adaptive treatment allocation frameworks to specific statistical objectives, and demonstrates the statistical guarantees for post-hoc estimation using martingale CLT. On the other hand, the treatment allocation objective, i.e., the semiparametric bound for right censored outcomes, has also been developed before (Westling et al., 2024). Therefore, this paper merely combines two established lines of research in a straightforward way, without introducing substantially new methodological or theoretical techniques.

  2. Some theoretical details appear to be inconsistent, which raises concern about the soundness of the paper. For example, the Optimal treatment allocation (line 241) assumes that the policy is clipped at fixed alpha and 1-alpha, yet equation 11 uses clipping bounds that approach 0 and 1 asymptotically. The effect of accumulated estimation bias is also not well justified. For example, in the proof of theorem 6.3, equation 69 does not generally holds without stronger assumptions.

  3. Some experiments claims appear to be inconsistent with figures, which also raises concern about the reliability of the experiments. For example, the authors claim that "Our ASE achieves a relative MSE of approximately 1.05×of the oracle at R= 2000, which nearly matches oracle performance despite using estimated nuisances," while the figure 3 (a) shows that the MSE of ASE (Ours) is much larger compared to the oracle. The experiments are also not sufficient. There are only two experiments, both use simulated censoring, and both set high censoring probability for the control group at first time point.

  4. The mathematical presentation is too dense for Section 5 and 6.

Reference

  1. Zhang, W., Zhang, Z., & Liu, A. (2023). Optimizing treatment allocation in randomized clinical trials by leveraging baseline covariates. Biometrics, 79(4), 2815–2829.
  2. Wei, W., Ma, X., & Wang, J. (2023). Fair adaptive experiments. Advances in Neural Information Processing Systems, 36, 19157–19169.
  3. Cook, T., Mishler, A., & Ramdas, A. (2024). Semiparametric efficient inference in adaptive experiments. Proceedings of the Third Conference on Causal Learning and Reasoning, 236, 1033–1064.
  4. Westling, T., Luedtke, A., Gilbert, P. B., & Carone, M. (2024). Inference for treatment-specific survival curves using machine learning. Journal of the American Statistical Association, 119(546), 1541–1553.
Questions for authors
  1. Please refer to my comments in Critical Weaknesses. The clipping bound of the optimal treatment allocation appears to be inconsistent. Which one should we use in practice and in theory? Should we assume that the policy is clipped at fixed α\alpha and 1−α1-\alpha, or at αr\alpha_r and 1−αr1-\alpha_r, where αr→0\alpha_r \to 0 as the sample size increases?

  2. Please refer to my comments in Critical Weaknesses, I found a potential inconsistency in the presentation of the experimental results. I encourage the authors to carefully revisit the experimental conclusions.

  3. I encourage the authors to do some real-world data analysis with real survival data, instead of using simulated censoring data.

4 comments
AuthorsOpenReview ↗

Response to reviewer 6ySG

Thank you for your valuable evaluation of our paper! We took all your comments to heart and improved our paper accordingly. All changes are highlighted in blue in the revised PDF.

> W1: Theoretical novelty beyond existing adaptive experimentation and survival inference

Thank you for this thoughtful assessment and for highlighting the connections to existing work. Our framework builds on adaptive designs for fully observed outcomes [1, 2, 3] and on semiparametric survival inference under a fixed assignment [4], from which we inherit the Neyman-type allocation structure and the martingale CLT. However, two key steps do not carry over directly under right censoring, which is why we derive a new method for our task.

Efficiency bound as a function of the allocation policy. The survival literature [4] characterizes efficiency under a fixed assignment, whereas experimental design requires the bound as a function of π\pi. Because of this, the results from the literature cannot be transferred to our case. As a remedy, we derive Theorem 4.1, which provides this in closed form, thereby recovering the Neyman allocation exactly without censoring (Section 4). With censoring, the optimal allocation shifts even when both arms share the same censoring hazard (Figure 2). Moreover, the variance target V^a\hat V_a is an explicit function of the hazard nuisances already estimated for the EIF, so no additional variance regression is needed (Remark 5.1).

Inference with an estimated censoring mechanism. In [1, 2, 3], the propensity is chosen by the experimenter, so the AIPW score is conditionally unbiased given the history. Under right censoring, the censoring mechanism is unknown even in a randomized trial and must be estimated. Hence, the estimated score is no longer conditionally unbiased. To deal with this, our analysis therefore shows that the first-order nuisance errors cancel after conditional expectation, yielding the bias bound ∣E[mt,r∣Hr−1]∣≲(erS)2+erSerG|\mathbb E[m_{t,r}\mid\mathcal H_{r-1}]|\lesssim(e^S_r)^2+e^S_re^G_r (Eq. (68)), and controls the remainder fluctuation under the shrinking truncation (Eq. (81)). This also leads to a robustness structure specific to censoring: ASE remains consistent under censoring-model misspecification (Corollary 6.6) and is doubly robust in the no-ties setting (Appendix G).

We also kindly refer to our response to Reviewer grf1 for a summary of which components are inherited and which are new.

Action: We have added [1] and [2] to the related work, and added more elaborate comparison to Appendix B.

References

[1] Zhang, W., Zhang, Z., & Liu, A. (2023). Optimizing treatment allocation in randomized clinical trials by leveraging baseline covariates. Biometrics, 79(4), 2815–2829.

[2] Wei, W., Ma, X., & Wang, J. (2023). Fair adaptive experiments. In Advances in Neural Information Processing Systems, 36, 19157–19169.

[3] Cook, T., Mishler, A., & Ramdas, A. (2024). Semiparametric efficient inference in adaptive experiments. In Proceedings of the Third Conference on Causal Learning and Reasoning, PMLR 236, 1033–1064.

[4] Westling, T., Luedtke, A., Gilbert, P. B., & Carone, M. (2024). Inference for treatment-specific survival curves using machine learning. Journal of the American Statistical Association, 119(546), 1541–1553.

> W2 + Q1: Clipping of the allocation policy and accumulated estimation bias

Thank you for these careful observations and for the opportunity to clarify the clipping rules and the control of the accumulated bias.

1. Clipping of the allocation policy. Both clipping rules are special cases of the truncation in Eq. (11). The fixed interval [α,1−α][\alpha,1-\alpha] corresponds to the constant schedule kr≡1/αk_r\equiv 1/\alpha, while the shrinking interval corresponds to kr→∞k_r\to\infty.

In theory, the target of the adaptive policy is the unclipped optimal ratio V1(x)/(V1(x)+V0(x))\sqrt{V_1(x)}/(\sqrt{V_1(x)}+\sqrt{V_0(x)}). In Proposition 6.4, we now assume that this ratio lies in [α,1−α][\alpha,1-\alpha] for some α>0\alpha>0, together with the variance non-degeneracy condition V0(x)+V1(x)≥v‾>0V_0(x)+V_1(x)\geq\underline v>0 (see also our response to Reviewer W5Qs, Q4). Under this condition, the clipping in Proposition 4.3 is inactive, so π⋆\pi^\star coincides with the unclipped ratio, and both truncation rules converge to the same π⋆\pi^\star as soon as 1/kr≤α1/k_r\leq\alpha.

In practice, a fixed α\alpha can be used whenever a lower bound on the optimal allocation is known, for example from design or ethical constraints. This is the choice in our experiments (α=0.05\alpha=0.05, Appendix K.2), where the oracle optimal ratio lies strictly inside [0.05,0.95][0.05,0.95] in both settings. When no such bound is available, the shrinking schedule ensures that the clipping interval eventually contains π⋆\pi^\star.

To cover both cases, we have relaxed Theorem 6.3 to any non-decreasing schedule with kr=o(r1/4)k_r=o(r^{1/4}), which includes constant krk_r. The proof is unchanged, since it only uses kr=o(r1/4)k_r=o(r^{1/4}).

2. Accumulated estimation bias. Thank you also for the careful reading of Eq. (69). To make the cumulative step fully rigorous, we have stated the rate conditions of Theorem 6.3 in expectation:

E[max⁡i∥λ^i,r−1S(a,⋅)−λiS(a,⋅)∥22]=o(r−1/2),E[kr∥πr−π∥2]=o(1),\mathbb E\Big[\max_i\|\hat\lambda^S_{i,r-1}(a,\cdot)-\lambda^S_i(a,\cdot)\|_2^2\Big]=o(r^{-1/2}),\qquad \mathbb E\big[k_r\|\pi_r-\pi\|_2\big]=o(1),

and analogously for λ^G\hat\lambda^G. By Eq. (68) and the Cauchy–Schwarz inequality, this yields E∣Δt,rA∣=o(r−1/2)\mathbb E|\Delta^A_{t,r}|=o(r^{-1/2}), and hence

E[∣ΔtA∣]≤1R∑r=1RE∣Δt,rA∣=o(1),\mathbb E\big[|\Delta^A_t|\big]\leq\frac{1}{\sqrt R}\sum_{r=1}^R\mathbb E|\Delta^A_{t,r}|=o(1),

since ∑r=1Rr−1/2=O(R)\sum_{r=1}^R r^{-1/2}=O(\sqrt R). By Markov’s inequality, the accumulated bias is therefore op(1)o_p(1). We have applied the same in-expectation conditions throughout the proof of Theorem 6.3, including the conditional variance step and the martingale fluctuation term.

Action: We have (i) clarified the relation between the two clipping rules after Theorem 6.3; (ii) relaxed the truncation schedule in Theorem 6.3 to include constant krk_r; (iii) stated the policy condition of Proposition 6.4 on the unclipped optimal ratio; and (iv) stated the rate conditions of Theorem 6.3 in expectation and updated Eq. (69) and the corresponding steps in Appendix F.4.

> W3 + Q2: Consistency between the text and Figure 3

Thank you for carefully checking our experimental results. The numbers stated in the text of Section 7 referred to an earlier version of our synthetic experiment, and we have updated them now. Figure 3 reports the final results, and we have corrected the text accordingly: at R=2000R=2000, ASE achieves a relative MSE of approximately 1.63×1.63\times the oracle, compared with 1.83×1.83\times for ASE-NA and 1.78×1.78\times for A2IPW-Naïve. Importantly, our conclusions are unchanged: ASE achieves the lowest error among all feasible estimators, and the censoring-aware allocation outperforms both uniform randomization and the censoring-agnostic A2IPW-Naïve.

Action: We have corrected the numerical statements in Section 7 to match Figure 3.

> W4: Clarity of Mathematical Presentation

Thank you for this helpful suggestion. We have expanded the math in Sections 5 and 6 to improve readability and provide clearer intuition behind the key mathematical formulations and theoretical results, while keeping the detailed technical derivations in the Appendix.

Action: We have revised Sections 5 and 6 to improve the clarity of the mathematical presentation.

> Q3: Real-world survival data

Thank you for this suggestion. To strengthen the empirical evaluation, we have added new experiments based on the ACTG 175 randomized clinical trial from SurvHTE-Bench [1], which combines the 2323 real covariates of 2,1392{,}139 trial participants with semi-synthetic survival outcomes generated by the benchmark. The table reports results over 500500 replications at n=3,000n=3{,}000, with MSE and coverage averaged over the benchmark’s evaluation horizons.

MethodMSE ×10−4\times10^{-4}Coverage
ASE2.202.2095.6%95.6\%
ASE-NA2.242.2494.9%94.9\%
ASE-MS2.202.2095.5%95.5\%
A2IPW-Naïve2.222.2295.6%95.6\%
Plug-in2.672.6727.7%27.7\%
Plug-in-NA2.402.4028.2%28.2\%

We highlight two findings: Inference remains valid under adaptive allocation with real covariates. ASE attains 95.6%95.6\% coverage, in line with Theorem 6.3. In contrast, the plug-in estimator severely undercovers, and ASE has a 17.6%17.6\% lower MSE under the same design. Note that the A-optimal allocation targets the variance of the efficient estimator and is therefore not designed to benefit the plug-in estimator, which relies solely on the fitted outcome model. The gain from adaptive allocation matches our theory quantitatively. In this setting, the optimal allocation is close to 1:1 (mean 0.520.52). Accordingly, the realized MSE ratio of ASE-NA to ASE is 1.021.02 [0.99,1.05][0.99, 1.05]. Settings with efficiency gains on the benchmark under differential dropout are discussed in our response to Reviewer W5Qs (W2 + Q1).

Action: We have added the new experiments based on ACTG 175 to Section 7, with further details in Appendix J.3.

References

[1] Noroozizadeh, S., Shen, X., Weiss, J., & Chen, G. H. (2026). SurvHTE-Bench: A benchmark for heterogeneous treatment effect estimation in survival analysis. In International Conference on Learning Representations.

Dear Reviewer,

The discussion period is nearing its end, and we haven’t yet seen a reply from you. We encourage you to respond to the authors’ comments, especially if you have further questions about points that may affect your assessment.

Thank you for taking the time to engage in the discussion.

Best,

Your AC

Reviewer_grf1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies efficient adaptive treatment allocation for right-censored survival outcomes. The proposed strategy adjusts treatment probabilities according to uncertainty arising from both survival and censoring processes, allocating more patients to the treatment arm with greater uncertainty conditional on patient characteristics. The authors derive a semiparametric efficiency bound for the average survival effect curve and an allocation rule that minimizes the sum of asymptotic variances across time horizons. They introduce the Adaptive Survival Estimator (ASE) and establish asymptotic normality and efficiency under stated conditions, as well as consistency under censoring-model misspecification when the event-hazard model is consistently estimated. Synthetic and semi-synthetic experiments provide evidence of efficiency gains over the evaluated baselines.

Critical strengths
  1. ASE extends adaptive experimental design to right-censored survival outcomes. The paper derives a censoring-aware allocation rule that minimizes the sum of asymptotic variances across survival horizons and provides guarantees of asymptotic normality and efficiency under stated assumptions. It also establishes consistency under censoring-model misspecification when the event-hazard model is consistently estimated.
  2. The paper is generally well organized and easy to follow. The progression from the efficiency bound to the allocation rule, estimator, and theoretical guarantees provides a clear logical structure. The synthetic and semi-synthetic experiments provide empirical support for ASE’s benefits, showing gains in estimation accuracy over uniform randomization and censoring-agnostic allocation in the evaluated settings.
Critical weaknesses

The incremental theoretical contribution relative to existing work could be clarified. Kato et al. (2020), “Efficient Adaptive Experimental Design for Average Treatment Effect Estimation”, already study adaptive treatment assignment based on conditional standard deviations to minimize the semiparametric efficiency bound, while Cook et al. (2024) establish asymptotic normality and efficiency for adaptive AIPW estimation. The proposed allocation rule and asymptotic analysis follow a closely related structure. In particular, a more explicit comparison of Theorem 4.1, Proposition 4.3, and Theorem 6.3 with these prior results would help assess the theoretical novelty.

Questions for authors

Could you clarify the main theoretical contributions beyond existing work on efficient adaptive treatment allocation, particularly Kato et al. (2020) and Cook et al. (2024)? Which technical challenges introduced by right censoring require new arguments, and which results follow from adapting existing methods?

4 comments
AuthorsOpenReview ↗

Response to reviewer grf1

Thank you for your positive evaluation of our paper! We took all your comments to heart and improved our paper accordingly. All changes are highlighted in blue in the revised PDF.

> W + Q: Theoretical contributions beyond existing adaptive experimentation methods

Thank you for this thoughtful question and for the opportunity to clarify how our results relate to and differ from Kato et al. [1] and Cook et al. [2]. Our framework builds on the variance-optimal allocation principle of [1], the martingale-based inference of [2], and the survival EIF [4, 5]. Still, the existing works are not directly applicable to our setting. The reason is that right censoring introduces new challenges at each stage, which we summarize below and which is why we need to derive a new framework.

1. Efficiency bound and optimal allocation (Theorem 4.1 and Proposition 4.3). Building on the survival EIF [4, 5], Theorem 4.1 characterizes the semiparametric efficiency bound of the average survival effect curve in closed form as a function of the allocation policy π\pi. This is the object needed for experimental design, and it is not studied in the survival literature, which considers a fixed treatment assignment. The resulting arm-specific variance Va(x)=∑t=0tmax⁡St(x,a)2∑i=0tλiS(x,a)Si(x,a) Gi−1(x,a)V_a(x)=\sum_{t=0}^{t_{\max}}S_t(x,a)^2\sum_{i=0}^{t}\frac{\lambda^S_i(x,a)}{S_i(x,a)\,G_{i-1}(x,a)} depends jointly on the event and censoring dynamics. Once VaV_a is available, the Neyman-type form of Proposition 4.3 follows from the same calculation as in [1]. Two properties are specific to censoring:

Strict generalization. Without censoring (Gt−1≡1G_{t-1}\equiv 1), Va(x)V_a(x) reduces to ∑tVar(1{T>t}∣X=x,A=a)\sum_t\mathrm{Var}(\mathbf 1\{T>t\}\mid X=x,A=a), and our policy recovers the Neyman allocation of [1] exactly (Section 4, “Neyman allocation as a special case”). With censoring, the allocation shifts even when both arms share the same censoring hazard (Figure 2), since Gt−1(x,a)G_{t-1}(x,a) inherits the arm-specific event dynamics.

Closed-form variance target. V^a\hat V_a is an explicit function of the hazard nuisances that are already estimated for the EIF (Eq. (10) and Remark 5.1). Hence, no additional regression for the conditional variance is required, and non-negativity holds by construction.

2. Asymptotic normality (Theorem 6.3). In [1, 2], the propensity πr\pi_r is chosen by the experimenter, so the AIPW score with outcome models fitted on past data is conditionally unbiased given the history Hr−1\mathcal H_{r-1}, and estimation error in the outcome model only affects the variance. Under right censoring, this property no longer holds: the censoring mechanism is unknown to the experimenter even in a randomized design and must be estimated, so the estimated survival score is not conditionally unbiased. Our analysis therefore requires new arguments:

Second-order bias under adaptive data. The first-order nuisance errors cancel only after taking the conditional expectation given Hr−1\mathcal H_{r-1}, leaving a bias of order (erS)2+erSerG(e^S_r)^2+e^S_re^G_r (Eqs. (64) and (68)), where erSe^S_r and erGe^G_r denote the L2L_2 errors of the event and censoring hazards. This requires rate conditions on both hazards, which are absent in [1, 2].

Remainder fluctuation under shrinking truncation. The centered nuisance remainder is scaled by the inverse propensity weights, so its conditional second moment is bounded by kr2k_r^2 times the squared hazard errors (Eq. (81)). This leads to the truncation schedule kr=o(r1/4)k_r=o(r^{1/4}).

Robustness to censoring misspecification. Since the censoring hazard enters the bias only through interactions with the event-hazard error, ASE remains consistent when the censoring model is misspecified (Corollary 6.6), and is doubly robust in the no-ties setting (Appendix G). These results have no counterpart for fully observed outcomes.

Empirically, this distinction is reflected in the comparison with A2IPW-Naïve [1], which applies Neyman allocation while ignoring censoring and shows a persistent efficiency gap relative to ASE (Section 7).

Action: We have revised the manuscript to make these distinctions explicit: (i) we have added the reference [1] to the related work and attributed the A2IPW baseline to it; (ii) we have stated after the Neyman special case in Section 4 that the allocation of [1] is recovered without censoring; and (iii) we added more elaborate comparison table to Appendix B.

References

[1] Kato et al. Efficient adaptive experimental design for average treatment effect estimation. arXiv, 2020.

[2] Cook et al. Semiparametric efficient inference in adaptive experiments. CLeaR, 2024.

[3] Waudby-Smith et al. Time-uniform central limit theory and asymptotic confidence sequences. Ann. Statist., 2024.

[4] Cui et al. Estimating heterogeneous treatment effects with right-censored data via causal survival forests. JRSS-B, 2023.

[5] Westling et al. Inference for treatment-specific survival curves using machine learning. JASA, 2024.

Dear Reviewer,

The discussion period is nearing its end, and we haven’t yet seen a reply from you. We encourage you to respond to the authors’ comments, especially if you have further questions about points that may affect your assessment.

Thank you for taking the time to engage in the discussion.

Best,

Your AC

Reviewer_grf1OpenReview ↗

Thank you for the clarification

Thank you for the detailed response and for clarifying the contributions beyond prior work, particularly the challenges introduced by estimating the censoring mechanism. I appreciate the revisions and the clearer positioning of the theoretical results. I will maintain my original rating of weak acceptance.

Thank you for your reply! We are happy to hear that your concerns are addressed, and thank you for maintaining your positive score.

Reviewer_W5QsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the adaptive treatment allocation in experiments with right censored survival outcomes where the authors derive a semiparametric efficiency bound for estimating average survival effect curve and use that to obtain an optimal allocation policy accounting for both surivival and censoring uncertainty. The authors propose an adaptive survival estimator that updates the probabilities of treatment as we get more data and estimates treatment effects using sequential nuisance estimation. The paper spends a great deal on providing theoretical results on asymptotic normality and robustness to nuisance misspecification and show that their estimator generally achieves lower estimation error compared to uniform allocation and some baselines that are agnostic to censoring through the presented synthetic and semi-synthetic experiments.

Critical strengths
  1. The paper is very well motivated to show the usefulness of the setup with clinical trials and censored outcomes and showing that ignoring censoring can be statistically inefficient.
  2. The paper goes beyond introducing a heuristic and connects the allocation mechanism to an explicit efficiency criterion.
Critical weaknesses
  1. If I am not mistaken in my understanding, there is a gap in the asymptotic-normality proof in the paper where the proof bounds quantities involve estimated inverse survival and censoring probability and the assumption in 6.1 is providing uniform lower bounds on the true hazards and survival probabilities but not necessarily their estimated versions.
  2. The other major concern of mine is that the experimental evaluation is very limited. This is especially true about the survival and censoring mechanism that is considered. As a suggestion the authors should consider evaluating their ASE under a much broader range of settings that is available in existing benchmarks that include different survival distributions and censoring levels and more realistic covariate structured. For instance SurvHTE-Bench [1] from last year's ICLR provides this comprehensive testbed that ASE should be adapted for evaluating. This is important to show whether the reported efficiency improvements generalize beyond the specific simulations considered in the paper.

[1] Noroozizadeh S, Shen X, Weiss J, Chen GH. SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis. In International Conference on Learning Representations 2026 Apr 20 (Vol. 2026, pp. 6049-6120).

Questions for authors
  1. How would the method and gains prove to be in varying experimental settings (see Weakness 2)?
  2. For theorem 6.3 I do not understand how the claim that the nuissance remainder is asymptotically negligible comes from so please elaborate on that?
  3. Shouldn't the estimated hazards require a positivity constraint too for Theorem 6.3?
  4. What assumption(s) ensure that the variance denominator is uniformly bounded away from 0 and without this condition how do the authors get the policy convergence rate?
4 comments
AuthorsOpenReview ↗

Response to reviewer W5Qs

Thank you for your valuable evaluation of our paper! We took all your comments to heart and improved our paper accordingly. We have incorporated all points marked with Action into the revised version of our paper.

> W1 + Q3: Boundedness of Estimated Inverse Probability Weights

Thank you for this careful observation and for highlighting the distinction between the true survival and censoring probabilities and their estimated counterparts. We appreciate the opportunity to clarify this point.

To ensure that the estimated inverse weights remain bounded, we revised Theorem 6.3 by clarifing that the positivity condition on the estimated nuisance functions must hold, analogous to the one in our confidence sequence result (Theorem H.2, Eq. (104)). Specifically, the fitted hazards are projected such that 1−λ^i,r−1S(x,a)≥c‾tS/21-\hat\lambda^S_{i,r-1}(x,a)\geq \underline c^S_t/2 and 1−λ^i,r−1G(x,a)/(1−λ^i,r−1S(x,a))≥c‾tG/21-\hat\lambda^G_{i,r-1}(x,a)/(1-\hat\lambda^S_{i,r-1}(x,a))\geq \underline c^G_t/2 for all i≤ti\le t, which keeps S^i,r−1\hat S_{i,r-1} and G^i−1,r−1\hat G_{i-1,r-1} uniformly bounded away from zero, without changing the rate conditions of Theorem 6.3.

Importantly, this clipping is already implemented in our code (see ase/nuisance.py and ase/stabilization.py in our anonymous repository). Hence, the added condition formalizes what our implementation already does, and all reported experimental results remain unchanged.

Action: We spelled out the positivity condition for the estimated nuisance positivity more carefully Theorem 6.3, and we updated the proof in Appendix F.4 accordingly.

> W2 + Q1: Additional experiments on SurvHTE-Bench

Thank you for this constructive suggestion and for pointing us to SurvHTE-Bench. We have now evaluated ASE on all five synthetic scenarios of the benchmark, which cover Cox, AFT, and Poisson event-time models with different censoring rates, and on the ACTG 175 setting, which combines the real covariates of 2,1392{,}139 trial participants with the benchmark's semi-synthetic outcomes. The table reports Scenarios A, B, C, and E, while Scenario D (9191\\% censoring) violates our overlap condition (Assumption 6.1) at late horizons by design and is therefore reported separately as a stress test in Appendix J.2 (MSE ratio 0.950.95 [0.90,1.00][0.90, 1.00]). We compare ASE with ASE-NA (same estimator, uniform 1:1 randomization) over 500500 replications. The MSE ratio is MSE(ASE-NA)/MSE(ASE)\mathrm{MSE}(\text{ASE-NA})/\mathrm{MSE}(\text{ASE}), where values above one favor adaptive allocation, and π⋆\pi^\star is the optimal allocation of each setting.

Settingπ⋆\pi^\star, mean (SD)MSE ratio [9595\\% CI]Coverage ASE / plug-in
A (Cox)0.490.49 (0.050.05)0.990.99 [0.96,1.02][0.96, 1.02]0.950.95 / 0.140.14
B (AFT)0.540.54 (0.030.03)1.011.01 [0.98,1.04][0.98, 1.04]0.940.94 / 0.140.14
C (Poisson)0.500.50 (0.0020.002)0.990.99 [0.97,1.02][0.97, 1.02]0.950.95 / 0.120.12
E (Poisson)0.490.49 (0.010.01)1.031.03 [1.00,1.06][1.00, 1.06]0.950.95 / 0.110.11
ACTG 1750.520.52 (0.100.10)1.021.02 [0.99,1.05][0.99, 1.05]0.960.96 / 0.280.28
Ours, synthetic (Sec. 7)0.180.18 (0.050.05)1.121.12 [1.02,1.23][1.02, 1.23]0.940.94 / 0.070.07
Ours, Twins (App. J.1)0.210.21 (0.050.05)1.111.11 [1.02,1.20][1.02, 1.20]0.940.94 / 0.200.20

Valid inference in every benchmark setting. ASE attains coverage between 0.940.94 and 0.960.96 across all event-time models and censoring levels, whereas the plug-in estimator covers in only 0.110.11 to 0.280.28 of the replications.

Efficiency recovers the efficiency of uniform randomization in balanced settings. In the benchmark, the optimal allocation is close to 1:1. Here, ASE and ASE-NA are statistically indistinguishable: all confidence intervals include one and exclude an efficiency loss larger than about 4%4\%. Hence, adaptivity comes at no measurable cost when there is little to exploit.

Efficiency gains on the benchmark under differential dropout. To examine whether the gains of ASE generalize beyond our own simulations, we kept Scenarios A and C of the benchmark unchanged and only added early dropout in the control arm with probability qq, so that Gt(x,0)G_t(x,0) becomes (1−q) Gt(x,0)(1-q)\,G_t(x,0) for all t≥0t\geq 0, while Gt(x,1)G_t(x,1) and the event-time distribution remain unchanged. Hence, a larger qq censors more control units and creates a stronger imbalance between the arms.

Scenarioq=0q=0q=0.75q=0.75q=0.90q=0.90
A (Cox)0.990.99 [0.96,1.02][0.96, 1.02]1.191.19 [1.07,1.30][1.07, 1.30]1.551.55 [1.37,1.73][1.37, 1.73]
C (Poisson)0.990.99 [0.97,1.02][0.97, 1.02]1.111.11 [1.05,1.18][1.05, 1.18]1.391.39 [1.29,1.49][1.29, 1.49]

The MSE ratio increases with qq in both scenarios A and C, and all confidence intervals for q>0q>0 exclude 11. At q=0.90q=0.90, uniform randomization incurs a 5555\\% (Scenario A) and 3939\\% (Scenario C) higher MSE than ASE. This is consistent with our own settings, where early control-arm dropout leads to an 1111\\% to 1212\\% higher MSE under uniform randomization.

Taken together, ASE automatically adapts to the setting: it recovers the efficiency of uniform randomization when the arms are balanced, and achieves substantial efficiency gains when the arms differ in their censoring dynamics.

Action: We have added the SurvHTE-Bench, ACTG 175, and differential dropout experiments to Section 7 and Appendices J.2 and J.3. and clarified in the Section 7 that the efficiency gains of ASE arise when the arms differ in their event or censoring dynamics.

> Q2: Asymptotic negligibility of the nuisance remainder

Thank you for this helpful comment. We would like to clarify how the nuisance remainder in Theorem 6.3 becomes asymptotically negligible. In Appendix F.4.3, we decompose the normalized remainder as

1R∑r=1Rmt,r=1R∑r=1RE[mt,r∣Hr−1]⏟ΔtA: conditional bias+1R∑r=1R(mt,r−E[mt,r∣Hr−1])⏟ΔtB: martingale fluctuation\frac{1}{\sqrt{R}}\sum_{r=1}^{R} m_{t,r}=\underbrace{\frac{1}{\sqrt{R}}\sum_{r=1}^{R}\mathbb{E}\left[m_{t,r}\mid\mathcal{H}_{r-1}\right]}_{\Delta_t^{A}\text{: conditional bias}}+\underbrace{\frac{1}{\sqrt{R}}\sum_{r=1}^{R}\bigl(m_{t,r}-\mathbb{E}\left[m_{t,r}\mid\mathcal{H}_{r-1}\right]\bigr)}_{\Delta_t^{B}\text{: martingale fluctuation}}

and show that both terms are op(1)o_p(1).

Conditional bias. After taking the conditional expectation, the first-order nuisance errors cancel, and only second-order products remain (Eq. (64)):

E[mt,r∣Hr−1]=E[Rt,r,1(Xr)−Rt,r,0(Xr)∣Hr−1],\mathbb{E}[m_{t,r}\mid\mathcal{H}_{r-1}] = \mathbb{E}\big[\mathcal{R}_{t,r,1}(X_r)-\mathcal{R}_{t,r,0}(X_r)\mid\mathcal{H}_{r-1}\big],

where each Rt,r,a\mathcal R_{t,r,a} is a weighted sum of products {λ^i,r−1S−λiS}{Gi−1−G^i−1,r−1}\{\hat\lambda^S_{i,r-1}-\lambda^S_i\}\{G_{i-1}-\hat G_{i-1,r-1}\}. Hence, ∣E[mt,r∣Hr−1]∣≲(erS)2+erSerG|\mathbb E[m_{t,r}\mid\mathcal H_{r-1}]|\lesssim (e^S_r)^2+e^S_r e^G_r (Eq. (68)), where erSe^S_r and erGe^G_r denote the L2L_2 errors of the event and censoring hazards. Under the hazard rate conditions of Theorem 6.3, this bias is of order o(r−1/2)o(r^{-1/2}) in expectation, so that ΔtA=op(1)\Delta_t^A=o_p(1) (see also our response to Reviewer 6ySG, W2).

Martingale fluctuation. Since ΔtB\Delta_t^B is a sum of martingale differences, its variance equals the average of the conditional second moments, which we bound as (Eq. (81))

\mathbb{E} [m_{t,r}^2\mid \mathcal{H}_{r-1}] \le C_t\, k_r^2 \big\\{(e^S_r)^2+(e^G_r)^2\big\\}.

The factor kr2k_r^2 comes from the truncated inverse propensity weights. Since kr=o(r1/4)k_r=o(r^{1/4}) and (erS)2+(erG)2=o(r−1/2)(e^S_r)^2+(e^G_r)^2=o(r^{-1/2}), the right-hand side vanishes, and hence ΔtB=op(1)\Delta_t^B=o_p(1).

Both bounds use that the estimated survival and censoring functions are bounded away from zero, which is guaranteed by the nuisance clipping condition we add to Theorem 6.3 (see our response to W1 + Q3).

Action: We added a new Remark 6.3a to state explicitly that the nuisance remainder is asymptotically negligible. We also expanded Appendix F.4.3 with a short roadmap of the decomposition and state the two bounds above explicitly.

> Q4: Boundedness of the variance denominator and policy convergence

Thank you for this insightful question and for highlighting the role of the variance denominator in the policy convergence rate.

Assumption 6.1 ensures that the inverse survival and censoring weights are bounded, and hence that Va(x)V_a(x) is bounded from above. However, it does not by itself prevent Va(x)V_a(x) from approaching zero. We have therefore stated explicitly in Proposition 6.4 the non-degeneracy condition that our analysis relies on: there exists v‾>0\underline v>0 such that V0(x)+V1(x)≥v‾V_0(x)+V_1(x)\geq \underline v for all xx. This is a uniform version of the condition V0(x)+V1(x)>0V_0(x)+V_1(x)>0 already imposed in Proposition 4.3.

Lower bound on the arm-specific variances. Together with the unclipped optimal ratio V1(x)/(V1(x)+V0(x))\sqrt{V_1(x)}/(\sqrt{V_1(x)}+\sqrt{V_0(x)}) lying in [α,1−α][\alpha,1-\alpha], the condition gives V1(x)≥α(V1(x)+V0(x))≥αv‾\sqrt{V_1(x)}\geq\alpha\big(\sqrt{V_1(x)}+\sqrt{V_0(x)}\big)\geq\alpha\sqrt{\underline v}, and analogously for V0(x)V_0(x). Hence Va(x)≥α2v‾V_a(x)\geq\alpha^2\underline v for both arms.

Policy convergence rate. Only the true variances need to be bounded away from zero, and no additional stabilization of the estimated variances is required. A direct calculation gives

∣π~r(x)−π⋆(x)∣≤∑a∣V^a,r−1(x)−Va(x)∣V1(x)+V0(x),∣V^a,r−1(x)−Va(x)∣≤∣V^a,r−1(x)−Va(x)∣Va(x),|\widetilde\pi_r(x)-\pi^\star(x)|\leq\frac{\sum_a|\sqrt{\hat V_{a,r-1}(x)}-\sqrt{V_a(x)}|}{\sqrt{V_1(x)}+\sqrt{V_0(x)}},\qquad |\sqrt{\hat V_{a,r-1}(x)}-\sqrt{V_a(x)}|\leq\frac{|\hat V_{a,r-1}(x)-V_a(x)|}{\sqrt{V_a(x)}},

so that ∣π~r(x)−π⋆(x)∣≤(αv‾)−1max⁡a∣V^a,r−1(x)−Va(x)∣|\widetilde\pi_r(x)-\pi^\star(x)|\leq(\alpha\underline v)^{-1}\max_a|\hat V_{a,r-1}(x)-V_a(x)|. Therefore, the rate ∥V^a,r−1−Va∥2=op(r−1/4)\|\hat V_{a,r-1}-V_a\|_2=o_p(r^{-1/4}) carries over directly to ∥π~r−π⋆∥2=op(r−1/4)\|\widetilde\pi_r-\pi^\star\|_2=o_p(r^{-1/4}), and since the truncation is non-expansive once 1/kr≤α1/k_r\leq\alpha, we obtain E[kr∥πr−π⋆∥2]=o(1)\mathbb{E}\bigl[k_r\|\pi_r - \pi^\star\|_2\bigr] = o(1) for kr=o(r1/4)k_r=o(r^{1/4}).

Action: We have stated the variance non-degeneracy condition explicitly in Proposition 6.4 and revised the proof in Appendix F.5 to explicitly derive the denominator bound and the resulting policy convergence rate.

Dear Reviewer,

The discussion period is nearing its end, and we haven’t yet seen a reply from you. We encourage you to respond to the authors’ comments, especially if you have further questions about points that may affect your assessment.

Thank you for taking the time to engage in the discussion.

Best,

Your AC

Comments on the paper

1 comment

Dear reviewers and AC,

Thank you for handling our submission and for the constructive reviews. In the past week, we have carefully addressed every concern raised by the reviewers. We provided detailed responses and uploaded a revised manuscript with new experimental results. Below, we summarize the main additions to the revised paper (highlighted in blue\color{blue}{\text{blue}}):

New experiments on SurvHTE-Bench and ACTG 175. We evaluated ASE on all five survival scenarios of SurvHTE-Bench (Cox, AFT, and Poisson event-time models with different censoring rates) and on ACTG 175 with real covariates. ASE maintains close-to-nominal coverage and automatically adapts to the setting: it recovers the efficiency of uniform randomization when the arms are balanced, and achieves substantial efficiency gains when the arms are imbalanced due to differential dropout. (Reviewers W5Qs and 6ySG)

More explicit theoretical guarantees. We spelled out the positivity condition for the estimated nuisances and stated the rate conditions in expectation in Theorem 6.3, unified the two clipping rules under a single truncation schedule kr=o(r1/4)k_r = o(r^{1/4}), and stated the variance non-degeneracy condition in Proposition 6.4. All experimental results remain unchanged. (Reviewers W5Qs and 6ySG)

Clearer positioning of our theoretical contributions. We added a paragraph after Theorem 6.3 and an extended comparison in Appendix B that separate what is inherited from existing work from what is new for right-censored outcomes, namely the efficiency bound as a function of the allocation policy, the closed-form censoring-aware variance target, and the inference theory with an estimated censoring mechanism. (Reviewers grf1 and 6ySG)

Improved presentation. We corrected the numerical statements in Section 7 to match Figure 3, with all conclusions unchanged, and expanded the math in Sections 5 and 6. (Reviewer 6ySG)

Taken together, these revisions provide new experimental evidence, more explicit theoretical guarantees, and a clearer understanding of the capabilities and scope of ASE. We believe they significantly strengthen the paper and its contribution to adaptive experimentation and causal inference for survival outcomes.

We are confident that, with the help of the reviews and our improvements, our work will be an important contribution to the community. Finally, we wanted to say "thank you" - we truly appreciate the help, which allowed us to revise our work and make important improvements.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.