Fig. 2

Flowchart of the multi-step method. In the first step, we use linear regression models to estimate the SNP effects on the known mediators (a^∗) and the direct effects between the SNPs and the outcome (c^∗). In the second step, we apply the EM algorithm to fit GMMs on a^∗ and c^∗. In the third step, a MCMC procedure is performed using the estimated GMM parameters from the last step in the priors to generate a posterior distribution for the hidden mediator’s effect size, b_H