Conformal prediction regions for dependent censored lifetimes

arXiv:2610.24911 · stat.ME · Submitted 2026-10-17 · 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: "Conformal prediction regions for dependent censored lifetimes".

Jane: Here's the gist of the paper. Joint-life insurance often involves several lifetimes subject to censoring at known dates We construct prediction regions by reducing the lifetime vector to a suitable scalar outcome Applying univariate weighted split-conformal calibration to the reduced outcomes then gives finite-sample marginal coverage with known retention probabilities under conditionally independent censoring The approach starts by representing a nested family of regions by the upper level sets of a scalar-valued reduction map s of the lifetime vector X.

Tom: A map s: R d → R¯ is defined as censoring-compatible if s(x c) = s(x) c for all x, c ∈ R d.

Jane: Theorem 1 establishes that this property is equivalent to the existence of nondecreasing transformations ψj: R → R¯ such that s(x) = min 1≤j≤d ψj (x(j)).

Tom: When all coordinates of an observation are censored at a common date, the admissible class also includes the first-death and last-survivor times.

Paper summary: Tom: We're diving into the details of this paper now. This is about constructing prediction regions for joint-life insurance data where multiple lifetimes are subject to known censoring dates Tom.

Jane: The core thesis is that they reduce the lifetime vector down to a single scalar outcome, and then they use univariate weighted split-conformal calibration on those reduced outcomes Jane.

Lu: They start by finding maps that make arbitrary componentwise censoring turn into ordinary right censoring, which are minima of coordinatewise nondecreasing transformations Lu.

Meng: That transformation is key because it defines what an upper orthant—basically the prediction region—actually looks like in this complex space Meng.

Tom: They claim that under a common censoring date, the admissible class also includes both the first-death and last-survivor times Tom.

Jane: And then they apply that univariate weighted split-conformal calibration to these reduced outcomes to give finite sample marginal coverage with known retention probabilities under conditionally independent censoring Jane.

Lu: They also mention that if you have estimated probabilities, the asymptotic validity follows from consistent weight or outcome-quantile estimation under the respective regularity conditions Lu.

Meng: It sounds mathematically dense, but what does that actually mean for deployment? What's the practical hurdle they overcame?

Tom: The hurdle seems to be making sure we get a reliable finite sample guarantee, which is what this method aims to provide Tom.

Jane: So, how do they handle the dependency structure within those lifetimes and censoring times when setting up the initial model? I want to understand the setup more clearly Jane.

Lu: The paper deals with joint distributions of lifetimes and covariates, and for coupled lives, they specify parametric or semiparametric copula models Lu.

Meng: Those copulas are complex things to fit. If we can't get the right dependence structure, the whole region prediction falls apart Meng.

Tom: They stress that the validity depends entirely on specifying both the margins and that dependence structure correctly Tom.

Jane: So, what’s a simpler way to think about what they are trying to achieve with these reduction maps and orthants? I want to get the big picture down for our listeners Jane.

Lu: Think of it like projecting a high-dimensional problem onto a single dimension where the geometry becomes manageable Lu.

Meng: Like simplifying a multi-variable risk assessment into something we can actually calculate and monitor at scale Meng.

Tom: That’s the mechanical goal, trying to tame the complexity so we can get some kind of reliable coverage guarantee back on our hands Tom.

Conclusion: Jane: So, wrapping up this discussion on "Conformal prediction regions for dependent censored lifetimes." The authors are Martin Bladt and their team at the University of Copenhagen Jane.

Tom: They’ve built a method that constructs prediction regions by reducing the lifetime vector to a suitable scalar outcome Tom.

Lu: This approach is mathematically grounded in finding those maps that simplify componentwise censoring into ordinary right censoring Lu.

Meng: The main implication for us is getting a concrete, finite sample guarantee on coverage, which means we aren't just guessing at the risk when data is sparse Meng.

Tom: It’s about moving from abstract mathematical structures to something with actual measurable inclusion probabilities in real-world insurance scenarios Tom.

Jane: So, the authors are showing that even with dependencies and censoring, you can get reliable prediction regions if you use this specific calibration method Jane.

Martin Bladt

University of Copenhagen

stat.ME

Submitted: 2026-10-17

Updated: 2026-11-03

Project page: https://dutangc.github.io/CASdatasets/reference/canlifins.html

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 82/100

The gist: Joint-life insurance often involves several lifetimes subject to censoring at known dates We construct prediction regions by reducing the lifetime vector to a suitable scalar outcome Applying…

Key concepts

Reduction Map s
A function that simplifies the high-dimensional lifetime vector into a single scalar outcome. This map must be 'censoring-compatible,' meaning it preserves the structure of censored data, which is key to defining the prediction regions.
Split-Conformal Calibration
A technique used to calibrate prediction regions by splitting data based on covariate values and applying conformal methods. It helps determine the appropriate thresholds for defining the boundaries of the prediction regions.
Retention Probability
The conditional probability that an observation's fully observed censoring vector falls within a specific orthant defined by the reduction map. This probability is estimated using binary regression on covariates and is crucial for ensuring coverage guarantees.
Double-Robustness
A property showing that the marginal coverage converges to 1 - alpha as the sample size grows, regardless of whether retention probabilities or conditional outcome quantiles are consistently estimated. This ensures asymptotic validity.

Terminology

Summary

Joint-life insurance often involves several lifetimes subject to censoring at known dates We construct prediction regions by reducing the lifetime vector to a suitable scalar outcome Applying univariate weighted split-conformal calibration to the reduced outcomes then gives finite-sample marginal coverage with known retention probabilities under conditionally independent censoring

Methodology for Prediction Regions

The approach starts by representing a nested family of regions by the upper level sets of a scalar-valued reduction map s of the lifetime vector X. A map s: R d → R¯ is defined as censoring-compatible if s(x c) = s(x) c for all x, c ∈ R d. p. 2 Theorem 1 establishes that this property is equivalent to the existence of nondecreasing transformations ψj: R → R¯ such that s(x) = min 1≤j≤d ψj (x(j)). p. 3 When all coordinates of an observation are censored at a common date, the admissible class also includes the first-death and last-survivor times. p. 1

Calibration and Coverage Guarantees

The construction involves selecting a region Rb(w) based on observations whose censoring vector lies in an orthant, which is determined by the reduction map. The retention probability is the conditional probability that the fully observed censoring vector lies in an orthant, and it can be estimated by binary regression on the covariates. p. 1 The overall construction gives finite-sample coverage with known retention probabilities, or without weights under censoring independent of lifetimes and covariates. p. 1 Asymptotic validity follows from consistent estimation of either the retention probabilities or the conditional outcome quantiles, subject to the corresponding regularity conditions. p. 1

Asymptotic Validity and Robustness

Theorem 4 establishes a double-robustness property, showing that under Assumption 3.1 and Assumption 3. p. 22, the marginal coverage converges to 1 − α as N, n → ∞. p. 8 This convergence holds for any score permitted by Algorithm 1. p. 8 The asymptotic validity of the conditional coverage statement in part (ii) depends on consistent outcome quantile estimation and the stated regularity conditions. p. 8

Application to Joint-Life Annuity Data

The application estimates lower prediction bounds for reported first- and second-death times in a joint-life annuity portfolio with a short observation period. The analysis uses the common observation period A = (A, A) as the observed administrative censoring vector. p. 0 Results show that bandwise calibration lowers the corresponding mean bounds and raises inclusion probabilities to 0. p. 9910 and 0.925 in certain age bands. p. 7

Simulation Findings

The simulation studies examine several aspects of the proposed method, including independent censoring, retention depending on covariates, and a short observation period. In setting S3, with the identity path, the fitted and oracle procedures have coverage 0. p. 1910 and 0.909 with mean bounds of 3. p. 685 and 3.87 years respectively. p. 9 In setting S2, using fitted weights, the procedure achieves a coverage of 0. p. 7907. The path estimation error decreases from 0. p. 7042 to 0.005 as the training size increases from 250 to 16,000. p. 5

Retention Probability and Comparison

The retention probability depends on the path only through the corner ψ−1(c0), and on the censoring distribution through its margins at the corner and its survival copula. Proposition 3 shows that for equal standardized censoring margins, the largest threshold with retention probability at least π0 is G¯−1ψ(π0) = (G¯−10(π0), if the coordinates of C are comonotone, G¯−10(π1/d0), if they are independent. p. 9 The comparison with Bonferroni bounds shows that Rψ(vFψ) ⊆ RBψ would follow from vFψ ≥ maxj vBψ, but the reverse inclusion follows from vFψ ≤ minj vBψ. p. 1

Path Selection and Estimation

The oracle criterion for selecting an orthant path is defined by maximizing the smallest standardised coordinate − log uj. The stability of the fitted path is shown in Theorem 5, which proves that Rψ(qψ(w) + 2εN) ⊆ Rψˆ(qψˆ(w)). p. 6 The final result shows that Lb(W) − qψ(W) → 0 in probability. p. 9

First-Death and Second-Death Bounds

The analysis reports lower bounds for the first-death times separately, with the primary bound equaling five years for 28% of retained observations. For the second-death analysis using s(X) = max(X(1), X(2)), pooling over all ages gives a median unweighted empirical inclusion probability of 0.983. Bandwise calibration produces lower bounds that are less often capped and empirical inclusion probabilities closer to 0. p. 695 in the oldest bands. p. 7

Conclusion on Bounds

The primary bound equals five years for 28% of retained observations. The analysis reports, by design, lower bounds of at most five years. p. 0 In conclusion, the primary procedure has empirical inclusion probability near the nominal level in aggregate, although it varies across entry-age bands. p. 7 The analysis reports lower bounds of at most five years by design. p. 0

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. Conformal prediction regions for dependent censored lifetimes can be used to construct prediction regions in R d that contain a vector of remaining lifetimes with probability at least 1 minus alpha, generalizing existing methods for single right-censored lifetimes.

  2. The system can perform finite-sample marginal coverage guarantees under conditionally independent censoring by using estimated retention probabilities and inverse-retention weights derived from binary regression on covariates, as shown in Algorithm 1 and Theorem 3.

  3. The AI can estimate lower prediction bounds for reported first- and second-death times in joint-life portfolios using the proposed method, including a common-date reduction map s(x) = max(x(1), x(2)) for the second death time analysis.

  4. The system can distinguish between nuisance models (e.g., Gompertz vs Weibull Accelerated Failure Time) by comparing coverage achieved with different score types and weight functions, as demonstrated in Setting S2 of the simulation study.

  5. The AI can estimate the conditional outcome quantile under covariate shift using a weighted split-conformal construction, allowing for calibration on retained observations whose covariate distribution differs from the target distribution, as described in Algorithm 1 and Theorem 3.

Related papers