OBJECTIVES: Restricted mean survival time (RMST) has been used in network meta-analyses (NMA) when the proportional hazards assumption does not hold. It has been argued that progression free (PFS) and overall survival (OS) should be modelled jointly to capture correlations. This research aims to compare the results for joint versus separate modelling of PFS and OS in RMST NMAs.

METHODS: Using a network of 5 studies of 3 treatments for operable stage IIIA-N2 non-small cell lung cancer, we jointly modelled PFS and OS using both additive and multiplicative fixed effects NMA models. NMA outcomes were treatment differences in restricted mean PFS, post-progression survival (PPS) and OS restricted to 5 years. For comparison, we performed NMAs on PFS and OS data separately to obtain estimates of treatment differences in restricted mean PFS and OS, from which we derived estimates for PPS. We estimated the bias in separate compared with joint modelling by calculating the percentage difference between restricted mean PFS, PPS, and OS.

RESULTS: The percentage difference between joint vs separate modelling of PFS ranged from 3.86%-4.29%, PPS ranged from 25.99%-37.36% and OS ranged from 10.11%-11.03%. Additionally, the direction and scale of these differences varies by treatment.

CONCLUSIONS: Separate modelling of PFS and OS in NMA using restricted means resulted in biased estimates of restricted mean PFS, PPS and OS compared with joint modelling. We recommend the joint synthesis of PFS and OS to capture correlations and ensure that OS is estimated to be greater than PFS when estimating restricted mean survival times. Further work is required to assess the impact of method for NMAs of RMST in health economic models to assess cost-effectiveness, including extrapolation beyond restriction time.