Abstract The attributable fraction is a common measure in epidemiological research, which quantifies the public health impact of a particular exposure on a particular outcome. Often, the exposure effect may be mediated through a third variable, which lies on the causal pathway between the exposure and the outcome. To assess the role of such mediators we propose a decomposition of the attributable fraction into a direct component and a mediated component. We show how these components can be estimated in cross-sectional, cohort and case-control studies, using either maximum likelihood or doubly robust estimation methods. We illustrate the proposed methods by an application to a study of physical activity, overweight and CVD. In an Appendix we provide R-code, which implements the proposed methods.