Abstract:
|
Group-based social dominance hierarchies are of essential interest in animal behavior research. Experimental studies often collect aggressive interactions data observed over time and researchers are interested in understanding how the underlying social hierarchy is established and dynamically evolves. Traditional ranking methods summarize interactions across the observation period and rely on aggregate counts. Instead, we take advantage of the timestamps of the interactions and propose a network point process model with latent ranks. We carefully motivate the form of this model so that it can incorporate important characteristics of animal interaction data, such as the winner effect, bursting and pair-flip phenomena. We apply the model to simulation and real data. With a suite of statistically developed diagnostic perspectives, we demonstrate that this model outperforms comparison models, in terms of recovering the underlying rankings, capturing relevant network structure and providing meaningful predictions.
|