Project
Advancing Relational Event Models Using Nonlinear Statistical Techniques

Given recent technological advancements, dynamic interaction data between two or multiple actors – so-called relational event history (REH) data – is increasingly becoming available. This fosters the opportunity to answer research questions about relationships of variables with the rate at which actors interact (i.e., the event rate) using so-called relational event models (REMs) which have proven very useful for the analysis of such data (Butts, 2008; DuBois et al., 2013; Stadtfeld et al., 2017). Therefore, it comes with no surprise that the application of REMs has received noticeable attention in various domains, ranging from topics like corporate networks (Valeeva et al, 2020) to interaction patterns of interactions in aviation (David and Schraagen, 2018).

REMs model a dyad’s event rate at a specific time as a log-linear function of so-called exogenous and endogenous predictors. While exogenous predictors describe actor or dyad attributes (e.g. actors’ age difference), endogenous predictors, such as inertia or reciprocity, summarize characteristics of previous relational events between certain actors at a specific point in time. Despite their flexibility, standard REMs generally assume linear and time-invariant effects, an assumption that is unlikely in empirically observed social networks (Leenders et al., 2016).

Over the course of 4 years, the project will explore several directions for extending the REM framework through the integration of Gaussian processes, aiming to capture nonlinear and time-varying effects in temporal networks. This ranges from the integration of Gaussian processes for the estimation of nonlinear effects of exogenous and endogenous predictors to the specification of Gaussian process priors for nonlinear memory decay functions, applying lower weights for the impact of events that occurred a long time ago.

This research aligns with the mission of IOPS by developing extensions to the REM framework, assessing their statistical properties and evaluating their value for real-world social network data.

Supervisors
Prof. Dr. Ir, Joris Mulder
Prof. Dr. Roger Leenders

Financed by
European Research Council (ERC)

Period
1 October 2025 – 30 September 2029