A0500
Title: A stochastic block prior for clustering in graphical models
Authors: Nikola Sekulovski - University of Amsterdam (Netherlands) [presenting]
Giuseppe Arena - University of Amsterdam (Netherlands)
Jonas Haslbeck - University of Amsterdam (Netherlands)
Karoline Huth - University of Amsterdam (Netherlands)
Nial Friel - University College Dublin (Ireland)
Maarten Marsman - University of Amsterdam (Netherlands)
Abstract: Existing statistical methods for analyzing graphical models in psychology often ignore the assumption of clustering, which refers to the grouping of variables that are more densely connected, despite its relevance in many psychological theories. The stochastic block model (SBM) is proposed as a prior distribution on the network structure in models for binary and ordinal data. The SBM assumes that variables belong to latent clusters and that the probability of an edge depends on cluster membership. Embedding the SBM in a Bayesian graphical modeling framework enables the formal incorporation of theoretical expectations about clustering, testing hypotheses about the number of clusters, and estimating cluster membership of the nodes from cross-sectional data. The benefits of this approach are demonstrated through a simulation study and a reanalysis of 30 empirical datasets. This method provides a principled approach to latent cluster inference in psychological network analysis by incorporating structural assumptions directly into the model through the prior.
