Model-Based Imputation Reduces Bias When Estimating Win Statistics in Survival Data

A simulation study has demonstrated that a model-based imputation method effectively reduces bias when estimating win statistics in survival analysis with non-informative censoring. The approach, evaluated across multiple prioritized composite outcomes, showed less estimation bias than both unadjusted analysis and inverse probability of censoring weighting while maintaining empirical variance comparable to weighting.

Researchers examined survival data involving a binary treatment variable and prioritized composite outcomes, where censoring occurs independently of the distribution of survival times. The objective centered on estimating key win statistics—specifically the win ratio, win odds, and net benefit—under conditions where non-informative censoring otherwise creates missing data and introduces bias into treatment evaluations.

To address this challenge, the proposed model-based imputation framework imputes the survival times of censored individuals using a conditional survival distribution derived from the Kaplan-Meier estimator. By directly reconstructing survival times through this estimator under non-informative censoring, the procedure adjusts for missing information across two or more prioritized composite endpoints.

In simulation evaluations varying both sample sizes and censoring rates, the model-based imputation strategy produced consistent results across diverse conditions. The investigators concluded that by mitigating censoring bias without inflating variability relative to inverse probability of censoring weighting, the imputation method provides a useful approach for calculating win statistics in prioritized composite survival settings.

The mathematical focus of the investigation targets three specific measures: the win ratio, the win odds, and the net benefit. In survival data involving a binary treatment variable and prioritized composite outcomes, non-informative censoring creates missing data that can distort the estimation of these win statistics. Under this condition, non-informative censoring operates independently of the distribution of survival times. When two or more prioritized composite endpoints are evaluated, censoring prevents the direct observation of survival times for censored individuals, introducing censoring bias into unadjusted comparative assessments between the treatment groups.

To overcome this missing data problem, the proposed model-based imputation methodology relies on a conditional survival distribution derived from the Kaplan-Meier estimator. Under non-informative censoring, this approach uses the Kaplan-Meier estimator to characterize the conditional survival distribution and directly impute the survival times of censored individuals. By replacing missing event times with values imputed from this conditional survival distribution, the framework adjusts for censoring bias across two or more prioritized composite outcomes, enabling the calculation of win statistics across prioritized endpoints.

The performance of the model-based imputation method was examined in a simulation study that compared the strategy against both the unadjusted method and the inverse probability of censoring weighting method. The simulation results showed that the proposed model-based imputation method achieved less bias than the unadjusted method and also demonstrated less bias than inverse probability of censoring weighting. In addition to reducing bias caused by missing data, the proposed method maintained empirical variance comparable to inverse probability of censoring weighting. Furthermore, the simulation study demonstrated consistent results across varying sample sizes and varying censoring rates, confirming that model-based imputation methods effectively adjust for censoring bias while maintaining comparable variability when estimating win statistics.

Subject of Research: Model-based imputation for estimating win statistics in survival analysis

Article Title: Model-based imputation methods for estimating win statistics in survival data with non-informative censoring

Article References: Jeong, Y., & An, H. (2026). Model-based imputation methods for estimating win statistics in survival data with non-informative censoring. BMC Medical Research Methodology. https://doi.org/10.1186/s12874-026-03085-8

Image Credits: AI Generated

DOI: 10.1186/s12874-026-03085-8

Keywords: win statistics, win ratio, win odds, net benefit, model-based imputation, survival data, non-informative censoring, prioritized composite outcomes, prioritized composite endpoints, Kaplan-Meier estimator, conditional survival distribution, inverse probability of censoring weighting, censoring bias, empirical variance, simulation study, binary treatment variable, survival time imputation