A Bayesian reanalysis of the THOR trial cohort 1 using reconstructed individual patient data: interval-specific...

The phase 3 THOR trial (cohort 1) showed improved overall survival (OS) and progression-free survival (PFS) with erdafitinib versus chemotherapy in FGFR2/3-altered advanced urothelial carcinoma after prior programmed cell death protein 1 (PD-1) / programmed death-ligand 1 (PD-L1) inhibitor therapy. When the assumption of proportional hazards (PH) is violated, a single hazard ratio (HR) may be difficult to interpret.

We derived reconstructed individual patient data (rIPD) for OS and PFS from published Kaplan-Meier curves. PH was assessed using scaled Schoenfeld residuals and log-log plots. We fit Bayesian piecewise-exponential models with administrative censoring at τ = 24 months (OS; cut-point 12 months) and τ = 12 months (PFS; cut-point 6 months). We reported interval-specific HRs, landmark survival differences (ΔOS, ΔPFS), number needed to treat (NNT), restricted-mean survival time (RMST), and posterior probabilities for clinically meaningful thresholds.

Statistical diagnostics indicated that the PH assumption was violated for both OS and PFS. For OS, HR was 0.59 (95% credible interval 0.41-0.84) in months 0-12 and 1.06 (0.48-2.48) in months 12-24. At 12 months, ΔOS(12) = 0.18 (0.06-0.30); conditional median NNT was 5.45. ΔRMST (0-24) = 3.02 months (0.74-5.23), with P(ΔRMST >0) = 99.5%. For PFS, the HR was 0.64 (0.45-0.89) in months 0-6 and 0.49 (0.25-1.00) in months 6-12; ΔPFS (6) = 0.16 (0.04-0.28) and ΔRMST (0-12) = 1.74 months (0.69-2.76).

This Bayesian rIPD-based reanalysis provides PH-robust estimands, indicating early OS benefit with uncertain later effects, and favorable PFS over 0-12 months with erdafitinib.