Separable and controlled direct effects for competing events: Estimation of component specific effects on sickness absence.

Abstract

In many settings, it is reasonable to think of treatment as consisting of a number of components, either because this is the case in practice or because it is conceptually possible to decompose treatment into separate components due to the way in which it exerts effects on the outcome of interest. For competing events, the treatment decomposition idea has recently been suggested to separate effects of treatments on the outcome of interest from effects mediated through competing events using so-called separable effects. Like the idea of separating effects of exposure, it has been pointed out that ideas from mediation analysis generally may help to clarify the interpretation of existing estimands used in competing events settings. One example is the use of the controlled direct effect, to conceptualize the effects of interventions preventing the competing event from occurring. In this article, we identify the controlled direct effect as a component specific effect and discuss the merits of this estimand when the prevented event is non terminal and other methods of effects separation are problematic. Our motivating example is the study of a policy initiative, introduced in 2001, aimed at reducing long term sickness absence (SA) in Norway. The initiative consists of different components, one being to encourage use of graded SA, which is considered a key tool in the Nordic countries to reduce long term SA. The analysis makes use of longitudinal registry data for 113 808 individuals, followed from the time of first SA.