Abstract
Producing accurate and reliable inference from animal social network analysis depends on the sampling strategy during data collection. An increasing number of studies now use large-scale deployment of GPS tags to collect data on social behaviour. However, these can rarely capture whole populations or sample at very high frequencies. To date, little guidance exists when making prior decisions about how to maximise sampling effort to ensure that the data collected can be used to construct reliable social networks. We use a simulation-based approach to generate a functional understanding of how the accuracy of various network metrics is affected by decisions along three fundamental axes of sampling effort: coverage, frequency and duration. Researchers often face trade-offs between these three sampling axes, for example due to resource limitations, and thus we identify which axes need to be prioritised as well as the effectiveness of different deployment and analytical strategies. We found that the sampling level across the three axes has different consequences depending on the social network metrics that are estimated. For example, global metrics are more sensitive than local metrics to the proportion of the population tracked, and that among local metrics some are more sensitive to sampling duration than others. Our research demonstrates the importance of establishing an optimal sampling configuration for drawing relevant and robust inferences and presents a range of practical advice for designing GPS based sampling strategies in accordance with the research objectives.