Synthesis of data from trials of interventions designed to change health behaviour; a case study

Many complex healthcare interventions aim to change the behaviour of patients or health professionals, e.g. stopping smoking or prescribing fewer antibiotics. This prompts the question of which behaviour change interventions are most effective. Synthesising evidence on the effectiveness of a particular type of behaviour change intervention can be challenging because of the high levels of heterogeneity in trial design. Here we use data from a published systematic review as a case study and compare alternative methods to address this heterogeneity. One important sources of heterogeneity is that compliance to a desired behaviour can be measured and reported in a variety of different ways. In addition, interventions designed to target behaviour can be implemented at either an individual or group level leading to trials with varying layers of clustering. To handle heterogeneous outcomes we can either convert all effect estimates to a common scale (e.g. using standardised mean differences) or have separate meta-analyses for different types of outcome measure (binary and continuous measures).To address the clustering structure, adjusted standard errors can be used with the inverse variance method, or weights can be assigned based on a consistent level of clustering, such as the number of healthcare professionals. A graphical method, the albatross plot utilises reported p-values only, and can synthesise data with both heterogeneous outcomes and clustering with minimal assumption and data manipulation. Based on these methods, we reanalysed our data in four different ways and have discussed the strengths and weaknesses of each approach.