PS06: Differential Censoring and Fragility Analysis for Impactful Decision Making in Survival Trials

Time-to-event endpoints play a central role in clinical trials, regulatory review, and health technology assessment. As trial conduct increasingly incorporates digital tools, remote follow-up, adaptive designs, and hybrid randomized–real-world evidence frameworks, censoring mechanisms have become more complex and more heterogeneous across treatment arms. A key challenge in this evolving field is differential censoring, namely systematic imbalance in censoring patterns between treatment groups. Differential censoring poses a serious but often underrecognized threat to valid survival inference and downstream decision-making. When left unexamined, it can distort treatment effect estimates, obscure assessments of data maturity, and undermine confidence in conclusions drawn from both primary and secondary endpoints. Beyond addressing threats to validity arising from informative censoring, this session also turns to a broader question: how robust are survival conclusions to plausible deviations in the observed data and modeling assumptions? Traditional reliance on p-values offers limited insight into the stability of trial conclusions under plausible perturbations of the data or deviations from underlying assumptions. The Fragility Index has gained traction as an intuitive robustness metric, and as noted by VanderWeele (2026 AJE) in a recent commentary, fragility can be reviewed as a form of sensitivity or bias analysis targeting specific data perturbations, such as misclassification. However, extensions of the fragility analysis to time-to-event endpoints remain limited, largely due to informative censoring, heterogeneous follow-up, and reliance on treatment-reassignment-based methodologies, which reduce clinical plausibility and hinder interpretability. The first speaker, YongGang Yao, Director of Biostatistics at Rapport Therapeutics, will present "Treating Censoring as Signal: Integrating Efficacy, Tolerability, and Treatment Duration in Clinical Trials." Dr. Yao will show how differential censoring, measured using the reverse hazard ratio (RevHR; Hsu et al., 2024, JNCI), can provide insight into treatment tolerability and discontinuation. He will connect efficacy, censoring, and time to treatment failure through a competing-risks framework and use a simulated trial to illustrate how RevHR diagnostics and causal methods for intercurrent events can provide a more complete interpretation of treatment benefit than efficacy analyses alone. To advance impactful and assumption-aware decision-making, the second and third speakers, Professor Jessie Tong (Johns Hopkins Bloomberg School of Public Health) and Professor Lifeng Lin (University of Arizona), will introduce two complementary methodological innovations. First, Professor Tong (Xing et al 2026 AJE) will present a Fragility Index for Survival outcomes (FIS), which preserves randomization by modifying outcome status, events or censoring, rather than treatment assignment, and quantifies the minimum number of outcome changes required to alter statistical significance in either direction. Second, Professor Lin will embed FIS within a structured tipping-point framework, SURvival Fragility Index Tipping-point (SURFIT) analysis, which systematically evaluates how alternative censoring assumptions shift trial conclusions and identifies explicit thresholds at which statistical significance is lost or the direction of treatment effects reverses. Collectively, the three speakers show how censoring diagnostics, fragility measures, and tipping-point analysis can be integrated into a unified and decision-focused robustness framework. By reframing statistical sensitivity or bias analysis as transparent, clinically interpretable decision thresholds, the session embodies the 2026 RISW theme of impactful transdisciplinary decision-making and equips statisticians, trialists, regulators, and health technology assessment stakeholders with practical tools for navigating survival evidence in an increasingly digital evidence landscape.