Fig 2. Methods for mediation analysis with high-dimensional DNAm data.

(A) Statistical methods for high-dimensional mediation analysis require a multivariable outcome model and multivariate mediator model. (B) Group 1 methods estimate the global mediation effect (α_a^Tβ_m) by fitting the outcome model and estimating the mediator-specific contributions; Group 2 methods estimate α_a^Tβ_m directly without fitting the original model; and Group 3 methods estimate the parameters of an alternative causal structure based on latent variables. (C) In Group 1, the methods HIMA, HDMA, and MedFix apply penalized regression to the outcome model and then linear regression to the mediator model; the method Pathway LASSO fits the outcome and mediator model simultaneously with a jointly penalized likelihood; and the Bayesian methods BSLMM and GMM use multivariate normal mixture models. (D) In Group 2, the methods PCMA and SPCMA use principal component analysis to replace the observed, correlated mediators with independent mediators that can be assessed one-at-a-time. The method HILMA uses a multi-step penalized regression procedure that estimates α_a^Tβ_m but not the mediation contributions. (C) In Group 3, the methods HDMM and LVMA construct latent mediators which replace the original mediators in the mediation model, and thus, they do not yield estimates of α_a^Tβ_m.