Semiparametric Inference for Marginal Win Odds with Censored Time-to-Event Outcomes

arXiv:2611.14785 · stat.ME · Submitted 2026-11-10 · 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: "Semiparametric Inference for Marginal Win Odds with Censored Time-to-Event Outcomes".

Jane: Here's the gist of the paper. Win statistics provide clinically interpretable measures for comparing treatment groups with prioritized composite outcomes, but efficient covariate-adjustment methods for win statistics with right censored time-to-event outcomes remain limited We develop a semiparametric framework for marginal win probability and win odds with prioritized time-to-event outcomes, including semi-competing risks, while also accommodating scalar survival outcomes as a special case.

Tom: The estimand compares independent draws from the marginal distributions of the potential outcomes under treatment and control, which preserves a marginal causal interpretation.

Jane: The resulting cross-fitted one-step estimator incorporates covariate adjustment and censoring martingale augmentation, allows flexible estimation of nuisance parameters, and attains the semiparametric efficiency bound under suitable conditions on consistency and convergence rate of nuisance parameters.

Tom: Simulation studies show that the estimator has small bias, accurate standard error estimation and confidence interval coverage, and substantial efficiency gain relative to the inverse-probability-of-censoring weighted (IPCW) estimator.

Paper summary: Tom: So we're looking at this paper today, "Semiparametric Inference for Marginal Win Odds with Censored Time-to-Event Outcomes". Basically, it tackles a problem where you want to compare treatment groups using win statistics, but your data has right censored time-to-event outcomes.

Jane: Right. So the main idea here is that win statistics are a good way to compare treatments when you have prioritized composite outcomes—like prioritizing death over just disease progression—but existing methods for adjusting for covariates and dealing with censoring in these scenarios are pretty limited, according to this paper <ref:2611.14785#pg1>.

Lu: What it claims is that they develop a semiparametric framework for both the marginal win probability and the win odds, which means they are focused on those two specific measures <ref:2611.14785#pg2>. They are trying to compare independent draws from the potential outcomes under treatment and control to keep that causal interpretation intact <ref:2611.14785#pg2>.

Meng: I wonder how this actually helps in a real-world setting, especially when we're thinking about deployment? Are we talking about a simple adjustment, or is this something that adds significant complexity to the pipeline?

Lalam: Lalam here. The paper introduces an estimand that compares those independent draws from the marginal distributions of the potential outcomes under treatment and control <ref:2611.14785#pg2>. They then derive an efficient influence function under conditional independent censoring, which lets censoring depend on baseline covariates or predictable event history <ref:2611.14785#pg1>.

Tom: That sounds like they are trying to build a better way to handle the data structure itself instead of just slapping on some post-hoc adjustment <ref:2611.14785#pg3>. It seems they've got this cross-fitted one-step estimator that incorporates both covariate adjustment and censoring martingale augmentation <ref:2611.14785#pg2>.

Jane: And they claim this estimator allows for flexible estimation of nuisance parameters, which is a big deal because you usually have to fix those parameters in simpler models <ref:2611.14785#pg3>. They also state it attains the semiparametric efficiency bound under suitable conditions on how well those nuisance parameters behave <ref:2611.14785#pg3>.

Meng: Efficiency bound is a strong claim. But what about the practical cost of estimating these nuisance functions? If they have to estimate a whole family of them, does that make it computationally expensive for us when we're shipping models at scale <ref:2611.14785#pg3>?

Lu: They show simulation studies where this estimator has small bias and accurate standard error estimation, which is important because you need reliable confidence intervals to trust the results <ref:2611.14785#pg3>. Plus, it shows a substantial efficiency gain relative to the inverse-probability-of-censoring weighted or IPCW estimator <ref:2611.14785#pg3>.

Paper summary: Tom: Substantial efficiency gain is what caught my attention, especially when compared to IPCW <ref:2611.14785#pg3>. The simulation results show that this new method has a relative efficiency between one point eight four and two point zero six across all scenarios, which means it cuts the empirical variance by about forty-six to fifty-one percent compared to IPCW <ref:2611.14785#pg3>.

Jane: That's a big reduction in variance for an estimation method that already handles censoring and covariates <ref:2611.14785#pg3>. The authors tested this on breast cancer clinical trial data comparing tamoxifen with radiotherapy against tamoxifen alone, where death was the prioritized outcome <ref:2611.14785#pg3>.

Meng: So, for that specific application, they got a win probability of zero point five three zero zero and a win odds of one point one two seven six with a standard error of zero point zero one three three and a ninety-five percent confidence interval between one point zero one five nine and one point two five one seven for the win odds at the five-year horizon <ref:2611.14785#pg3>. That sounds pretty concrete, but what about the limitations they mentioned?

Lu: The paper does flag that their framework handles scalar survival outcomes as a special case, but it's important to note that identifying the parameter psi h is done by assuming sequential independent censoring, which is called coarsening-at-random <ref:2611.14785#pg1>.

Tom: And they also have some flexibility with the estimation of those nuisance functions, but they recommend we look closely at the distribution of those estimated censoring probabilities and weights <ref:2611.14785#pg3>. That's a practical thing for us engineers to consider when we try to deploy this <ref:2611.14785#pg3>.

Jane: So, to wrap up the summary of "Semiparametric Inference for Marginal Win Odds with Censored Time-to-Event Outcomes", the authors propose a new framework that uses a cross-fitted one-step estimator to estimate marginal win probability and odds under prioritized time-to-event outcomes <ref:2611.14785#pg2>. This approach incorporates covariate adjustment and censoring martingale augmentation to improve precision <ref:2611.14785#pg2>.

Meng: So, when we think about the broader meaning of this for medical research or even just complex decision-making, it seems like they are pushing for a way to get more precise comparisons when the outcomes are complicated and censored <ref:2611.14785#pg3>. It moves beyond just using the standard inverse-probability-of-censoring weighting, IPCW <ref:2611.14785#pg3>.

Lu: The implication is that if you're in a situation where you need to compare groups based on a hierarchy of outcomes, this method gives you an estimand that preserves a marginal causal interpretation <ref:2611.14785#pg2>. It’s about getting the most precise comparison possible given the data structure <ref:2611.14785#pg3>.

Tom: I think what this paper really does is show how to get better numbers when you can't just use a simple weighting scheme, but you need that extra layer of complexity from covariate adjustment and martingale augmentation <ref:2611.14785#pg2>. It’s about improving the statistical power for these specific types of win statistics <ref:2611.14785#pg3>.

Paper summary: Jane: And it gives us a way to get those standard errors directly from the estimated influence function without needing to do bootstrap resampling, which is a nice operational advantage <ref:2611.14785#pg3>. So, for anyone listening who isn't reading the math, it’s about creating a more robust statistical tool when you are analyzing survival data with complex priorities <ref:2611.14785#pg2>.

Lu: The paper is really pushing the idea that you can achieve efficiency gains by flexibly estimating nuisance functions within a semiparametric setup <ref:2611.14785#pg3>. It shows how baseline covariates can improve precision without actually changing the marginal win probability or win odds estimand <ref:2611.14785#pg3>.

Meng: From a deployment standpoint, if we could implement this kind of efficient influence function directly into our production system, it would mean better performance on real traffic data where censoring is common <ref:2611.14785#pg3>. It's about making the inference more stable when the assumptions about how the data was censored might not be perfectly met <ref:2611.14785#pg3>.

Tom: So, to summarize for you, "Semiparametric Inference for Marginal Win Odds with Censored Time-to-Event Outcomes" presents a semiparametric framework to estimate marginal win probability and odds under prioritized time-to-event outcomes <ref:2611.14785#pg2>. They claim their cross-fitted one-step estimator is more efficient than IPCW by about forty percent in simulation studies <ref:2611.14785#pg3>.

Jane: It’s about taking a method that compares outcomes based on a rule, like the one described in page two, and making it statistically tighter by incorporating covariate adjustment and augmenting the censoring mechanism <ref:2611.14785#pg2>.

Lu: The core contribution is this estimand comparing independent draws from the marginal distributions under treatment versus control which keeps that marginal causal interpretation intact <ref:2611.14785#pg2>. That's a fundamental property they are preserving <ref:2611.14785#pg3>.

Meng: It seems like the real impact here is showing a way to improve the precision of estimation in randomized trials when baseline covariates are prognostic for the outcome, which is something we always struggle with <ref:2611.14785#pg3>.

Tom: And they provide a concrete example on page three where they show how Martinussen and Vansteelandt developed debiased estimation of causal net benefit using flexible nuisance estimation based on efficient influence functions <ref:2611.14785#pg3>. That’s a direct link to what we do in practice with survival data analysis <ref:2611.14785#pg3>.

Jane: So, the implication is that for researchers dealing with these complex prioritized outcomes, this framework gives them a way to achieve better precision than current methods like IPCW <ref:2611.14785#pg3>. It’s about getting more reliable estimates when the data has censoring and we care about specific event priorities <ref:2611.14785#pg2>.

Conclusion: Tom: So we're wrapping up this look at "Semiparametric Inference for Marginal Win Odds with Censored Time-to-Event Outcomes." Basically, these authors have built a new statistical framework to compare treatment groups when you have complex survival data and specific outcome priorities.

Jane: Right, so they're taking something called win statistics—which are good for comparing treatments if you care about a specific sequence of events—and making sure that the way they adjust for things like patient characteristics and censored data is as smart as possible.

Lu: They've developed this semi-parametric setup to estimate the marginal win probability and those win odds, which keeps that causal interpretation intact by comparing independent potential outcomes.

Meng: So, what does this actually mean when we're deploying models? Is it just another fancy formula for adjusting for covariates, or is there something fundamentally new in how we handle the uncertainty from censoring?

Lalam: It's about getting a more precise comparison than the standard inverse probability of censoring weighted method because they use an efficient influence function that combines covariate projection and a martingale augmentation term.

Tom: That sounds like a big technical lift for the estimation side, but what's the payoff in terms of actual results? Do we get better numbers or just more complexity?

Jane: They show simulation studies where this estimator actually performs well, with small bias and standard errors that are accurate, and it shows a substantial efficiency gain over the IPCW method.

Lu: The numbers show a relative efficiency between one point eight four and two point zero six across all scenarios, which means they cut the empirical variance by about forty-six to fifty-one percent compared to IPCW.

Meng: Forty percent reduction in variance is significant when you're dealing with real traffic data where you don't know exactly how the censoring mechanism is behaving across different patient groups.

Lalam: And they tested this on a breast cancer trial, and they got a win odds of one point one two seven six with a standard error of zero point zero one three three at the five-year horizon.

Tom: So, the authors are showing that if you use this cross-fitted approach, you can get much tighter confidence intervals around your treatment comparisons without needing to do more complex resampling methods like bootstrap.

Jane: Exactly. The authors stress that this framework lets us flexibly estimate nuisance functions, which means we don't have to fix those parameters in a simple model and guess what they are doing.

Lu: It really focuses on the idea that you can get better precision by incorporating information about baseline covariates into the estimation process itself without changing what you're trying to estimate, which is a key theoretical point.

Meng: I just hope this flexibility doesn't introduce new failure modes when we try to scale it up. We need to know exactly how sensitive the results are if our model assumptions on those nuisance functions are slightly off.

Tom: That’s the practical worry, Meng. But from what I'm seeing, they provide a concrete way to get better numbers in these tricky survival scenarios compared to what we currently have available.

Jane: So, if you're working with prioritized outcomes and censored data, this paper gives you a new tool for inference that is much more robust than just the basic weighting schemes we use now.

Zhiguo Li

Duke University

stat.ME

Submitted: 2026-11-10

Updated: 2026-11-13

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

Importance score: 92/100

The gist: Win statistics provide clinically interpretable measures for comparing treatment groups with prioritized composite outcomes…

Key concepts

Marginal Win Probability
This is the probability of winning a comparison between two treatment groups, calculated by considering the independent marginal distributions of potential outcomes under each treatment scenario. It's used to compare treatments directly based on these underlying distributions.
Win Odds
Win odds are a measure derived from the marginal win probability, expressed as the ratio of winning probability to losing probability. This metric is preferred because it offers a direct probabilistic-index interpretation, making causal comparisons between treatment arms more intuitive.
Efficient Influence Function
This mathematical tool describes how the estimator reacts to changes in the data. The paper derives an efficient influence function that combines components from covariate adjustment, inverse-probability-of-censoring weighting, and a martingale augmentation to ensure the estimator achieves its theoretical maximum efficiency.
Semi-competing Risks
This advanced outcome setting involves two distinct events where one event precludes the other from occurring. The framework is designed to handle this structure while still providing accurate estimates for the win odds, which is crucial for complex clinical trial data.

Terminology

Summary

Win statistics provide clinically interpretable measures for comparing treatment groups with prioritized composite outcomes, but efficient covariate-adjustment methods for win statistics with right censored time-to-event outcomes remain limited We develop a semiparametric framework for marginal win probability and win odds with prioritized time-to-event outcomes, including semi-competing risks, while also accommodating scalar survival outcomes as a special case. p. 0 The estimand compares independent draws from the marginal distributions of the potential outcomes under treatment and control, which preserves a marginal causal interpretation. p. 0 The resulting cross-fitted one-step estimator incorporates covariate adjustment and censoring martingale augmentation, allows flexible estimation of nuisance parameters, and attains the semiparametric efficiency bound under suitable conditions on consistency and convergence rate of nuisance parameters. p. 0 Simulation studies show that the estimator has small bias, accurate standard error estimation and confidence interval coverage, and substantial efficiency gain relative to the inverse-probability-of-censoring weighted (IPCW) estimator. p. 0

Framework Development

The framework is developed considering two outcome settings: a scalar right-censored time-to-event outcome and a two-component prioritized semi-competing risk outcome. For the two-component setting, the restricted potential outcomes are defined as Yτ(a) = Tτ1(a), Tτ2(a), where Tτ2(a) reflects the semi-competing risk structure. p. 3 The parameter of interest is the win odds, defined as θh(τ) = ψh(τ)/ (1 − ψh(τ)). p. 4

Identification of Estimands

The marginal win estimand is defined through restricted potential outcomes independently of the censoring mechanism. The parameter ψh(τ) is identified by an inverse-probability-of-censoring identity, which requires assuming sequential independent censoring (coarsening-at-random). p. 6 This identity identifies the marginal distributions Fτa of Yτ(a), and hence ψh(τ) = Z Z h(y1, y0) dFτ1(y1) dFτ0(y0). p. 7

Efficient Influence Function

The observed-data efficient influence function for the marginal win probability is characterized by Theorem 1, which involves a term combining an inverse-probability-of-censoring weighted component, a baseline covariate projection, and a censoring martingale augmentation. The efficient influence function for ψ is given by ϕobs(O; ψ, η) = µ1(X) + µ0(X) − 2ψ + A π(X) ϕ1(O; η) + 1 − A 1 − π(X) ϕ0(O; η). p. 8 This function decomposes into a covariate projection term, an inverse-probability-of-censoring weighted complete-outcome contribution, and a censoring-martingale augmentation term. p. 9

Estimation Procedure

A cross-fitted one-step estimator is constructed by combining the fold-specific plug-in win probability from (6) with an empirical correction based on the efficient influence function in Theorem 1. This estimator allows flexible estimation of nuisance functions and attains the semiparametric efficiency bound under suitable conditions. p. 0 The resulting asymptotic normality is established by Theorem 2, showing that the cross-fitted one-step estimator is asymptotically linear. p. 9

Finite Sample Performance

Simulation studies demonstrate that the proposed estimator has small empirical bias, and its standard errors based on the influence function are close to the empirical standard deviations. For prioritized semi-competing risks, the proposed estimator had relative efficiency between 1. p. 284 and 2.06 across all scenarios, representing a reduction of approximately 46%–51% in empirical variance compared with IPCW. p. 0

Real Data Application

The method was applied to breast cancer clinical trial data comparing tamoxifen combined with radiotherapy against tamoxifen alone, where death was prioritized over disease progression or other nonterminal events. The proposed one-step estimator gave a win probability of 0. p. 95300, which gave a win odds of 1. p. 01276 with a standard error of 0. p. 00133 and a 95% confidence interval of (1. p. 00159, 1.2517) for the win odds at the five-year horizon. p. 0 The estimated variance ratio comparing IPCW with the one-step estimator was 1. p. 034, corresponding to an approximately 25% reduction in estimated variance. p. 0

Discussion

The proposed framework focuses on the marginal win probability and the corresponding win odds, which is motivated by the causal interpretation of the estimand. The marginal comparison between independent draws from the distributions of the two treatment arms is identifiable from the randomized trial and has a direct probabilistic-index interpretation. p. 0 The proposed estimator allows flexible estimation of nuisance functions, and its standard errors can be obtained directly from the estimated influence function without bootstrap resampling. p. 9 The authors recommend examining the distribution of estimated censoring probabilities and weights and assessing the sensitivity of the results to the specification of the models for the nuisance functions. p. 9 Code and Data Availability. p. 0 The R package winoddsEIF, implementing the proposed method, together with code for reproducing the simulation studies and real-data analysis, is publicly available at https://github. p. 3com/zhiguoli919/winoddsEIF. The breast cancer trial data used in the application are available in the R package collett. 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. Improve estimation of marginal win probability by using a cross-fitted one-step estimator that incorporates both baseline covariate adjustment and a censoring martingale augmentation, which allows flexible estimation of nuisance functions.

  2. Enable direct variance estimation based on the influence function, as stated in Section 4, allowing researchers to obtain standard errors without bootstrap resampling.

  3. Enhance precision for prioritized outcomes by accounting for dependence between event processes and censoring that may depend on the observed event history, which is achieved through a censoring martingale augmentation involving predictable event history.

  4. Reduce the empirical variance of win statistics by incorporating baseline prognostic information and observed event history, as demonstrated by the finding that the proposed estimator was consistently more efficient than IPCW.

  5. Improve inference for win odds by using log-scale construction, which yields a positive confidence interval for theta, and provide standard deviations estimated via the sample variance of cross-fitted influence-function contributions.

Sources

Related papers