Background: Time-to-event outcomes are commonly used in survival analyses to describe the duration of time until a given event (e.g. death) and are usually measured as hazard ratios (HR). The GRADE approach to calculating absolute effects is to establish a baseline (event rate at a particular point in time) in the control arm and then apply the HR to calculate the event rate in the intervention arm. The challenge arises around the uncertainty of what the event is that the outcome describes and how absolute-effect size is interpreted.
Objectives: To assess how time-to-event outcomes are presented in Summary of Findings (SOF) tables.
Methods: Based on an a priori protocol we systematically identified all Cochrane Cancer Reviews that reported at least one outcome measured as HR and provided a SOF table (published 2011-2016). Six authors performed all steps in duplicate and disagreements were solved by discussion. We extracted data regarding the calculation of absolute effects, consistency between outcomes in abstract, methods, results and SOF table, and assessment of censoring.
Results: 77 reviews met our inclusion criteria. In 21 (27%) no absolute effect for HR outcomes was calculated. In 14 (18%) absolute effects in SOF tables were correctly calculated and labelled and no confusion occurred between positive (people alive) and negative (deaths) events throughout the review. 12 reviews (16%) provided wrong results by entering positive event-control risk into GRADE software, leading to less instead of more people alive in the favoured arm. In 22 (29%) reviews absolute effects were correctly calculated, but confusing, as there is no link between outcomes in the review (e.g. survival) and outcomes in SOF (e.g. mortality, negative event). For eight (10%) reviews it is completely unclear how authors assumed control risk and whether results are correct. Only 5 reviews reported censoring in survival curves and discussed potential impact.
Conclusions:There is an urgent need for author guidance on how to calculate absolute effects based on HR data and how to present data. Moreover, censoring in individual trials should be taken into account.