BISoN: A Bayesian Framework for Inference of Social Networks

Abstract

Social networks are often constructed from point estimates of edge weights. In many contexts, edge weights are inferred from observational data, and the uncertainty around point estimates can be affected by various factors. Though this has been acknowledged in previous work, methods that explicitly quantify uncertainty in edge weights have not yet been widely adopted, and remain undeveloped for common types of data. Furthermore, existing methods are unable to cope with some of the complexities often found in observational data, and do not propagate uncertainty in edge weights to subsequent analyses. We introduce a unified Bayesian framework for modelling social networks based on observational data. This framework, which we call BISoN, can accommodate many common types of observational social data, can capture confounds and model effects at the level of observations, and is fully compatible with popular methods of social network analysis. We show how the framework can be applied to common types of data and how various types of downstream analyses can be performed, including non-random association tests and regressions on network properties. Our framework opens up the opportunity to test new types of hypotheses, make full use of observational datasets, and increase the reliability of scientific inferences. We have made example R code available to enable adoption of the framework.

Cost and benefits of being social: examining the influence of sociality on faecal parasites in free-ranging rhesus macaques

“…We generated undirected weighted Bayesian networks using the BISoN framework and bisonR package [58]. This framework allowed us to account for uncertainty in the edges connecting individuals in the network based on how often they were sampled and, more importantly, propagate this uncertainty to subsequent analyses.…”

“…For the HH grooming network, edges represented the time a pair of individuals engaged in grooming interactions relative to the total time that dyad was observed. Given that there is no natural statistical model for duration data [58], the time spent grooming and the sampling effort for a dyad were converted to counts by dividing each of these terms by the length of a focal period (5-mins) to make sure each count represented independent sampling events.…”

Modelling human and non-human animal network data in R using STRAND

“…While this approach of weighting counts by an exposure variable is significantly better than ignoring variation in risk of observation (Farine and Whitehead 2015), construction of a simple ratio divides out sample size information, leading dyadic observations based on little data to carry disproportionate weight in downstream analyses (see McElreath 2020; Hart et al 2021b, for a review of this issue). Moreover, zeros arising from censoring (i.e., due to members of a dyad being unavailable; “denominator zeros”) are often confounded with true zeros (i.e., members of a dyad being present but not interacting; “numerator zeros”).…”

“…For example, it is common for researchers to regress outgoing ties on incoming ties to estimate reciprocity (e.g., Carter and Wilkinson 2013), but such regressions are known to suffer from residual confounding (see Koster and Leckie 2014). Similarly, it is common for researchers to correct for sampling effort by creating a Simple Ratio Index (SRI; Cairns and Schwager 1987; Whitehead and James 2015; Farine and Whitehead 2015), but such indices are well-know to divide out sample size and give the weakest data-points disproportionate weight in downstream analyses (Hart et al 2021b).…”

“…Our approach to network modeling here complements that of the team (Hart et al 2021b), who have also developed a suite of network analysis models using Bayesian methods. Our R package provides additional functionality to typical end-users, however, in that it integrates new Bayesian network analysis models on the ‘back-end’ with a user-friendly ‘front-end’ interface, that allows even casual R users to specify complex network models (which may include a variety of block-, individual-, and dyad-level covariate data) on-the-fly, using nothing more than base R syntax.…”