Networks as mediating variables: a Bayesian latent space approach

Abstract The use of network analysis to investigate social structures has recently seen a rise due the high availability data and numerous insights it can provide into different fields. Most analyses focus on topological characteristics networks estimation relationships between nodes. We adopt perspective by considering whole as random variable conveying effect an exposure response. This point view represents classical mediation setting, where interest lies in estimating indirect effect, that is, propagated through mediating variable. introduce latent space model mapping smaller dimension hidden positions units network. coordinates each node are used mediators relationship further extend framework using Generalised Linear Models instead linear ones, previously done literature, adopting approach based derivatives obtain effects interest. A Bayesian allows us get entire distribution generally unknown, compute corresponding highest density interval, which gives accurate interpretable bounds for mediated effect. Finally, application interactions among group adolescents their attitude toward substance is presented.