Résumé (EN)

We know that settings are important when modeling networks (Feld, 1981; Pattison & Robins, 2002; Preciado and Snijders, 2011) and that network endogeneities may play out subtly differently in different contexts (for example Lubbers and Sniders, 2007), yet much of sna is concerned with explaining the structure of particular networks. There are obvious reasons why this should be. Firstly computational limitations of currently available best methods for analysis require that networks are of manageable sizes. Secondly, there are a host of issues surrounding drawing the boundary of a network and this is most easily done by delineating the networks node set according to some common affiliation, setting or context. Thirdly, there has up until recently been no apparent way of taking heteregonenity across settings into account. Fourth, the approaches that conceptually have been available for handling heterogeneity across contexts, models for multilevel analysis of networks (MAN), are not well suited for handling partially overlapping settings nor are they well equipped to detail the kind of dependencies that stem from multiple and overlapping affiliations. With the advent of the multilevel network ERGM (MLNERGM) (Wang et al., 2012; Wasserman and Iacobucci, 1991) we now have a tool for encompassing all of these issues as well as having at our disposal the additional benefit of being able to properly specify relational dependencies between settings. We conceive of multilevel network analysis (MNA) as being analyses for data where we have a people set P and an affiliation set A, with ties in PxP, PxA, AxP, and AxA. The purpose of this paper is to seek to provide a consistent framework for MAN in terms of MNA, and to illustrate the limitations of the former. We illustrate several aspects of how hierarchical ERGMs for MAN relates to MNA by way of empirically informed Monte Carlo studies as well as a brief empirical analysis. Most researchers also associate the term “multilevel” in a network context with MAN and, in light of computation issues, MAN might appear an appealing technique for taking heterogeneity into account.