One-step individual participant data network meta-analysis of time-to-event data

Network meta-analysis (NMA) combines direct and indirect evidence from trials to calculate and rank treatment effect estimates. While modelling approaches for continuous and binary outcomes are relatively well developed, less work has been done with time-to-event outcomes. Such outcomes have usually been analysed using Cox proportional hazard (PH) models, but in oncology, with longer follow-up of trials and time-dependent effects of targeted treatments, this may no longer be appropriate. Alongside this, NMA conducted in the Bayesian setting has been increasing in popularity. In this thesis I extend the work of Royston and Parmar to the NMA setting, showing that Royston-Parmar models, fitted in WinBUGS, provide a flexible, practical approach for Bayesian NMA with time-to-event data and can accommodate non-PH. Inconsistency in NMA occurs when the direct and indirect evidence are not in agreement with each other and can result in biased treatment effect estimates. It is therefore important that attempts are made to identify, understand and, where appropriate, adjust for inconsistency. In this thesis I consider four increasingly complex methods of assessing inconsistency in NMA, proposed (relatively) recently in the literature. Motivated by individual participant data (IPD) from 42 trials comparing radiotherapy, sequential and concomitant chemotherapy from 7531 people with lung cancer, I illustrate why one of these approaches may be misleading and propose an alternative approach. Stratified medicine aims to identify groups of patients most likely to respond to treatment. However, many trials are underpowered to detect clinically meaningful differences in subgroups. NMA models fitted with treatment-covariate interactions potentially have greater power to identify such differences. In the final part of this thesis I extend the one-step IPD NMA Royston-Parmar model to include treatment-covariate interactions, providing practical guidance on how to deal with missing covariate data and how to combine or separate within and across trial information