How can social network analysis improve the study of primate behavior?

American Journal of Primatology 73:703–719 (2011)
REVIEW ARTICLE
How Can Social Network Analysis Improve the Study of Primate Behavior?
CE´DRIC SUEUR1– 3�, ARMAND JACOBS3,4, FRE´DE´RIC AMBLARD5, ODILE PETIT2,4, AND ANDREW J. KING6
1Department of Ecology and Evolutionary Biology, Princeton University, Princeton, New Jersey
2Unit of Social Ecology, Free University of Brussels, Brussels, Belgium
3Primate Research Institute, Kyoto University, Kyoto, Japan
4Department of Ecology, Physiology, and Ethology, IPHC, CNRS-UDS, Strasbourg, France
5IRIT, Institute of Research in Computer Science of Toulouse, Paul Sabatier University, Toulouse, France
6Structure and Motion Laboratory, Royal Veterinary College, University of London, Hertfordshire, United Kingdom
When living in a group, individuals have to make trade-offs, and compromise, in order to balance the
advantages and disadvantages of group life. Strategies that enable individuals to achieve this typically
affect inter-individual interactions resulting in nonrandom associations. Studying the patterns of this
assortativity using social network analyses can allow us to explore how individual behavior influences
what happens at the group, or population level. Understanding the consequences of these interactions
at multiple scales may allow us to better understand the fitness implications for individuals. Social
network analyses offer the tools to achieve this. This special issue aims to highlight the benefits of social
network analysis for the study of primate behaviour, assessing it’s suitability for analyzing individual
social characteristics as well as group/population patterns. In this introduction to the special issue, we
first introduce social network theory, then demonstrate with examples how social networks can
influence individual and collective behaviors, and finally conclude with some outstanding questions for
future primatological research. Am. J. Primatol. 73:703–719, 2011. r 2011 Wiley-Liss, Inc.
Key words: interaction; association; social system; social structure; methodology; behavioral
sampling
INTRODUCTION
Many animal species live in groups. Commonly
cited benefits to sociality include decreased predation
risk through better detection of predators or
‘‘selfish herd’’ effects, and increased foraging
efficiency through better acquisition and defense of
food resources [Alexander, 1974; Hamilton, 1971;
Wrangham, 1980]. Conversely, living in group gener-
ates competition where resources are limited in space
or time [Janson & Goldsmith, 1995] that can result
in increased rates of disease transmission through
closer/more frequent contact among individuals
[Freeland, 1976, 1979; Huffman & Chapman, 2009;
Nunn et al., 2006]. Moreover, living in group can be
difficult to maintain where individuals differ in their
morphological–physiological state [Krause & Ruxton,
2002]. Thus, individuals face a trade-off and have to
balance both the advantages and disadvantages of
group life [Conradt & Roper, 2005].
Trade-offs associated with group living can
emerge at the individual, group, or population level,
and the costs and benefits associated with these
trade-offs can be understood by scrutinizing the
associations and interactions between individuals at
each of these levels. In this review, we use ‘‘associa-
tion’’ to describe a situation where two or more
individuals share the same space at the same time,
and ‘‘interaction’’ to describe behavior directed from
one individual to another (e.g. grooming). We use
relationship when we wish to refer to either associa-
tion, interaction, or both. In almost all social species,
individuals are known to associate in a nonrandom
way [honeybees Apis mellifera: Naug, 2009; guppies
Poecilia reticulata: Morrell et al., 2008; Colombian
ground squirrel Spermophilus colombianus: Manno,
2008; African elephants Loxodonta Africana:
Wittemyer et al., 2005; bottlenose dolphins Tursiops
truncatus: Lusseau, 2003; Japanese macaques
Macaca fuscata: Koyama, 2003; yellow baboons
Papio cynocephalus: Silk et al., 2004], resulting in
unevenly distributed social interaction. This can lead
to distinct patterns of subgrouping, and these
subgroups can be nested within larger collectives at
the group or population level.
Published online 22 December 2010 in Wiley Online Library
(wileyonlinelibrary.com).
DOI 10.1002/ajp.20915
Received 26 July 2010; revised 19 November 2010; revision
accepted 20 November 2010
Contract grant sponsors: Franco-American Commission; The
Fyssen Foundation; The Japan Society for the Promotion of Science.
�Correspondence to: Ce´dric Sueur, IPHC– DEPE, Ethologie
des Primates, 23 rue Becquerel, Strasbourg 67000, France.
E-mail: cedric.sueur@c-strasbourg.fr
r 2011 Wiley-Liss, Inc.

Primates—the subjects of this review and this
special issue—are known to display nonrandom
associations and interactions. Studies of these rela-
tionships can help us to understand how ecological
factors shape social behavior [Crook & Gartlan,
1966; Gray, 1987; Thierry et al., 2004; Wrangham,
1987], and ultimately, the evolution of sociality
[Hinde, 1976; Krause & Ruxton, 2002; Whitehead,
2008]. Patterns of association can emerge as a result
of individuals sharing identical or similar motiva-
tions (levels of hunger, thirst, or response to threat).
Such assortativity can reduce some of the costs of
sociality [Ramos-Fernandez et al., 2006; Ward et al.,
2008], since group cohesion can be ‘‘easier’’ to
achieve where individuals have fewer conflicts of
interest [Conradt & Roper, 2000; King & Cowlishaw,
2009a]. However, in cases in which sociality results
in asymmetric payoffs for individuals (where indivi-
duals are required to group to reduce predation risks
[Hill & Lee, 1998], but experience competition over
resources [Janson & Goldsmith, 1995], for example),
repeated interactions that result in stable social
bonds may buffer the effects of competition [de Waal,
1986]. Indeed, social interaction can have a profound
influence on patterns of association [Sterck et al.,
1997; Thierry et al., 2004; Wrangham, 1987]. In this
context, variability in social structure can be under-
stood in terms of an individual adapting to its social
and ecological environment.
The problem is that traditional primate research
focuses on local dyadic associations and interactions,
and we have little knowledge of how these dyadic
relationships scale to a more global social structure.
In order to understand the complexity of a social
structure, and link the behavior of individuals with
the functioning and efficiency of the dynamic group-
level properties [Camazine et al., 2003; Couzin &
Krause, 2003], we need to consider and analyze all
relationships linking all group members [Croft et al.,
2005; Flack et al., 2006; Wey et al., 2008; Whitehead,
2008]. Only then can we begin to explore variation in
individual fitness in the context of social behavior.
Analyzing all relationships linking all group
members is not a new concept. Hinde [1976] defined
the social structure of a group as composed by the
nature, quality, and patterning of relationships of its
group members, where the relationship between two
individuals captures their repeated interactions over
time. Nevertheless, most of the studies focusing on
animal groups—especially primates—concern only the
analyses of dyadic relationships, even though the
context of these relationships extends beyond the dyad
of just two individuals [Arnold & Barton, 2001;
Kutsukake & Castles, 2004; Silk, 1999], and specific
individuals (or specific relationships) can be the
mainstay of a social structure [Flack et al., 2006;
Whitehead, 2008]. For instance, if an old and/or
dominant female in a primate system characterized
by matrilineal relationships dies, the relationships
among all the remaining individuals may change
dramatically, and the group may even permanently
split as a result [e.g. Beisner et al., 2010; Koyama,
2003; Lefebvre et al., 2003]. Thus, to reiterate,
relationships need to be studied in the context in
which they evolve, i.e. the social group, and Social
Network Analysis (SNA) offers a framework to do this.
SNA methodology originated from mathematical
graph theory and was soon applied in the social
sciences [see Newman, 2010; Scott, 2000, for a
review]. Early researchers envisaged studying hu-
man social groups in a physics framework, studying
people as if they were ‘‘atoms’’ and relationships
between people as ‘‘social gravitation’’ [see Borgatti
et al., 2009 for details]. Although this representation
did not persist, even metaphorically, SNA remained
an important tool in the social sciences, and is the
basis of some extremely important concepts. For
instance, the concept of ‘‘small worlds’’—networks
whose structure allows a connection between two
random individuals in the network via only a few key
individuals—has resulted in the famous expression
of ‘‘six degrees of separation’’ [Watts, 2004]. More
recently, the availability of large data sets that
describes the interactions between hundreds of
thousands or millions of components (e.g. commu-
nication networks from mobile phones, or the World
Wide Web, or transport networks of roads or
airports) have allowed common signatures to be
identified. Specifically, many of these networks
correspond to a ‘‘scale-free network,’’ where the
distribution of the relationships per individual
follows a power-law, and we discuss the implication
of the scale-free distribution in primate systems later
in this review.
In biology, SNA has contributed to our under-
standing of gene, protein, and cell relationships
[Barabasi & Oltvai, 2004; Laughlin & Sejnowski,
2003], and in the past 10 years, this methodology has
been increasingly applied to the study of animal
behavior [Borgatti et al., 2002; Flack et al., 2006;
Krause et al., 2009; Lusseau, 2003; Whitehead,
2008]. Although this approach is being used more
and more by primatologists [Chepko-Sade & Sade,
1979; Flack et al., 2006; Voelkl & Kasper, 2009;
Table I], studies applying SNA to explain social
relationships (see Table II for a list of methodological
papers) are still surprisingly rare.
Our aim is to highlight the benefits that can be
gained from using SNA to study primates, and
specifically primate behavior [Brent et al., 2011,
provide a more detailed historical perspective on the
use of SNA in primatology]. We deliberately base the
structure of our review on that of previously
successful reviews and books that explore the use
of network approaches to study animal behavior
[Croft et al., 2008; Wey et al., 2008; Whitehead,
2008], updating the information and tailoring our
examples for a primatology audience. We also hope to
Am. J. Primatol.
704 / Sueur et al.

emphasize how SNA methodology can be applied to
analyze both individual social characteristics and
group/population patterns.
BUILDING SOCIAL NETWORKS
The Social Network Framework
An adjacency matrix or sociomatrix (Table III)
is classically the most suitable representation of
social association or interaction data, long used by
primatologists [for recent examples see: Silk et al.,
2002a,b; Sueur & Petit, 2008; Thierry et al., 2004].
The adjacency matrix is also the basis for SNA.
A matrix contains rows and columns defining specific
individuals, subgroups of individuals, groups of a
TABLE I. Examples of Studies on Social Networks in Primates
Study Topic Species Edge attribute Measures
Chepko-Sade and
Sade [1979]
Patterns of group
splitting
Macaca mulatta Fission Fission events, individual
‘‘connector’’
Dow and de
Waal [1989]
Analyzing network
subgroup interactions
Macaca arctoides Grooming, aggression Compactness, isolation
Chepko-Sade
et al. [1989]
Analyzing network of
pre-fission group
M. mulatta Grooming Cluster analysis
Sade [1989] Analyzing centrality in
networks
M. mulatta Grooming Path length
Watts [2000] Grooming reciprocity Pan troglodytes Grooming Matrices correlations
Flack et al. [2006] Policing and network stability Macaca nemestrina Grooming, play Degree, clustering
Sueur and Petit
[2008]
Organization during
collective movements
M. mulatta,
Macaca tonkeana
Proximity,
following
Eigenvector, modularity,
density
McCowan et al.
[2008]
Detecting group
instability
M. mulatta Grooming,
aggression
Degree, reciprocity
Voelkl and Noe¨
[2008]
Propagation of social
information
M. mulatta Grooming Path length, diffusion
analysis
Henzi et al. [2009] Cyclicity in social network Papio cynocephalus Grooming, proximity Clustering
Ramos-Fernandez
et al. [2009]
Analyzing associations
patterns
Atteles geoffroyi Proximity Strength, eigenvector
Franz and Nunn
[2009]
Detecting social learning Macaca fuscata Co-feeding Diffusion analysis
Voelkl and Kasper
[2009]
Emergence of cooperation 30 species Grooming,
proximity
Probability to cooperate
Kasper and Voelkl
[2009]
Global analysis 30 species Grooming,
proximity
Several group and
individual measures
Sueur et al. [2010] Sub-grouping patterns
during fission
M. mulatta,
M. tonkeana
Proximity,
sub-groups
Matrix correlation,
multidimensional scaling
TABLE II. Examples of Methodological Publications
on Social Network Analysis
Study Topics
Cairns and
Schwager [1987]
Comparing associations indices
Whitehead [1997] Analyzing animal social structure
Bejder et al. [1998] Testing associations patterns
Girvan and
Newman [2002]
Properties of community structure
Lusseau and
Newman [2004]
Individual role in network
DeJordy et al. [2007] Visualizing proximity data
Lusseau et al. [2009] Incorporing uncertainty in
social network analysis
TABLE III. Binary Matrix of Relationships Between
Individuals a to s
a b c d e f g h i j k l m n o p q r s
a 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
b 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
c 0 1 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0
d 0 1 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0
e 0 0 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0
f 0 0 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0
g 0 0 0 0 1 1 1 0 1 0 0 0 0 0 0 0 0 0
h 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 0 0
i 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
j 0 0 0 0 0 0 1 0 0 1 0 1 1 0 0 0 0 0
k 0 0 0 0 0 0 0 0 0 1 0 0 1 1 0 0 0 0
l 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
m 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0
n 0 0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 0
o 0 0 0 0 0 0 0 0 0 0 1 1 0 1 1 1 0 0
p 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0
q 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1
r 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 1
s 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1
The network (graph shown in Fig. 1) is unweighted: either individuals are
connected (value of relationships is 1), or they are not connected (value of
relationships is 0).
Am. J. Primatol.
Social Network Analysis in Primates / 705

same species or populations, or different species,
which are commonly referred as actors. The data
contained within the matrix describes relationships
between these actors. Graph theory represents these
actors (called vertices or nodes) and the relationships
between them (called edges) and represents them as
a network (Fig. 1). Attributes such as age, sex, size,
hierarchical rank, species, or even categories such as
‘‘predator’’ or ‘‘prey’’ can also be assigned to the
nodes in a network. For instance, black nodes in
Figure 1 might represent males and white ones
females, or nodes of the same color might belong to
the same matriline, thus providing a visual repre-
sentation of how sex or kinship influences social
structure. The ties between the nodes—the edges—
can be ‘‘positive,’’ ‘‘negative’’ (e.g. affiliative or
aggressive behavior respectively in the case of
individuals), or represent disease transmission or
information flow. This framework is therefore extre-
mely flexible, and can be used to address a variety of
questions concerning sociality in primates (Tables I
and II). We will now describe how networks can be
constructed and analyzed from matrix data, and
what software is available to do this.
Global Network Properties
In an ‘‘undirected network,’’ edges only show
that two nodes are connected (Fig. 2A). Indeed, as
shown in Table III and Figure 1, relationships
between nodes of an undirected network are only
represented by binary values, 1 for connected nodes,
0 for non connected nodes, and the network is
therefore symmetrical. In a ‘‘directed network,’’ an
edge is oriented from an emitter node to a receptor
node, and the matrix is therefore unsymmetrical
(Fig. 2B). Whether a network is directed or undirected,
it is largely a consequence of the type of
edges (associations or interactions) represented,
and Figure 2B provides an example of both directed
and undirected networks. The relationship between
individuals B and C in Figure 2B is symmetrical
since the frequency of behavior directed from B
toward C is equivalent to those from C toward B.
This is typical of association data where an observer
will score the occasions individuals B and C
were together at some defined spatial–temporal
criterion. However, the relationship between A and
C in Figure 2B is asymmetrical since only C
emits behaviors toward A. This asymmetry is
more typical of interaction data where one
individual tends to direct behavior toward another
(e.g. aggression).
Networks can also be ‘‘unweighted’’ or
‘‘weighted.’’ In an unweighted network (Figs. 1 and
3A), each edge is assigned a binary value: 0 if there is
no relationship (and then no edge) and 1 when two
nodes have a relationship (whatever the strength
of relationship between these two individuals).
However, where the frequency and/or duration of
interactions can be recorded, the networks built from
these interactions are weighted (Fig. 3B). When
constructing weighted networks, a normalized value
is often used, since this allows researchers to more
easily compare the relative strength of interaction
across nodes. Otherwise, a filter is applied to a
Fig. 1. Graphical representation of a theoretical social network. A node represents a group member. Nodes of same color belong to the
same subgroup. An edge (line) represents the relationship between two individuals. This social network corresponds to the social
structure of a group of 19 individuals (labeled from a to s: see Table III for the sociomatrix corresponding to this graph). Individual
network measures are presented in Table IV.
Fig. 2. Graphical representation of (A) an undirected and
unweighted three-node network, and (B) a directed and
weighted three-node network. (B) Gives information about the
direction and the strength of interactions between individuals,
while (A) does not.
Am. J. Primatol.
706 / Sueur et al.

weighted network to categorize associations, which is
especially useful when a network is ‘‘full mesh’’ i.e.
all nodes are interconnected. This filter is most often
applied to generate ‘‘above average’’ and ‘‘below
average’’ relationships (but can be set at any level),
and thus reduce the network to different clusters or
(sub)groups of nodes of different relationship
strength [Franks et al., 2010]. Presenting spatial
information in this way can be extremely informative
where researchers are interested in relational data
for animal systems that display high fission–fusion
dynamics [Aureli et al., 2008], and we discuss how to
apply a filter in our next section.
Software to Build Networks
Software used to undertake SNA are SOCPROG
[Whitehead, 2009] (http://myweb.dal.ca/hwhitehe/
social.htm), Ucinet [Borgatti et al., 2002] (http://
www.analytictech.com/ucinet/), or Pajek [De Nooy
et al., 2005] (http://vlado.fmf.uni-lj.si/pub/networks/
pajek/). Each of these tools allows calculation of SNA
statistics (discussed below) and visualization of
networks. Matman [de Vries et al., 1993] (http://
www.noldus.com/animal-behavior-research/products/
matman) is also a useful piece of software for matrix
analyses, and while it does not allow the calculation
of individual network statistics, it is useful to compare
matrices or to test reciprocity of interactions.
When presented with a weighted network, as
described above, Whitehead’s SOCPROG [2008,
2009] makes it possible to test for preferred/avoided
associations (i.e. relationships) between individuals
and also the opportunity to apply a filter to the
network to transform the weighted network (Fig. 3A)
into unweighted (i.e. 0–1) network (Fig. 3B). The
level of preferred relationships (i.e. the filter) can be
set either at a normal significant value (a5 0.05) or
at a value where all individuals are directly or
indirectly connected. In each case, SOCPROG gen-
erates random matrices and assesses whether the
value of an association in the observed matrix is
higher than the generated association in the random
matrices (in 95% of cases, for instance if a5 0.05). If
the observed association is significant, it is named as
preferred association.
NETWORK STATISTICS
Measuring Attributes and Roles of Individuals
Although primatologists have long been inter-
ested in the ‘‘roles’’ that certain individuals have in a
social context, e.g. measuring agonistic interactions,
and assessing dominance rank [Aureli & de Waal,
2000; Chepko-Sade et al., 1989; Schino, 2001], we are
now in a position to rigorously quantify and describe
these roles using SNA. Brent et al. [2011] provide a
discussion of the similarities and differences between
traditional and modern methods of SNA, and there
are now multitudes of statistics that represent an
individual ‘‘influence’’ upon a network. The simplest
of these is the node degree [Newman, 2003]. The
node degree is a statistic that describes the number
of edges (relationships) that a node (an actor) is
involved in whatever the edges’ values (see Table IV
for degree of each node represented in Fig. 1). For
directed networks representing asymmetrical rela-
tionships, one can distinguish between the in-degree
(the number of edges for which the node is the
receptor) and the out-degree (the number of edges for
which the node is the emitter). This distinction may
be important when studying the reciprocity of
interactions such as grooming or aggression where
Fig. 3. Networks of leader–follower associations during collective
movements for a group of rhesus macaques [see Sueur & Petit,
2008, for details]: (A) all relationships are represented (weighted
networks); (B) only preferred relationships are represented
(unweighted networks). Preferred relationships were obtained
using the avoided/preferred associations test in SOCPROG
[Whitehead, 2008, 2009]. Network graphs were drawn using
Netdraw in UCINET 6.0 [Borgatti et al., 2002]. Nodes represent
individuals. The number represents the hierarchical rank of the
individual. Distance between individuals represents the Half
Weight Index (HWI: the more closely associated two individuals
are, the more frequently they associated they are during
collective movements) and the graph was drawn using multi-
dimensional scaling [Borgatti et al., 2002; Whitehead, 2008,
2009]. Similar shapes characterize individuals having the same
matriline and identical colors define individuals belonging to
the same subgroup during collective movements, which was
defined by the modularity clustering method [Newman, 2004;
Whitehead, 2008, 2009]. The size of a node is directly related to
the individual eigenvector centrality coefficient (the higher the
centrality coefficient is, the greater the influence of the
individual in the joining of group members is, that is, central
individuals are more often followed).
Am. J. Primatol.
Social Network Analysis in Primates / 707

nodes represent individuals.
Another simple measure taking into account the
edges’ values is the strength, which is the sum of a
node’s edges’ values (see Table IV for strength of
each node represented in Fig. 1). For unweighted
networks, the number of edges a node possesses can
be used as an indication of node strength.
Although node degree and strength take into
account the relationships of a node with its direct
neighbors, other indices allow computation of the
impact of a single node on the overall network
structure. For instance, the node n in Figure 1 might
be considered as the most central one because it is
linked to many other nodes (and thus has got high
score for the node degree measure, see above). The
most central node can also be considered to be j, or g,
as a consequence of their relative importance in the
global structure of the network: even though these
nodes (j and g) have a smaller node degree than n,
the network would split into two subgroups if either
was removed.
‘‘Centrality’’ is another concept of node impor-
tance. There are a number of different statistics that
allow the calculation of individual centralities, which
have different meanings and purpose. According to
the kind of data collected and to the network
structure, the different measures of centralities
might be correlated or not. Thus, researchers should
consider their choice of centrality coefficient care-
fully, and choose according to the scoring methods
used and the questions in hand. Two common
statistics are the betweenness centrality coefficient,
defined as the number of shortest paths that pass
through the considered individual (with the shortest
path being the shortest distance, i.e. number of
edges, between two nodes), and the eigenvector
centrality coefficient [Newman, 2004]. The eigenvec-
tor centrality coefficient represents the connectivity
of an individual within its network according to the
number and strength of connections and considering
the centrality of the individuals it is connected to
[see Whitehead, 2008 for details about the calcula-
tion of this index]. Both coefficients can be extremely
informative, and we discuss them in turn with
illustrative examples.
The betweenness centrality coefficient, or ‘‘bet-
weenness,’’ can be important to measure whether
the edges in the network represent some active
transfer of information, or the transmission of a
disease, thus identifying nodes that play a key role in
diffusion process. Following this definition, the
removal of the most central individuals, in term of
the highest betweenness, quickly results in a
disconnection of the network into several discon-
nected parts. However, the consequences of such a
removal will depend on the inter-individual between-
ness. Indeed, if the differences in betweenness are
weak (low variance between individuals), the re-
moval of central individuals will have little impact on
network structure. In contrast, if the variance of
betweenness is large (typical scale-free networks,
where the distribution of betweenness coefficients
follows a power-law, i.e. one or two individuals are
very central compared to their conspecifics), then
removing one individual can have a substantial
impact on the network structure. For example, one
might expect that the removal of the central
individual in a tolerant, egalitarian species (such as
TABLE IV. Individual Measures of Nodes of Figure 1
Node/individual Degree Strength� Betweenness Eigenvector
Clustering
coefficient
a 1 1 0 0 –
b 3 3 17.5 0.02 0
c 3 3 19.667 0.04 0.33
d 3 3 10.333 0.04 0.33
e 4 4 12.833 0.06 0.5
f 5 5 41.667 0.07 0.4
g 4 4 81 0.11 0.33
h 3 3 17 0.05 0.33
i 1 1 0 0.01 –
j 4 4 81.833 0.27 0.33
k 3 3 16.5 0.29 0.67
l 1 1 0 0.1 –
m 2 2 8.333 0.19 1
n 7 7 46.333 0.51 0.33
o 5 5 6.667 0.39 0.4
p 3 3 0 0.32 1
q 5 5 8 0.4 0.5
r 3 3 9.333 0.26 0.67
s 2 2 0 0.16 1
�Note that since the network shown in Figure 1 is unweighted, strength is equal to the node degree.
Am. J. Primatol.
708 / Sueur et al.

Tonkean macaques, Macaca tonkeana or other
Sulawesi macaques) where all individuals are con-
nected together will not have the same impact
as the removal of the central individual in a despotic
species (such as rhesus macaques, M. mulatta or
Japanese macaques, M. fuscata) where kinship and
dominance highly constrain relationships [Thierry
et al., 2004]. We illustrate this point in Figure 1,
where the shortest path between the individuals a and
s is 8 while the shortest path between j and g is 1.
Nodes (i.e. individuals) having the most important
betweenness coefficient are j and g (see Table IV
for betweenness coefficient of each node represented
in Fig. 1).
Lusseau and Newman [2004] used the between-
ness centrality coefficient to determine the centrality
of individuals in a bottlenose dolphin population.
They found that some individuals had significantly
higher betweenness coefficients than others. These
central individuals were suggested to be linking
subgroups in this highly dynamic fission–fusion
system, and acting as ‘‘brokers’’ who mediated fusion
in the population. Centrality can then be linked to
other individual traits. In the case of the dolphins,
Lusseau and Conradt [2009] showed that central
individuals appear to have a greater knowledge about
their environment, and as a consequence precipitate
shifts in behavior of their conspecifics. This study
raises the intriguing possibility that for certain
animal systems (like those that show high fission-
fusion dynamics), greater knowledge, or innovation
may be inferred from the centrality of an individual
in its social network. Since individuals can improve
their performance with experience, i.e. better infor-
mation [Corning & Lahue, 1972; Schneirla, 1943;
Thorpe, 1963], it will be interesting to explore the
directionality of such correlation in future. That is,
are individuals central because they possess knowl-
edge, or they are knowledgeable because they are
central in the network, a position which they may
have acquired through some other process?
The eigenvector centrality coefficient is another
measure of centrality. Individuals will score highly
on the eigenvector centrality coefficient either if they
are highly connected to other network members or if
they are connected to individuals who are highly
connected themselves (see Table IV for eigenvector
coefficient of each node represented in Fig. 1).
This coefficient ranges from 0 (least central) to 1
(most central), and unlike the betweenness coefficient
takes into account edges’ direction and value. There-
fore, it is a useful index to determine an individual’s
centrality within weighted networks [Bonacich,
2007], and to provide a useful tool to assess an
individual’s ‘‘influence’’ in a network. We provide
three examples here, concerning group coordination,
group cohesion, and grooming interaction.
First, in the case of group coordination, Sueur
and Petit [2008] and Sueur et al. [2009] studied
collective movement in rhesus macaques (Fig. 3) and
Tonkean macaques. They created matrices describ-
ing the frequency with which individuals followed
one another when moving after a resting period.
Calculating an eigenvector centrality coefficient for
the resulting networks, they found that in both
species, some individuals were more central than
others, and thus had a disproportionate influence on
group coordination. Specifically, when these central
individuals departed, whatever their position in the
movement (i.e. first to leave, first to follow, second to
follow, etc.), they were joined by a great number of
individuals. Sueur and Petit [2008] labeled these
‘‘determinant individuals’’ and found that they were
dominants or individuals linking different matrilines
in rhesus macaques (Fig. 3), and the most affiliated
individuals in Tonkean macaques. These differences
could be attributed to differences in each species
respective social systems [see Sueur and Petit, 2008;
Sueur et al., 2009, for a discussion].
Second, in the case of group cohesion, Ramos-
Fernandez et al. [2009] used the same eigenvector
centrality coefficient to understand the spatial
association of spider monkeys (Ateles geoffroyi). They
not only showed that females had a higher eigen-
vector than males, but also reported that young adult
males played the role of ‘‘brokers’’ between the
female and the male clusters, similar to that seen in
a dolphin population [Lusseau & Newman, 2004].
Finally, in the case of grooming, eigenvector
centrality can also be informative, especially where
two individuals have both the same number and
strength of grooming relations. Two such individuals
would not necessarily have the same eigenvector
centrality, and Kanngiesser et al. in this issue
provide an example of this in a captive chimpanzee
(Pan troglodytes) community. They found that two
subjects had the same number of grooming interac-
tions, but found that one scored much higher in their
eigenvector centrality. It turned out that the high
scoring individual was the alpha male, who was
connected to individuals who were themselves highly
connected, unlike his lower ranked group-mate.
Between the Individual and the Group Level:
Searching for Cliquishness
In our discussion of individual measures (above),
we have presented several examples where an
individuals centrality was related to their ability to
‘‘mediate’’ or ‘‘join’’ two subgroups, and where
groups divide into subgroups of individuals sharing
a similar characteristics [Aureli et al., 2008; Sueur
et al., 2010]. In primates, it is well known that sex,
age, or kinship influence subgrouping patterns of
individuals (see Fig. 1). Individuals of same age or
same sex, for instance, may interact more frequently
and groups may show higher levels of synchrony of
activity where individual interests converge as a
Am. J. Primatol.
Social Network Analysis in Primates / 709

result of similar energetic or reproductive state
[King & Cowlishaw, 2009b]. This specific clustering
is expected to increase individual fitness [Conradt &
Roper, 2000; Krause & Ruxton, 2002]. Nevertheless,
while some individual/centrality measures allow us
to identify subgrouping, more specific tools have
been developed to assess how a group is clustered
(i.e. the cliquishness of the group).
One such statistic is the clustering coefficient
[Newman, 2003]. This measures the amount which a
node tends to cluster with other nodes, and captures
the level of cohesion of the network (see Table IV for
clustering coefficient of nodes in Fig. 1). Specifically,
if individuals connected to the focal subject of
interest are fully connected to all other individuals,
the clustering coefficient will be equal to 1; if no
individual is connected with another, the clustering
coefficient will equal to 0 [Whitehead, 2008]. This
local clustering coefficient has the potential for broad
application in primatology. For example, in a
pioneering network study, Flack et al. [2006] studied
the policing and group stability in pig-tailed maca-
ques (M. nemestrina) using such an indicator. They
identified specific individuals with a high clustering
coefficient and, using experimental and theoretical
knockout, removed these individuals from the group.
As a consequence of the removal, the aggressive
interactions between the remaining group members
were more frequent and grooming interactions were
less frequent, and less diverse, than before the
removal. The authors therefore concluded that the
individuals with a high clustering coefficient were
important in conflict management and group stabi-
lity. We can illustrate this with one of our example
networks: if we remove individuals 9 and 6 from the
network displayed in Figure 3B, the network divides
into two subgroups. Thus, the higher the clustering
coefficient of an individual, the higher the prob-
ability that the group splits after this individual is
removed. This may be particularly important for the
management of captive primate populations, where
individuals and same age cohorts are moved between
zoological collections.
Another method to identify subgroups within a
network is to compare observed relationships of
individuals to an expected or theoretical probability
that these individuals would be connected. The
modularity method developed by Newman [2004]
for community detection on a network uses this type
of calculation and can be found in SOCPROG
[Whitehead, 2009]. Sueur and Petit [2008] used this
modularity method to identify subgroups of rhesus
macaques during collective movement, and found
that rhesus macaques were associated according to
kinship during collective movements (Fig. 3A).
We should also mention at this point that when
producing network graphs or diagrams, it is also
possible to infer something about patterns of
subgrouping in a network, before calculating any
network statistics. In Figure 3A and B, individuals
are positioned in the graph according to their
associations using the multidimensional scaling
method. This method plots individuals with strong
relationships close to one another in the network
space [see Whitehead, 2008, 2009 for details about
calculation].
A more quantitative approach that allows
visualization of sub-grouping is Hierarchical cluster
analysis. Using this method, one can obtain a
dendrogram with individuals on one axis and the
degree of associations between these individuals on
the other axis (Fig. 4A). Chepko-Sade et al. [1989]
applied this methodology to study and visualize the
distribution of grooming interactions in a group of
rhesus macaques. Coupled with the modularity
method (see Fig. 4B) and the knot diagram (see
Fig. 4C), cluster analysis enables researchers to
assess the amount of subdivision or ‘‘communities’’
that exist in a network. Subgrouping patterns of
African elephants, known for their high degree of
fission-fusion, offer a classic example. Wittemyer
et al. [2005] were able to show that their elephant
population was multitiered with different subgroups
and different strength of relationship, and that these
network properties were influenced by ecological
constraints; with more frequent fission with in-
creased ecological pressures.
Subgrouping patterns can also be identifiable at
a much smaller scale (than population level).
Chimpanzees, like elephants, show a fission–fusion
structure and Clark [2011] showed distinct sub-
grouping in the affiliative network of captive chim-
panzees, and these subgrouping patterns were
aligned with maternal kinship. As default, clusters
are defined at 95% (i.e. significance), but it may be
appropriate to set two individuals as belonging to the
same cluster when they spend at least 50% of their
respective time together, or when they spend time
together above the population mean, depending on
the rationale for using cluster analyses. For example,
the choice of cluster thresholds is particularly important
where researchers are interested in the spread of
disease or of information in a network, and Godfrey
et al. [2009] found that higher levels of connectivity
between individuals increased the risk of parasite
infection in a lizard population (Egernia stokesii).
Another intermediate analysis using a social
networking approach would be to assess whether the
interactions between individuals are symmetrical or
asymmetrical. We decided to introduce this symme-
try measure in this section, because symmetry of
relationships is not an attribute of individuals or of a
group. Rather, it is a more a specific measure of
overall relationships. Symmetry of relationships can,
by definition, only be carried out on directed net-
works (i.e. actor and receiver interactions, such as
grooming, or unidirectional aggression) but not on
undirected networks (e.g. simple associations defined
Am. J. Primatol.
710 / Sueur et al.

by contacts or proximities). Tests of symmetry
will also advance the outstanding question of the
function of grooming relationships, and whether
individuals groom an individual in return for grooming,
or for other commodities, e.g. tolerance at feeding
sites or coalition support [Dufour et al., 2009; Fruteau
et al., 2009; Kanngiesser et al., 2011; Pele´ et al., 2010].
In this way, the term ‘‘symmetrical’’ can be replaced
by the term ‘‘reciprocal’’ [Hemelrijk, 1990; White-
head, 2008]. Tests of symmetry normally compare
one-half matrix (emitter to receptor for instance) to
the other one (receptor to emitter) but can also
compare data collected to random matrices. This
comparison with random networks was introduced by
Hemelrijk [1990] and are variants of the Mantel test
(parametrical) or the Kendall test
(nonparametrical). These tests use randomization
and a predefined amount of permutations (e.g.
1,000) allowing the calculation of accurate and stable
P-values [Hemelrijk, 1990; Lusseau et al., 2009; Vries
et al., 1993; Whitehead, 2008].
Symmetry tests (or more precisely the intensity
of asymmetry) are often a reflection of social system
properties. We know for instance that in despotic
species, agonistic interactions will be more asym-
metric than in tolerant species [Thierry et al., 2004].
Asymmetry in a system is however not limited to
agonistic interaction. Sueur and Petit [2008] as-
sessed the symmetry of leader–follower interactions
in rhesus and Tonkean macaques when studying
the process of collective movements, as mentioned
earlier. They found that in the egalitarian Tonkean
macaque, leader–follower interactions (i.e. the fre-
quency with which individual A followed individual
B and vice versa) was symmetrical, whereas in the
despotic and nepotistic rhesus macaque, these inter-
actions were asymmetrical: among adults subordi-
nate individuals more often followed dominant
individuals, and juveniles followed their mothers.
Group Structure Properties
Group measures concern the overall group
structure, and are especially useful for questions
about group cohesion. Average, maximum, or var-
iance values of individual measures also provide
useful information on group-level properties, and we
begin our group structure section with a discussion
of these. For instance, we can calculate the mean
clustering coefficient (i.e. a global clustering coeffi-
cient) or the mean node degree, that each give an
indication of how individuals are connected and
consequently a group’s cohesion or stability. For
instance, if the mean clustering coefficient for a
group network is high, then the group may be
unstable because the removal or the death of specific
individuals can lead to the group splitting. In
contrast, low clustering coefficients (and then a high
robustness of the social network because individuals
are all similarly connected) can result in stability in
the network, and facilitate the transmission of
information of disease. For example, Naug [2009]
showed that a low clustering coefficient was corre-
lated to the transmission of disease in a honeybee
colony.
The cumulative distribution of these group
measures, and specifically the mean node degree,
Fig. 4. Hierarchical cluster analysis of associations of the
theoretical group in Figure 1. (A) Dendrogram of associations
[average linkage, cophenetic correlation coefficient: 0.71, see
Croft et al., 2008; Whitehead, 2008, 2009 for details about
calculation]; gray lines suggest one subgroup, black lines suggest
the other one. (B) Modularity of the dendrogram suggesting a
division into clusters with an association index of 0.11 (solid
line). (C) Knot diagram giving the cumulative number of
bifurcations at different bifurcation distances and suggesting a
knot at an association index at 1 (dashed line, direct associations
for this case). A knot is a level of association such that the rate of
cluster formation suddenly changes with the association index.
Am. J. Primatol.
Social Network Analysis in Primates / 711

also provide an indication of whether a network is
random (linear distribution) or scale-free (power
distribution). In bottlenose dolphins and in Colum-
bian ground squirrels (Spermophilus columbianus),
exploration of cumulative distributions suggested
that both species have social structures more like
scale-free than random networks [Lusseau, 2003;
Manno, 2008, respectively]. Earlier in this review
(Measuring Attributes and Roles of Individuals, above),
we discussed how variance in betweeness may have
implications for the stability of the network. In a
scale-free network, if an individual with a high
betweenness dies, then the network may be likely
to split, because such individuals play the role of
intermediary between other network members.
Two simpler and more commonly applied statis-
tics to measure group cohesion are group density and
diameter. The group density is defined as the number
of observed edges divided by the number of possible
edges in the network. For instance, the group density
of the network in Figure 1 is 0.18 (31 observed edges
divided by (n2�n)/2 possible ones). Thus, the higher
the density, the more cohesive and stable the group.
The diameter is defined as the longest path length in
the network and is also a good measure of the group
cohesion [Wasserman & Faust, 1994]. The lower the
diameter is, the higher the group cohesion is and
then the higher the transmission of information and/
or of disease between two individuals would be. For
instance, the diameter of the network in Figure 1 is 8
(i.e. the number of edges between the individual a
and the individual l or between the individual a and
the individual s). A low diameter—high cohesion—
may be due to social structure constraints [Thierry
et al., 2004] or because individuals that share similar
characteristics (e.g. nutritional needs) tend to cluster
together in time and space [Krause & Ruxton, 2002;
Ramos-Fernandez et al., 2006]. The calculation of
density and diameter is however currently suitable
only for unweighted networks, and methods to
calculate these two statistics for weighted networks
are still under development [Wasserman & Faust,
1994; Whitehead, 2008]. One solution is to transform
weighted networks to unweighted networks (using
the preferential/avoided associations test in SOC-
PROG [Whitehead, 2008, 2009], Fig. 3).
COMPARING NETWORKS
An underdeveloped application of SNA is the
comparison of a single network, different networks
(groups) of a same species, or different species of a
same genus through time. Comparing matrices in
this way is not a new method, but the manner in
which data are represented for network analyses
lends itself to matrix correlations and thus comprise
an important part of the modern SNA toolkit (see
Brent et al. [2011] for a discussion on this point).
These tests can be carried out in Matman software
[de Vries et al., 1993] or SOCPROG [Whitehead,
2008, 2009], and are used to correlate matrices
within and across groups.
Comparing Different Relationship Matrices
Within a Group
This special issue provides a number of examples
of comparing different kinds of relationships or
interactions within a group. The first concerns social
foraging. King et al. [2011] compared co-feeding
networks (i.e. occasions where individual A is seen
foraging in a food patch with individual B), to
grooming and kinship networks in wild chacma
baboons (Papio ursinus), while controlling for dom-
inance relationships. They found that co-feeding was
significantly correlated to grooming relationships, but
not to relatedness among individuals, and suggest
that this may be a result of grooming relationships
affording tolerance in competitive foraging situations.
The other two examples in the special issue concern
captive chimpanzees. Composite indices are often
used as a proxy for the strength of social relationships
in primates [Silk et al., 2006a,b], and Clark [2011]
shows that the grooming network of a captive
chimpanzee community was positively correlated with
the association network, whereas Kanngiesser et al.’s
[2011] study found that grooming networks were
correlated to both age and kinship.
Comparing Individual Centralities
of Different Relationships Within a Group
Individual measures (discussed above) for dif-
ferent interaction types can also be used to explore
similarities in the properties of the network at a
group level. For instance, the eigenvector centrality
coefficient can be calculated for two types of inter-
actions (e.g. grooming and third party support in
aggressive interactions). Then, the different eigen-
vector coefficient for each type of interactions can be
compared in order to assess whether the most central
individual for grooming interactions is also the most
central one for support during conflicts. This can
provide insight into the selection pressures that
individuals face, and Sih et al. [2009] used a
correlation of ‘‘degrees’’ (see Measuring attributes
and roles of individuals, above) between males and
females for mating and found that males mating with
many females tend not to mate with females who
themselves mate with a lot of other males. Such a
pattern would be extremely difficult to uncover by
any other methodological approach.
Comparing Relationship Matrices Over Time
Individual relationships can also be explored
over time: during seasons, over years, or before and
after specific events, e.g. the death of an individual or
the replacement of an alpha individual. Studies of
Am. J. Primatol.
712 / Sueur et al.

nonprimates again offer insight of how useful this
approach can be for the study of primate behavior.
Wittemyer et al. [2005] studied an African elephant
population during a 4-year period and observed that
the high-order social units (metagroup) changed
over time, according to the ecological constraints.
In the same way, Hansen et al. [2009] used SNA to
study changes in social interaction of river otters
(Lontra canadensis) captured and held in captivity
for 10 months; they found that the strength of
social interactions declined over time. However,
such analyses are currently restricted to the
comparison of two static networks at different
‘‘snapshots’’ in time and very few indicators
currently exist to characterize the dynamics of social
networks at multiple time points (or continuously)
over a given period [Lahiri & Berger-Wolf, 2008;
Tantipathananandh & Berger-Wolf, 2009].
Comparing Relationship Matrices of Different
Groups
Comparing group networks from the same
species in different environmental conditions, or for
different group sizes can form the basis of compara-
tive studies. For instance, Sueur et al. [2010]
compared subgrouping patterns after short-term
fissions between Tonkean and rhesus macaques
and showed that rhesus macaques had higher
fission–fusion dynamics [Aureli et al., 2008] than
Tonkean macaques. Comparisons of the densities or
the diameters of two different groups (see Group
structure properties, above) would also provide an
indication of both network cohesion and indication of
individual network members level of tolerance or
propensity to cooperate [Anderson, 2007; Aviles,
1999; de Waal & Luttrell, 1989; Thierry et al.,
2004]. This approach would clearly be useful for
researchers interested in cooperation (e.g. grooming,
support during conflicts, vigilance). Networks of
observed groups may also be compared with net-
works of artificial groups in which all parameters are
fixed and known. This comparison can help to
understand mechanisms underlying some social or
collective phenomenon [Sueur et al., 2009; Voelkl &
Noe¨, 2008]. Adopting this approach, Franz and Nunn
[2009] studied how social learning opportunities can
be affected by the group social structure.
Basic descriptive network measures, such as the
average, variance, or distribution of interactions can
also be compared in order to assess the importance
and diversity of the interactions in one group relative
to another. Even if the size or the sampling
period differs between groups, indices of individuals
(centrality statistics) or indices about the group
(density, mean path length) can be comparable
using SNA (to a certain extent). For instance, Faust
and Skvoretz [2002] compared 42 networks from
diverse species and contexts. When comparing
networks, however, we have to be careful to compare
data on relationships collected/calculated in a
similar manner [Cairns & Schwager, 1987;
Whitehead, 2008, 2009]. Hopefully, as more and
more SNA studies are carried out, a common
methodology will emerge based on the success of
previous investigations and early comparisons of
these data [Franks et al., 2009; James et al., 2009;
Whitehead, 2008].
PERSPECTIVES AND FUTURE
RESEARCHES
We have outlined the SNA toolkit available to
primatologists and presented some examples of how
these can be applied at many levels (individual,
group, and population) of social behavior. In this
section, we identify five key research questions that
we believe are ripe for exploitation and hope that our
discussion of these issues will act as a springboard
for future primate social network research.
What is a Social Group?
Primates are commonly described as living in
stable groups since many live all their lives in the
same social group and maintain stable and long-term
relationships with conspecifics [Kummer, 1971;
Thierry et al., 2004; Wrangham, 1987]. This level
of stability in sociality differs to that of many other
gregarious species, such as shoals of fishes or flocks
of birds. The composition of such ephemeral aggre-
gations may change by the minute, hour, or day, and
individuals may not have opportunity to develop and
maintain stable relationships [Couzin & Krause,
2003]. Primatologists therefore define a group
according to when individuals in proximity to one
another tend to interact in a positive (e.g. affiliative)
rather than negative (e.g. agonistic) manner, as well
as their activities being somehow synchronized in
time and space [Kummer, 1971]. We suggest how
SNA can help in both cases.
Throughout this review, we have used the term
relationship to describe both associations between
individuals—which may be passive sharing of time
and space—and interactions—which are more com-
plex interactions that are typically directed—such as
grooming behavior, reconciliation, or support during
conflict. When defining a group, it could then be
argued that one should use only interaction data, and
not association data. Nevertheless, association data
can be extremely informative, especially where these
associations correlate to interaction data [Clark,
2011; Silk et al., 2002a,b], since association data
are typically easier to collect. How can SNA of
relationship data be used to define a group?
Instead of categorizing groups as ‘‘despotic’’ or
‘‘egalitarian’’ (with respect to agonistic interaction),
or ‘‘cohesive’’ or ‘‘fission–fusion’’ (with respect to
Am. J. Primatol.
Social Network Analysis in Primates / 713

patterns of spatial-temporal association) [Aureli
et al., 2008; Thierry et al., 2004], SNA provides a
suite of statistics for describing primate group
structure [Kasper & Voelkl, 2009]. Sublevels of
association/interaction can be defined within larger
levels of organization [Aureli et al., 2008; Couzin,
2006; Wittemyer et al., 2005], and where groups split
(in space and time) daughter groups may still have
positive interactions during their next encounters,
and even foraging on the same resources. We
therefore have to understand the ecological and
social drivers (or constraints) underlying formation
and stability of social groups [Gero et al., 2008]. One
way to do this is to calculate intermediate network
measures such as the modularity method or hier-
archical cluster analysis (Between the individual and
the group level: searching for cliquishness, above).
These approaches will identify different levels of
associations, which can then be used to determine
group composition. One can then consider what
social and ecological factors may be influencing the
composition of these subunits/groups. For example,
Wolf et al. [2007] studied the social structure of a
Galapagos sea lion (Zalophus wollebaeki) population.
Analyses of their proximity network revealed that
the population could be divided in different sub-
groups, defined both by the topography of the island
on which they lived, and by male territoriality.
Who are the Most Important/Central
Individuals in a Network?
Traditionally, the most central individual in a
primate group was defined as the highest ranking
individual, as defined by dominance hierarchy
derived from agonistic interaction data, or the most
affiliated individual defined by the frequency of
affiliative interactions. This is a kind of centrality,
but the most central or important individuals may
also be considered the individuals facilitating group
stability. These individuals are not necessary the
most affiliated (or those with higher node degrees),
but rather those individuals with the highest
clustering coefficients (i.e. the broker individuals,
which can act as a bridge to two subunits, as already
discussed). It is therefore important to understand
that centrality can have different meanings: the
central individual can be defined either as the
individual allowing group cohesion by linking two
subgroups, or as the individual allowing group
stability by managing conflicts, or, it may represent
the individual having the highest frequency of
grooming interactions. The interpretation of the
centrality will depend on your question and of the
index used to calculate this centrality.
We also need to understand how behavioral
traits or characteristics, such as sex, age, dominance,
boldness, knowledge, or experience, may influence
the network centrality of an individual. As already
mentioned, it has been suggested that individuals
may develop relationships based on sharing similar
needs and motivations [Couzin, 2006] and Ramos-
Fernandez et al. [2006] proposed that nonrandom
associations among group members can emerge
solely as a consequence of the way in which
individuals forage. How individual characteristics
may confer an adaptive advantage and enhance
the fitness of individuals via their network
position (through the network centrality of indivi-
duals displaying this trait, for example) remains to
be explored.
Can Network Properties Influence Individual
Fitness?
Primates, but also all other animal groups, show
an assortativity in the way group members associate
and interact together, as we have already discussed.
From a network perspective, these relationships
define the structure of the group network, which in
turn impact on the way an individual will behave
with its partners. There is now growing support for
the concept that the structure of a network is linked
to its functioning, and is thus selected in order to
increase the fitness of its members [Krause et al.,
2007]. In baboons, Silk et al. [2009] showed that
females having more relationships, i.e. a greater
centrality, have more offspring than females with
fewer relationships. In the wasp, Ropalidia marginata,
the network structure of the colony changes drama-
tically, from tolerant to despotic after the removal of
the queen. This change is a result of increases in
aggressive interaction and competition for centrality
and control of colony reproduction [Bhadra et al.,
2009]. These examples not only emphasize how an
individual behavior can influence a network, but also
how the network influences individual behavior via a
feedback loop [Krause et al., 2007; Naug, 2009].
King et al. [2011] provide a further example of
how network structure may impact directly on
individuals in the network. Their SNA of foraging
associations in baboons showed that dominance
interactions and affiliative patterns were crucial to
shaping the co-feeding network. Examination of how
the co-feeding network was structured suggested
that individuals arranged themselves to increase
foraging benefits and reduce aggression at the level
of the individual. This arrangement was expected to
reinforce the structure of grooming and dominance
interactive networks. Traditional dyadic analyses of
these patterns would not allow them to have come to
such an interpretation.
The question of whether network structure can
influence individual fitness can also be investigated
at the population level. Animal population structure
and dynamics have important implications for the
transmission of genes, diseases, and information
[Krause et al., 2007; Whitehead, 2008]. All these
Am. J. Primatol.
714 / Sueur et al.

transmissions will have an impact on the main-
tenance or the changes in genetic or cultural
diversity, and where costs of sociality vary across
individuals in a population. This can result in
different behavioral strategies at the group level
[Aureli et al., 2008; Conradt & Roper, 2005; Krause
& Ruxton, 2002; Wrangham, 1987]. For example,
varying inter-individual distances (dispersion of
group members) and temporary or irreversible group
fission events can allow individuals to temporarily or
permanently reduce the costs they experience.
Some species of primates, fish, cetaceans, in-
sects, and bats, all display a degree of fission–fusion
dynamics [see Aureli et al., 2008 for a review]. SNA
can allow researchers to characterize these dynamics
at the population level. Relationships between
groups can be measured not only by using direct
interactions, such as dispersal of individuals or
group encounters [Drewe et al., 2009], but also by
using indirect interactions [Formica et al., 2010].
Indirect interaction can be important where disease
or information can be transmitted among groups or
individuals even if they do not share the same
space at the same time. For several primate
species, group home ranges are overlapping and
scent marking behavior (for instance) can allow
groups or individuals to exchange information.
Analyzing intergroup relationships (direct or
indirect) using SNA might improve our understand-
ing of intergroup information exchange, group
interactions, population structure, and disease
transmission.
What Determines Network Stability?
One important question in primatology is why
some groups are more stable than others with similar
size, and in similar ecological conditions. SNA, once
again, should be an important tool for understanding
these differences. One way to understand network
stability over time is to analyze how relationships are
modified, or alter after the death of an individual,
which can act as a kind of ‘‘natural experiment’’.
One can also study the effect of removal of
individuals from a network theoretically, using an
experimental knockout in SNA software such as in
Ucinet. Results of such simulations should however
be interpreted with caution since natural groups can
have a great inertia to keep cohesion due to external
or internal forces not taken into account in a
simulated experiment. Experimental knockouts like
that devised by Flack et al. [2006], are surely the way
to go where this is practically possible. Comparisons
of simulated and real knockouts will allow us to
better understand the mechanisms underlying group
cohesion, especially where simulations are not
supported by data [James et al., 2009].
Most of the studies on the influence of
social relationships upon group stability sum the
interactions of individuals over a defined observation
period. Then, they attempt to understand how the
interactions observed during this observation influ-
ence patterns of fission [e.g. van Horn et al., 2007].
This approach therefore considers relationships,
and the network of these relationships as static.
However, relationships between individuals change
over time, and the group network is dynamic. King
et al. [2011] suggest that to compliment their
analysis of static networks, data over a finer
temporal scale may provide insight into variation in
the value of grooming and food resources, and if
these translate to changes in co-feeding networks.
Other studies support this notion, since Henzi et al.
[2009] showed cyclicity in the proximity and groom-
ing networks of female baboons and suggest that this
cyclicity can be attributed to changes in food
availability (and thus value of grooming). However,
it will be important for future studies on this topic to
disentangle the influence of simple changes spatial
distribution of individuals as a consequence of the
configuration of food resources (e.g. when food is
patchy, individuals will be spread out) that therefore
may have nothing to do with the value of social
relationships.
Although it is currently difficult to define
communities and subgroups when a network changes
over time, a number of new approaches makes this
possible. For instance, Tantipathananandh and
Berger-Wolf [2009] proposed two different algo-
rithms for identifying communities in dynamic social
networks. By applying these algorithms to several
data sets of human and animal societies, they
revealed novel patterns of associations, which were
previously obscured by static network approaches.
These approaches can allow researchers to tackle
new questions. For example, two dyads might show
the same cumulative frequency of interactions after
a certain time period of observation. However, the
two dyads interactions may differ markedly in their
temporal distribution. For instance, each dyad may
have interacted at the same low frequency each day,
or one dyad may have had all their interactions in a
single day. Thus, the frequency of interactions is the
same across the two dyads, but this tells us nothing
about the stability of social relationships. This
requires a dynamic analysis and this kind of
modeling approach and its impact on information
or disease transmission is discussed in this issue by
Jacobs and Petit [2011] and Hoppitt and Laland
[2011].
Can SNA be Used to Manage and Conserve
Primates?
SNA can be useful in several ways for the
management and conservation of wild and captive
primates. Sometimes, both in the wild and in
captivity, group size may increase so that it has
Am. J. Primatol.
Social Network Analysis in Primates / 715

serious consequences for the well-being of animals;
increasing within-group food competition and
aggression which may cause injury. In the wild, this
may result in permanent fission, and under these
conditions in captivity, it may be necessary to divide
the group into two or more subgroups. Similarly,
where opportunity for immigration in wild popula-
tions is limited due to physical barriers created by
human populations, natural fission events may be
prevented, potentially escalating human–primate
conflicts. Using cluster analysis on grooming inter-
actions, researchers could predict the best group
division that would result in preservation of the most
affiliated individuals (see Clark [2011] for use of
cluster analysis on captive chimpanzees). Using
centrality measures on aggressive interactions, one
can also identify the individual most disruptive to
the network and then remove it in order to
potentially decrease aggression rates and enhance
group stability.
Clark [2011] provides an informative discussion
of how SNA can be used from a captive management
perspective, and two empirical examples are pre-
sented in this special issue. Dufour et al. [2011] used
SNA to study the effect of relocation on stress and
proximities in a captive capuchin and squirrel
monkeys, and Beisner et al. [2011] used SNA to
demonstrate the degree of integration of high-
ranking natal adult males into social network of
rhesus macaques. Beisner et al. find that the
presence of adult natal males facilitates high levels
of intense aggression and fragmentation of captive
macaque social networks. Although we have known
for a long time that the presence of natal males in
captive social groups is associated with high risks of
aggression, this analysis adds insight into the social
dynamics involved. Thus, understanding interactions
between group members in such situations can
enable managers to monitor and respond to network
changes in captive groups by removing some indivi-
duals. Individuals, subgroups, and whole groups are
also frequently moved in captive and wild situations.
SNA can inform institutions not only on the stability
of networks pre- and post-relocation, but also
about the predictors of group instabilities [McCowan
et al., 2008].
Principles of individual centrality in a network
might also facilitate management practices. Specifi-
cally, if there is a requirement for a new training
technique/behavior to be introduced, a group for
sanitary reasons or for medical, or behavioral
research [Laule et al., 2003; Savastano et al., 2003;
Schapiro et al., 2003], informing the most central
individual first should allow a faster diffusion of the
information inside the group than informing another
individual. Concerning disease transmission, we
might also expect that, in case of infection, isolating
or first treating central individuals would limit the
transmission of the disease.
CONCLUSIONS
Despite primate social interactions and resulting
relationships typically involving more than two
individuals, most previous studies focus only on
dyadic interactions. SNA broadens our perspective,
allowing researchers to quantify and explore not only
the overall social structure of primate groups or
populations, but also to examine the relationships a
specific individual has with all other group members
[Whitehead, 2008], i.e. the ‘‘social niche’’ [Flack
et al., 2006].
The examples that are contained within this
special issue showcase the potential of social network
analyses to the study of primatology. Taking these
works alongside books by Whitehead [2008] and Croft
et al. [2008], and other review articles by Krause
et al. [2007] and Wey et al. [2008], we hope that this
special issue will help to stimulate other researchers
to explore the questions we have identified and more,
in a variety of primate species and study populations;
we eagerly anticipate their findings.
ACKNOWLEDGMENTS
The authors are grateful to Lauren Brent,
Bernhard Voelkl, and Claudia Kasper for their
helpful comments. C.S. was supported by the Fran-
co-American commission, the Fyssen foundation,
and the Japan Society for the Promotion of Science.
A.J. was supported by a grant from the French
Ministry of Education and Research. This research
adheres to the American Society of Primatologists
principles for the ethical treatment of primates.
REFERENCES
Alexander R. 1974. The evolution of social behavior. Annual
Review of Ecology and Systematics 5:325–383.
Anderson JR. 2007. Animal behavior: tolerant primates
cooperate best. Current Biology 17:R242–R244.
Arnold K, Barton RA. 2001. Postconflict behavior of spectacled
leaf monkeys (Trachypithecus obscurus). II. Contact with
third parties. International Journal of Primatology 22:
267–286.
Aureli F, de Waal FBM. 2000. Natural conflict resolution.
Berkeley, CA: University of California Press.
Aureli F, Schaffner CM, Boesch C, Bearder SK, Call J,
Chapman CA, Connor R, Fiore AD, Dunbar RIM,
Henzi SP, Holekamp K, Korstjens AM, Layton R, Lee P,
Lehmann J, Manson JH, Ramos-Fernandez G, Strier KB,
van Schaik CP. 2008. Fission–fusion dynamics: new
research frameworks. Current Anthropology 49:627–654.
Avile´s L. 1999. Cooperation and non-linear dynamics: an
ecological perspective on the evolution of sociality. Evol-
utionary Ecology Research 1:459–477.
Barabasi A, Oltvai ZN. 2004. Network biology: understanding
the cell’s functional organization. Nature Reviews. Genetics
5:101–113.
Beisner B, Jackson M, Cameron A, McCowan B. 2010.
Detecting instability in animal social networks: genetic
fragmentation is associated with social instability in rhesus
macaques. PLoS One.
Beisner BA, Jackson ME, Cameron A, McCowan B. 2011.
Effects of natal male alliances on aggression and power
Am. J. Primatol.
716 / Sueur et al.

dynamics in Rhesus Macaques. American Journal of
Primatology 73:790–801.
Bejder L, Fletcher D, Bra¨ger S. 1998. A method for testing
association patterns of social animals. Animal Behaviour 56:
719–725.
Bhadra A, Jordan F, Sumana A, Deshpande SA, Gadagkar R.
2009. A comparative social network analysis of wasp
colonies and classrooms: linking network structure to
functioning. Ecological Complexity 6:48–55.
Bonacich P. 2007. Some unique properties of eigenvector
centrality. Social Networks 29:555–564.
Borgatti S, Everett M, Freeman L. 2002. UCINET 6 For
Windows. Software for Social Network Analysis.
Borgatti SP, Mehra A, Brass DJ, Labianca G. 2009. Network
analysis in the social sciences. Science 323:892–895.
Brent LJN, Lehmann J, Ramos-Ferna´ndez G. 2011. Social
network analysis in the study of nonhuman primates: a
historical perspective. American Journal of Primatology 73:
720–730.
Cairns SJ, Schwager SJ. 1987. A comparison of association
indices. Animal Behaviour 35:1454–1469.
Camazine S, Deneubourg J, Franks NR, Sneyd J, Theraula G,
Bonabeau E. 2003. Self-organization in biological systems.
Princeton, NJ: Princeton University Press.
Chepko-Sade BD, Sade DS. 1979. Patterns of group splitting
within matrilineal kinship groups. Behavioral Ecology and
Sociobiology 5:67–86.
Chepko-Sade BD, Reitz KP, Sade DS. 1989. Sociometrics of
Macaca Mulatta IV: network analysis of social structure of a
pre-fission group. Social Networks 11:293–314.
Clark FE. 2011. Space to choose: network analysis of social
preferences in a captive chimpanzee community, and
implications for management. American Journal of Prima-
tology 73:748–757.
Conradt L, Roper T. 2005. Consensus decision making in
animals. Trends in Ecology and Evolution 20:449–456.
Conradt L, Roper TJ. 2000. Activity synchrony and social
cohesion: a fission-fusion model. Proceedings of the Royal
Society B: Biological Sciences 267:2213–2218.
Corning WC, Lahue R. 1972. Invertebrate strategies in
comparative learning studies. American Zoologist 12:455–469.
Couzin ID. 2006. Behavioral ecology: social organization in
fission-fusion societies. Current Biology 16:R169–R171.
Couzin ID, Krause J. 2003. Self-organization and collective
behavior in vertebrates. Advances in the Study of Behavior
32:1–75.
Croft DP, James R, Ward AJW, Botham MS, Mawdsley D,
Krause J. 2005. Assortative interactions and social networks
in fish. Oecologia 143:211–219.
Croft DP, James R, Krause DJ. 2008. Exploring animal social
networks. Princeton, NJ: Princeton University Press.
Crook JH, Gartlan JS. 1966. Evolution of primate societies.
Nature 210:1200–1203.
DeJordy R, Borgatti SP, Roussin C, Halgin DS. 2007.
Visualizing proximity data. Field Methods 19:239–263.
de Waal FBM. 1986. The integration of dominance and social
bonding in primates. The Quarterly Review of Biology 61:
459–479.
de Waal FBM, Luttrell LM. 1989. Toward a comparative
socioecology of the genus Macaca: different dominance
styles in rhesus and stumptail monkeys. American Journal
of Primatology 19:83–109.
Dow MM, de Waal FBM. 1989. Assignment methods for the
analysis of network subgroup interactions. Social Networks
11:237–255.
Drewe JA, Madden JR, Pearce GP. 2009. The social network
structure of a wild meerkat population: 1. Inter-group
interactions. Behavioral Ecology and Sociobiology 63:
1295–1306.
Dufour V, Sueur C, Whiten A, Buchanan-Smith HM. 2011.
The impact of moving to a novel environment on social
networks, activity and wellbeing in two new world primates.
American Journal of Primatology 73:802–811.
Dufour V, Pele´ M, Neumann M, Thierry B, Call J. 2009.
Calculated reciprocity after all: computation behind token
transfers in orang-utans. Biology Letters 5:172–175.
Faust K, Skvoretz J. 2002. Comparing networks across space
and time, size and species. Sociological Methodology 32:
267–299.
Flack JC, Girvan M, de Waal FBM, Krakauer DC. 2006.
Policing stabilizes construction of social niches in primates.
Nature 439:426–429.
Formica VA, Augat ME, Barnard ME, Butterfield RE,
Wood CW, Brodie ED. 2010. Using home range estimates
to construct social networks for species with indirect
behavioral interactions. Behavioral Ecology and Sociobio-
logy 64:1199–1208.
Franks D, James R, Noble J, Ruxton G. 2009. A foundation for
developing a methodology for social network sampling.
Behavioral Ecology and Sociobiology 63:1079–1088.
Franks DW, Ruxton GD, James R. 2010. Sampling animal
association networks with the gambit of the group.
Behavioral Ecology and Sociobiology 64:493–503.
Franz M, Nunn CL. 2009. Network-based diffusion analysis: a
new method for detecting social learning. Proceedings of the
Royal Society B: Biological Sciences 276:1829–1836.
Freeland WJ. 1976. Pathogens and the evolution of primate
sociality. Biotropica 8:12–24.
Freeland WJ. 1979. Primate social groups as biological islands.
Ecology 60:719–728.
Fruteau C, Voelkl B, van Damme E, Noe¨ R. 2009. Supply
and demand determine the market value of food providers
in wild vervet monkeys. Proceedings of the National
Academy of Sciences of the United States of America 106:
12007–12012.
Gero S, Engelhaupt D, Whitehead H. 2008. Heterogeneous
social associations within a sperm whale, Physeter macro-
cephalus, unit reflect pairwise relatedness. Behavioral
Ecology and Sociobiology 63:143–151.
Girvan M, Newman MEJ. 2002. Community structure in social
and biological networks. Proceedings of the National
Academy of Sciences of the United States of America 99:
7821–7826.
Godfrey S, Bull C, James R, Murray K. 2009. Network
structure and parasite transmission in a group living lizard,
the gidgee skink, Egernia stokesii. Behavioral Ecology and
Sociobiology 63:1045–1056.
Gray JA. 1987. The psychology of fear and stress. Cambridge,
England: Cambridge University Press [Archive].
Hamilton WD. 1971. Geometry for the selfish herd. Journal of
Theoretical Biology 31:295–311.
Hansen H, McDonald DB, Groves P, Maier JAK, Ben-David M.
2009. Social networks and the formation and maintenance
of river otter groups. Ethology 115:384–396.
Hemelrijk CK. 1990. Models of, and tests for, reciprocity,
unidirectionality and other social interaction patterns at a
group level. Animal Behaviour 39:1013–1029.
Henzi S, Lusseau D, Weingrill T, van Schaik C, Barrett L.
2009. Cyclicity in the structure of female baboon
social networks. Behavioral Ecology and Sociobiology 63:
1015–1021.
Hill RA, Lee PC. 1998. Predation risk as an influence on group
size in cercopithecoid primates: implications for social
structure. Journal of Zoology 245:447–456.
Hinde RA. 1976. Interactions, relationships and social struc-
ture. Man 11:1–17.
Hoppitt W, Laland KN. 2011. Detecting social learning using
networks: a users guide. American Journal of Primatology
73:834–844.
Huffman MA, Chapman CA. 2009. Primate parasite ecology:
the dynamics and study of host-parasite relationships.
Cambridge: Cambridge University Press.
Am. J. Primatol.
Social Network Analysis in Primates / 717