Combined test versus logrank/Cox test in 50 randomised trials

Area of Science:

• Biostatistics
• Clinical Trials Methodology

Background:

• Logrank test and Cox proportional hazards model are standard for time-to-event outcomes in randomized controlled trials (RCTs).
• Sample size and power calculations typically assume proportional hazards (PH), where the hazard ratio remains constant.
• Failure of the PH assumption can significantly reduce the power of standard statistical tests.

Purpose of the Study:

• To assess the frequency and impact of non-proportional hazards (non-PH) in published phase 3 clinical trials.
• To compare the performance of the standard logrank/Cox test with a recently proposed combined test under non-PH conditions.

Main Methods:

• Systematic search of four leading medical journals for phase 3 clinical trials with time-to-event outcomes.
• Digitization of Kaplan-Meier curves to reconstruct individual patient-level data.
• Testing for non-PH and comparing results from the logrank/Cox test and the combined test.

Main Results:

• The PH assumption was checked in only 28% of the 50 included trials.
• Evidence of non-PH was detected in 31% of comparisons.
• The combined test showed higher significance rates (55%) compared to the Cox test (49%), with interpretation changes in 4 out of 5 discordant trials.
• Non-PH was often characterized by early treatment effects that diminished over time.

Conclusions:

• Non-PH is likely under-identified in RCTs but may be prevalent.
• The combined test demonstrated superior performance over the Cox test in this reanalysis.
• The combined test represents a more robust approach for the design and analysis of clinical trials with time-to-event data.