Clusterwise p* models for social network analysis



Clusterwise p* models are developed to detect differentially functioning network models as a function of the subset of observations being considered. These models allow the identification of subgroups (i.e., clusters) of individuals who are ‘structurally’ different from each other. These clusters are different from those produced by standard blockmodeling of social interactions in that the goal is not necessarily to find dense subregions of the network; rather, the focus is finding subregions that are functionally different in terms of graph structure. Furthermore, the clusterwise p* approach allows for local estimation of network regions, avoiding some of the common degeneracy problems that are rampant in p* (e.g., exponential random graph) models. © 2011 Wiley Periodicals, Inc. Statistical Analysis and Data Mining 4: 487–496, 2011