Limited Assessment of Proportional Hazards Despite More Frequent Violations Identified in Updated Oncology Trial Analyses: Frequency, Patterns, and Statistical Implications

Prior studies suggest that proportional-hazards violations (PHVs) occur in oncology randomized controlled trials (RCTs) and may affect conclusions, but evidence has largely focused on initial reports rather than updated analyses. We examined PH testing practices, frequency, patterns, and statistical implications of PHVs in updated oncology RCTs. We focused on oncology RCTs supporting US Food and Drug Administration approvals. Phase II-III oncology RCTs supporting FDA approvals between 2006 and 2025 were identified. Individual patient data were reconstructed from Kaplan-Meier curves. PHVs were assessed using Schoenfeld residuals. Treatment effects were analyzed with Cox proportional-hazards models, log-rank tests, MaxCombo, and restricted mean survival time difference (RMSD). Two hundred seventy endpoint comparisons (150 overall survival [OS], 120 non-OS) from 168 RCTs were included. PH testing dropped from 24.1% in initial reports to 11.5% in updates (McNemar p < 0.001). Although PH testing increased over the past decade in initial publications (Wald p < 0.001), no trend in updates (Wald p = 0.29). PHV frequency was numerically higher in updated than initial analyses (38.9% vs. 33.0%). Non-OS endpoints, metastatic settings, and immunotherapy or targeted-therapy regimens were associated with PHVs. Among updated analyses, nominal significance differences between PH-dependent and alternative methods were more frequent with PHVs (MaxCombo: 11.4% vs. 3.6%, chi-square p = 0.012; RMSD: 8.6% vs. 2.4%, chi-square p = 0.021). PHV patterns were diminishing effects (35.2%), delayed separation (27.6%), and crossing hazards (27.6%). PH testing declined in updated analyses despite numerically more frequent PHVs. Although updates are descriptive and do not invalidate original conclusions, unrecognized non-proportionality may affect long-term benefit interpretation, supporting PH diagnostics and prespecified complementary analyses.