Constructing, conducting and interpreting animal social network analysis

Abstract

1. Animal social networks are descriptions of social structure which, aside from their intrinsic interest for understanding sociality, can have significant bearing across many fields of biology. 2. Network analysis provides a flexible toolbox for testing a broad range of hypotheses, and for describing the social system of species or populations in a quantitative and comparable manner. However, it requires careful consideration of underlying assumptions, in particular differentiating real from observed networks and controlling for inherent biases that are common in social data. 3. We provide a practical guide for using this framework to analyse animal social systems and test hypotheses. First, we discuss key considerations when defining nodes and edges, and when designing methods for collecting data. We discuss different approaches for inferring social networks from these data and displaying them. We then provide an overview of methods for quantifying properties of nodes and networks, as well as for testing hypotheses concerning network structure and network processes. Finally, we provide information about assessing the power and accuracy of an observed network. 4. Alongside this manuscript, we provide appendices containing background information on common programming routines and worked examples of how to perform network analysis using the r programming language. 5. We conclude by discussing some of the major current challenges in social network analysis and interesting future directions. In particular, we highlight the under-exploited potential of experimental manipulations on social networks to address research questions.

Keywords: fission-fusion dynamics; group living; methods; social behaviour; social dynamics; social network analysis; social organisation.

Figure 1

Primary steps and key considerations in the collection and analysis of animal social networks (see also Table 1).

Figure 2

Thresholding networks can have significant impact on the structure and statistical properties of a network. (a) This network of bottlenose whales (Hyperoodon ampullatus), based on observations of groups of animals made at sea off Nova Scotia from 1988 to 2003 (see Gowans, Whitehead & Hooker 2001 for methodology), was calculated using the half‐weight index (hwi), to account for potentially missed observations of individuals in groups. This network was then thresholded at (b) half the mean hwi, (c) at the mean hwi (d) and at twice the mean hwi. Nodes are coloured by community (detected using leading eigenvector communities; Newman 2006) and sized by their degree (strength in the original network, binary degree in the others). This figure highlights how thresholding can lead to unpredictable results, such as individual ‘x’ changing communities, and varying relationships between node properties (such as correlations between node measures, e). For example, the correlation of individuals’ rank in terms of strength between the networks (a) and (b) is only 0·57, and the relationship has an R ² of only 0·28.