Determining Temporal Stability in Dominance Hierarchies

The importance of social hierarchies has led to the development of many techniques for inferring social ranks, leaving researchers with an overwhelming array of options to choose from. Many of our research questions involve longitudinal analyses, so we were interested in a method that would provide reliable ranks across time. But how does one determine which method performs best?\n\nWe attempt to answer this question by using a training-testing procedure to compare 13 different approaches for calculating dominance hierarchies (seven methods, plus 6 analytical variants of these). We assess each methods performance, its efficiency, and the extent to which the calculated ranks obtained from the training dataset accurately predict the outcome of observed aggression in the testing dataset.\n\nWe found that all methods tested performed well, despite some differences in inferred rank order. With respect to the need for a \"burn-in\" period to enable reliable ranks to be calculated, again, all methods were efficient and able to infer reliable ranks from the very start of the study period (i.e., with little to no burn-in period). Using a common 6-month burn-in period to aid comparison, we found that all methods could predict aggressive outcomes accurately for the subsequent 10 months. Beyond this 10-month threshold, accuracy in prediction decreased as the testing dataset increased in length. The decay was rather shallow, however, indicating overall rank stability during this period.\n\nIn general, a training-testing approach allows researchers to determine the most appropriate method for their dataset, given sampling effort, the frequency of agonistic interactions, the steepness of the hierarchy, and the nature of the research question being asked. Put simply, we did not find a single best method, but our approach offers researchers a valuable tool for identifying the method that will work best for them.\n\nHighlightsO_LIAll ranking methods tested performed well at predicting future aggressive outcomes, despite some differences in inferred rank order.\nC_LIO_LIAll ranking methods appear to be efficient in inferring reliable ranks from the very start (i.e., with little to no burn-in period), but all showed improvement as the burn-in period increased.\nC_LIO_LIUsing a common 6-month burn-in period, we found that all methods could predict aggressive outcomes accurately for the subsequent 10 months. Beyond this threshold, accuracy in prediction decreased as the testing dataset increased in length.\nC_LIO_LISwitching to a data-driven approach to assign k-values, via the training/validation/testing procedure, resulted in a marked improvement in performance in the modified Elo-rating method.\nC_LI