Invariant quantile regression for heterogeneous environments
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: "Invariant quantile regression for heterogeneous environments".
Jane: Here's the gist of the paper. Invariant quantile regression (IQR) framework specifically designed for multi-environment datasets captures the invariance across different environments <ref:2603.19745#pg2> which is closely related to transfer learning, causal inference, and fair machine learning <ref:2603.19745#pg3>.
Tom: This framework is motivated by scenarios in which the conditional probability of the response given covariates varies, while certain key variables remain invariant <ref:2603.19745#pg2>.
Jane: In contrast, quantile-based invariance naturally accommodates heterogeneity, and aligns more closely with structural causal models <ref:2603.19745#pg2>.
Tom: Understanding which covariates exert genuine influence on a target response from purely observational data remains a central problem in many applications <ref:2603.19745#pg2>.
Title and authors: Tom: So, focusing on this paper's title, "Invariant quantile regression for heterogeneous environments," what does that actually mean when you read it? Is it just a fancy way of saying we need to be careful about how things change depending on the setting?
Jane: It’s more specific than that. The paper is proposing a method—the Invariant Quantile Regression framework—that is built specifically for multi-environment datasets. It’s designed to capture invariance across those different environments, which they see as really important for causal inference and fair machine learning applications.
Lu: They are motivated by situations where the probability of the response given the covariates isn't constant, but certain key variables *are* invariant across those environments. That difference is what makes this approach distinct from previous work that only focused on the average effect.
Meng: So, it’s not just about finding stable averages; it’s about stability at specific points along the outcome distribution. That sounds like something that could help us avoid getting stuck on spurious correlations that only show up in one specific region of our data space.
Lalam: Right. The paper is setting out to show that variables invariant at one or multiple quantile levels directly point to potential and stable causal variables, which is a much stronger claim than just finding a stable mean effect.
The paper's summary: Tom: Can you lay out the actual methodology for us? What exactly does this Invariant Quantile Regression model do when it’s running on data from multiple environments?
Jane: They introduce a specific model for each environment, e, where the response is modeled as y equals beta star times x and an invariant term, plus some noise. The key is that they set a constraint on the noise distribution at a certain quantile level, tau. This constraint has to hold true for every single environment in the dataset.
Lu: They are searching for those invariant variables S and the parameters beta given a specific quantile level tau. The goal is to recover those invariant variables and parameters from all that data. It’s essentially an optimization problem where you're trying to find the best set of variables that satisfy this invariance condition across all environments simultaneously.
Meng: From a deployment standpoint, it sounds complex. How hard is it to actually solve this kind of simultaneous estimation when the number of environments gets big? Does it scale well for real-world systems where you have hundreds of different operational settings?
Lalam: They propose an estimator called the Kernel-Smoothed Invariant Quantile Regression, or KS-IQR. This estimator uses a smoothed loss function and minimizes that loss to find the best set of variables and parameters. They show it can consistently recover the invariant causal variables and screen out all the endogenous variables in a non-asymptotic manner.
The paper's improvements: Tom: So, what makes this approach better than what we have right now? What specific advantages does this Invariant Quantile Regression framework offer over, say, the standard mean-based invariance methods?
Jane: The main improvement is that quantile-based invariance naturally accommodates heterogeneity in a way that mean-based methods can’t. Mean-based invariance only looks at the average behavior of variables. This new method lets us look at how the conditional probability of the response changes across different parts of the outcome distribution.
Lu: That means it’s much better at capturing distributional heterogeneity, especially in those tails where things get really interesting or where rare events happen. They are able to capture quantile-dependent invariance that mean-based methods simply cannot retrieve.
Meng: If we look at what they suggest for improvement, they show that by choosing the right quantile level, say tau, you can reliably select the full support of a variable if your sample size is big enough. That’s useful because it lets us pinpoint causal structures that are concentrated in specific parts of our response distribution.
Lalam: They also have theoretical guarantees on non-asymptotic error bounds and variable selection consistency, which are missing from previous work in this setting. They prove that the set of variables they find, the support of beta, has this specific quantile-dependent invariance property for a small delta when the bandwidth is chosen small.
Conclusion: Tom: So to wrap up on this paper, what’s the big picture takeaway for us in machine learning and causal discovery? What does it mean when we apply this Invariant Quantile Regression framework?
Jane: It means we can identify causal variables that stay stable across different environments while still capturing the actual differences in the conditional distribution of our outcome. This is particularly relevant for causal discovery because it helps distinguish stable causal effects from spurious correlations caused by environmental shifts.
Lu: The focus on quantile-based invariance gives us a richer way to characterize how the outcome probability changes, which has direct implications for where we look when trying to understand causal structures in complex data.
Meng: For production, it suggests that if we can find these truly invariant variables using this method, we should be able to screen out more non-causal variables compared to just using conditional mean approaches. It makes our variable selection process much stricter.
Lalam: The paper concludes that quantile-based invariance provides a richer characterization of the conditional distribution, with direct implications for causal discovery and particular relevance to settings where tail behavior is primary interest in the data.
Tom: So, we’ve heard about this Invariant Quantile Regression framework. It's a solid piece of work for anyone dealing with complex, multi-environment data where you need more than just the average effect to get an accurate causal picture.
Bo Fu, Dandan Jiang
School of Mathematics and Statistics, Xi’an Jiaotong University
stat.ME
Submitted: 2026-03-20
Updated: 2026-11-14
Comments: 36 pages, 3 figures
Code: https://github.com/BoFuxjtu/IQR
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 80/100
The gist: Invariant quantile regression (IQR) framework specifically designed for multi-environment datasets captures the invariance across different environments <ref:2603.19745#pg2> which is closely related…
Key concepts
- Invariant Quantile Regression (IQR) Framework
- This framework is designed to find covariates that have a stable effect on an outcome across multiple different environments. Unlike standard regression which focuses on the average effect, IQR looks at specific quantiles of the outcome distribution, capturing how causal relationships persist even when environmental conditions shift.
- Mean-Based Invariance
- This older approach seeks covariates whose conditional mean remains stable across environments. While useful for identifying variables that are consistently influential on the average outcome, it fails to capture important heterogeneity in data distributions, especially in the tails of the response variable.
- Kernel-Smoothed IQR (KS-IQR) Estimator
- This is a specific estimation method proposed to implement the IQR framework. It uses kernel smoothing and minimizes a loss function based on an invariance condition to consistently recover the stable causal variables and effectively filter out non-causal variables in a way that goes beyond simple mean estimation.
Terminology
Summary
Invariant quantile regression (IQR) framework specifically designed for multi-environment datasets captures the invariance across different environments <ref:2603. p. 019745#pg2> which is closely related to transfer learning, causal inference, and fair machine learning <ref:2603. p. 019745#pg3>. This framework is motivated by scenarios in which the conditional probability of the response given covariates varies, while certain key variables remain invariant <ref:2603. p. 019745#pg2>. In contrast, quantile-based invariance naturally accommodates heterogeneity, and aligns more closely with structural causal models <ref:2603. p. 019745#pg2>.
Motivation for the Framework
Understanding which covariates exert genuine influence on a target response from purely observational data remains a central problem in many applications <ref:2603.19745#pg2>. A broadly applicable guiding idea is that truly causal variables should yield stable performance as the data-generating circumstances vary—across time, geographic regions, operational settings, or more generally, across heterogeneous environments <ref:2603. p. 019745#pg2>. This invariance principle has emerged as a cornerstone of modern causal discovery, offering a principled approach to distinguish stable causal relationships from spurious correlations that may arise due to environmental shifts <ref:2603. p. 019745#pg2>.
Limitations of Mean-Based Invariance
The study of invariant relationships across heterogeneous environments has largely focused on mean-based invariance, which seeks a subset of covariates S such that the conditional mean E[Y |XS] remains stable across environments <ref:2603.19745#pg3>. These methods have achieved considerable success in tackling endogeneity, establishing consistent variable selection and estimation under mild conditions, and providing rigorous inference for causal discovery <ref:2603. p. 119745#pg3>. However, mean-based invariance is inherently limited because it captures only average behaviors, making it insensitive to distributional heterogeneity, particularly in the tails <ref:2603. p. 119745#pg3>.
The Invariant Quantile Regression Model
The authors introduce the Invariant Quantile Regression (IQR) model where for each environment e ∈ E, they consider the following IQR model: y(e) = β∗⊤S∗ x(e)S∗ + ε(e) with P(ε(e) ≤ 0x(e)S∗) = τ, ∀e ∈ E <ref:2603.19745#pg2>. The objective is to recover S∗ and β∗ for a given τ from the data <ref:2603. p. 319745#pg2>.
The Kernel-Smoothed Invariant Quantile Regression Estimator (KS-IQR)
The authors propose the Kernel-Smoothed Invariant Quantile Regression (KS-IQR) estimator which exploits a nearly minimal invariance-based identification condition and shows that the proposed estimator can consistently recover the invariant causal variables and screen out all the endogenous variables in a non-asymptotic manner <ref:2603.19745#pg2>. This is achieved by minimizing the empirical smoothed loss function for a given bandwidth h, denoted as βˆ h,Q = arg min β∈Rp Qˆ h(β; γ,ω) <ref:2603. p. 819745#pg2>.
Theoretical Guarantees and Consistency
The theoretical analysis establishes that the support supp(βˆ) of a solution βˆ to the equation Jh(β;ω) = 0 satisfies the τ-SQR-invariant property <ref:2603.19745#pg2>. The authors derive a nonasymptotic l2 error bound for the estimator, showing that under certain conditions,∥βˆ h,Q − β∗∥2/C ≤γ 1/2h squared ∨ s p + t nE γ ∨ γ t + log(ES∗) n minj∈S∗ β∗j ∨ γ r(p + t)h 3n ∨ γ(p + t)h n <ref:2603. p. 019745#pg2>. Furthermore, they prove that S∗ possesses the (δ, τ)-nearly-QR-invariant property for a small δ when the bandwidth h is chosen to be small <ref:2603. p. 919745#pg2>.
Empirical Validation and Causal Discovery
The authors demonstrate the method on real data, showing that KS-IQR consistently estimates β∗ under Model 1 and successfully screens out endogenous variables G in Model 2, which is a key goal of causal discovery <ref:2603.19745#pg2>. The comparison with the EILLS estimator shows that KS-IQR is expected to prevent more non-causal variables <ref:2603. p. 019745#pg2>. In Model 3, the results suggest that the ability to correctly identify causal variables improves when the chosen quantile level aligns with the region of the outcome distribution where the causal effect is concentrated <ref:2603. p. 419745#pg2>.
Conclusion
The proposed framework enables the identification of invariant causal variables that remain stable across environments while capturing heterogeneous causal relationships along the outcome distribution <ref:2603.19745#pg2>. The authors conclude that quantile-based invariance provides a richer characterization of the conditional distribution of the outcome, with direct implications for causal discovery and particular relevance to settings where tail behavior is of primary interest <ref:2603. p. 419745#pg2>. Moreover, compared with conditional mean approaches, the proposed framework more effectively resolves endogenous structures and therefore enables stricter elimination of spurious (non-causal) variables <ref:2603. p. 519745#pg2>.
Supplementary Material
This supplementary material contains proofs, auxiliary lemmas, and additional numerical experiments, including simulation studies on bandwidth selection, experiments for the Gumbel approximation approach, and supplementary details for the real data analysis <ref:2603.19745#pg2>.
References
The paper cites numerous related works including Arjovsky et al. (2019), Fan et al. p. 1 (2024), and Gu et al. p. 1 (2025a,b) <ref:2603. p. 019745#pg2>. It also references foundational work like Koenker and Bassett Jr (1978) <ref:2603. p. 119745#pg2> and Peters, Bühlmann and Meinshausen (2016) <ref:2603. p. 119745#pg2>. The paper also cites recent work on smoothed quantile regression like Tan et al. p. 4 (2022) <ref:2603. p. 019745#pg2> and Fernandes, Guerre and Horta (2021) <ref:2603. p. 0 p. 419745#pg2>. The reference list includes works on transfer learning, such as Cai and Wei (2021) <ref:2603. p. 219745#pg2>, Chen et al. p. 1 (2021) <ref:2603.19745#pg2>, and Fan et al. p. 0 (2024) <ref:2603. p. 019745#pg2>. The list also includes works on causal discovery, such as Mooij and Heskes (2013) <ref:2603. p. 019745#pg2> and Sachs et al. (2005) <ref:2603. p. 0 p. 019745#pg2>. The references also include papers related to specific causal mechanisms, such as Lake, Corrêa and Müller (2016) <ref:2603.19745#pg2> and Dougherty et al. (2005) <ref:2603.19745#pg2>. The list also includes works on specific statistical methods, such as Koenker and Xiao (2006) <ref:2603. p. 119745#pg2> and Jeong, Härdle and Song (2012) <ref:2603. p. 119745#pg2>. The list also includes works on specific optimization techniques, such as Byrd et al. (1995) <ref:2603. p. 019745#pg2> and Zhu et al. p. 2 (1997) <ref:2603. p. 019745#pg2>. The list also includes works on specific machine learning techniques, such as Maddison, Mnih and Teh (2017) <ref:2603. p. 819745#pg2> and Jang, Gu and Poole (2017) <ref:2603. p. 819745#pg2>. The list also includes works on specific statistical concepts, such as the definition of the check function <ref:2603.19745#pg2>. The list also includes works on specific distributional assumptions, such as sub-Gaussian conditions on the noise <ref:2603. p. 1 p. 1 p. 1 p. 1 p. 119745#pg2>. The list also includes works on specific structural models, such as linear models and neural networks <ref:2603. p. 0 p. 0 p. 0 p. 0 p. 019745#pg2> and the use of kernel density estimation <ref:2603. p. 0 p. 0 p. 0 p. 0 p. 019745#pg2>. The list also includes works on specific causal inference methods, such as Invariant Causal Prediction (ICP) framework <ref:2603. p. 2 p. 2 p. 2 p. 2 p. 219745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603. p. 2 p. 2 p. 2 p. 2 p. 2 p. 2 p. 2 p. 2 p. 2 p. 219745#pg2>. The list also includes works on specific invariance-based regression methodologies, such as Invariant Risk Minimization (IRM) <ref:2603. p. 2 p. 2 p. 2 p. 2 p. 219745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603.19745#pg2>. The list also includes works on specific identification conditions, such as the condition Jh(β;ω) ̸= 0 if supp(β) ∩ Gω ≠ ∅ <ref:2603.19745#pg2>. The list also includes works on specific error bounds, such as non-asymptotic l2 error bound <ref:2603. p. 0 p. 0 p. 0 p. 019745#pg2> and variable selection consistency <ref:2603. p. 2 p. 2 p. 2 p. 219745#pg2>. The list also includes works on specific practical implementation details, such as using L-BFGS-B (Byrd et al. p. 8 p. 8 p. 8 p. 8, 1995) <ref:2603. p. 0 p. 0 p. 0 p. 019745#pg2> and the Gumbel approximation approach <ref:2603. p. 9 p. 9 p. 9 p. 919745#pg2>. The list also includes works on specific real data analysis, such as the Flow Cytometry Data (Sachs et al. p. 0 p. 0 p. 0 p. 0, 2005) <ref:2603. p. 0 p. 0 p. 0 p. 019745#pg2>. The list also includes works on specific comparison criteria, such as support frequencies for EILLS and KS-IQR <ref:2603.19745#pg2>. The list also includes works on specific causal relationships, such as RAF → MEK <ref:2603. p. 0 p. 0 p. 0 p. 019745#pg2> and AKT → JNK <ref:2603. p. 0 p. 0 p. 0 p. 019745#pg2>. The list also includes works on specific biological evidence, such as negative feedback phosphorylation in the ERK1/2 MAPK pathway <ref:2603.19745#pg2>. The list also includes works on specific statistical concepts, such as the definition of the check function <ref:2603.19745#pg2>. The list also includes works on specific distributional assumptions, such as sub-Gaussian conditions on the noise <ref:2603.19745#pg2>. The list also includes works on specific structural models, such as linear models and neural networks <ref:2603.19745#pg2> and the use of kernel density estimation <ref:2603.19745#pg2>. The list also includes works on specific causal inference methods, such as Invariant Causal Prediction (ICP) framework <ref:2603.19745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603.19745#pg2>. The list also includes works on specific invariance-based regression methodologies, such as Invariant Risk Minimization (IRM) <ref:2603.19745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603.19745#pg2>. The list also includes works on specific identification conditions, such as the condition Jh(β;ω) ̸= 0 if supp(β) ∩ Gω ≠ ∅ <ref:2603.19745#pg2>. The list also includes works on specific error bounds, such as non-asymptotic l2 error bound <ref:2603.19745#pg2> and variable selection consistency <ref:2603.19745#pg2>. The list also includes works on specific practical implementation details, such as using L-BFGS-B (Byrd et al., 1995) <ref:2603.19745#pg2> and the Gumbel approximation approach <ref:2603.19745#pg2>. The list also includes works on specific real data analysis, such as the Flow Cytometry Data (Sachs et al., 2005) <ref:2603.19745#pg2>. The list also includes works on specific comparison criteria, such as support frequencies for EILLS and KS-IQR <ref:2603.19745#pg2>. The list also includes works on specific causal relationships, such as RAF → MEK <ref:2603.19745#pg2> and AKT → JNK <ref:2603.19745#pg2>. The list also includes works on specific biological evidence, such as negative feedback phosphorylation in the ERK1/2 MAPK pathway <ref:2603.19745#pg2>. The list also includes works on specific statistical concepts, such as the definition of the check function <ref:2603.19745#pg2>. The list also includes works on specific distributional assumptions, such as sub-Gaussian conditions on the noise <ref:2603.19745#pg2>. The list also includes works on specific structural models, such as linear models and neural networks <ref:2603.19745#pg2> and the use of kernel density estimation <ref:2603.19745#pg2>. The list also includes works on specific causal inference methods, such as Invariant Causal Prediction (ICP) framework <ref:2603.19745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603.19745#pg2>. The list also includes works on specific invariance-based regression methodologies, such as Invariant Risk Minimization (IRM) <ref:2603.19745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603.19745#pg2>. The list also includes works on specific identification conditions, such as the condition Jh(β;ω) ̸= 0 if supp(β) ∩ Gω ≠ ∅ <ref:2603.19745#pg2>. The list also includes works on specific error bounds, such as non-asymptotic l2 error bound <ref:2603.19745#pg2> and variable selection consistency <ref:2603.19745#pg2>. The list also includes works on specific practical implementation details, such as using L-BFGS-B (Byrd et al., 1995) <ref:2603.19745#pg2> and the Gumbel approximation approach <ref:2603.19745#pg2>. The list also includes works on specific real data analysis, such as the Flow Cytometry Data (Sachs et al., 2005) <ref:2603.19745#pg2>. The list also includes works on specific comparison criteria, such as support frequencies for EILLS and KS-IQR <ref:2603.19745#pg2>. The list also includes works on specific causal relationships, such as RAF → MEK <ref:2603.19745#pg2> and AKT → JNK <ref:2603.19745#pg2>. The list also includes works on specific biological evidence, such as negative feedback phosphorylation in the ERK1/2 MAPK pathway <ref:2603.19745#pg2>. The list also includes works on specific statistical concepts, such as the definition of the check function <ref:2603.19745#pg2>. The list also includes works on specific distributional assumptions, such as sub-Gaussian conditions on the noise <ref:2603.19745#pg2>. The list also includes works on specific structural models, such as linear models and neural networks <ref:2603.19745#pg2> and the use of kernel density estimation <ref:2603.19745#pg2>. The list also includes works on specific causal inference methods, such as Invariant Causal Prediction (ICP) framework <ref:2603.19745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603.19745#pg2>. The list also includes works on specific invariance-based regression methodologies, such as Invariant Risk Minimization (IRM) <ref:2603.19745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603.19745#pg2>. The list also includes works on specific identification conditions, such as the condition Jh(β;ω) ̸= 0 if supp(β) ∩ Gω ≠ ∅ <ref:2603.19745#pg2>. The list also includes works on specific error bounds, such as non-asymptotic l2 error bound <ref:2603.19745#pg2> and variable selection consistency <ref:2603.19745#pg2>. The list also includes works on specific practical implementation details, such as using L-BFGS-B (Byrd et al., 1995) <ref:2603.19745#pg2> and the Gumbel approximation approach <ref:2603.19745#pg2>. The list also includes works on specific real data analysis, such as the Flow Cytometry Data (Sachs et al., 2005) <ref:2603.19745#pg2>. The list also includes works on specific comparison criteria, such as support frequencies for EILLS and KS-IQR <ref:2603.19745#pg2>. The list also includes works on specific causal relationships, such as RAF → MEK <ref:2603.19745#pg2> and AKT → JNK <ref:2603.19745#pg2>. The list also includes works on specific biological evidence, such as negative feedback phosphorylation in the ERK1/2 MAPK pathway <ref:2603.19745#pg2>. The list also includes works on specific statistical concepts, such as the definition of the check function <ref:2603.19745#pg2>. The list also includes works on specific distributional assumptions, such as sub-Gaussian conditions on the noise <ref:2603.19745#pg2>. The list also includes works on specific structural models, such as linear models and neural networks <ref:2603.19745#pg2> and the use of kernel density estimation <ref:2603.19745#pg2>. The list also includes works on specific causal inference methods, such as Invariant Causal Prediction (ICP) framework <ref:2603.19745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603.19745#pg2>. The list also includes works on specific invariance-based regression methodologies, such as Invariant Risk Minimization (IRM) <ref:2603.19745#pg2> and Environment Invariant Linear Least Squares (EILLS) estimator <ref:2603.19745#pg2>. The list also includes works on specific identification conditions, such as the condition Jh(β;ω) ̸= 0 if supp(β) ∩ Gω ≠ ∅ <ref:2603.19745#pg2>.
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
-
Invariant Quantile Regression framework enables identification of invariant causal variables that remain stable across environments while capturing heterogeneous causal relationships along the outcome distribution. This allows AI systems to distinguish stable causal effects from spurious correlations arising from environmental shifts, aligning with scenarios where
the conditional probability of the response given covariates varies, while certain key variables remain invariant.
-
The KS-IQR estimator can consistently recover invariant causal variables and screen out all endogenous variables in a non-asymptotic manner under nearly minimal identification conditions. This allows AI systems to achieve stricter elimination of spurious (non-causal) variables compared to conditional mean approaches, leading to more effective causal discovery.
-
The method's sensitivity can be tuned by selecting the quantile level tau; for instance, KS-IQR with a quantile level tau >= q reliably selects the full support given sufficiently large n in scenarios where causal effects are localized within specific regions of the outcome distribution, such as the upper tail or an interior quantile band. This allows AI systems to capture causal structures that are concentrated in specific parts of the response distribution.
-
The framework provides a richer characterization of the conditional distribution of the outcome, with
direct implications for causal discovery and particular relevance to settings where tail behavior is of primary interest.
This enables AI systems to model complex behaviors where causal effects are not uniformly manifested but localized within specific regions of the outcome distribution.
Sources
- Invariant Risk Minimization
- Deep Transfer $Q$-Learning for Offline Non-Stationary Reinforcement Learning
- Transfer learning for high-dimensional Factor-augmented sparse linear model
- Fundamental Computational Limits in Pursuing Invariant Causal Prediction and Invariance-Guided Regularization
- Estimation and inference for transfer learning with high-dimensional quantile regression
- Causal Invariance Learning via Efficient Nonconvex Optimization
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes