PS10: Strategies to Assess the Impact of Informative Missingness in Regulatory Submissions for Time-to-Event Trials

Missing data is a key challenge in clinical trials. While often viewed as an operational nuisance, its impact is more far reaching, often impacting the statistical validity of the trial's results. The ability to demonstrate the robustness of trial results in the face of such incomplete data, whether this is due to handling of intercurrent events or due to reasons other than intercurrent events, is fundamental to maintaining the integrity of a regulatory submission and building confidence with agencies. Since the introduction of the ICH E9 R1 addendum, there is an increasing recognition that the assumptions of missing data in clinical trials needs to be addressed in a systematic fashion, and that the trial conclusions are robust to various missing data assumptions (ICH E9 R1 addendum on estimands and sensitivity analyses in clinical trials [1] National Research Council, Panel on Handling Missing Data in Clinical Trials [2], Torres et al [7], and Permutt et al [3]). Any statistical analysis will implicitly or explicitly make unverifiable assumptions about missing data, and hence accompanying sensitivity analyses are used to address the robustness of these assumptions. However, sensitivity analyses proposed generally address a specific aspect of the missing data assumptions. Thus, regulatory agencies have recommended sponsors to conduct at least a two-dimensional tipping point analysis to more extensively explore the assumptions about the treatment and standard of care. Over the past years, significant progress has been made in developing statistically rigorous methods to implement clinically interpretable sensitivity analyses to address the missing data assumptions in clinical trials for continuous and (to a lesser extent) for binary or categorical endpoints. Studies with time-to-event outcomes have received less attention. In time to event analyses, typical statistical model such as Kaplan Meier or Cox proportional hazards handles missing follow-up in the form of censoring, with assumption that reason for not observing the event of interest is unrelated to their instantaneous risk of experiencing an event, i.e., non-informative. However, when patients are naïvely censored at the time of intercurrent events, in many situations, the underlying assumption of these methods are questionable. This can lead to biased comparisons particularly when such censoring mechanisms are informative. In these studies, challenges with respect to the robustness and integrity of primary analysis conclusions arise when there is any incomplete follow-up of the participants' clinical event of interest. In oncology trials, informative missingness occur when patients who switch to other anti-cancer therapies prior to disease progression (PFS) are censored at time of switching which then may mask any underlying deterioration of the disease or other comorbidities. Similarly, when patients are censored due to study discontinuation with reasons due to local investigator progression but prior to it being confirmed by blinded central reviewer, a naïve censoring of the primary event based on central PFS can be informative. In renal outcome trials, when a biomarker is part of the composite endpoint (e.g., sustained 50% decline in estimated glomerular filtration rate from baseline with confirmation of at least X months), typical strategies using censoring of the primary endpoint when a biomarker assessment is missed can introduce missing follow-up. In cardiovascular outcome trials, when the endpoint is a composite of cardiovascular deaths, or heart failure, censoring at time of withdrawal from study may not be reasonable. Over the years, sensitivity analyses to missing data assumptions in survival endpoints have played a central role in the regulatory decision making in public FDA advisory meetings. During the FDA Cardiovascular and Renal Drugs Advisory Committee meeting (CRDAC) in 2012, the committee voted against recommending the approval of rivaroxaban proposed for patients with acute coronary syndrome (ACS) due to missing data concerns observed in ATLAS ACS 2-TIMI Trial. Specifically, tipping point analyses conducted by the FDA reviewers revealed that a deviation from ignorable censoring resulted in the reversal of significance of the treatment effect. More recently, in the FDA Oncology Advisory Committee (ODAC) in 2024, the committee voted 10 to 2 against approval of sotorasib mainly because progression free survival (PFS) benefit of sotorasib vs docetaxel in second-line treatment for non-small cell lung cancer that has a KRAS gene mutation (2L KRASm NSCLC) cannot be reliably interpreted due to the asymmetric early dropouts among the patients in the study. There were concerns on informative censoring that affected the interpretability of positive trial findings. Recent work to assess the impact of such informative missingness include methods that extend the reference-based sensitivity analyses methods, originally designed for continuous outcomes, to survival settings as well as delta adjustments to assess the impact on departures from non-informative censoring by Atkinson et al. [4]. Other authors such as He et al 2023 discussed targeting a treatment policy like estimand to handle missing follow-up as a potential starting point for delta adjustment. There have also been frameworks proposed to conducting tipping point analyses, using model based approaches by Lipkovich et al [5] and model free methods by Oodally et al [6], which assess how extreme the departures are from the primary assumptions to tip results of the primary findings, and whether such departures are clinically plausible. While an increasing use of such sensitivity analyses and tipping point analyses have been noted in the recent regulatory submissions, there is still a lack of a clear and consistent framework of how such sensitivity analyses should be constructed, e.g., handling of intercurrent events in time to event endpoints, or meaningful hypothetical scenarios of interest. Often the scenarios considered relevant by the sponsor for such analyses may not align with what the FDA considers most appropriate to evaluate, e.g., the starting point for the tipping point analysis. This session is designed to bring together the statisticians from Industry and FDA to hopefully align on strategies to handle intercurrent events in the time to event studies, share their perspectives on the role of potential sensitivity analyses, tipping point analysis, and how they can be interpreted in the context of the clinical plausibility of the scenarios being considered. Presentations will include case studies outlining how such sensitivity analyses are developed, together with key insights on how to interpret the results in the context of the clinical plausibility of the assumptions of missing follow-up of time to event outcomes. The session will also include presenter(s)/discussants from FDA who will share insights on the impact of missing data on the interpretability of the regulatory submissions in time to event outcomes, potential ways to assess the robustness of the results when there are departures from the underlying assumptions. References 1. ICH E9 (R1) addendum on estimands and sensitivity analysis in clinical trials to the guideline on statistical principles for clinical trials. 05 August 2021 2. National Research Council, Panel on Handling Missing Data in Clinical Trials, Committee on National Statistics, Division of Behavioral and Social Sciences and Education. 2010. The prevention and treatment of missing data in clinical trials. Washington (DC): National Academies Press 3. Permutt, T. 2016. Sensitivity analysis for missing data in regulatory submissions. Statistics in Medicine 35 (17):2876–2879 4. Atkinson A, Kenward MG, Clayton T, Carpenter JR. Reference-based sensitivity analysis for time-to-event data. Pharmaceutical Statistics. 2019; 18: 645–658. 5. Lipkovich, I., Ratitch, B., and O'Kelly, M. (2016) Sensitivity to censored-at-random assumption in the analysis of time-to-event endpoints. Pharmaceut. Statist., 15: 216–229 6. Oodally A, Wang C, Li Z, Chakravartty A. Tipping Point Sensitivity Analysis for Missing Data in Time-to-Event Endpoints: Model-Based and Model-Free Approaches. Arxiv:2506.19988 7. Torres, C., Levin, G., Rubin, D., Koh, W., Chiu, R. and Permutt, T. (2025), A Tipping Point Method to Evaluate Sensitivity to Potential Violations in Missing Data Assumptions. Pharmaceutical Statistics, 24: e70002 8. He J, Crackel R, Koh W, et al. Retrieved-Dropout-Based multiple imputation for time-to-event data in cardiovascular outcome trials. J Biopharm Stat. 2023;33(2):234-252. ttps://doi.org/10.1080/10543406.2022.2118763