Assessment of proportional hazard assumption in aggregate data: a systematic review on statistical methodology in clinical trials using time-to-event endpoint

Background: The evaluation of the proportional hazards (PH) assumption in survival analysis is an important issue when Hazard Ratio (HR) is chosen as summary measure. The aim is to assess the appropriateness of statistical methods based on the PH assumption in oncological trials. Methods: We selected 58 randomised controlled trials comparing at least two pharmacological treatments with a time-to-event as primary endpoint in advanced non-small-cell lung cancer. Data from Kaplan–Meier curves were used to calculate the relative hazard at each time point and the Restricted Mean Survival Time (RMST). The PH assumption was assessed with a fixed-effect meta-regression. Results: In 19% of the trials, there was evidence of non-PH. Comparison of treatments with different mechanisms of action was associated (P = 0.006) with violation of the PH assumption. In all the superiority trials where non-PH was detected, the conclusions using the RMST corresponded to that based on the Cox model, although the magnitude of the effect given by the HR was systematically greater than the one from the RMST ratio. Conclusion: As drugs with new mechanisms of action are being increasingly employed, particular attention should be paid on the statistical methods used to compare different types of agents.