|
Abstract:
|
Mediation analysis assesses relative contributions of direct and indirect effects to explore underlying processes by which exposures affect outcomes. Although methods to test for mediation effects have been widely used, relatively little research has focused on developing effect size measures, especially in settings with multiple exposures where standard use of proportion mediated is not applicable. We propose an effect size measure based on variance decompositions which summarize the overall effect of multiple exposures. We further develop multistage constrained least squares estimators to effectively constrain the range of the proposed measure in the range of 0 to 1, a desirable feature for relative effect size measures. Closed forms and asymptotic properties are derived and evaluated through numerical simulation. We apply the proposed measures to the Agricultural Health Study to explore the effect of multiple smoking related exposures on lung functions through methylation at various CpG sites.
|
