Pith. sign in

REVIEW 2 major objections 4 minor 128 references

Clipped regression outcomes need not break conformal prediction: a new nonconformity score yields tight marginal coverage of the latent outcome, and a two-threshold variant recovers conditional coverage on the fully observed cases.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 09:30 UTC pith:TKA43NXL

load-bearing objection Finite-sample results are solid and worth building on, but the main asymptotic theorem in the paper has a genuine proof gap and is not established as written. the 2 major comments →

arxiv 2607.20736 v1 pith:TKA43NXL submitted 2026-07-22 stat.ME

Conformal Prediction for Regression with Clipped Outcomes

Matteo Sesia , Vladimir Svetnik This is my paper
classification stat.ME MSC 62G1562G0862N01
keywords conformal predictionclipped outcomesdouble censoringpredictive inferenceregressionconditional coveragemarginal coverageuncertainty quantification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper studies conformal prediction for regression when outcomes are doubly censored, or clipped, at known fixed thresholds — the situation in drug-discovery assays where activity is recorded only as below the lower limit or above the upper limit. The authors show that the natural extension of standard conformalized quantile regression to this setting is unsatisfactory: it over-covers the true latent outcome on average while under-covering precisely the easier-to-predict cases whose outcomes are fully observed. They diagnose a genuine conflict between marginal and conditional coverage under clipping, and build a method for each end of the trade-off: ClipCQR, which attains tight finite-sample marginal coverage, and ClipCQR+, which additionally guarantees conditional coverage on the internal, non-extrapolating region of feature space. If correct, the paper establishes that clipped calibration data do not void conformal guarantees, and that the real design decision is which coverage target to calibrate for.

Core claim

The central claim is that the correct way to calibrate conformal prediction on clipped outcomes is to define the nonconformity score on the latent outcome so that it is constant on each censored region: the score of a clipped observation equals the score of the unobserved latent value that produced it (s(x, y) = s(x, Π(y))). Because of this clipping identity, calibration scores computed from clipped data are exchangeable with the test score even though the test target is the latent Y, and standard conformal arguments apply. The paper proves that ClipCQR then achieves marginal coverage within [1−α, 1−α+1/(n+1)] (Theorem 4), that ClipCQR+ additionally achieves P(Y ∈ Ĉ(X) | X ∈ int(f̂)) ≥ 1−α i

What carries the argument

The engine is the nonconformity score of Eq. (6), the smallest threshold τ at which the snapped interval family φ(x;τ) = ψ([f̂_lo(x)−τ, f̂_up(x)+τ]) first contains the outcome, where ψ is the snapping function that sends any endpoint past a censoring threshold all the way to the support boundary. For censored observations the score charges only the distance from the base interval to the threshold, floored at the negative half-width, rather than taking the clipped value at face value. Its decisive property is the clipping identity s(x,y) = s(x,Π(y)): the score reads the latent outcome through the same clip the data suffered, which restores exchangeability between calibration scores and the te

Load-bearing premise

The asymptotic headline rests on the base model estimating the latent quantiles consistently inside the observable range and placing no probability mass at the censoring thresholds — assumptions that rule out the common practice of training directly on the clipped outcomes, although the finite-sample guarantees in Theorems 4 and 5 do not depend on them.

What would settle it

Train a base quantile model on the clipped outcomes in the way many practitioners would — ignoring the censoring — and run ClipCQR+ at growing n on a synthetic problem with a known latent distribution. If worst-slab conditional coverage stays visibly below the nominal level even at large n, then Assumption 11 is violated in exactly the setting the paper warns about, and the oracle-convergence claim has no practical purchase there; conversely, if a censoring-aware base model converges to the snapped oracle's conditional coverage, the central mechanism is confirmed. The same experiment settles w

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • ClipCQR gives finite-sample marginal coverage of the latent outcome within [1−α, 1−α+1/(n+1)] with no assumptions on the base model beyond exchangeability — the same tight guarantee as standard conformal prediction in the uncensored case.
  • ClipCQR+ additionally guarantees conditional coverage P(Y ∈ Ĉ(X) | X ∈ int(f̂)) ≥ 1−α in finite samples, so a user who can tell when the base model is not extrapolating obtains a conditional guarantee that ordinary marginal calibration cannot provide.
  • When the base model consistently estimates the latent conditional quantiles inside the observable range and places no point mass at the censoring thresholds, ClipCQR+ converges to the snapped oracle and matches its conditional coverage on a set of features with probability tending to one.
  • The naive extension eCQR is provably conservative — its slack equals the probability that the raw interval lies entirely beyond the censoring threshold on the censored side — and its asymptotic limit lacks conditional coverage, confirming that the marginal-versus-conditional trade-off under clipping is structural, not an artifact of a particular method.
  • The finite-sample guarantees of Theorems 4 and 5 hold regardless of the base model, so even a poorly calibrated predictor can be wrapped in valid clipped-outcome intervals; only the oracle-matching asymptotics require base consistency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The clipping identity suggests a general design principle: any conformal score used with censored calibration data should be invariant to the unobservable part of the outcome. A testable generalization is that a score obeying s(x,y) = s(x,Π(y)) for any measurable clipping map Π yields the same finite-sample marginal guarantee, even when censoring thresholds vary by observation.
  • The paper's Assumption 11 excludes base models trained directly on the clipped outcomes, which is the default in much applied work. An immediate practical question is whether a censoring-aware training loss is required for ClipCQR+ to deliver its conditional-coverage gains, or whether the finite-sample guarantees alone justify switching from eCQR to ClipCQR or ClipCQR+; the paper's own experiments
  • The snapped oracle's behavior — collapsing extrapolated endpoints onto the thresholds — implies that under clipping the natural output is a three-state object: 'below the limit,' 'above the limit,' or a bounded interval. This suggests a decision-oriented framing in which the practitioner acts on the three states directly, without needing the interval to be informative inside the unobservable regio

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. This paper studies split conformal prediction for regression when the calibration outcomes are doubly censored (clipped) at known fixed thresholds. It introduces a new nonconformity score, Eq. (6), that is invariant to clipping, leading to ClipCQR with tight finite-sample marginal coverage (Theorem 4), and a two-group procedure ClipCQR+ that calibrates internal and extrapolation regions separately to improve conditional coverage (Theorem 5). The paper also analyzes asymptotic oracle convergence for ClipCQR+ (Theorem 13), related methods (eCQR, peCQR), and reports synthetic and QSAR experiments.

Significance. The score identity s(x,y)=s(x,Π(y)) is a clean and useful observation; it yields the first tight finite-sample marginal coverage guarantee in the CQR family under clipped outcomes, and the proofs of Theorems 4 and 5 are rigorous and convincing. The conditional-calibration idea is practically motivated and the experiments support the main qualitative claims. However, the central asymptotic result, Theorem 13, currently rests on a proof gap that is load-bearing: if repaired, the paper would be a solid contribution; as written, the asymptotic oracle-convergence claim is not established. The finite-sample results are a valuable contribution independently of the asymptotic part.

major comments (2)
  1. [Appendix A6.4, proof of Theorem 13, Step 1(a)] The proof defines τ̂(α; bI) as the threshold calibrated on the n_int := |{i : X_i ∈ U}| points in U, then invokes Theorem A22 for the conditional distribution given X∈U. But Algorithm A5 defines bI = {i : X_i ∈ int(f̂_n)} with int(f̂_n) = {x : C_L < f̂_lo,n(x) ≤ f̂_up,n(x) < C_R}, while U = {x : C_L < q_lo(x) ≤ q_up(x) < C_R}. These sets differ at every finite n. The proof later shows P(int(f̂_n) Δ U | f̂_n) → 0, but this set convergence is never used to bound the difference between the empirical score distribution on bI and the conditional distribution on U. Without such a transfer, the convergence τ̂(α; bI) → 0 is not established, and parts (i)–(iii) of Theorem 13 do not follow. This gap is likely fixable by adding a stability argument for conditional conformal quantiles under vanishing set mismatch, but as written the proof is incomplete.
  2. [Appendix A6.4, proof of Theorem 13, Step 1(a)] The sentence 'there is no censoring within U' is imprecise and potentially misleading. For x∈U, the outcome Y can still fall below C_L or above C_R with probability up to α/2. What the argument actually needs is the quantile-level condition π_L=π_R=0 for the conditional distribution on U, which does hold because q_lo(x) > C_L and q_up(x) < C_R. Additionally, the invocation of Lemma A32 to equate Assumption 10 (or A29) with L1-consistency on U requires stating that the conditional distribution of (X,Y) given X∈U satisfies the assumptions of that lemma. The proof should make these conditions explicit rather than relying on an imprecise 'no censoring' statement.
minor comments (4)
  1. [Algorithm A5] In the 'Evaluate interval' line, τ̂(α, X_{n+1}) should be τ̂+(α, X_{n+1}) to match Eq. (8).
  2. [Eq. (4)] The symbol ψ is used for both the vector-valued snapping map and its coordinate functions; this is a minor notational overload that can be clarified, e.g., by writing ψ(ℓ,u) = [ψ_L(ℓ), ψ_U(u)].
  3. [Section 4 and Appendix A7] A few software-name capitalizations are inconsistent: 'R package xgboost' and 'python package PyTorch' should be 'R package XGBoost' and 'Python package PyTorch'.
  4. [Appendix A6.6, proof of Theorem A26] In the sandwich display, bGm^{-1} is introduced but the symbol bGm is not defined; the intended empirical CDF of the oracle scores should be named consistently.

Circularity Check

0 steps flagged

No significant circularity: finite-sample guarantees follow from exchangeability and the proved score identity; asymptotic results rest on stated consistency assumptions rather than fitted targets.

full rationale

The paper's main finite-sample claims are self-contained. Lemma 3 derives the nonconformity score from the nested interval family phi(x;tau) = psi(phi_CQR(x;tau)), and the key identity s(x,y)=s(x,Pi(y)) (Eq. A18) is proved directly. This identity is what allows the scores evaluated on clipped calibration responses to be exchangeable with the latent test score, so Theorem 4 is a standard split-conformal argument, not a conclusion built into the score's definition. Theorem 5 similarly follows from exchangeability of the scores inside the internal calibration set bI; the conditional event is the same set used for calibration, but the coverage guarantee is a genuine consequence of conformal calibration on that random subset, not an assumption. The asymptotic results, including Theorem 13, are stated under explicit assumptions (Assumptions 10, 11, and 12) that concern the base quantile estimator's consistency away from the censoring thresholds and the non-extremity of the censoring; they do not assume the target oracle convergence. O0 is defined independently from the population quantiles and the algorithm, and the proof attempts to show the conformal threshold vanishes and the band approaches O0. A skeptical concern is that the proof of Theorem 13 substitutes the population region U for the data-dependent internal set int(fhat_n) and does not fully bridge the finite-sample discrepancy; however, this is a proof-gap/correctness issue rather than a circular reduction of the result to its inputs. Self-citations, such as Sesia and Candès (2020), are background references for uncensored oracle properties and are not load-bearing in the derivation. No fitted parameter is renamed as a prediction, and no central claim is equivalent to an input by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 8 axioms · 0 invented entities

There are no free parameters fitted to data: conformal thresholds are data-driven, and the jitter scale δ is a standard technical device. The finite-sample result rests on exchangeability plus the score identity; the asymptotic claims buy their oracle limit with the listed consistency and regularity assumptions. No new physical or probabilistic entities are introduced.

axioms (8)
  • domain assumption Exchangeability (or i.i.d.) of (X_i, Y_i, Ẽ_i) for i=1..n+1, and independence of calibration data from the pretrained base predictor.
    Section 2.1: 'We assume throughout that {(X_i, Y_i, Ẽ_i)} are exchangeable random samples...'; all split-conformal coverage proofs rely on this.
  • domain assumption Known, fixed censoring thresholds C_L < C_R and a bounded support [Y_min, Y_max] with finite range R.
    Section 2.1 defines Ẽ = Π(Y) and R = Y_max − Y_min > 0; the score and snapping maps (4)–(6) use these quantities. Section 5 states heterogeneous thresholds are future work.
  • domain assumption Base prediction interval Ф(x)=[Ф_lo(x),Ф_up(x)] is fixed, pretrained, independent of calibration data, and satisfies Ф_lo(X)≤Ф_up(X) a.s.
    Section 2.1: 'We make no other assumptions on this model...'; the nested-family construction in (5) requires ordered endpoints.
  • domain assumption Assumption 6 regularity: continuous conditional density of Y|X, positive on the support, uniformly bounded; q_lo(X), q_up(X) have continuous distributions without point masses.
    Stated in Section 3 before Proposition 7 and used throughout the oracle and asymptotic arguments.
  • domain assumption Assumption 10 consistency: the base predictor estimates conditional quantiles within the observable range, i.e. E|ΠФ_lo(X)-Πq_lo(X)| → 0 and analogously for the upper endpoint.
    Stated in Section 3; this is the key stated condition for Theorem 13's convergence to O_0.
  • domain assumption Assumption 11 no threshold concentration: Ф_lo(X) does not concentrate at C_L and Ф_up(X) does not concentrate at C_R.
    Stated in Section 3; the paper itself notes it excludes models trained directly on clipped outcomes, which have point masses at C_L and C_R.
  • domain assumption Assumption 12 non-extreme censoring: P(C_L < q_lo(X) ≤ q_up(X) < C_R) > 0.
    Stated in Section 3; used to drive the internal-group calibration in Theorem 13.
  • standard math Standard split-conformal and order-statistic lemmas: exchangeability of scores plus distinctness implies coverage within [1−α, 1−α+1/(n+1)]; quantile-inflation lemma for ties.
    Used in proofs of Theorems 4, A16, and A22; referenced to Romano et al. (2019) and Angelopoulos et al. (2024).

pith-pipeline@v1.3.0-alltime-deepseek · 42913 in / 16519 out tokens · 149147 ms · 2026-08-01T09:30:59.964056+00:00 · methodology

0 comments
read the original abstract

We study conformal prediction for regression using calibration data with outcomes that are doubly censored (clipped) at known fixed thresholds. We show that existing methods are unsatisfactory in this setting, as they yield intervals that may have higher marginal coverage than desired and yet lose conditional coverage precisely for the easier-to-predict cases whose outcomes are typically fully observed. This reveals that marginal coverage, the usual target of conformal prediction, may not be the ideal goal under clipping. We address this challenge by introducing a new nonconformity score and calibration methods at both ends of this trade-off: one for tight marginal coverage, and a two-step method that prioritizes conditional coverage. We characterize their finite-sample coverage and oracle-like asymptotic behavior under suitable consistency of the underlying model, and we compare them to more direct adaptations of existing approaches.

Figures

Figures reproduced from arXiv: 2607.20736 by Matteo Sesia, Vladimir Svetnik.

Figure 1
Figure 1. Figure 1: Performance of conformal prediction intervals on doubly censored synthetic data with uni [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

128 extracted references · 9 linked inside Pith

  1. [1]

    2025 , journal=

    Conformal Inference for Open-Set and Imbalanced Classification , author=. 2025 , journal=

    2025

  2. [2]

    Journal of chemical information and modeling , volume=

    Deep neural nets as a method for quantitative structure--activity relationships , author=. Journal of chemical information and modeling , volume=. 2015 , publisher=

    2015

  3. [3]

    Advances in neural information processing systems , volume=

    Classification with valid and adaptive coverage , author=. Advances in neural information processing systems , volume=

  4. [4]

    Stat , volume=

    Knowing what you know: valid confidence sets in multiclass and multilabel prediction , author=. Stat , volume=

  5. [5]

    arXiv preprint arXiv:2307.09302 , year=

    Conformal prediction under ambiguous ground truth , author=. arXiv preprint arXiv:2307.09302 , year=

    Pith/arXiv arXiv

  6. [6]

    Adaptive conformal classification with noisy labels , author=. J. R. Stat. Soc. Ser. B Methodol. , volume=. 2025 , publisher=

    2025

  7. [7]

    Journal of Machine Learning Research , volume=

    Label noise robustness of conformal prediction , author=. Journal of Machine Learning Research , volume=

  8. [8]

    Pattern Recognition , volume=

    Nested conformal prediction and quantile out-of-bag ensemble methods , author=. Pattern Recognition , volume=. 2022 , publisher=

    2022

  9. [9]

    Information and Inference: A Journal of the IMA , volume=

    The limits of distribution-free conditional predictive inference , author=. Information and Inference: A Journal of the IMA , volume=. 2021 , publisher=

    2021

  10. [10]

    European conference on machine learning , pages=

    Inductive confidence machines for regression , author=. European conference on machine learning , pages=. 2002 , organization=

    2002

  11. [11]

    2024 , publisher=

    O’Neill, Eoghan , journal=. 2024 , publisher=

    2024

  12. [12]

    arXiv preprint arXiv:2203.04676 , year=

    SparseChem: Fast and accurate machine learning model for small molecules , author=. arXiv preprint arXiv:2203.04676 , year=

    Pith/arXiv arXiv

  13. [13]

    Journal of Banking & Finance , volume=

    Grabit: Gradient tree-boosted Tobit models for default prediction , author=. Journal of Banking & Finance , volume=. 2019 , publisher=

    2019

  14. [14]

    Pattern Anal

    A deep learning approach to censored regression , author=. Pattern Anal. Appl. , volume=. 2024 , publisher=

    2024

  15. [15]

    Econometrica , pages=

    Regression analysis when the dependent variable is truncated normal , author=. Econometrica , pages=. 1973 , publisher=

    1973

  16. [16]

    Econometrica , pages=

    Estimation of relationships for limited dependent variables , author=. Econometrica , pages=. 1958 , publisher=

    1958

  17. [17]

    2005 , publisher =

    Nondetects and Data Analysis: Statistics for Censored Environmental Data , author =. 2005 , publisher =

    2005

  18. [18]

    Pesticide science , volume=

    The use of artificial neural networks in QSAR , author=. Pesticide science , volume=. 1992 , publisher=

    1992

  19. [19]

    Drug discovery today , volume=

    The rise of deep learning in drug discovery , author=. Drug discovery today , volume=. 2018 , publisher=

    2018

  20. [20]

    Nature reviews Drug discovery , volume=

    Applications of machine learning in drug discovery and development , author=. Nature reviews Drug discovery , volume=. 2019 , publisher=

    2019

  21. [21]

    Drug discovery today , volume=

    Uncertainty quantification in drug design , author=. Drug discovery today , volume=. 2021 , publisher=

    2021

  22. [22]

    Drug discovery with explainable artificial intelligence , author=. Nat. Mach. Intell. , volume=. 2020 , publisher=

    2020

  23. [23]

    Enhancing uncertainty quantification in drug discovery with censored regression labels , author=. Artif. Intell. Life Sci. , volume=. 2025 , publisher=

    2025

  24. [24]

    2026 , journal=

    Elements of Conformal Prediction for Statisticians , author=. 2026 , journal=

    2026

  25. [25]

    2000 , publisher=

    Asymptotic statistics , author=. 2000 , publisher=

    2000

  26. [26]

    Stat , volume=

    A comparison of some conformal quantile regression methods , author=. Stat , volume=. 2020 , publisher=

    2020

  27. [27]

    arXiv preprint arXiv:2501.10117 , year=

    Prediction Sets and Conformal Inference with Interval Outcomes , author=. arXiv preprint arXiv:2501.10117 , year=

    arXiv

  28. [28]

    Boosting e-

    Lee, Junu and Ren, Zhimei , journal=. Boosting e-

  29. [29]

    arXiv preprint arXiv:2506.16229 , year=

    Diversifying Conformal Selections , author=. arXiv preprint arXiv:2506.16229 , year=

    arXiv

  30. [30]

    Estimating diagnostic uncertainty in artificial intelligence assisted pathology using conformal prediction , author=. Nat. Comm. , volume=. 2022 , publisher=

    2022

  31. [31]

    Conformal prediction in clinical medical sciences , author=. J. Healthc. Inform. Res. , volume=. 2022 , publisher=

    2022

  32. [32]

    Uncertainty in lung cancer stage for survival estimation via set-valued classification , author=. Stat. Med. , volume=. 2022 , publisher=

    2022

  33. [33]

    Biometrics , volume=

    Conformal predictive intervals in survival analysis: a resampling approach , author=. Biometrics , volume=. 2025 , publisher=

    2025

  34. [34]

    Model-assisted cohort selection with bias analysis for generating large-scale cohorts from the

    Birnbaum, Benjamin and Nussbaum, Nathan and Seidl-Rathkopf, Katharina and Agrawal, Monica and Estevez, Melissa and Estola, Evan and Haimson, Joshua and He, Lucy and Larson, Peter and Richardson, Paul , journal=. Model-assisted cohort selection with bias analysis for generating large-scale cohorts from the

  35. [35]

    Comparison of population characteristics in real-world clinical oncology databases in the

    Ma, Xinran and Long, Lura and Moon, Sharon and Adamson, Blythe JS and Baxi, Shrujal S , journal=. Comparison of population characteristics in real-world clinical oncology databases in the. 2020 , publisher=

    2020

  36. [36]

    An enhanced prognostic score for overall survival of patients with cancer derived from a large real-world cohort , author=. Ann. Oncol. , volume=. 2020 , publisher=

    2020

  37. [37]

    Random survival forests , author =. Ann. Appl. Statist. , year =

  38. [38]

    Distribution-free predictive inference for regression , author=. J. Am. Stat. Assoc. , volume=. 2018 , publisher=

    2018

  39. [39]

    arXiv preprint arXiv:2411.11824 , year=

    Theoretical foundations of conformal prediction , author=. arXiv preprint arXiv:2411.11824 , year=

    Pith/arXiv arXiv

  40. [40]

    International conference on machine learning , pages=

    SAFFRON: an adaptive algorithm for online control of the false discovery rate , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  41. [41]

    Online rules for control of false discovery rate and false discovery exceedance , author=. Ann. Stat. , volume=. 2018 , publisher=

    2018

  42. [42]

    Constructing inverse probability weights for marginal structural models , author=. Am. J. Epidemiol. , volume=. 2008 , publisher=

    2008

  43. [43]

    Survival model predictive accuracy and

    Heagerty, Patrick J and Zheng, Yingye , journal=. Survival model predictive accuracy and. 2005 , publisher=

    2005

  44. [44]

    Time-dependent

    Heagerty, Patrick J and Lumley, Thomas and Pepe, Margaret S , journal=. Time-dependent. 2000 , publisher=

    2000

  45. [45]

    Conformal Risk Control , volume =

    Angelopoulos, Anastasios and Bates, Stephen and Fisch, Adam and Lei, Lihua and Schuster, Tal , booktitle =. Conformal Risk Control , volume =. 2024 , organization =

    2024

  46. [46]

    International Conference on Learning Representations , year=

    Conformalized Survival Analysis for General Right-Censored Data , author=. International Conference on Learning Representations , year=

  47. [47]

    Package `

    Blanche, P and Blanche, MP , journal=. Package `

  48. [48]

    Biometrics , volume=

    Semiparametric models of time-dependent predictive values of prognostic biomarkers , author=. Biometrics , volume=. 2010 , publisher=

    2010

  49. [49]

    Gaussian approximations and multiplier bootstrap for maxima of sums of high-dimensional random vectors , author=. Ann. Stat. , pages=. 2013 , publisher=

    2013

  50. [50]

    Evaluating prediction rules for t-year survivors with censored regression models , author=. J. Am. Stat. Assoc. , volume=. 2007 , publisher=

    2007

  51. [51]

    Application of the time-dependent

    Zheng, Yingye and Cai, Tianxi and Feng, Ziding , journal=. Application of the time-dependent. 2006 , publisher=

    2006

  52. [52]

    Adapting machine learning techniques to censored time-to-event health record data: A general-purpose approach using inverse probability of censoring weighting , author=. J. Biomed. Inform. , volume=. 2016 , publisher=

    2016

  53. [53]

    A threshold-free summary index of prediction accuracy for censored time to event data , author=. Stat. Med. , volume=. 2018 , publisher=

    2018

  54. [54]

    Biometrical Journal , volume=

    A novel nonparametric time-dependent precision--recall curve estimator for right-censored survival data , author=. Biometrical Journal , volume=. 2024 , publisher=

    2024

  55. [55]

    JCO Clin

    Creating a Proxy for Baseline Eastern Cooperative Oncology Group Performance Status in Electronic Health Records for Comparative Effectiveness Research in Advanced Non--Small Cell Lung Cancer , author=. JCO Clin. Cancer Inform. , volume=. 2025 , publisher=

    2025

  56. [56]

    Biometrika , volume=

    Estimating subject-specific survival functions under the accelerated failure time model , author=. Biometrika , volume=. 2003 , publisher=

    2003

  57. [57]

    Time-dependent predictive values of prognostic biomarkers with failure time outcome , author=. J. Am. Stat. Assoc. , volume=. 2008 , publisher=

    2008

  58. [58]

    arXiv preprint arXiv:0907.3740 , year=

    Empirical bernstein bounds and sample variance penalization , author=. arXiv preprint arXiv:0907.3740 , year=

    Pith/arXiv arXiv

  59. [59]

    PloS one , volume=

    Weight trimming and propensity score weighting , author=. PloS one , volume=. 2011 , publisher=

    2011

  60. [60]

    2013 , publisher=

    Counting processes and survival analysis , author=. 2013 , publisher=

    2013

  61. [61]

    The control of the false discovery rate in multiple testing under dependency , author=. Ann. Stat. , pages=. 2001 , publisher=

    2001

  62. [62]

    Testing for outliers with conformal p-values , author=. Ann. Stat. , volume=. 2023 , publisher=

    2023

  63. [63]

    The positive false discovery rate: a Bayesian interpretation and the q-value , author=. Ann. Stat. , volume=. 2003 , publisher=

    2003

  64. [64]

    A direct approach to false discovery rates , author=. J. R. Stat. Soc. Ser. B Methodol. , volume=. 2002 , publisher=

    2002

  65. [65]

    Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications , pages =

    Conformal Survival Bands for Risk Screening under Right-Censoring , author =. Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications , pages =. 2025 , volume =

    2025

  66. [66]

    Multiple testing in clinical trials , author=. Stat. Med. , volume=. 1991 , publisher=

    1991

  67. [67]

    Learn then test: Calibrating predictive algorithms to achieve risk control , author=. Ann. Appl. Stat. , volume=. 2025 , publisher=

    2025

  68. [68]

    Controlling the false discovery rate: a practical and powerful approach to multiple testing , author=. J. R. Stat. Soc. Ser. B Methodol. , volume=. 1995 , publisher=

    1995

  69. [69]

    Selection by prediction with conformal p-values , author=. J. Mach. Learn. Res. , volume=

  70. [70]

    Foundations and trends in machine learning , volume=

    Conformal prediction: A gentle introduction , author=. Foundations and trends in machine learning , volume=. 2023 , publisher=

    2023

  71. [71]

    JCO Clin

    Machine learning in oncology: methods, applications, and challenges , author=. JCO Clin. Cancer Inform. , volume=

  72. [72]

    Artificial Intelligence Review , volume=

    Deep learning for survival analysis: a review , author=. Artificial Intelligence Review , volume=. 2024 , publisher=

    2024

  73. [73]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Deephit: A deep learning approach to survival analysis with competing risks , author=. Proceedings of the AAAI conference on artificial intelligence , volume=. 2018 , organization =

    2018

  74. [74]

    Survival outcome prediction in cervical cancer: Cox models vs deep-learning model , author=. Am. J. Obstet. Gynecol. , volume=. 2019 , publisher=

    2019

  75. [75]

    Biometrika , volume =

    Gao, Zijun , title =. Biometrika , volume =. 2024 , month =

    2024

  76. [76]

    Strong control, conservative point estimation and simultaneous conservative consistency of false discovery rates: a unified approach , author=. J. R. Stat. Soc. Ser. B Methodol. , volume=. 2004 , publisher=

    2004

  77. [77]

    Electron

    Is distribution-free inference possible for binary regression? , author=. Electron. J. Statist. , volume=

  78. [78]

    Artificial Intelligence Applications and Innovations: AIAI 2014 Workshops: CoPA, MHDW, IIVC, and MT4BD, Rhodes, Greece, September 19-21, 2014

    Aggregated conformal prediction , author=. Artificial Intelligence Applications and Innovations: AIAI 2014 Workshops: CoPA, MHDW, IIVC, and MT4BD, Rhodes, Greece, September 19-21, 2014. Proceedings 10 , pages=. 2014 , organization=

    2014

  79. [79]

    Conformal and probabilistic prediction and applications , pages=

    On the calibration of aggregated conformal predictors , author=. Conformal and probabilistic prediction and applications , pages=. 2017 , organization=

    2017

  80. [80]

    Bernoulli , volume=

    Conformal prediction: a unified review of theory and new challenges , author=. Bernoulli , volume=. 2023 , publisher=

    2023

Showing first 80 references.