Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes

arXiv:2501.04615 · stat.ME, math.ST, stat.TH · Submitted 2025-01-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes".

Jane: Here's the gist of the paper. Our objective is to construct well-calibrated prediction sets for a time-to-event outcome subject to right-censoring with guaranteed coverage.

Tom: This paper proposes doubly robust and efficient methods, including inverse-probability-of-censoring weighting (IPCW) and augmented inverse probability of censoring weighting (AIPCW), for constructing well-calibrated prediction bounds for right-censored time-to-event outcomes.

Jane: Problem Setup The paper sets up the problem by defining the fully observed data as a i.i.d.

Tom: sample D of pXi, Yi, and ∆i1tT ď Cu.

Title and authors: Tom: So we're looking at this paper called Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes. Jane, what's the big picture here?

Jane: Well, essentially the title tells you they are building prediction sets for survival time data that handle right censoring in a way that gives you a guaranteed coverage rate. It’s not just guessing where the next event will happen; it’s about creating a bound around that guess, and making sure that bound actually works reliably.

Lu: They are tackling the problem of constructing these prediction sets without needing to rely on a perfectly specified parametric or semiparametric survival model, which is usually required in older conformal inference methods.

Meng: So they are trying to make a more robust way to get those bounds, right? That sounds good if it means I don't have to spend weeks tuning a model just because my data doesn't fit the exact distribution they assume.

Lalam: It seems like they are proposing new ways to do that calibration, moving away from methods that make strong assumptions about how the censoring happens.

Tom: Right, so we’ve established what it is: prediction sets for survival time with guaranteed coverage under right-censoring conditions. That's the core idea of this paper by Farina, Tchetgen, and Kuchibhotla.

The paper's summary: Jane: So, what’s the actual approach they are taking to get these sets? I see they are moving away from assuming Type-I censoring with fully observed censoring times.

Tom: That’s a key point, Jane. They are considering the more common right-censoring setting where you only observe either the event time or the censoring time, whichever comes first.

Lu: They propose two main approaches for calibrating these prediction intervals based on inverse probability of censoring, IPCW and its augmented version called AIPCW.

Meng: Inverse probability of censoring weighting sounds complicated to implement in a production system; what exactly does that weighting do for the actual estimate? Does it just fix the bias from the missing information?

Jane: The IPCW method involves weighting observations with observed failure times by the reciprocal of an estimated conditional survival function related to the censoring mechanism given your covariates.

Tom: And then they have this augmented version, AIPCW, which adds an extra term that uses a model for the outcome of interest to recover information from those censored observations.

Lalam: So if IPCW is about weighting things based on what we *do* see, AIPCW is about using a separate model for the outcome to fill in what we *don't* see.

The paper's improvements: Tom: Now, the authors highlight several improvements they’ve made compared to what was out there before. One big thing is achieving double robustness for these prediction sets.

Jane: Double robustness means that the coverage error comes out in a mixed bias product form when you estimate the nuisance parameters. That’s a strong property because it lets you get valid bounds even if one of those underlying survival curves isn't perfectly specified.

Lu: They also introduce methods to provide asymptotically PAC calibrated lower predictive bounds, and Theorem two point one zero shows that the error term in their coverage guarantee vanishes whenever the product of errors converges to zero in probability without needing a specific rate of convergence <ref:2501.04615#pg1>.

Meng: That’s good for robustness, but what about computation? If we're on a high-throughput system, does this complexity translate to something manageable when we are running these inverse probability weighting estimators?

Tom: They also mention outcome regression methods like Outcome Regression and Calibrated Outcome Regression as computationally efficient alternatives.

Jane: But the authors say those regression-based methods have a heavy dependence on getting the model specification exactly right, which is a trade-off for speed.

Conclusion: Tom: So, to wrap up, this paper by Farina and colleagues shows that AIPCW provides a good balance between accuracy and sharpness for prediction bounds in this context.

Jane: The double robustness property of AIPCW ensures consistency even if you have partial model misspecifications, which is crucial for real-world use.

Lu: They successfully handle standard right-censoring by focusing on whether the event time or the censoring time is observed first, rather than assuming Type-I censoring.

Meng: If we look at this from a deployment standpoint, AIPCW seems like the recommended approach because it offers stability across different settings, especially in complex ones like Setting four.

Lalam: They even applied these methods to real data on rheumatoid arthritis patients to check their reliability, and the empirical coverage showed that AIPCW did better than the outcome regression-based methods in terms of reliability.

Tom: So, we’ve looked at how this paper on Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes works. It gives us a more solid tool for building reliable prediction bounds when dealing with censored survival data.

Department of Statistics & Data Science, Carnegie Mellon University · Department of Statistics & Data Science, University of Pennsylvania

stat.ME, math.ST, stat.TH

Submitted: 2025-01-08

Updated: 2026-06-06

Comments: 48 pages, 11 figures

Journal ref: Biometrika 113(3), asag036 (2026)

DOI: 10.1093/biomet/asag036

Code: https://github.com/rebyfa98/DR_conformal_censored

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 100/100

The gist: Our objective is to construct well-calibrated prediction sets for a time-to-event outcome subject to right-censoring with guaranteed coverage.

Key concepts

Inverse-Probability-of-Censoring Weighting (IPCW)
This method weights observed data points by the reciprocal of the estimated conditional survival function related to censoring. It helps correct for the bias introduced by right-censoring, making predictions more accurate without needing to assume perfect knowledge of the censoring mechanism.
Augmented Inverse Probability of Censoring Weighting (AIPCW)
AIPCW enhances IPCW by adding an extra term that uses a model for the outcome of interest. This augmentation allows the method to recover crucial information from censored observations, improving prediction accuracy while maintaining robustness against model misspecification.
Outcome Regression (OR) and Calibrated Outcome Regression (COR)
These are regression-based alternatives that estimate quantiles directly. OR estimates the primary event time quantile, while COR adjusts this level to match desired coverage. They are computationally fast but require a very accurate specification of the survival function models to perform reliably.

Terminology

Summary

Our objective is to construct well-calibrated prediction sets for a time-to-event outcome subject to right-censoring with guaranteed coverage. p. 0

Where the paper says it

Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes Rebecca Farina1 , Eric J. Tchetgen Tchetgen2 , and Arun Kumar Kuchibhotla1 1Department of Statistics & Data Science, Carnegie Mellon University 2Department of Statistics & Data Science, University of Pennsylvania Abstract Our objective is to construct well-calibrated prediction sets for a time-to-event outcome subject to right-censoring with guaranteed coverage.

This paper proposes doubly robust and efficient methods, including inverse-probability-of-censoring weighting (IPCW) and augmented inverse probability of censoring weighting (AIPCW), for constructing well-calibrated prediction bounds for right-censored time-to-event outcomes. p. 0
Where the paper says it

Unlike existing conformal methods for survival data, which assume Type-I censoring with fully observed censoring times, we consider the more common right-censoring setting in which only the censoring time or only the event time is observed, whichever comes first. Under a standard conditional independence censoring condition, we propose and analyze several lower prediction bounds for the survival time of a future observation, including inverse-probability-of-censoring weighting, and its augmented version based on the semiparametric efficient influence function for the relevant marginal quantile of the outcome accounting for dependent censoring.

Problem Setup

The paper sets up the problem by defining the fully observed data as a i.i.d. sample D of pXi, Yi, and ∆i1tT ď Cu. p. 1

Where the paper says it

Thus, the fully observed data is given by the i. sample D “ pXi , Yi , ∆iq n i“1 .

It introduces notation for the underlying (unobserved) full data Z and defines predictive bounds LppXq and UppXq with a marginal coverage guarantee. p. 1
Where the paper says it

For any γ P p0, 1q, the conditional quantile function of T | X is defined as qT|Xpγ | Xq “ inftt P R` : ST|Xpt | Xq ď 1 ´ γu. We aim to obtain predictive bounds LppXq and UppXq for time-to-event outcome T leveraging the covariate vector X with a marginal coverage guarantee, i.

The oracle lower prediction bound is defined as the conditional quantile of T X at level α, qTXpα Xq. p. 5
Where the paper says it

We refer to β as pseudo-quantile level as the approach in fact does not rely on qpT|Xp¨ | ¨q to be consistent for the true conditional quantile function in order for the resulting prediction interval to be well-calibrated; rather, any working model, possibly incorrect, for the conditional quantile function is guaranteed to produce a prediction set which has correct marginal coverage. However, the optimal or oracle choice of qpT|Xpβ | Xq, assuming the survival curve of censoring defining the weights matches the truth, entails the true conditional quantile function, because then, β ” α by definition of the conditional quantile, ensuring not only correct marginal coverage but also correct conditional coverage, a clear improvement in coverage control.

IPCW and AIPCW Methods

The paper proposes two main approaches based on inverse-probability-of-censoring weighting (IPCW) and its augmented version (AIPCW). The IPCW method involves weighting observations with observed failure time by the reciprocal of the estimated conditional survival function of the censoring mechanism given covariates. p. 5

Where the paper says it

We propose a novel approach to construct valid PAC type LPBs for the survival time of future observations, satisfying Definition 1. The key idea behind our method is to calibrate the coverage of prediction sets for the underlying time-to-event outcome in view, by incorporating IPCW to complete-cases to account for selection bias due to censoring, using for weight the reciprocal of the estimated conditional survival function of the censoring mechanism given covariates, evaluated at the observed failure time; therefore allowing us to properly estimate and thus calibrate the coverage probability.

The AIPCW method extends this by introducing an augmentation term that leverages a model for the outcome of interest to recover information from censored observations. p. 8
Where the paper says it

To address these limitations, we extend our framework by introducing an augmentation to IPCW that recovers information from censored observations by leveraging a model for the outcome of interest. The AIPCW method has attractive theoretical appeal over the IPCW approach, by producing a doubly robust estimator, which means that the estimator remains consistent if either the survival model or the censoring model is correctly specified (but not necessarily both).

Outcome Regression Methods

Two outcome regression-based methods are also introduced: Outcome Regression (OR) and Calibrated Outcome Regression (COR). OR directly estimates the conditional quantile of the primary event time at the desired nominal level, while COR adjusts this quantile level to match the desired marginal coverage. p. 4

Where the paper says it

As a complementary contribution, we introduce two simple yet effective approaches for constructing prediction bounds based on outcome regression: Outcome Regression (OR), and Calibrated Outcome Regression (COR). The first estimates the conditional quantile function of the primary event time at the desired nominal level, while the second calibrates the estimated conditional quantile function of the survival time to achieve the desired marginal coverage guaranty.

These methods offer a computationally efficient alternative but rely on consistent estimation of survival function models. p. 4
Where the paper says it

The first estimates the conditional quantile function of the primary event time at the desired nominal level, while the second calibrates the estimated conditional quantile function of the survival time to achieve the desired marginal coverage guaranty. Although the validity of this method is contingent on a consistent estimator of the survival function model, it provides a computationally efficient alternative to the IPCW and AIPCW approaches.

Simulation and Empirical Results

The authors conduct comprehensive simulation studies across six synthetic settings to evaluate the performance of their proposed methods against various benchmarks, including naive quantile-based baselines and recent conformal survival methods. The AIPCW-SL method is highlighted as the recommended approach, showing stable performance and near-nominal coverage across challenging scenarios like Setting 4. p. 1

Where the paper says it

Although IPCW and AIPCW methods are well developed for estimation and inference in standard survival analysis settings, (a) and (b) appear to provide their first use for obtaining well-calibrated prediction intervals. In addition, we consider two so-called outcome regression-based methods that obviate the need for estimating the censoring mechanism: the Outcome Regression method (OR), which entails estimating the conditional quantile function for the outcome at the desired nominal level; and the Calibrated Outcome Regression method (COR) which adjusts the quantile level to match the desired nominal marginal coverage.

Real Data Application

The methods are applied to real data from a cohort of patients with rheumatoid arthritis to estimate predictive bounds for survival time to death. The empirical coverage is assessed using both IPCW/AIPCW and outcome regression-based approaches, showing that AIPCW outperforms regression-based methods in terms of coverage reliability. p. 1

Where the paper says it

Although IPCW and AIPCW methods are well developed for estimation and inference in standard survival analysis settings, (a) and (b) appear to provide their first use for obtaining well-calibrated prediction intervals. In addition, we consider two so-called outcome regression-based methods that obviate the need for estimating the censoring mechanism: the Outcome Regression method (OR), which entails estimating the conditional quantile function for the outcome at the desired nominal level; and the Calibrated Outcome Regression method (COR) which adjusts the quantile level to match the desired nominal marginal coverage.

Discussion

The work demonstrates that the AIPCW approach provides a favorable trade-off between validity and bound sharpness across a range of scenarios. The double robustness property of AIPCW ensures consistency even with partial model misspecifications, enhancing its reliability in practical applications. p. 4

Where the paper says it

Using AIPCW also leads to a desirable double robustness property, which ensures that the error in the coverage bound of the AIPCW prediction sets, due to estimation of nuisance parameters (mainly of the conditional survival curve for the primary outcome and for the censoring mechanism given covariates, respectively), is of mixed bias product form (Rotnitzky et al. This ensures that our method can provide valid bounds even if one of the required survival curves is misspecified, offering enhanced flexibility and reliability in practical applications.

Future work is suggested to extend the framework to handle other coarsened data settings like interval censoring. p. 3
Where the paper says it

In the Type-I censoring setting, Wan and Yi (2025) propose distributionfree lower predictive bounds by combining hybrid quantile learning with a data-filtered, adaptive conformal calibration scheme. Moving beyond Type-I censoring, Sesia and Svetnik (2025b) extend the conformalized survival analysis framework of Cand`es et al.

References

Basu, D., An essay on the logical foundations of survey sampling, part 1. p. 6

Where the paper says it

β P r0, 1s : PpHT´IPCWpβq ě 1 ´ α ) , and obtain the LPB as Lpp¨q “ qpT|XpβpHT´IPCW | ¨q. The classical IPCW estimator can be viewed as a form of Horvitz–Thompson estimator (HT-IPCW), a well-known inverse probability weighted estimator commonly used in the survey sampling literature, which computes the sample average normalizing by the sample size (Horvitz and Thompson, 1952).

Bostr¨om, H. p. 2

Where the paper says it

3 Related work Recent conformal prediction methods for survival data include Cox regression-based (Teng et al. , 2021) and random survival forests-based (Bostr¨om et al.

p. 2
Where the paper says it

3 Related work Recent conformal prediction methods for survival data include Cox regression-based (Teng et al. , 2021) and random survival forests-based (Bostr¨om et al.

, Asker, L., Gurung, R., Karlsson, I., Lindgren, T., and Papapetrou, P. (2017). Conformal prediction using random survival forests. p. 2
Where the paper says it

3 Related work Recent conformal prediction methods for survival data include Cox regression-based (Teng et al. , 2021) and random survival forests-based (Bostr¨om et al.

Bostr¨om, H., Johansson, U., and Vesterberg, A. (2019). Predicting with confidence from survival data. p. 3

Where the paper says it

(2025) proposes a bootstrap resampling approach to build conformal predictive intervals for general right-censored data, achieving asymptotic marginal validity under conditionally independent censoring. Building on our work, Si and Qiu (2025) study lower prediction bounds with training-set conditional APAC-type guarantees, meaning that the approximation is placed on the confidence level, rather than on the miscoverage probability as in our asymptotic PAC formulation, while conditioning on the training sample; they adopt the efficient influence function correction introduced by our approach.

Breslow, N. E., Contribution to discussion of paper by dr cox. p. 0

Where the paper says it

(2008); Collett (2023)). While classical tools such as the Kaplan–Meier estimator of the marginal survival curve (Kaplan and Meier, 1958), and the Cox proportional hazards regression model (Cox, 1972) are well established, there remains a lack of flexible frameworks that provide rigorous uncertainty quantification for predictive inference of rightcensored time-to-event outcomes.

Cand´es, E. p. 2

Where the paper says it

Lpp¨q is an asymptotic pα, ϵq-PAC LPB for T if P ´ PpT ě LppXq|Dq ě 1 ´ α ´ opp1q ¯ ě 1 ´ ϵ as n Ñ 8, where the opp1q random variable can depend on ϵ. As noted in Cand`es et al.

, Lei, L., and Ren, Z. (2023). Conformalized survival analysis. p. 3
Where the paper says it

In the Type-I censoring setting, Wan and Yi (2025) propose distributionfree lower predictive bounds by combining hybrid quantile learning with a data-filtered, adaptive conformal calibration scheme. Moving beyond Type-I censoring, Sesia and Svetnik (2025b) extend the conformalized survival analysis framework of Cand`es et al.

Chen, W., Wang, Z., Ha, W., and Barber, R. F. (2016). Trimmed conformal prediction for highdimensional models. p. 1

Where the paper says it

, 2024), time series analysis (Stankeviciute et al. , 2021), highdimensional regression (Chen et al.

Chernozhukov, V. p. 8

Where the paper says it

In short, the procedure outlined in this section constructs an asymptotically pα, ϵq-PAC LPB Lpp¨q for the right-censored survival time T. Moreover, two-sided bounds can also be obtained if the non-conformity score is chosen appropriately: RpX, Tq “ |2Sp T|XpT|Xq ´ 1| (Chernozhukov et al.

, W¨uthrich, K., and Zhu, Y. (2021). Distributional conformal prediction. p. 0
Where the paper says it

(2005), provides a general blueprint for constructing well-calibrated prediction intervals, with marginal coverage guarantees under relatively mild conditions. The main advantage of conformal prediction is that it provides valid finite-sample prediction intervals for a new observation exchangeable with the training sample, without imposing strong distributional assumptions on the underlying data generating mechanism.

Choi, H. K., Hern´an, M. A., Seeger, J. D., Robins, J. p. 1

Where the paper says it

Notably, to calibrate our prediction intervals, we propose two separate approaches based on inverse-propability-of-censoring. (Robins, 1993; Robins and Rotnitzky, 1992; Robins and Finkelstein, 2000).

M., and Wolfe, F. (2002). Methotrexate and mortality in patients with rheumatoid arthritis: a prospective study. p. 0
Where the paper says it

1 Introduction 1. 1 Background A fundamental challenge of survival analysis is the potential for a person’s event time to not be completely observed if, either she were to drop out, or the study ended before she experienced the event of primary interest: a phenomenon commonly known as right-censoring (see, e.

Collett, D. p. 0

Where the paper says it

, Kleinbaum and Klein (1996); Hosmer Jr et al. (2008); Collett (2023)).

(2023). Modelling survival data in medical research. p. 3
Where the paper says it

While these papers substantially advance conformal uncertainty quantification for censored outcomes, they do not simultaneously meet our specific desiderata (i)–(iii): in particular, for the setting and goal we consider, we provide a doubly robust procedure which attains semiparametric efficiency (under standard regularity conditions) and yields second-order coverage error control. In a separate research strand, inverse-probability-of-censoring weighting (IPCW) and augmentedinverse-probability-of-censoring weighting (AIPCW) have played a pivotal role in the development of robust and efficient methodologies for censored data; primarily in the context of inference for a finite dimensional functional of interest.

Cox, D. R., Regression models and life-tables. p. 8

Where the paper says it

2 and otherwise unrestricted. The tangent space for a semiparametric model is technically defined as the closed linear span of scores of all regular parametric sub-models contained in the model.

Datta, J., and Polson, N. p. 6

Where the paper says it

The H´ajek IPCW estimator normalizes by the sum of the inverse probability weights, such that the coverage probability is estimated as PpIPCWpβq “ ř iPI2 ∆i1tTi ě qpT|Xpβ | Xiqu{Sp C|XpTi | Xiq ř iPI2 ∆i{Sp C|XpTi | Xiq , and the optimal β is given by βp IPCW “ suptβ P r0, 1s : PpIPCWpβq ě 1 ´ αu. The H´ajek estimator is approximately unbiased and often yields lower variance than the Horvitz-Thompson estimator (Datta and Polson, 2022).

(2022). Inverse probability weighting: from survey sampling to evidence estimation. p. 6
Where the paper says it

β P r0, 1s : PpHT´IPCWpβq ě 1 ´ α ) , and obtain the LPB as Lpp¨q “ qpT|XpβpHT´IPCW | ¨q. The classical IPCW estimator can be viewed as a form of Horvitz–Thompson estimator (HT-IPCW), a well-known inverse probability weighted estimator commonly used in the survey sampling literature, which computes the sample average normalizing by the sample size (Horvitz and Thompson, 1952).

Fleming, T. p. 8

Where the paper says it

In the special case of completely independent censoring, where the censoring time C is independent of pT, Xq, the estimation of the censoring mechanism simplifies significantly. In this scenario, the censoring survival function S ‹ C ptq depends only on t and can be consistently estimated at a root-n rate using standard nonparametric methods, such as the Kaplan-Meier estimator (Kaplan and Meier, 1958; Fleming and Harrington, 2013).

R., Counting processes and survival analysis. p. 0
Where the paper says it

1 Introduction 1. 1 Background A fundamental challenge of survival analysis is the potential for a person’s event time to not be completely observed if, either she were to drop out, or the study ended before she experienced the event of primary interest: a phenomenon commonly known as right-censoring (see, e.

Golmakani, M. K., Super learner for survival data prediction. p. 3

Where the paper says it

(2025) construct conformal prediction intervals employing inverse censoring weighting and quantile regression as a flexible base learner. A different line of work focuses on survival curves rather than on the time-to-event: Sesia and Svetnik (2025a) construct conformal bands for individual survival curves for screening-type tasks by combining inverse censoring weighting with multiple-testing FDR control.

Gui, Y., Hore, R., Ren, Z., and Barber, R. F. (2023). Conformalized survival analysis with adaptive cut-offs. p. 3

Where the paper says it

In the Type-I censoring setting, Wan and Yi (2025) propose distributionfree lower predictive bounds by combining hybrid quantile learning with a data-filtered, adaptive conformal calibration scheme. Moving beyond Type-I censoring, Sesia and Svetnik (2025b) extend the conformalized survival analysis framework of Cand`es et al.

H´ajek, J. p. 6

Where the paper says it

However, if estimated survival probabilities defining the weights Sp C|XpTi | Xiq are relatively small, the estimator may become unstable, leading to increased variability. To address this issue, we consider the H´ajek version of IPCW estimator (H´ajek, 1971; Basu, 1970), a well-known ratio-adjusted version of the Horvitz–Thompson estimator.

, Comment on “an essay on the logical foundations of survey sampling, part 1. p. 6
Where the paper says it

β P r0, 1s : PpHT´IPCWpβq ě 1 ´ α ) , and obtain the LPB as Lpp¨q “ qpT|XpβpHT´IPCW | ¨q. The classical IPCW estimator can be viewed as a form of Horvitz–Thompson estimator (HT-IPCW), a well-known inverse probability weighted estimator commonly used in the survey sampling literature, which computes the sample average normalizing by the sample size (Horvitz and Thompson, 1952).

Holt, Rinehart and Winston.

Horvitz, D. p. 6

Where the paper says it

β P r0, 1s : PpHT´IPCWpβq ě 1 ´ α ) , and obtain the LPB as Lpp¨q “ qpT|XpβpHT´IPCW | ¨q. The classical IPCW estimator can be viewed as a form of Horvitz–Thompson estimator (HT-IPCW), a well-known inverse probability weighted estimator commonly used in the survey sampling literature, which computes the sample average normalizing by the sample size (Horvitz and Thompson, 1952).

G., A generalization of sampling without replacement from a finite universe. p. 0
Where the paper says it

(2005), provides a general blueprint for constructing well-calibrated prediction intervals, with marginal coverage guarantees under relatively mild conditions. The main advantage of conformal prediction is that it provides valid finite-sample prediction intervals for a new observation exchangeable with the training sample, without imposing strong distributional assumptions on the underlying data generating mechanism.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Double robustness for prediction sets ensures that coverage error is of mixed bias product form, meaning "the error in the coverage bound of the AIPCW prediction sets, due to estimation of nuisance parameters (mainly of the conditional survival curve for the primary outcome and for the censoring mechanism given covariates, respectively), is of mixed bias product form." This allows the system to provide valid bounds even if one required survival curve is misspecified.

  2. The method can account for standard right-censoring by considering only the censoring time or only the event time is observed, whichever comes first, moving beyond assumptions made by existing conformal methods that assume Type-I censoring with fully observed censoring times.

  3. The system can provide asymptotically PAC calibrated lower predictive bounds (LPBs) that are robust to model misspecification, as demonstrated by Theorem 2.10 showing the opp1q remainder term in our coverage guarantee vanishes whenever the same product of errors converges to zero in probability, without requiring any specific rate of convergence.

  4. The system can utilize inverse-probability-of-censoring weighting (IPCW) to adaptively account for the unknown censoring mechanism, which completely obviates the need for artificial censoring and the required tuning of a cutoff parameter.

  5. The system can leverage outcome regression methods like Outcome Regression (OR) and Calibrated Outcome Regression (COR) to obtain computationally efficient alternatives, though these are noted as less robust due to their heavy dependence on correct model specification.

Abstract

Our objective is to construct well-calibrated prediction sets for a time-to-event outcome subject to right-censoring with guaranteed coverage. Inspired by modern conformal inference, our approach avoids the need for a well-specified parametric or semiparametric survival model. Unlike existing conformal methods for survival data, which assume Type-I censoring with fully observed censoring times, we consider the more common right-censoring setting in which only the censoring time or only the event time is observed, whichever comes first. Under a standard conditional independence censoring condition, we propose and analyze several lower prediction bounds for the survival time of a future observation, including inverse-probability-of-censoring weighting, and its augmented version based on the semiparametric efficient influence function for the relevant marginal quantile of the outcome accounting for dependent censoring. We formally establish asymptotic coverage guarantees of the proposed methods, and demonstrate both theoretically and through empirical experiments, that the augmented approach substantially improves efficiency over all other proposed methods. Specifically, its coverage error bound is doubly robust, and therefore of second order, thus ensuring that it is asymptotically negligible relative to the coverage error of the other methods.

Sources

Related papers