Improving methods for identifying pair bonds in animal social networks

In social animals, maintaining strong bonds with conspecifics can be important for reproduction and survival. However, it remains unclear how best to quantify these bonds. Existing social network techniques typically focus on building robust overall networks of co-occurrences but approaches that maximize whole-network accuracy may not be optimal for detecting the most important bonds between individuals. We used data from winter social interactions in a population of spotless starlings, Sturnus unicolor , fitted with passive integrated transponder tags and detected at feeders using radio frequency identification loggers to evaluate existing methods and develop new methods for extracting co-occurrence data from temporal streams of detections. We aimed to develop a method capable of extracting key social bonds: pairs of males and females that bred in consecutive springs. Our results show that although all approaches can differentiate most known pair bonds from other social relationships, methods that focus on capturing fine-scale associations are most effective at delineating them. We then applied this method to birds that were found breeding in spring with a new partner, revealing that in over half of these newly formed pairs, individuals within a pair were each other's strongest social associate during the preceding winter. Applying a discriminant function analysis to the best performing network method demonstrated good sensitivity for identifying true pair bonds. However, we also found substantial variation in social association patterns among pairs, including differences in the formation and dissolution of bonds across the nonbreeding season. Our method improves on previous approaches while also adding to the growing evidence that pair bonding is a complex process that can begin well before the onset of breeding.