Large-scale national and multi-institutional patient safety improvement programmes are being developed in the health care systems of several countries to address problems in the reliability of care delivered to patients. Drawing upon popular collaborative improvement models, these campaigns are ambitious in their aims to improve patient safety in macro-level systems such as whole health care organisations. This article considers the methodological issues involved in conducting research and evaluation of these programmes. Several specific research challenges are outlined, which result from the complexity of longitudinal, multi-level intervention programmes and the variable, highly sociotechnical care systems, with which they interact. Organisational-level improvement programmes are often underspecified due to local variations in context and organisational readiness for improvement work. The result is variable implementation patterns and local adaptations. Programme effects span levels and other boundaries within a system, vary dynamically or are cumulative over time and are problematic to understand in terms of cause and effect, where concurrent external influences exist and the impact upon study endpoints may be mediated by a range of organisational and social factors. We outline the methodological approach to research in the United Kingdom Safer Patients Initiative, to exemplify how some of the challenges for research in this area can be met through a multi-method, longitudinal research design. Specifically, effective research designs must be sensitive to complex variation, through employing multiple qualitative and quantitative measures, collect data over time to understand change and utilise descriptive techniques to capture specific interactions between programme and context for implementation. When considering the long-term, sustained impact of an improvement programme, researchers must consider how to define and measure the capability for continuous safe and reliable care as a property of the whole care system. This requires a sociotechnical approach, rather than focusing upon one microsystem, disciplinary perspective or single level of the system.