Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Behav Ecol Sociobiol

DOI 10.1007/s00265-009-0865-8

METHODS

Sampling animal association networks with the gambit


of the group
Daniel W. Franks & Graeme D. Ruxton & Richard James

Received: 15 May 2009 / Revised: 14 September 2009 / Accepted: 16 September 2009


# Springer-Verlag 2009

Abstract Ecologists increasingly use network theory to sampling regimes, and choice of filter on the accuracy
examine animal association patterns. The gambit of the and precision with which measures are estimated. We find
group (GoG) is a simple and useful assumption for strong support for the use of weighted, rather than filtered,
accumulating the data necessary for a network analysis. network measures and show that different filters have
The gambit of the group implies that each animal in a group different effects depending on the nature of the sampling.
is associating with every other individual in that group. We make several practical recommendations for ecologists
Sampling is an important issue for networks in wild planning GoG sampling.
populations collected assuming GoG. Due to time, effort,
and resource constraints and the difficulty of tracking Keywords Association networks . Gambit of the group .
animals, sampled data are usually a subset of the actual Network sampling
network. Ecologists often use association indexes to
calculate the frequency of associations between individuals.
These indexes are often transformed by applying a filter to Introduction
produce a binary network. We explore GoG sampling using
model networks. We examine assortment at the level of the Ecologists are often interested in understanding social
group by a single dichotomous trait, along with many other associations and interactions between animals. To examine
network measures, to examine the effect of different these associations, ecologists are increasingly turning to
network theory, within which they represent animals as
nodes and associations as links between nodes. Network
Communicated by J. Krause
theory offers a set of statistical measures to quantify,
Electronic supplementary material The online version of this article describe, and compare the structure of social relations.
(doi:10.1007/s00265-009-0865-8) contains supplementary material, The availability of network analysis software has provided
which is available to authorized users.
statistical tools that allow biologists to readily apply
D. W. Franks (*) network theory to their study populations (Newman 2003;
YCCSA,
Department of Biology and Department of Computer Science,
Croft et al. 2008; Whitehead 2008). A network approach is
University of York, often applied as a descriptive tool to help understand the
York 105, UK social structure of a population (e.g., dolphins, Lusseau
e-mail: df525@york.ac.uk 2003; fish, Croft et al. 2004; African buffalos, Cross et al.
G. D. Ruxton
2004; and sea lions, Wolf et al. 2007). A key aim of animal
Division of Ecology and Evolutionary Biology, social network analysis is often to look for a biological
Faculty of Biomedical and Life Sciences, University of Glasgow, signal or biological meaning in the potentially tangled mess
Glasgow 128, Scotland of observed associations between individual animals.
R. James
The gambit of the group (GoG) is a relatively simple
Department of Physics, University of Bath, method for accumulating relational data (Whitehead and
Bath 27, UK Dufault 1999). When using this technique, ecologists must
Behav Ecol Sociobiol

make the assumption that all individuals seen grouping Dufault (1999). Additionally, appropriate randomization of
together, during an observation census, are associating with the relational data can help to compensate for some types of
every other individual in that group. On observing groups, biases; see Croft et al. (2008) and Whitehead (2008).
the ecologist notes the number of times each individual is In many of the early applications of network analysis to
observed in the same group as each other individual to group-derived animal associations, researchers transformed
construct indexes of associations between individuals. After their association indexes by applying an edge filter to
numerous censuses and numerous group observations in produce a binary network representing (usually) just the
each census, the ecologist will be left with relational data stronger associations (e.g., Croft et al. 2004; Sundaresen et
representing the association strength between animals based al. 2007; and Wolf et al. 2007). If a filter threshold of 0.5
on group association. In a fission–fusion animal society, a was used, for example, then connections with an associa-
main parent group can fracture into smaller groups (such as tion index greater than or equal to 0.5 would be retained as
fish shoals forming each day), and the smaller groups, later, a network connection and connections with a weight less
coalesce to rejoin as the parent group (such as fish for than 0.5 would be removed. By transforming the data in
several daytime shoals collecting at a bank at night, Croft et this way, the researcher is able to calculate a multitude of
al. 2003). The fission–fusion process is likely to affect the network metrics often developed in the social sciences
social structure of many animal societies. The GoG is a (Wasserman and Faust 1994), but there are potential issues.
potentially convenient way of sampling such systems. The choice of appropriate value for the filter threshold is a
Ecologists capture group associations over a number of difficult problem. Filtering a network discards data that
days, weeks, or years, which allow them to collect could improve the accuracy of any analysis (and maybe, the
relational data regarding associations between individuals. weak connections are important—see Granovetter 1974).
Sampling is an important issue for GoG networks in wild This reduction in data resolution could prove to be
populations (Croft et al. 2008). Due to time, effort, and important, especially when working with limited sample
resource constraints and the practical difficulty involved in sizes. If the threshold is too high, then we lose more data. If
tracking animals, the sampled relational data produced are the threshold is too low, then we strengthen the weak
usually only a subset of the actual network. That is, not all connections. Because of this problem, ecologists are
individuals from the population will be sampled at each beginning to move away from the filtering approach by
census point. Incorrect sampling protocols risk creation of keeping the initial association indexes and conducting a
artifactual results. For example, assortment by degree weighted analysis of the network (Lusseau et al. 2008).
(where the degree of an individual is the number of Even then, we cannot be sure how much of the biological
connections from that individual to distinct other individ- signal of interest is being measured. For example, some of
uals and assortment by degree examines whether individ- what looks like a signal might be accounted for by
uals are generally connected to other individuals with sampling bias (see e.g., Croft et al. 2008). Here, we
similar degree scores) is typically identified by a positive illustrate and explore this issue using model networks in
Pearson correlation coefficient between a focal individual's which we insert a known biological signal. To ground the
degree of connectedness and the mean degree of connect- model, we make use of calculations of Croft et al. (2008)
edness of the individuals to which it is directly connected who looked at a sample of long-term red deer data and
(Lusseau and Newman 2004; Croft et al. 2005). Ecologists select parameters that are more or less consistent with this
often use association indexes to calculate the frequency of population (in addition to further explorations). The
associations between individuals, rather than a simple count biological signal is a controlled assortment at the level of
of the number of times individuals have been sampled in a the group by a single dichotomous trait (e.g., sex). We
same group. A sensible choice of association index can go examine this, along with other network measures, to
some way to alleviating certain types of sampling bias (see examine how well network measures pick up on this signal,
Whitehead 2008 and Croft et al. 2008, for more informa- given different sampling regimes, choice of weighted or
tion). In situations where there is a sampling bias such that unweighted measures, and choice of filter. The aim is to
it is more likely that two individuals will be recorded in a help guide field biologists in planning their data collection,
given census if they are in separate groups, for example, so as to collect data that makes the most effective use of
then the half-weight index may help to compensate by network theory.
weighting separate group observations at half the weight of
other observations (Cairns and Schwager 1987). Various
indexes have been developed to account for different The model
sampling biases. More detail can be found in papers by
Cairns and Schwager (1987); Ginsberg and Young (1992); We assume a population of n animals, where each animal is
Bejder et al. (1998); Whitehead (1999); and Whitehead and either male or female (where f is the proportion of females).
Behav Ecol Sociobiol

l
That is, each animal is allocated its trait/sex independently allocate all animals to a group Σ gl ( n. If once we define
and randomly, with the probability of an individual being the size of the final group, the total capacity is greater than
female, the value of the parameter f. Although we use male the number of animals, l then, the size of the final
Σ gl > n,
and female as traits for the animals, our model equally group is reduced, such that the capacity equals the number
applies to a variety of traits (e.g., infant/adult). In addition to l
of animals, Σ gl ¼ n. Once the groups and their respective
preference by trait, each animal, i, is allocated a preference capacities are defined, the animals are allocated to a group
toward grouping with each other individual animal, j. These as follows: Each group l with remaining capacity (that is, to
biases are stored as a matrix Bij, where Bij =Bji. The greater which we have previously allocated fewer individuals than
the value of Bij, the higher the relative probability that i and the group’s capacity, so all groups start with remaining
j will be allocated to the same group. We draw each Bij, i<j capacity) is allocated a scores sli pertaining to each
from a uniform distribution ½0 ! d; 0 þ d $ for some param- unallocated (i.e., not in a group) animal i defined as:
eter d ∈ [0,1], with increasing d, signifying increasing levels
j
X
of preference. Each value of Bij is also increased by b for
sli ¼ Bij þ 0:001 ð2Þ
individuals of the same trait (in this case by sex,
al
representing a sex bias in preferential associations) and
decreased by b if they are the opposite trait, where b ∈ [0,1] where ai is the number of individuals that have previously
is a constant. Thus, the greater the value of b, the stronger the been allocated to group l. A small value (0.001) is added to
preference for assorting with same sex individuals. Note that each score to give each group a chance of being selected,
for our purposes, we do not need to accurately model a even if there are no animals in it. If a group has filled its
mechanism by which real animals decide to join groups; we capacity then it cannot be selected (ski =0, for each group k
simply need to capture preferential associations and assorta- with no remaining capacity). An animal to allocate i is then
tive mixing in the groups along with grouping constraints selected from all unallocated animals with the roulette
governed by the group size distribution. Our algorithm wheel selection algorithm (Mitchell 1998), where their
allows us to capture these desirable properties. proportion of the roulette wheel is the maximum score
We simulate sampling by recording associations over a for that animal to be allocated to any group. We select
number of censuses, m. In each census we define new individuals in this manner, rather than from a uniform
groups, allocate the animals to groups, and sample the distribution, because it results in a stronger affect of the
groups. The preference matrix (B) remains fixed through- bias to be grouped with other individuals. This means
out; that is, we make the assumption that the social that animals with stronger association preferences will,
dynamics of the group are at some sort of equilibrium. generally, be allocated to groups earlier than those with
This process is repeated until all censuses are complete. weak preferences. A group is then selected with the
roulette wheel selection algorithm, where the group’s
Forming groups proportion of the roulette wheel is their score for the
selected animal sli, and the selected animal is then allocated
In each census, for m censuses, new groups are formed as to the selected group (and its remaining capacity is
follows: A group l is defined, and its size, g (representing updated). This process is repeated until all animals are
its capacity), is selected from a Pareto distribution: allocated to a group (and thus, all groups are full). When
b=0, then allocation to groups is not assorted by trait and as
g ¼ ð1 ! U Þ!1=K ð1Þ b increases towards b=1, there is an increased bias for the
group to be assorted by trait (because b is reflected in each
where K is the shape of the distribution, and U is a random group’s score for an individual through Bij). Likewise,
number between zero and one selected from a uniform increasing d would affect Bij and thus, the probability of
distribution. Group sizes are rounded to the nearest whole individuals being allocated to a group with other particular
number, and when a group size of zero is drawn, we draw individuals.
again from the distribution until a nonzero rounded group
size is drawn. Many group living animals have a group size Sampling the groups
distribution, which is right skewed and approximately,
Pareto or exponential (Okubo 1986; Krause and Ruxton After the groups have been allocated, we simulate gambit
2002). Our group size distribution was selected to match of the group sampling. We simulate the sampling process
the general shape distribution in the red deer data (see where we have complete information in that we sample all
below). Additional groups (along with their group size), groups and when we only have partial information where
continue to be defined in this manner until we have enough we sample a proportion p of the animals (when p=1.0 we
total capacity (in the sum of the capacity of all groups) to have complete information). Our assumption that groups
Behav Ecol Sociobiol

reform for each census corresponds to the sampling of a red deer group distribution. We examine sampling for a trait
fission–fusion system where successive samples of a preference (b=1) and no trait preference (b=0). Unless
population are taken at time intervals long enough to allow, otherwise stated, there is no individual pairing bias (d=0).
in principle, complete mixing. Thus, we do not model For all cases, we vary m (the number of censuses) from one
timescale issues related to short-term switching of individ- to 23, and p (the proportion of individuals sampled in each
uals between groups during a census or the fracturing of census) from 0.1 to one and average the results over 30
groups. We simply capture biases in the allocation of runs.
individuals to groups driven by their association prefer-
ences (captured in the matrix B). Network analysis
Sampling proceeds as follows: A group l is selected at
random (from groups we have not previously sampled in We first analyze the network using the raw SRI values and
this census, each with the same probability of being use weighted network measures that take into account the
selected) and added to our list of groups to observe g. We values of the association indexes. Next, we examine the
repeat this process until we l have observed our sample impact of selecting different threshold values and compare
proportion of the population Σ gl ( p:n. the results to our weighted analysis. Although biologists
We then model gambit of the group sampling by usually collect weighted association data (as in our model),
recording the group allocations of observed individuals in they often apply a threshold to the data so that they can
each census. This models the way that biologists usually analyze the network using binary measures (the process of
record this information. From these data, we then construct filtering discussed in the “Introduction” section). For
our association index. We used the simple ratio index (SRI) example, if the threshold is set to 0.5, any association
(Ginsberg and Young 1992) in the standard manner (Croft strength greater than or equal to 0.5 will become one, while
et al. 2008) to calculate the association weight wab between any other association strength becomes zero. Thus, this
animal a and animal b: process eliminates weaker network connections and (usu-
ally) strengthens stronger ones. What values should be used
X
wab ¼ : ð3Þ for the threshold is an open question, and may vary
X þ Yab þ Ya þ Yb
depending on the type of system being examined. Typical-
where X is the number of times a pair of animals were ly, ecologists will experiment with the threshold until it
observed in the same group, Ya is the number of times a produces a network that matches their intuition. Implement-
(but not b) was observed in a census, Yb is the number of ing both a weighted and binary analysis allows us to
times b (but not a) was observed in a census, and Yab is the compare the two approaches and to look for any differences
number of times both a and b were observed in different in sampling effects between them.
groups in a census.

Parameter values Results: weighted networks

Croft et al. (2008) analyzed a small slice of the long-term Figure 1 shows the average (over 30 replications for each data
investigation of the red deer (Cervus elaphus) population point) of a number of weighted network measures taken from
on the island of Rum (Clutton-Brock et al. 1982). One of our model. Each figure shows the consequence of the number
the pieces of information gathered in this study has been the of censuses and the proportion of individuals sampled in each
identity of members of groups of deer during regular visual census. The value of the measure with the maximum number
censuses. This is exactly the sort of data for which we can of censuses and all individuals sampled can be taken as the
construct networks via the gambit of the group. Croft et al most reliable indicator of the true value of the measure. The
(2008) used as an illustrative data set 26 censuses of group results, generally, show (with the exception of the mean
membership between January and May 1990. There were degree) that the number of censuses is much more important
342 deer (212 female and 130 male) in their network with than the proportion of individuals sampled for most measures
all links included. Croft et al. (2008) used this data set to (although, both affect the results). Figure 1g shows a measure
explore the effect of filtering on simple node-based binary of trait assortativity where there is no assortativity in the
network measures and whether these measures were population being sampled (b=0). Conversely, Fig. 1h shows
sensitive to trait segregation. For the present study, we use the same measure where there is assortativity by trait in the
similar population parameters to the earlier study. We, population (b=1). For all sampling levels (even very low
therefore, set n=400 (the number of individuals), f=0.6 (the levels), we do not find assortativity by trait when it does not
proportion of individuals that are female), and set both the exist. Additionally, we always find assortativity by trait when
mode and shape of the Pareto distribution to one to fit the it does exist. However, if the sampling is inadequate—
Behav Ecol Sociobiol

Fig. 1 Mean network measures


for weighted networks over the
number of censuses and propor-
tion sampled. Each data-point is
averaged over 30 replications
taken from networks each with a
different random seeds. a Mean
degree. b Mean path length. c
The clustering. d The degree
correlation. e The betweenness.
f Mean C.V. of the betweenness.
g The mean weighted trait
assortativity where there is no
trait bias (b=0). h The trait
assortativity where there is a
strong trait bias (b=1). There is
no individual bias (d=0)
Behav Ecol Sociobiol

particularly, if not enough censuses are taken—we record a Results: biased grouping
level of assortativity that is very much artificially high (note
that the estimate with most samples settles on ∼0.3, while the We tested the above condition with an individual bias
estimate with the least samples is ∼0.7). A similar (d = 1) after all censuses with the entire population
observation applies to most of the other measures. We can, sampled. We checked that the individual bias worked by
therefore, see that without enough samples, researchers comparing the association value (SRI) of individuals with
sampling real populations are unlikely to find assortativity the top ten highest biases, with ten randomly selected
where it does not exist, but are likely to find highly associations, replicated over 10 runs and found that the
exaggerated values where a population is moderately group allocation did assign biased individual pairings into
assorted. For most measures, however, increasing the groups more often (a 2-tailed t test revealed p<0.001).
number of censuses gives reasonable results. However, we found that individual pairing, biases make
no qualitative difference to the overall network statistics
(thus, no graphs shown).
Results: binary networks

Figures 2, 3, and 4 show the results of setting the threshold Further analyses
to 0.2, 0.5, and 0.8, respectively. We select these threshold
values to show the impact of selecting high medium and In the model above, we sample a proportion of the
low values. It is easy to see from comparing these figures individuals in the population (still, always sampling an
that the choice of threshold can drastically change the entire group when encountered). We have also examined
resulting network measures. Additionally, the different the effect of basing our sampling on a proportion of
threshold levels have different effects depending on the groups, rather than proportion the population. We found
level of sampling (conversely, the level of sampling has a similar results, except that with low proportions sampled,
different effect depending on the threshold). the measure, generally, became less accurate, meaning that
A threshold of 0.2 gives results that a qualitative when we base our sampling on the proportion of groups
visual inspection shows are, easily, most similar to our sampled, rather than the proportion of the population
weighted analysis (Fig. 2). However, the level of trait sampled, the proportion sampled becomes more important
assortativity remains high (∼0.7; Fig. 2h) even with our even for our weighted measures (Electronic appendix
best sampling effort. We know from our weighted analysis figure A1). We examined our findings when varying the
that assortativity is closer to 0.3. Thus, there is nothing we exponent K of the group size distribution and also with a
can do with our sampling effort to correct this: it is simply much higher population size (1,000 animals). In both cases,
a result of our threshold selection. Assortativity becomes we found no significant difference in the results (e.g.,
falsely exaggerated further as we increase the threshold to Electronic appendix figure A2).
0.5, where our best sampling effort shows assortativity of We tested a situation where the population comprises
one (Fig. 3h). This strongly indicates that no males or 80% females and found that a small amount of
females group together, although, this is untrue. Addition- sampling is robust (Electronic appendix figure A3).
ally, this threshold has the highest variation in assortativ- However, this is not surprising because we know (from
ity, for different levels of sampling, for the case where the Fig. 1h) that there is an initial sampling bias for high
population actually has no assortativity (including an levels of assortativity. A sexually assorted population
average false positive of ∼0.2 at our best sampling effort). comprising of mainly females is clearly going to give a
Only when we increase the threshold to 0.8 (Fig. 4h) does high level of assortativity. Thus, the bias towards high
increased sampling effort improve our assortativity estimate assortativity at low levels of sampling is unlikely to provide
(although, the best estimate gives a slightly lower measure us with a discrepancy when the actual assortativity is high.
than our weighted analysis). However, even here, we need a However, we are still unable to have confidence in the
very high sampling effort (in terms of both censuses and the assortment measure at low levels of sampling because, with
proportion sampled) before we get a realistic measure real populations, we would not know a priori whether our
(otherwise, we get a highly exaggerated value). population is highly assorted by trait (and so, would not
Compared to the weighted analysis, there is a clear know if the high assortativity of our small sample is a true
difference in the type of sampling effort required to get representation).
reasonable results. Whereas, the number of censuses was Although animal group size distributions are typically
more important than the proportion sampled in the right skewed (such as our red deer group size distribution),
weighted analysis, both become equally important for the we also configured our model with group-sizes selected
binary analysis. from a normal distribution (μ=10, σ=4) to examine the
Behav Ecol Sociobiol

Fig. 2 Mean network measures


for binary networks with a
threshold of 0.2. Each data-point
is averaged over 30 replications
taken from networks each with a
different random seeds. g Shows
the mean weighted trait assorta-
tivity where there is no trait bias
(b=0) and h shows the trait
assortativity where there is a
strong trait bias (b=1). There is
no individual bias (d=0)
Behav Ecol Sociobiol

Fig. 3 Mean network measures


for binary networks with a
threshold of 0.5. Each data-point
is averaged over 30 replications
taken from networks each with a
different random seeds.
g Shows the mean weighted trait
assortativity where there is no
trait bias (b=0) and h shows the
trait assortativity where there is
a strong trait bias (b=1). There is
no individual bias (d=0)
Behav Ecol Sociobiol

Fig. 4 Mean network measures


for binary networks with a
threshold of 0.8. Each data-point
is averaged over 30 replications
taken from networks each with a
different random seeds. g Shows
the mean weighted trait assorta-
tivity where there is no trait bias
(b=0) and h shows the trait
assortativity where there is a
strong trait bias (b=1). There is
no individual bias (d=0)
Behav Ecol Sociobiol

significance of the distribution. We generally found through a network, then chance encounters with many
qualitatively similar results when we used the Pareto individuals might be as (or more) important than capturing
distribution (Electronic appendix figure A4). When group preferential association. Filtered networks are also useful for
formation was biased by sex, however, we found a much looking for trait assortment by association type. For example,
higher level of assortment (even with the highest level of Croft et al. (2009) studied associations in a wild population
sampling). Given that we used the same algorithm for of Trinidadian guppies, finding that strong ties were
group allocation, this suggests that assortment by a positively assorted and weak ties were negatively assorted
dichotomous trait such as sex may be more likely to occur by behavioral trait. This finding is only possible by thresh-
strongly when group sizes are normally distributed, rather olding the network to preserve only strong associations or
than when they are right skewed. only weak associations. However, when doing this, sampling
issues become more pertinent.
Our simulations show that little can be deduced from
Discussion gambit of the group networks with limited sampling.
Comparison to a randomized null model—that controls
We offer strong, independent support to the argument made for group size distributions, the type of sampling, and the
by Lusseau et al. (2008) advocating the use of weighted number of samples—would greatly help matters. However,
network measures, rather than filtering data prior to even when using an appropriate null model, an inadequate
analysis. We show that the value of the threshold used to level of sampling means it will only reveal whether
filter association data can drastically change the resulting observed measures are higher or lower than we might
network measures. Crucially, we show that different filters expect. We must be more cautious regarding our interpre-
have different effects depending on the level of sampling. tation of the absolute structure of the network. We might,
Researchers commonly use mid-valued thresholds to filter for example, find that our data show a higher positive trait
network association data (see Croft et al. 2008). Often, this assortativity (in dichotomous traits) than is shown to be
value is arrived at using an intuitive approach: exploring likely by a null model where we keep the connections
network plots with different threshold levels and selecting constant and randomly shuffle the traits. However, in our
the filter that appears to give the best core structure. model, trait assortativity is artificially high with inadequate
However, it is often difficult to determine which individuals sampling, showing positive trait assortativity where there is
or connections are important (or which connections are no assortment. Although our null model comparison would
collectively important). Our comparison of binary thresh- reveal that the assortativity is higher than expected, it does
olded networks with weighted networks shows that for our not tell us the extent of the assortativity. Our weighted
simulated gambit of the group associations, lower values of analysis for the assorted animals, for example, shows a
the threshold are a better match to our weighted analysis. much higher level of assortativity (for our dichotomous
This does not mean that lower thresholds are always going trait) for a low level of sampling than for a (more accurate)
to be the best option, but it does mean that threshold high level of sampling (Fig. 1g). Yet, the null model
selection may not be as simple as picking out core (essentially, Fig. 1h) would, on average, show zero
structures manually. Given that weighted analyses over- assortativity for each. So, although, we would have
come such issues, they seem to be a sensible choice. confidence that we have assorted data, we do not know
Filtered networks are, however, a sensible choice for the the extent of the assortment. For this, there is little
graphical display of network structures as they can provide substitute for a good level of sampling.
researchers with a clear view of the core structure of the We can cautiously derive several further points of advice
network, rather than a typically visually incomprehensible for ecologists sampling animal association networks—at
mass of weighted connections. Weighted analyses also have least with the SRI associations—from our simulation study.
the advantage of increasing data resolution without increasing We recommend that when sampling effort is limited and a
sampling effort. When we filter a network, we are discarding weighted analysis will be used, then effort is better invested
data and therefore, reducing the resolution of our data, more heavily in the number of censuses (the number of
meaning that we might need to sample for slightly longer sampling events after fission–fusion is expected to occur)
when performing a binary analysis. Another point to keep in than the proportion sampled in each census (although, this
mind is that rare (how rare depends on the threshold level) or should not be neglected). This approach can typically
random associations might be important (Granovetter 1974) resolve most sampling issues quite easily. If, however,
depending on the purpose of our network investigation. This sampling effort is limited and a binary analysis will be
might particularly be the case if there are more of these performed, then caution needs to be taken when selecting a
discarded connections than there are common connections. threshold and much more effort needs to be placed into
If, for example, we are interested in disease transmission sampling. With a binary analysis, both the number of
Behav Ecol Sociobiol

censuses and the proportion of individuals observed in each Croft DP, James R, Krause J (2008) Exploring animal social networks.
Princeton University Press, Princeton
census become equally crucial (i.e., you need a high level
Croft DP, Krause J, Darden SK, Ramnarine IW, Faria JJ, James R
of each). Further research is needed to fully assess the (2009) Behavioural trait assortment in a social network: patterns
generality of our results. For example, the inclusion of a and implications. Behav Ecol and Sociobiol 63:1495–1503
few well-connected individuals in the network might Cross PC, Lloyd-Smith JO, Bowers JA, Hay CT, Hofmeyr M, Getz
WM (2004) Integrating association data and disease dynamics in
change the relative importance of the number of censuses
a social ungulate: bovine tuberculosis in African buffalo in the
versus the proportion of the network sampled. It will also Kruger National Park. Ann Zool Fenn 41:879–892
be useful to explore the biases that different association Ginsberg JR, Young TP (1992) Measuring association between
indexes might cause (and biases they might solve). individuals or groups in behavioural studies. Anim Behav
44:377–379
As animal association data collected with the gambit of Granovetter M (1974) Getting a job. Harvard University Press,
the group is weighted, it is possible for researchers who Cambridge, MA
have previously conducted a binary analysis to conduct a Krause J, Ruxton GD (2002) Living in groups. Oxford University
weighted analysis on their current data. Our simulation Press, Oxford University
Lusseau D (2003) The emergent properties of a dolphin social
results suggest that this would be a particularly useful
network. Proc R Soc Biol Sci Ser B 270:186–188
exercise if the proportion of individuals sampled in each Lusseau D, Newman MEJ (2004) Identifying the role that individual
census is estimated to have not been very high when the animals play in their social network. Proc R Soc Biol Lett 271:
data were originally collected. S477–S481
Lusseau D, Whitehead H, Gero S (2008) Incorporating uncertainty in
the study of animal social networks. Anim Behav 75:1809–1815
Acknowledgments We would like to thank two anonymous referees, Mitchell M (1998) An introduction to genetic algorithms. MIT,
Darren Croft, Jason Noble, and Jon Pitchford for their valuable input. Cambridge, Massachusetts
DWF is funded by NERC grant NE/E016111/1. Newman MEJ (2003) The structure and function of complex
networks. SIAM Review 45:167–256
Okubo A (1986) Dynamic aspects of animal grouping: swarms,
schools, flocks, and herds. Adv Biophys 22:1–94
References
Sundaresen SR, Fischhoff IR, Rubenstein DI (2007) Network metrics
reveal differences in social organization between two fission-
Bejder L, Fletcher D, Bräger S (1998) A method for testing fusion species, Grevy’s xebra and onager. Oecologia 151:140–149
association patterns of social animals. Anim Behav 56:719–725 Wasserman S, Faust K (1994) Social network analysis: methods and
Cairns SJ, Schwager SJ (1987) A comparison of association indexes. applications. Cambridge University Press, New York
Anim Behav 35:1454–1469 Whitehead H (1999) Testing Association Patterns of Social Animals.
Clutton-Brock TH, Guiness FE, Albon SD (1982) Red deer: behaviour Anim Behav 57:F26–F29
and ecology of two sexes. University of Chicago Press, Chicago Whitehead H (2008) Analyzing animal societies: quantitative methods
Croft DP, Arrowsmith BJ, Bielby J, Skinner K, White E, Couzin ID, for vertebrate social analysis. University of Chicago Press,
Magurran AE, Ramnarine I, Krause J (2003) Mechanisms Chicago
underlying shoal composition in the Trinidadian guppy, Poecilia Whitehead H, Dufault S (1999) Techniques for analyzing vertebrate
reticulata. Oikos 100:429–438 social structure using identified individuals: review and recom-
Croft DP, Krause J, James R (2004) Social networks in the guppy mendations. Adv Study Behav 28:33–74
(Poecilia reticulata). Proc R Soc London 271:S516–S519 Wolf JBW, Mawdsley D, Trillmich F, James R (2007) Social structure
Croft DP, James R, Ward AJW, Botham MS, Mawdsley D, Krause J in a colonial mammal: unravelling hidden structural layers and
(2005) Assortative interactions and social networks in fish. their foundations by network analysis. Anim Behav 74:1293–
Oecologia 143:211–219 1302

You might also like