Incidence and impact of proportional hazards violations in phase 3 cancer clinical trials.

1561 Background: Hazard ratio (HR)-based analyses used in oncology trials rely on the assumption of proportional hazards, i.e. a HR that is constant over time. Proportional hazards violations (PHVs) may lead to misinterpretation of trial results. Restricted mean survival time (RMST) is valid with non-proportional hazards and has received recent attention specifically for immunotherapy (IO) trials but has not been routinely adopted in oncology trial design as a whole. We aimed to comprehensively characterize the incidence and factors associated with PHVs among phase 3 oncology trials and assess RMST as an alternate measure of treatment effect in survival analysis. Methods: We used Clinicaltrials.gov to identify all superiority-design, 2-arm phase 3 cancer trials with time-dependent endpoints with published results through February 2020. We manually reconstructed patient-level data from published Kaplan-Meier (KM) curves, assessed PHVs with the Schoenfeld residual test (p <.05) and analyzed the RMST. To assess reconstruction accuracy, reported and reconstructed HRs were compared. Univariable logistic regression was used to assess the likelihood of PHVs by trial characteristic, with statistically significant factors (p <.05) included in a multivariable analysis. Concordance of RMST-based and HR-based analysis was established when both tests agreed as to the statistical significance of the comparison. Results: Of 342 KM comparisons eligible for reconstruction, 318 comparisons across 315 trials, enrolling 347,538 patients from 1989-2017, were accurately reconstructed and analyzed. PHVs were identified in 76/318 (23.9%) trials. There was no difference in likelihood of PHVs among IO vs non-IO trials (LR 2.31, 95% CI 0.30 – 17.85, P =.37), nor by disease site, year of trial initiation, or sample size. Few trials with PHVs (16/76) pre-specified a plan to account for non-proportional hazards in statistical design. Trials with an overall survival (OS) primary endpoint (PEP) were less likely to have PHVs than trials with a non-OS PEP (LR: 0.50, 95% CI 0.28 - 0.90, P =.02). Trials whose PEP was non-significant were more likely to have PHVs (LR 1.73, 95% CI 1.01 - 2.97, P =.047). No factor remained significantly associated with PHV in multivariable analysis. Overall, 291/318 (91.5%) KM comparisons were concordant. Among trials with PHVs, 5/76 were significant by RMST but not HR, and 5/76 were significant by HR but not RMST. Of these, 1 led to FDA drug approval, and 2 others are cited in NCCN guidelines. Conclusions: PHVs are common across all phase 3 cancer clinical trials. Attempts to account for PHVs in trial design are lacking despite the potential for trial misinterpretation in the event of non-proportional hazards. RMST-based analysis is broadly concordant with HR-based analysis and may aid in interpretation of trials with PHVs. Hence, we recommend that prospective trials include a priori a statistical plan to account for PHVs.