Keywords: Mediation analysis; Graphical model; Regularized regression
Intermediate variables mediate pathways from genetic predisposition to disease phenotype. The mediators themselves may form networks due to shared pathways or common disease mechanism. Most of the current literature focus on low-dimensional mediators. It is challenging to consider a large number of network-structured mediators simultaneously. Additionally, the strength of connections in the networks may vary across subjects and the connections may also mediate the pathways. In this work, we propose a two-stage method for mediation analysis. In the first stage, we propose a Gaussian graphical model with covariate-dependent mean and precision matrix to estimate subject-specific networks. In the second stage, we model the connection strength between mediators estimated from the first stage jointly with mediators and covariates to identify the important variables via regularized regression. The performance of our proposed method is assessed by extensive simulation studies. We apply the method to a Huntington’s disease (HD) study to investigate the effect of HD causal gene on the change of motor symptom phenotype as mediated through brain subcortical and cortical grey matter atrophy.