Bayesian mixed-effect models for independent dynamic social network data

Relational event or time-stamped social network data have become increasingly available over the years. Accordingly, statistical methods for such data have also surfaced. These techniques are based on log-linear models of the rates of interactions in a social network via actor covariates and network statistics. Particularly, the use of survival analysis concepts has stimulated the development of powerful methods over the past decade. These models mainly focus on the analysis of single networks. To date, there are few models that can deal with multiple relational event networks jointly. In this paper, we propose a new Bayesian hierarchical model for multiple relational event sequences. This approach allows inferences at the actor level, which are useful in understanding which effects guide actors' preferences in social interactions. We also present Bayes factors for hypothesis testing in this class of models. In addition, a new Bayes factor to test random-effect structures is developed. In this test, we let the prior be determined by the data, alleviating the issue of employing improper priors in Bayes factors and thus preventing the use of ad-hoc choices in absence of prior information. We use data of classroom interactions among high school students to illustrate the proposed methods.