OBJECTIVES: Indirect treatment comparisons (ITC) of survival outcomes are typically derived in connected networks, with or without population-adjustment, just as any other continuous outcome, provided that the assumption of proportional hazards (PH) holds. However, when PH does not hold, the hazard ratio (HR) cannot be considered constant through time, and a time-varying relative effect should be used. This research explored the derivation of time-varying indirect HRs, in the context of an anchored and unanchored ITC, in a case study in late-stage oncology where PH did not hold for any of the constituting direct comparisons.

METHODS: Population-adjusted Kaplan-Meier (KM) curves for overall survival (OS) and progression-free survival (PFS) were obtained from trial A (active-treatment-A vs placebo) and trial B (active-treatment-B vs placebo) and digitized using the Guyot algorithm. The PH assumption was tested within each trial and shown to be violated in PFS. Standard independent parametric models were fit to all trials’ PFS arms and the time-varying hazard profiles of each parametric model were obtained. For each model, the anchored and unanchored time-varying HR of active-treatment-A vs active-treatment-B were calculated. The anchored approach was derived from the ratio of the time-varying HRs of active-treatment-A vs placebo and active-treatment-B vs placebo based on each arm’s hazard profile. The unanchored approach was derived from the ratio of the active-treatment-A and active-treatment-B arm’s hazard profiles.

RESULTS: The time-varying HRs were similar between approaches with the anchored approach showing wider confidence intervals. The Gompertz and generalised gamma models lacked face validity as extreme values were generated when hazards tended to zero.

CONCLUSIONS: The anchored approach may be beneficial since it theoretically can compensate for prognostic factors, but the approach yielded additional uncertainty due to larger number of arms included in the hazard extrapolations and HRs. Neither approach suited models with hazards tending to zero.