Restricted mean survival time: an alternative to the hazard ratio for the design and analysis of randomized trials with a time-to-event outcome.
Royston P, Parmar MK. · 2013
Read Restricted mean survival time: an alternative to the hazard ratio for the design and analysis of randomized trials with a time-to-event outcome. by Royston P, Parmar MK. (2013) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc3922847.txt.
Introduction
Patrick Royston and Mahesh K. B. Parmar propose a different way to measure treatment benefit in trials where the outcome is time until an event, in BMC Medical Research Methodology in 2013. Such trials usually report a hazard ratio, which quietly assumes that the relative risk between treatment groups stays constant over time. Recent trials show that assumption cannot be trusted.
Their alternative is the restricted mean survival time: the average time lived up to a chosen horizon, read as the area under the survival curve. The paper shows how to design trials and size samples around this measure, using flexible survival curves that allow the treatment effect to grow or fade. Worked examples show the approach stays readable for clinicians while avoiding the constant-risk assumption.
Catalog record
| CREATOR | Royston P, Parmar MK. |
|---|---|
| YEAR | 2013 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 58 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Restricted mean survival time: an alternative to the hazard ratio for the design and analysis of randomized trials with a time-to-event outcome By Patrick Royston, Mahesh KB Parmar Published in: BMC Medical Research Methodology DOI: 10.1186/1471-2288-13-152 ABSTRACT Background Designs and analyses of clinical trials with a time-to-event outcome almost invariably rely on the hazard ratio to estimate the treatment effect and implicitly, therefore, on the proportional hazards assumption. However, the results of some recent trials indicate that there is no guarantee that the assumption will hold. Here, we describe the use of the restricted mean survival time as a possible alternative tool in the design and analysis of these trials. Methods The restricted mean is a measure of average survival from time 0 to a specified time point, and may be estimated as the area under the survival curve up to that point. We consider the design of such trials according to a wide range of possible survival distributions in the control and research arm(s). The distributions are conveniently defined as piecewise exponential distributions and can be specified through piecewise constant hazards and time-fixed or time-dependent hazard ratios. Such designs can embody proportional or non-proportional hazards of the treatment effect. Results We demonstrate the use of restricted mean survival time and a test of the difference in restricted means as an alternative measure of treatment effect. We support the approach through the results of simulation studies and in real examples from several cancer trials. We illustrate the required sample size under proportional and non-proportional hazards, also the significance level and power of the proposed test. Values are compared with those from the standard approach which utilizes the logrank test. Conclusions We conclude that the hazard ratio cannot be recommended as a general measure of the treatment effect in a randomized controlled trial, nor is it always appropriate when designing a trial. Restricted mean survival time may provide a practical way forward and deserves greater attention. BACKGROUND Most randomized controlled trials (RCTs) with a time-to-event outcome are designed and analyzed with a target hazard ratio (HR) for the treatment effect in mind. By convention, the HR is usually taken as the hazard function in the research arm divided by that in the control arm, with values < 1 representing a ‘positive’ treatment effect. In advanced cancers with a mortality outcome, for example, a popular choice of target HR is 0.75. This implies a reduction of 25 percent in the instantaneous mortality rate at all times after randomization. According to a standard sample size calculation based on the logrank test, about 510 events are needed to attain power 90 percent to detect such a treatment effect at a two-sided significance level of 5 percent in a trial with equal allocation to control and research arms. For a single HR to make scientific sense, we must assume that proportional hazards (PH) of the treatment effect holds, at least approximately. We have argued previously [1] that when the PH assumption fails, it is misleading to report the treatment effect through the estimated HR, since it depends on follow-up time. A simple example of departure from PH occurs when one group is assigned to immediate surgical treatment and the other to medical treatment. Suppose time of randomisation is the origin of the survival time. When surgery increases short-term mortality but confers long-term benefit on the survivors – a reasonable alternative hypothesis – PH does not apply and the HR is a misleading and inappropriate summary. More technically, we are unconvinced by papers such as Schemper et al [2] where an overall estimate of the HR is regarded as an average of time-dependent HRs over the event times, nor by proposed variants based on different and arbitrary weighting schemes. The main issue is that an average HR is uninterpretable. Under PH, for example, the HR can usefully be applied to the survival function in the control arm to obtain an impression of the survival curve in the research arm. When PH is breached, this property no longer holds. Furthermore, the HR depends on the follow-up time. It has become apparent in some recently reported trials, e.g. IPASS [3] and ICON7 [4], that gross breaches of the PH assumption can and do occur—even to the extent of observing crossing survival curves, where a local estimate of the log HR changes sign over time. Non-PH may be due to different biological modes of action of the treatments being compared, or as identified in IPASS, to the presence of differentially responsive sub-populations. As noted before [1], we are dissatisfied with the HR as a universal summary measure. For example, even when PH holds, the HR is not as meaningful clinically as some type of difference in average survival times or proportions at a fixed time-point, obscuring the absolute difference between the treatments and failing to convey the clinical value of a treatment. (By ‘survival time’ we mean generic time to event, for whatever event is of interest.) Furthermore, early stopping rules that assume PH can generate inappropriate decisions if the HR later changes substantially. Also, no single summary of HR or risk difference can adequately describe cases in which the treatment effect changes in direction as follow-up increases. In our earlier paper [1], we suggested an approach to the analysis of an RCT in which the PH assumption is breached. We proposed to estimate and report the restricted mean survival time (RMST) [5], expressing the treatment effect as the difference in RMST between the randomized arms at a suitable follow-up time, t∗. We constructed confidence intervals through the standard error of the difference in RMST. Further experience with the RMST measure in a larger number of trials has given us the impression that when the PH assumption is approximately satisfied, the test of the null hypothesis based on RMST difference often has operating characteristics similar to the logrank test. Specifically, the significance level and power of the two tests appear to be similar. An advantage of the RMST is that it is valid under any distribution of the time to event in the treatment groups, of which PH models are a (small) sub-class. Furthermore, it is readily interpretable as the ‘life expectancy’ between randomization (t = 0) and a particular time horizon (t = t∗). Owing to the current dominance of the HR and its presumed time independence, trial reports often ignore the possibility of non-PH and typically place little emphasis on the extent of follow-up, which should be a key aspect of the trial design and analysis. For example, a treatment effect can exhibit PH in the short term but non-PH over a longer period (e.g. the GOG111 trial, see Reference [1]). It is particularly important to ensure sufficient follow-up when there may be good biological or other reasons to expect the effect of a treatment to vary over time. The primary estimate of the RMST is specifically aligned to a chosen t∗ and this must be made explicit. Obviously, though, as part of the analysis, the treatment effect can be explored over a range of alternative t∗ values. In a previous report [6], we described the implementation of a general method, Assessment of Resources for Trials or