Abstract
Social network analysis provides a useful lens through which to view the structure of animal societies, and as a result its use is increasingly widespread. One challenge that many studies of animal social networks face is dealing with limited sample sizes, which introduces the potential for a high level of uncertainty in estimating the rates of association or interaction between individuals. We present a method based on Bayesian inference to incorporate uncertainty into network analyses. We test the reliability of this method at capturing both local and global properties of simulated networks, and compare it to a recently suggested method based on bootstrapping. Our results suggest that Bayesian inference can provide useful information about the underlying certainty in an observed network. When networks are well sampled, observed networks approach the real underlying social structure. However, when sampling is sparse, Bayesian inferred networks can provide realistic uncertainty estimates around edge weights. We also suggest a potential method for estimating the reliability of an observed network given the amount of sampling performed. This paper highlights how relatively simple procedures can be used to estimate uncertainty and reliability in studies using animal social network analysis.
Highlights
Social networks are widely used to investigate the social structure of animal populations [1,2,3,4,5]
We calculated the mean weighted degree across all individuals as a function of sample size. We found that both the Bayesian inference method and the bootstrapped SRI (b-SRI) methods produced useful information about the uncertainty in an association network
Incorporating uncertainty at the level of an edge value enabled us to address two key questions: (i) how well do b-SRI and Bayesian network methods perform at capturing the structure of real underlying networks, and (ii) how can we determine from observation data alone whether a network has been sampled enough to provide accurate results?
Summary
Social networks are widely used to investigate the social structure of animal populations [1,2,3,4,5]. There are no existing methods for estimating whether a set of observations reliably captures the structure of the real set of interactions (or the real network). We develop an approach based on Bayesian inference to estimate rates of association or interaction between pairs of individuals, as well as their uncertainty. We compare our method to an existing approach of computing index-based social networks from observed data and estimating uncertainty using bootstrapping [9,10,11]. We demonstrate how both these methods can be used to determine how accurately the associations among individuals and social network properties have been captured in empirical studies
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
