OBJECTIVES:

In 2013, Nicolas R. Latimer published guidelines regarding the extrapolation of survival analysis data for cost-effectiveness models. Tremblay et al completed his work by highlighting the need to perform a statistical test for proportional hazard (PH), as visual inspection of the log-cumulative hazard curves is insufficient to determine if the PH assumption is justified. The goal of this work was to estimate the risk of error on efficacy estimation (percentage of error on AUC) by wrongly accepting the PH assumption.

METHODS:

We digitized PFS or OS Kaplan-Meier curves of 30 recently published phase III trials of innovative oncological drugs. For each trial, we generated individual patient data (IPD) using Guyot’s algorithm, and the PH assumption was tested using the Schoenfeld test. The hazard ratio published was applied directly on the comparator/placebo arm to create the treatment curve. The risk of error between the “real” treatment curve and the “created” one was calculated as the percentage of error between the areas under the curve (AUC) and plotted against the p-value of the Schoenfeld test.

RESULTS:

In 73.3% (22/30) of the trials, the PH assumption was verified at a 0.05 threshold. For these trials, the difference of the AUC was less than 5% in 72.7% (16/22) of cases. However, for the 8/30 trials where the PH assumption fails, the difference on the AUC was greater than 5% for 100% of cases, with a maximum difference of 19.4%.

CONCLUSIONS:

Based on this preliminary work, extrapolation using HR method is not adapted if PH assumption failed, considering the high risk of error on efficacy estimation. We plan to incorporate more trial results in order to explore a threshold level on the Schoenfeld test which would guarantee safe extrapolation.