Mediation analysis with multiple mediators: An application to the study of adolescent eating disorders

Bianca De Stavola, Ph.D.
University College London
(formerly of London School of Hygiene and Tropical Medicine)

ABSTRACT:

The statistical estimands that decompose the effect of an exposure on an outcome along different pathways have been the focus of much recent interest in the epidemiological literature. These contributions however have focussed almost entirely on settings with a single mediator, or a set of mediators considered en bloc, while many applications ideally involve decompositions into path-specific effects through many, separate, mediators. 

In this presentation we give counterfactual definitions of such path-specific estimands in settings with multiple mediators. We focus on the situation where earlier mediators may affect later ones, and show that there are many ways in which such decomposition can be done. Webriefly review the strong assumptions under which the effects are identified and suggest a sensitivity analysis approach when a particular subset of the assumptions cannot be justified.

These ideas are illustrated using data from the Avon Longitudinal Study of Parents and Children (ALSPAC), a birth cohort of children born between April 1991 and December 1992 in the South-West of the United Kingdom.  We investigate whether the presumed causal effect of maternal pre-pregnancy BMI (classified as: underweight, normal, overweight/obese) on offspring adolescent eating disorders behaviours is mediated via birth weight, childhood size, and growth trajectories. The results show interesting opposite decompositions for the effect of maternal underweight and overweight. Sensitivity analyses indicate that the results are robust to variations in some of the identifying assumptions.