Meta-analysis methods for combining information from different sources in evaluating health interventions

This thesis considers the quantitative synthesis of evidence from different study types in order to assess the effectiveness of health interventions. Bayesian MCMC methodology is used extensively, but not exclusively, for the analyses described herein. The thesis commences with consideration of different study designs used in health and related disciplines together with consideration of the validity of these sources. Existing synthesis methods for combining information, first, from a single study design (often referred to as meta-analysis), and then from multiple sources of evidence are then reviewed. A meta-analysis of the randomised evidence on cholesterol lowering observations is presented. This analysis is then extended to a more generalised synthesis by including data from aetiological cohort studies in the analysis using hierarchical modelling methods. Such models allow for heterogeneity between study types. A second generalised synthesis considers evidence from three sources relating to the use of electronic fetal heart rate monitoring during labour. The particular problem of publication bias, and how it can be addressed in a generalised synthesis framework, where there are potentially differential levels of publication bias for the different sources of evidence, is discussed. Adverse events from interventions are often rare, and hence, difficult to detect and quantify using randomised controlled trials. The use of generalised synthesis to quantify adverse events is illustrated using data relating to adverse events of hormone replacement therapy and breast implants. The sparseness of the event data in these examples presents specific statistical problems which are explored. A sensitivity analysis framework for assessing the robustness of results to under-reported adverse events is outlined. A final example, the use of warfarin to prevent strokes in patients with atrial fibrillation, illustrates how disparate sources of data can be synthesised to construct a net-clinical-benefit model where potential benefits of treatment are weighed up against potential harm due to adverse events. This analysis synthesises clinical event data from randomised controlled trials, observational cohort studies for both benefit and harms as well as quality of life data. The net-clinical-benefit of the treatment is expressed, together with corresponding uncertainty measures, for patients with different underlying risks. This thesis illustrates that with the increase in computer power and development of software to fit complex models using Bayesian MCMC methodology, it is now possible to think beyond the models currently used to synthesise medical data. It is hoped that such efforts will be seen as tentative first steps in a future where quantitative models are created routinely to summarise the totality of evidence, and inform models to make decisions for future patients.