Abstract

Many models of dynamic network data assume that the set of nodes (social actors) remains constant over time. By contrast, we propose a framework that allows for models where both the set of nodes and the social ties connecting them change over time.  We depart from the conventional setup where the distribution of the vertex set is a point mass, and instead model its evolution via a stochastic process.  Our approach is modular, in the sense that joint distribution of the vertex sets over time can be modeled marginally and then the joint distribution of the edge sets can be specified, conditionally upon it. The conditional independence statements implied by our approach then mean that posterior sampling for the parameters corresponding to these factors (joint distribution of the vertex sets, joint distribution of the edge sets) can be performed separately. We illustrate our methodology via both simulation studies and data analysis. 
Joint work with Sofia Olhede and Patrick Wolfe.