Recent advance in design and analysis in clinical trials when the proportional hazard assumption is impossible to hold

Restricted mean survival time (RMST)-based analysis is becoming more popular as an alternative to the conventional logrank-hazard ratio approach when comparing the survival benefits of two interventions in randomized clinical trials. When the logrank test is the primary analysis, an event-driven design is used to achieve the target power to detect a specified treatment effect measured by the hazard ratio. However, this approach is not applicable when employing the RMST-based test because the power does not depend solely on the number of observed events and the between-group difference measured by RMST but by parameters such as event and censoring time distributions. If these parameters are not correctly specified in calculating the required sample size in the design stage, the study may not have the targeted power to detect a specified difference in RMST. To avoid this, we propose a blind sample size re-estimation method for the RMST-based test. We consider the survival function estimated with the pooled sample in a blind review as constraints for the survival functions of the two groups. Under the constraints, the maximum variance for the RMST difference and the corresponding sample size are calculated using non-linear programming. The resulting sample size will represent the worst-case scenario, or the upper bound of the required sample size needed to achieve the target power to detect the specified RMST difference. Our comprehensive numerical studies indicated that the proposed method effectively circumvents the potential underestimation of the required sample size due to the misspecification of survival functions in the design stage. 

In randomized clinical trials, survival analyses without risk stratification, or with stratification based on pre-selected factors revealed at the end of the trial to be at most weakly associated with event risk, are surprisingly common. We show that such analyses can unwittingly induce non-proportional hazard conditions and potentially deliver hazard ratio estimates that dilute the evidence of benefit for the test relative to the control treatment. To safeguard against this, we draw attention to 5-STAR, a novel methodology in which a treatment-blinded algorithm is applied to the survival times from the trial to partition patients into risk strata based on covariates observed to be jointly prognostic for event risk. A treatment comparison is subsequently done within each identified risk stratum and stratum-level results are averaged for overall inference. We illustrate the utility of 5-STAR using a reanalysis of data for the primary and key secondary endpoints from three published cardiovascular outcomes trials. 

Effectiveness of immune-oncology chemotherapies has been presented in recent clinical trials. In immune-oncology chemotherapies, it has often been suggested the presence of the lag-time until the immune therapy began to act. It implies the use of hazard ratio under the proportional hazards assumption would not be appealing, and many alternatives have been investigated such as the restricted mean survival time. In addition to such overall summary of the treatment contrast, the lag-time is also an important feature of the treatment effect. Identical survival functions up to the lag-time implies patients who are likely to die before the lag-time would not benefit the treatment and identifying such patients would be very important. We propose the semiparametric piecewise accelerated failure time model and its inference procedure based on the semiparametric maximum likelihood method. It provides not only an overall treatment summary, but also a framework to identify patients who have less benefit from the immune-therapy in a unified way. 

The limitations of the traditional log-rank/hazard-ratio test/estimation approach for comparing time-to-event outcomes between groups have been widely discussed. While theories supports that this traditional approach offers the most powerful test when the proportional hazards assumption is correct, it loses this advantage when the assumption is incorrect. Also, interpreting hazard ratios under non-proportional hazards conditions becomes challenging, as they deviate from their foundational definition within this assumption. Methods to derive hazard ratios under non-proportional hazards conditions often involve weighted averages of time-specific hazard ratios, where weights may lack intuitive meaning or depend on study-specific censoring time distributions. To address these limitations, alternative approaches have been proposed. One such approach is using average hazard with survival weight (AHSW), which offers a robust alternative recently proposed. However, the use of AHSW with group sequential designs remain underexplored. This talk examines the implementation of AHSW-based analysis in group sequential designs, focusing on methodology and practical considerations.
 

The rapid advancement of cancer immunotherapy heralds an exciting new era in cancer treatment, but also presents unique challenges in the design and analysis of clinical trials. One such challenge is the frequent discrepancy between progression-free survival (PFS) and overall survival (OS) findings. While OS remains the gold standard primary endpoint for regulatory approval in oncology trials, serving as a clinically meaningful objective measure of both safety and efficacy, PFS is often used as the primary endpoint to accelerate drug approvals for with life-threatening malignancy. The rationale is that PFS can be assessed earlier than OS, and its effects are presumed to predict future OS outcomes. However, recent trials have highlighted a disconnect between PFS and OS findings. In some instances, PFS improvements have even been associated with potential OS detriment, often manifesting as complex non-proportional hazards (NPH) patterns.

In this talk, we explore the underlying causes of these conflicting findings, particularly the NPH patterns observed in OS data, and develop a proper strategy to detect and estimate the true treatment effect in the presence of such discrepancies.