OBJECTIVES: There are currently no clear guidelines from healthcare authorities on how to conduct network meta-analysis (NMA) on survival endpoints. This research aims to compare two different methods to conduct network meta-analyses of survival outcomes and explore a number of approaches for extrapolation beyond the follow-up time and give recommendations on when each method should be implemented.
METHODS: NMAs based on (1) hazard ratios and (2) reconstructed patient-level data using fractional polynomials were fitted to two different networks of evidence for both overall survival (OS) and progression-free survival (PFS). Assessments were made of the number of statistically significant non-proportional hazard ratios and also the loss of information when only those studies reporting Kaplan-Meier curves were analyzed. Anchoring methods were also conducted using external data to ensure plausible long-term predictions.
RESULTS: In the networks studied, 11% of trials reporting OS contained significant non-proportional hazard ratios compared to 29% for PFS. There was a 10% loss of trials from OS and 26% from PFS when only trials that reported Kaplan-Meier charts were included in the networks of evidence. The NMA based on fractional polynomials with anchoring produced survival estimates that fitted most of the study arms well and produced plausible long-term predictions. The NMA based on hazard ratios fitted the data relatively less well and did not produce plausible long-term survival estimates; additional techniques could be used to improve long-term predictions.
CONCLUSIONS: NMAs based on reconstructed survival data have some clear advantages over simpler methods based on hazard ratios. However, due to the loss of information and potential selection bias when only studies reporting Kaplan-Meier charts are included in a network of evidence, the fractional polynomial method may not always be sufficient if used alone. In some cases the hazard ratio approach may still be the preferred option.
