Which test for crossing survival curves? A user's guideline

Area of Science:

• Biostatistics
• Clinical Trials
• Survival Analysis

Background:

• Effective knowledge exchange is crucial between statisticians and clinicians in clinical trials.
• Time-to-event endpoints are common in clinical trials, with survival distribution comparisons being a frequent question.
• The log-rank test is standard for comparing survival distributions but can be underpowered with non-proportional hazards, necessitating advanced methods.

Purpose of the Study:

• To facilitate the selection of appropriate statistical tests for detecting survival differences in two-arm clinical trials with crossing hazards.
• To evaluate the performance of various statistical tests designed for non-proportional or crossing hazards in oncology clinical trials.

Main Methods:

• A review of recent two-arm clinical oncology trials with crossing survival curves was conducted.
• Data from selected trials were reconstructed using a state-of-the-art algorithm, focusing on publications with reported numbers at risk.
• The study compared the p-values from the log-rank and Peto-Peto tests against nine alternative tests for detecting survival differences.

Main Results:

• Fifteen trials were selected from 1400 reviewed, with three additional individual patient data sets included, totaling 18 studies.
• Significant survival differences were detected in nine of the 18 studies using the investigated tests.
• A critical finding was that 28% of studies lacked essential 'number at risk' data, hindering reproducibility and plausibility checks.

Conclusions:

• Inference methods designed for non-proportional hazards are beneficial for survival analysis in clinical trials.
• These advanced methods provide valuable guidance for selecting alternatives to the standard log-rank test.
• Improving data reporting standards, particularly 'number at risk', is essential for robust clinical trial analysis.