Calibration of treatment effects in network meta-analysis using individual patient data

Health technology assessments require reliable estimates of relative treatment effects for a given patient population, to inform decision making. Standard network meta-analysis (NMA) and indirect comparisons combine aggregate data (AgD) from multiple studies on treatments of interest, assuming that any treatment effect modifiers are balanced across populations. This assumption can be relaxed if individual patient data (IPD) are available from all studies, using an IPD network meta-regression (NMR). However, in many cases IPD are only available from one or a subset of studies. Recently proposed methods for population-adjusted indirect comparisons aim to adjust for differences between one IPD study and one AgD study. However, the resulting comparison is only valid in the AgD study population without additional assumptions, and the methods cannot be extended to larger treatment networks. Meta-regression approaches can be used in larger networks, but typically incur aggregation bias. In this thesis, webeginby reviewing the literature onpopulation adjustment and related problems, giving a critique of current methods. We review applications of current methods in the published literature and in National Institute for Health and Care Excellence technology appraisals. Motivated by these reviews we propose a general method, Multilevel Network MetaRegression (ML-NMR), that overcomes some of the disadvantages of current approaches and reduces to AgD NMA and IPD NMR as special cases. We discuss the computational aspects of implementingML-NMR, before applying to a real example of plaque psoriasis treatments. The ML-NMR framework is then extended to handle more general likelihoods, illustrated with an artificial example of survival outcomes and a reanalysis of the plaque psoriasis example incorporating multiple outcomes. An extensive simulation study is conducted to assess the performance of ML-NMR and current methods in a range of scenarios and under various failures of assumptions. We conclude with a discussion and suggestions for future research.