Skip to main content


The feasibility and reliability of using restricted mean survival time in aggregate data meta-analysis of time-to-event outcomes

Date and Location




Sunday 22 September 2013 - 10:30 - 12:00


Presenting author and contact person

Presenting author

Yinghui Wei

Contact person

Yinghui Wei
Abstract text
Background: Meta-analyses of time-to-event trial outcomes commonly use the hazard ratio (HR) as the treatment effect measure. For aggregate data reviews, this relies on extracting or estimating HR from published analyses. Use of the HR implicitly assumes proportional hazards (PH), which may be violated for some or all trials in a meta-analysis. An alternative treatment effect measure is the between-arm difference in the restricted mean survival time (RMST). For a given arm, the RMST is the expected time-to-event up to t* and may be estimated as the integrated survival function S(t), RMST = integral_0^ t* (S(t)) dt , where t* must be chosen by the analyst. The RMST relaxes the need for PH assumption and allows treatment effect to vary with time. However, the RMST analysis relies on having IPD or reconstructing survival data from published curves. Objectives & Methods: We aim to assess the feasibility and reliability of estimating the restricted mean survival time for aggregate data meta-analysis of randomized trials. We compare the non-PH test results and estimates of the RMST differences from the individual participant data (IPD), survival curves reconstructed from this IPD, and from published survival curves. Results: Based on a meta-analysis of neo-adjuvant chemotherapy for invasive bladder cancer (6 trials), testing of the non-PH assumption, the estimated HR and difference in RMST are similar for IPD and curves reconstructed from it. However, the ability to use published survival curves to assess non-PH and estimation of the RMST was limited to 3 trials. Conclusions: Reconstructed survival curves enable non-PH testing and good approximations of the difference in RMST, offering an alternative to the HR for meta-analysis of time-to-event outcomes. Better approximations could be achieved if good quality Kaplan Meier curves, including the number of participants at risks were regularly available.