Few fully developed statistical approaches exist for jointly modeling and comparing multiple networks. Here, we develop a fully hierarchical Bayesian model for network data in order to understand how network structure varies across observed networks, as well as how some exogenous mechanisms might drive these observed differences.
We consider latent space models for network data (Hoff et al. 2002), which model nodes who are closer together in a (typically Euclidean) latent social space as more likely to be tied. These models are sometimes discounted because network structure is not explicitly incorporated (such as in the ERGM) and have been criticized for their inability to account for clustering within a network. Instead of explicitly adding more structure to the model, we propose taking advantage of the negative curvature of a hyperbolic latent space to smoothly grow the complexity of this model and to accommodate the modeling of more complicated network structure, such as degree heterogeniety and clustering. We then extend this model to the multilevel setting and demonstrate its utility in a social science application.
|