Title: Hierarchical multilevel analysis of network dynamics: Bayesian estimation and prior sensitivity
Authors: Tom Snijders - University of Groningen, University of Oxford (Netherlands) [presenting]
Johan Koskinen - University of Manchester (United Kingdom)
Abstract: Multilevel longitudinal network data sets, by which are meant longitudinal network data sets that were collected according to the same design in multiple, disconnected groups, are starting to be available more and more. This offers new possibilities for generalization and requires new methods of analysis. A multilevel version of the stochastic actor-oriented model (SAOM) is presented. In a hierarchical model for such data structures, there are two sets of parameters: parameters at the highest level, the population of groups, which may be called the population parameters; and parameters at the group (or network) level. For the joint analysis of group- and population-level, a fully Bayesian approach is followed in which the network in each group evolves according to a SAOM; the groupwise parameters are drawn from a multivariate normal distribution; and the parameters of this multivariate normal have a conjugate prior distribution. A special case of the prior is that some of the variances of the multivariate normal are 0. In many applications the number of groups is rather small, so that sensitivity for the prior is an issue. Prior sensitivity is compared for various weakly informative and non-informative priors.