Dr. Jennifer Rider was the second presenter in this session. She discussed clinically interpretable estimates, moving beyond median survival and hazard ratios. She began with a, self-admittedly cynical, description of the traditional approach to clinical trial design and reporting.
While median survival and hazard ratios are the most commonly reported effect measures in clinical research, Dr. Rider began highlighting the limitations of using median survival. First, in studies with limited follow-up or relatively low event rates, it is not possible to estimate median survival. Second, medians are relatively insensitive to long-term survivors. Additionally, comparisons of median survival have lower statistical precision than the hazard ratio. Thus, there are differences interpreting group differences. However, there are also limitations to the use of the hazard ratio. First, in the absence of proportional hazards, a single hazard ratio is an uninterpretable measure. Despite this, proportionality is often not assessed in clinical trials reporting and, where violations seem obvious based on survival curves, a single cumulative hazard ratio is often reported. Even further, it is often difficult to determine the validity of the use of Cox models using standard lack-of-fit tests.
In place of, or complimenting, these measures, Dr. Rider suggested the use of restricted mean survival time (RMST). While this is not a new effect measure, it is relatively uncommonly used in the clinical literature. The RMST is defined as the area under the survival curve at some landmark time. She touted benefits of the RMST including an intuitive, clinically meaningful interpretation, the ability to assess either relative (ratio) or absolute (difference) effect measures, and the ability to evaluate changes over time.
She then highlighted data from one of her colleagues, Dr. Trinquart, who compared the hazard ratio and RMST difference in Phase III clinical trials in oncology. As highlighted in the figure below, there was an agreement in the direction of effect between hazard ratio and RMST difference in the vast majority of cases. However, the hazard ratio very often indicated a larger treatment effect than the RMST ratio. In some cases, the effect was more than two-fold larger. These data suggest that we should perhaps consider the hazard ratio an optimistic estimate of treatment effect.
Dr. Rider then highlighted further data demonstrating that hazard ratios are often misinterpreted by clinicians, often resulting in an overestimation of treatment benefit.
She then used a case example of the SPCG-4 trials showing the use of RMST in this dataset. As highlighted in the figures below, while the direction of effect was the same between the hazard ratio and RMST difference, the RMST difference may offer a more clinically meaningful estimate to facilitate patient discussion.
Presented by: Jennifer Rider, ScD, MPH, Assistant Professor, Department of Epidemiology, Boston School of Public Health, Boston, Massachusetts
Written by: Christopher J.D. Wallis, MD, PhD, FRCSC, Twitter: @WallisCJD at the 20th Annual Meeting of the Society of Urologic Oncology (SUO), December 4 - 6, 2019, Washington, DC