Cluster analysis of dyad distributions in networks

Existing statistical models for network data that are easy to estimate and fit are based on the assumption of dyad independence or conditional dyad independence if the individuals are categorized into subgroups. We discuss how such models might be overparameterized and argue that there is a need for subgrouping methods to find appropriate models. We propose clustering of dyad distributions as such a method and illustrate it by analyzing how cooperative learning methods affect friendship data for school children.Work for this report was initiated while Frank and Komaska visited the Department of Statistics at the University of California, Riverside. Partial support was provided by the Swedish Council for Research in the Humanities and Social Sciences.