Post provided by Damien Farine, Sebastian Sosa, David Jacoby, Mathieu Lihoreau and Cédric Sueur
Here at Methods in Ecology & Evolution and the Journal of Animal Ecology we are excited by the new directions that the next decade of research into animal social networks will bring. We hope to encourage new advances in the study of animal social networks by calling for high-quality papers for a cross-journal Special Feature on animal social networks. Below, Damien Farine and the Special Feature Guest Editors have reviewed some areas of animal social network research that deserve particular attention.
Generating Network Structures
There is plenty of social network science to keep empiricists busy. Despite hundreds of papers having been published on animal social networks, we still know almost nothing about the processes that generate network structure. How do relationships among individuals form? How much of a role do social factors play in the transmission of traits from one generation to the next? What are the evolutionary explanations for the growing evidence that animals seem to alter their behaviour to conform with their social environment? Our general lack of understanding of such processes limits our ability to draw meaningful conclusions from the growing literature describing patterns of social structure. An illustrative example is the question of how limited individuals are by their cognitive processing. For example, can comparing network complexity in insects and primates tell us about the cognitive capacities necessary to express efficient and complex networks? The number of relationships individuals can track will undoubtedly be much lower than the number of connections captured in the structure of their social network.
One type of interaction that has long been central to the study of social behaviour, and to which networks should have a lot to offer, is cooperation. A key prediction to explain how cooperation can be maintained in populations is assortativity. If cooperators interact predominantly with other cooperators, then the presence of non-cooperators in a population would have little effect on their fitness. Assortment is easily measured using networks but much more work is needed to understand mechanisms that shape patterns of assortativity in networks, including the influence of actions such as cooperation on the network. We should strive to identify the mechanistic rules that can explain emergent network structure in animal populations. For example, in populations where cooperators interact more often with one another, is assortativity the result of partner choice (or grouping decisions), or can other types of social interactions generate opportunities to passively develop reciprocal relationships?
In the broader context of social evolution, networks should play a central role in the study of social selection. Social networks are particularly useful in characterising the unique social environment that each individual experiences, thereby allowing these to be linked to their fitness. Global properties of social networks also can fundamentally shape where and how selection operates. Much more could be learnt from the formal integration of network approaches into models of social evolution. For example, social networks could allow us to generate robust estimates of the effective group each individual experiences, which is a contentious issue in the debate on the relative merits of kin-selection versus multi-level selection. On key empirical question is whether global network properties can play a role in evolutionary or co-evolutionary dynamics. For example, whilst network structure can clearly generate selection on virulence, can parasites also predictably shape global properties of a network?
One question that often emerges in discussions of spread dynamics through networks is the relative importance of the social position of the initial source of information or a disease. That is, does it matter whether a newly infected node is at the core or at the periphery of a network? To date there has been no formal model of this seemingly foundational question. Intuitively, it seems obvious that emergence in a peripheral node will result in a slower spread. But what is the expected size of such an effect (i.e. core vs periphery in exactly the same network)? Probabilistically, given all nodes are more likely to be connected to a core (i.e. central) node in the network, then transmission is expected to rapidly spread to the core of the network, at which point the peripheral node will no longer have any influence on the global outcome of an epidemic.
A second key question is whether social transmission, particularly of information, through animal networks follows a process of simple versus complex contagion (the latter requiring reinforcement by several nodes before a transmission event occurs). If this is the case, then the influence of a node being at the core versus the periphery of the network on spread dynamics could change dramatically. Complex contagion has further interesting properties, including through positive feedbacks that can cause hysteresis effects.
There are likely to be many types of feedback mechanisms operating in networks. One obvious feedback is between catching a disease (usually more likely at the centre of a network) and network position (becoming less central when sick). Such feedbacks have the potential to either strengthen the evolutionary potential of a selective agent or negate its effects completely. Identifying feedbacks is becoming increasingly possible, facilitated by new techniques for long-term tracking of individuals, experimental manipulations of network structure, and manipulations of individual state. However, examples remain rare in the literature.
Response to Disturbances
Finally, tracking the response of social networks to disturbances ranging from habitat changes (e.g. fire) to disease, should be a key research priority. Social networks have the potential to play an important role in understanding the resilience of biological systems. Yet, we know almost nothing about how animal social networks adapt, whether they have tipping points, or how populations might change when faced with demographic changes. Similarly, very little consideration has been made on how changes in network structure could shape other biological processes, such as life history evolution. The role of networks in evolution generally remains unexplored.
There are a wide variety of network metrics (node-based, dyadic, and global) and the application and development of new metrics continue to evolve. It is crucial to consider how the values generated by a network metrics (new and old) are interpreted biologically and recognize their limitations. It would be useful to have manuscripts that address questions about:
- How mathematical definitions of different network metrics translate to biological processes;
- Which metrics provide similar, redundant, or unique information relative to other metrics.
Null models that control for data dependency are part of many network analyses. These are used to evaluate whether an observed network metric is extreme relative to that expected in randomly-assembled networks. However, despite some refinements, such as how to apply null models to data collected using biologgers, the design and application of Bejder et al.’s approach to generating random networks using permutations has changed little in two decades. Bejder et al.’s method was shaped by limitations in computational power that no longer exist; we can now use available computational power to develop and use more robust permutation-based methods for generating random networks. Further, their application to different models for describing data, such as exponential random graph models (ERGMs) or network-based diffusion analyses (NBDA) also warrants exploration. Some suggested directions that could be explored include:
- Does combining are permutation tests with existing statistical models (such as ERGMs or NBDA) increase the robustness of hypothesis testing?
- Data-stream permutation tests can be used for almost any type of network, but for many types of data (e.g. directed networks) the algorithms are not yet described.
- Would running more replications of a null model to draw probability distributions or using a burn-in phase for the null model to generate more robust P values?
- Given recent advances enabling the collection of relatively complete records of social behaviour, when do we still need null models?
- Is it possible to use information-theoretical approaches to test hypotheses using network data?
- How do we deal with multiple hypothesis testing using social network data?
Methods to construct networks have also changed little over the past decades. Yet with developments in statistical inference methods, and greater acknowledgement of the role that sampling uncertainty plays in shaping data collection and the inferences drawn from data, much could be done to improve on these methods. While some recent suggestions appear promising, they require extensive testing and translation into useable techniques for animal social networks. We are currently under-using the information we have when constructing networks. A clear advance would be to use global information about observations to generate more accurate predictions of edge weights (connection strength between two individuals) or to identify erroneous observations (which have a large impact on network robustness). Questions to start with include:
- Can we use data from unmarked individuals to quantify the uncertainty of individual-level network metrics?
- Can we infer population network structure when identities of individuals are only known temporarily?
- Can we identify the directionality of interactions based on a sequential stream of observed interactions?
- Can we construct networks from an observed process?
- Can data from the observation of other individuals tell us more about the relationship between two focal individuals?
- How do we move beyond using networks to describe patterns and instead use networks to identify causal relationships?
- Can networks collected with different protocols can be compared?
Network comparisons and limitations of data collection
A major limitation to testing questions about the evolution of social networks is the inability to conduct comparative network studies (but see this article by Sah et al.). How data are collected (both the number and kind of observations) fundamentally affects the properties of networks, which has made comparison of networks almost impossible. But to what extent can we control for data collection methods and other factors that impact network metrics (such as group size and population density)?
Simulating different observational datasets from the same ‘real’ data or experimental comparisons would be important steps in testing the extent to which networks vary according with observation methods and effort. This could be a simple way to start getting some idea of the scale of the comparison problem. Null models are also an avenue worth exploring as they could provide an expected distribution of differences between networks that accounts for factor such as data collection methods.
One application of networks, especially network comparison, is to the study of social complexity. In many animal groups, individuals have repeated social interactions, often including different types of interactions or across different modalities. However, we are still limited by relying on descriptive approaches rather than attempted to uncover causal mechanisms underlying complexity in animal societies. Recent calls have been made to make use of multi-layered (or multiplex) network methods to better capture the role of such complexity on biological structure and processes. But are multi-layered networks, and the software to analyse them, really more robust? Studies are needed that robustly test the application of these methods under realistic levels of uncertainty in data collection and network construction.
Many animal societies also exhibit complex patterns scaling from individuals to social units to population structure, or so-called ‘multi-level societies’. Networks could be a ‘go-to’ tool to identify and statistically test for the presence of multiple levels of non-random structural patterns. To date there is no well-established method for doing this.
A general motivation for constructing, analysing, and comparing networks is to understand population processes such as the spread of information or diseases. Network tools, such as network-based diffusion analysis, have facilitated the study of information transmission, but to date there is no easy way to infer from network data alone whether the process of transmission is active (e.g. individuals are recruited) or passive (two individuals are more likely to be in same vicinity). Networks also can be used to test questions about other population process, such as the spread of physiological states. Developing appropriate mathematical models to capture such processes warrants greater attention.
In the context of disease transmission, there is also little consensus about the best ways to use network data to identify the relative importance of different types of network connections. Further, individual animals are likely to change behaviour given a change of state, causing a rather large headache when using networks collected either before or after a spreading process. However, this presents not only a challenge but also an opportunity: can we infer a change of state from changes in the network properties?
Change in-and-of-itself can be challenging to detect and quantify, and changes in one individuals’ behaviour will by definition change the properties of other individuals’ networks. Such fundamental properties of dynamic networks warrant much greater methodological considerations.
Social Networks and Adaptation
Major challenges also remain about how to use social networks in the study of adaptation. Even with several recent technological advances now allowing for the long-term collection of fine-scale interactions, there remain methodological problems with relating fitness outcomes with social network histories if a fitness measure is the outcome of two or more individuals’ histories.
For example, in species with biparental care, how do we determine the importance of one parent’s network position relative to the others? Some approaches from population ecology, such as ‘de-lifing’, could warrant exploration in this context. This points to a broad problem of how to derive appropriate measures of fitness in multi-generational interaction populations. For example, not every interaction an individual has had will matter, and interactions might not have the same meaning for all actors involved. A substantial methodological gap is methods for determining at which times or spatial scales social interactions have greatest impact on fitness.
Moving Towards Network Prediction
Ultimately, the power of network analysis is as a predictive tool. A well-developed prediction study can allow causal inferences to be made and tested. Being able to predict the dynamics of networks will also provide a greater understanding of the building blocks of animal societies, and therefore allow us to estimate their resilience to disturbances. However, we are still very much limited in our ability to quantitatively describe how a network changes over time. Network prediction therefore remains possibly the largest methodological gap in the animal social network toolbox, and one of some importance given the ever-increasing amount of disturbance of animal populations.
In this blog post, I have highlighted just some of the exciting opportunities and challenges that lay ahead in our continued discovery of the building blocks of animal societies. I hope these, and many more I have undoubtedly overlooked, will stimulate new lines of enquiry and improved methodological approaches. We welcome submissions addressing the methodological gaps to Methods in Ecology & Evolution and submissions addressing the empirical gaps to Journal of Animal Ecology to form a joint special feature.
This Special Feature was proposed following the first symposium on Animal Behavioural Network Analysis at the XXXVIII Sunbelt Conference in Utrecht in 2018 organized by Sebastian Sosa, Mathieu Lihoreau, David Jacoby and Cédric Sueur. This symposium presented new concepts and tools for social network analysis applied within different areas of animal research.
Joint Special Feature Details
Manuscripts should be submitted in the usual way through the Journal of Animal Ecology or Methods in Ecology and Evolution websites. Submissions should clearly state in the cover letter accompanying the submission that you wish the manuscript to be considered for publication as part of this Special Feature. Full details on the Open Call are here, pre-submission enquiries are not necessary, but any questions can be directed to: firstname.lastname@example.org or email@example.com
The deadline for submission is: Monday 26 August.