Title: Maximum pseudolikelihood estimation for models of social network data
Authors: Jonathan Stewart - Florida State University (United States) [presenting]
Michael Schweinberger - Rice University (United States)
Abstract: The statistical analysis of social network data requires both methods and theory for dependent data in some of the most challenging scenarios. Often, we obtain only a single observation of dependent relationships or interactions in a social network and wish to estimate and infer statistical models with parameter vectors of (possibly) increasing dimension. On the methodological side, the estimation of statistical models for dependent network data gives rise to significant computational challenges. We revisit the familiar maximum pseudolikelihood estimation paradigm and demonstrate how it offers both a scalable and accurate alternative to the gold-standard Monte-Carlo maximum likelihood estimation. On the theoretical side, we demonstrate that many statistical models of social network data possess important conditional independence properties, and discuss how naturally occurring structure in social networks facilitate statistical estimation of complex models. An example of a model for social network data capturing brokerage is given. Theoretical guarantees which highlight the key points are also presented.