The causal analysis of animal social networks

Behavioural ecologists aim to understand the causes of animal social network structure. To do this, they formulate theoretical models—whether verbally or formally—proposing causal mechanisms that explain the observed structure. These mechanisms can then be tested empirically, by assessing their compatibility with social network data collected in wild or captive populations of animals. This inferential task is, however, extremely challenging. Observed social interactions are often noisy, and the causal mechanisms of interest can be confounded by biological factors, or by the sampling process. Recent research has highlighted that common methods in the field (e.g., network permutations, covariate selection based on predictive criteria) fail to effectively address these challenges, often leading to wrong conclusions. More broadly, these issues reflect a growing disconnect between theoretical and empirical research in the field. In this talk, I will propose an alternative inferential framework. My framework integrates tools from the field of formal causal inference (e.g., Directed Acyclic Graphs) and probabilistic modelling (e.g., Bayesian multilevel models), while drawing on models from network science and behavioural ecology. I provide a workflow for empiricists to first translate their theoretical domain expertise into formal causal assumptions, and second, derive and evaluate statistical models from these assumptions. I showcase how doing so addresses the challenges posed by the inherent noise and confounding factors in animal social network analysis, and explain why causal inference in such systems cannot be achieved without an explicit theoretical grounding. More generally, my framework lays the groundwork for a stronger, more transparent, and more rigorous bridge between theoretical and empirical research in behavioural ecology, as well as in the broader context of the social, ecological and evolutionary sciences.