A causal framework for the drivers of animal social network structure

A major goal of behavioural ecology is to explain how phenotypic and ecological factors shape the networks of social relationships that animals form with one another. This inferential task is notoriously challenging. The social networks of interest are generally not observed, but must be approximated from behavioural samples. Moreover, these data are highly dependent: the observed network edges correlate with one another, due to biological and sampling processes. Failing to account for the resulting uncertainty and biases can lead to dysfunctional statistical procedures, and thus to incorrect results. Here, we argue that these problems should be understood—and addressed—as problems of causal inference. For this purpose, we introduce a Bayesian causal modelling framework that explicitly defines the links between the target interaction network, its causes, and the data. We illustrate the mechanics of our framework with simulation studies and an empirical example. First, we encode causal effects of individual-, dyad-, and group-level features on social interactions using Directed Acyclic Graphs and Structural Causal Models. These quantities are the objects of inquiry, our estimands. Second, we develop estimators for these effects—namely, Bayesian multilevel extensions of the Social Relations Model. Third, we recover the structural parameters of interest, map statistical estimates to the underlying causal structures, and compute causal estimates from the joint posterior distribution. Throughout the manuscript, we develop models layer by layer, thereby illustrating an iterative workflow for causal inference in social networks. We conclude by summarising this workflow as a set of seven steps, and provide practical recommendations.Author summary: Behavioural ecologists ask mechanistic questions about behaviour—causal questions. When studying animal societies, these questions often concern the drivers of social network structure. Addressing causal questions from observed social interactions, whether in wild or captive settings, poses serious inferential challenges. Social network data are often noisy and biased, and causal effects may be confounded. As a result, estimating the effects of interest requires careful causal and probabilistic modelling—tools that most empiricists in the field are not trained to use. By integrating techniques from causal inference and Bayesian statistics, we introduce a practical framework for researchers to conduct causal inference in their own study system. We start by distinguishing three levels of abstractions for any social network under scrutiny. We then outline an iterative workflow, built around a few key steps: (i) defining the causal effect of interest; (ii) translating one’s domain expertise into qualitative, then quantitative causal assumptions; (iii) building a statistical model designed to estimate that effect. Throughout, we emphasise the justification and validation of statistical models, while offering guidance for readers who are unfamiliar with formal modelling. More broadly, our framework lays the groundwork for a stronger and more transparent bridge between theoretical and empirical research in behavioural ecology.