Differences in emergent survival plateaus between PFS and OS may imply clinically unintuitive dichotomy between the resulting proportions of long-term survivors (LTS) when they are analyzed separately via mixture cure models (MCM). We present a novel Bayesian hierarchical (BH) MCM framework assuming a dependency between PFS and OS to estimate LTS rates in CheckMate 067 and demonstrate its practical utility over frequentist MCMs in long-term QALY estimations.
Frequentist and BH MCMs were fitted to PFS and OS data from the trial with minimum 60-months of follow-up. In the frequentist MCMs, PFS and OS were modelled separately whereas in BH MCMs both endpoints were modelled jointly with a shared LTS rate. In both approaches, background mortality rates were taken from World Health Organisation’s age, gender and country-specific lifetables and time-to-event outcomes for the non-LTS were modeled using a range of standard parametric distributions. Estimated incremental QALYs gains for nivolumab containing therapies versus ipilimumab under both approaches were compared using local tariffs from US, Canada, UK, France, Sweden, Belgium, Netherlands, Portugal and Australia.
Among all combinations of distributions considered for the BH MCM, the exponential-exponential fit adequately captured the observed survival trends for both endpoints with reasonable goodness-of-fit measures and shared LTS rates which were (95% credible intervals) 46.3% (32.8%, 62.6%) for nivolumab+ipilimumab, 37.8% (21.6%, 55.7%) for nivolumab, and 15.1% (6.8%, 26.0%) for ipilimumab. For each arm, shared LTS-rates were in between individually-estimated LTS rates from the OS and PFS data in the frequentist MCMs. Compared to frequentist MCMs, over 20-years BH MCMs produced higher incremental QALY gains for nivolumab+ipilimumab and nivolumab versus ipilimumab with differences ranging from 0.30-0.48 and 0.17-0.26, respectively.
Our BH MCM framework can alleviate the disparity between individually estimated OS- and PFS-based LTS rates, and allow for more robust clinical inference and extrapolations.
© 2021 Published by Elsevier Inc.