The identification of mediating effects using genome-based restricted maximum likelihood estimation

Abstract

Mediation analysis is commonly used to identify mechanisms and intermediate factors between causes and outcomes. Studies drawing on polygenic scores (PGSs) can readily employ traditional regression-based procedures to assess whether traitMmediates the relationship between the genetic component of outcomeYand outcomeYitself. However, this approach suffers from attenuation bias, as PGSs capture only a (small) part of the genetic variance of a given trait. To overcome this limitation, we developed MA-GREML: a method for Mediation Analysis using Genome-based Restricted Maximum Likelihood (GREML) estimation.Using MA-GREML to assess mediation between genetic factors and traits comes with two main advantages. First, we circumvent the limited predictive accuracy of PGSs that regression-based mediation approaches suffer from. Second, compared to methods employing summary statistics from genome-wide association studies, the individual-level data approach of GREML allows to directly control for confounders of the association betweenMandY. In addition to typical GREML parameters (e.g., the genetic correlation), MA-GREML estimates (i) the effect ofMonY, (ii) thedirect effect(i.e., the genetic variance ofYthat is not mediated byM), and (iii) theindirect effect(i.e., the genetic variance ofYthat is mediated byM). MA-GREML also provides standard errors of these estimates and assesses the significance of the indirect effect.We use analytical derivations and simulations to show the validity of our approach under two main assumptions,viz., thatMprecedesYand that environmental confounders of the association betweenMandYare controlled for. We conclude that MA-GREML is an appropriate tool to assess the mediating role of traitMin the relationship between the genetic component ofYand outcomeY. Using data from the US Health and Retirement Study, we provide evidence that genetic effects on Body Mass Index (BMI), cognitive functioning and self-reported health in later life run partially through educational attainment. For mental health, we do not find significant evidence for an indirect effect through educational attainment. Further analyses show that the additive genetic factors of these four outcomes do partially (cognition and mental health) and fully (BMI and self-reported health) run through an earlier realization of these traits.