The problem of unmeasured variables in animal social network analysis: can edge-based multilevel models provide a reliable solution?

Abstract

Recent approaches to analysing social networks suggest that modeling the edges of the network and using multilevel models will produce more informative estimates. These recent methods have been proposed as a way to better handle the dependency structures of social networks, account for biases in data collection, and retain uncertainty when making inferences about social network structures. We find that they have the potential to also effectively handle unmeasured variables that act as statistical confounds in social network analysis. Using simulated data, we highlight that static social network analyses can be used to identify patterns in social networks, but generally cannot be used to identify the underlying mechanisms behind the patterns. To identify mechanisms, we show that taking a dynamic approach and using edge-based models with additive and multiplicative random effects provides a means to identify mechanisms even when statistical confounds are present. Additive and multiplicative random effects also provide information about social structures not captured by the predictors, facilitating exploratory analysis. We suggest that a keep-it-maximal approach for random effects structures is beneficial for edge-based multilevel models of social networks, and that such approaches can be particularly effective when there are unmeasured variables that are not captured by model predictors.