Article type
Abstract
Background: Complex interventions consist of multiple interacting components whose effect on the outcome is not easily discernible. Therefore, such interventions may be better investigated within a network of trials that allows sharing information across studies. Several models have been suggested for the analysis of complex interventions in network meta-analysis (NMA). Lumping interventions may result in increased heterogeneity whereas splitting leads to lack of precision and ignores the sharing components across interventions. Other approaches assume additivity of effects of components or regress one component on the other.
Objectives: To present a new, more pragmatic, method for disentangling the effects of components in NMAs of complex interventions.
Methods: We borrow methodology from mediation analysis to model the pathway leading from one component to the outcome both directly and via its combination with other components. In this way, we allow the effect of each component to differ depending on the combinations in which it appears. Unlike previous approaches that assume interaction between two components at a time, our model aims to identify causal relationships among all components simultaneously. We illustrate our method using a NMA of psychological interventions for heart coronary disease. The dataset involves 36 studies measuring all-cause mortality.
Results: We found that no component has an important benefit compared to usual care but the addition of behavioural and relaxation components on the top of educational improves the performance of the latter significantly. Our model suggested that the assumption of additivity on the effects of components might not be plausible. The difference between the sum of the effects of the aforementioned components and the effect of their combinations was 0.81 (-1.26,3.79).
Conclusions: NMAs of complex interventions should try to answer two questions: a) which components work and, b) how do they work. Our approach targets at both questions. Finding a reasonable pathway across components, though, is often challenging and clinical input from experts in the field is necessary.
Objectives: To present a new, more pragmatic, method for disentangling the effects of components in NMAs of complex interventions.
Methods: We borrow methodology from mediation analysis to model the pathway leading from one component to the outcome both directly and via its combination with other components. In this way, we allow the effect of each component to differ depending on the combinations in which it appears. Unlike previous approaches that assume interaction between two components at a time, our model aims to identify causal relationships among all components simultaneously. We illustrate our method using a NMA of psychological interventions for heart coronary disease. The dataset involves 36 studies measuring all-cause mortality.
Results: We found that no component has an important benefit compared to usual care but the addition of behavioural and relaxation components on the top of educational improves the performance of the latter significantly. Our model suggested that the assumption of additivity on the effects of components might not be plausible. The difference between the sum of the effects of the aforementioned components and the effect of their combinations was 0.81 (-1.26,3.79).
Conclusions: NMAs of complex interventions should try to answer two questions: a) which components work and, b) how do they work. Our approach targets at both questions. Finding a reasonable pathway across components, though, is often challenging and clinical input from experts in the field is necessary.