Estimating complex intervention effects as risk differences in cluster randomized trials

Complex interventions are defined as series of inter-related events occurring within a broader system they are in constant interaction with. Health service organization, health promotion interventions or health educational interventions are complex interventions. They are expected to be adapted to the context in which they are implemented. The aim of population health intervention research is to assess complex health interventions. Among the methodological approaches used for assessment is the cluster randomized design. In this latter design, intact social units such as geographic areas, schools, families or medical practices are randomized rather than individuals themselves. Such a design raises two issues. The first is to specify the population to be analyzed. Classically, the intention-to-treat approach prevails. However, the per-protocol and as-treated populations are also of interest. Indeed, the complementary results obtained considering these latter populations may allow for identifying components of the complex interventions that are useless or need to be modified. They may also allow for understanding whether a lack of efficacy is due to a truly inefficacious intervention or to a lack of adherence to the intervention. However, identifying participants who are adherent to the protocol is not easy. The boundary between lack of compliance and adaptation of the intervention to the context is indeed tenuous. The second issue is a statistical one. When the outcome is binary, intervention effects are expected to be expressed on both a relative and an absolute scale. Cluster randomized trials are classically analyzed by modelling approaches using marginal or mixed models because of the hierarchical data structure (participants are embedded in clusters). When the outcome is binary, this generally leads to expressing the intervention effect as an odds ratio, but we also want to estimate a risk difference. However, to date, we lack clear recommendations on the optimal statistical method for estimating such an adjusted risk difference.

The final objective of the ESCIENT project consists of guidelines on how a cluster randomized trial assessing a complex health intervention should be planned, conducted, and above all analyzed.