Content uploaded by Russell Hill
Author content
All content in this area was uploaded by Russell Hill on Oct 26, 2015
Content may be subject to copyright.
Content uploaded by Russell Hill
Author content
All content in this area was uploaded by Russell Hill
Content may be subject to copyright.
SOCIAL NETWORK SIZE IN HUMANS
R. A. Hill
University of Durham
and
R. I. M. Dunbar
University of Liverpool
This paper examines social network size in contemporary Western society
based on the exchange of Christmas cards. Maximum network size aver-
aged 153.5 individuals, with a mean network size of 124.9 for those indi-
viduals explicitly contacted; these values are remarkably close to the group
size of 150 predicted for humans on the basis of the size of their neocortex.
Age, household type, and the relationship to the individual influence net-
work structure, although the proportion of kin remained relatively con-
stant at around 21%. Frequency of contact between network members was
primarily determined by two classes of variable: passive factors (distance,
work colleague, overseas) and active factors (emotional closeness, genetic
relatedness). Controlling for the influence of passive factors on contact
rates allowed the hierarchical structure of human social groups to be de-
limited. These findings suggest that there may be cognitive constraints on
network size.
KEY WORDS:Frequency of contact; Group size; Humans; Neocortex size;
Social networks
Copyright 2003 by Walter de Gruyter, Inc., New York
Human Nature, Vol. 14, No. 1, pp. 53–72. 1045-6767/03/$1.00+.10
53
Received March 22, 2002; accepted July 19, 2002; revised version received October
30, 2002.
Address all correspondence to Evolutionary Anthropology Research Group, Department of
Anthropology, University of Durham, 43 Old Elvet, Durham DH1 3HN, U.K. Email:
r.a.hill@durham.ac.uk
Analyses of human social networks have a long history in both the socio-
logical and anthropological literature (Milardo 1988). However, relatively
few studies have attempted to investigate complete social networks in hu-
mans (McCarty et al. 1997), primarily due to the difficulty in estimating
and defining an individual’s “network” from the range of interactions that
exist within everyday life. As a result, studies have tended to focus on de-
termining total network size (Johnson et al. 1995; Killworth et al. 1990,
1998; McCarty et al. 2001; Pool and Kochen 1978), with relatively little at-
tention paid to the interactions within these networks. However, depend-
ing on the definition of personal networks, and the required relationships
or ties between individuals to warrant inclusion in a network, previous
studies have found total networks to run from about 250 individuals
(Killworth et al. 1984) to approximately 5,000 (Pool and Kochen 1978; Kill-
worth et al. 1990). There is thus little consensus as to what constitutes a so-
cial network in humans.
In primates, social networks are more easy to define (see Kudo and Dun-
bar 2001) and are often delimited by the size of the social group. Further-
more, primate social relationships are generally characterized by intense
social grooming (Dunbar 1991), and both primate group size (Dunbar 1992)
and grooming clique size (Kudo and Dunbar 2001) are a function of relative
neocortical volume. Similar relationships have been reported for carnivores
(Dunbar and Bever 1998) and cetaceans (Marino 1996; Tschudin 1997).
These findings suggest that there may be a cognitive constraint on the size
of social networks in those species that live in intensely social groups (as op-
posed to simple aggregations), perhaps because the number or volume of
neocortical neurons limits an organism’s information processing capacity,
and hence the number of social relationships that an individual can moni-
tor simultaneously (Dunbar 1992, 1998; Barton and Dunbar 1997).
Since the size of the human neocortex is known, the relationship be-
tween group size and neocortex size in primates can be used to predict the
cognitive group size for humans. Dunbar (1993) utilized this approach to
predict that humans should live in social groups of approximately 150 in-
dividuals. Evidence from the ethnological literature provides some sup-
port for this, since census data from a range of tribal and more traditional
societies indicate that groups of about this size are in fact a common com-
ponent of human social systems (see data collated by Dunbar 1993; see
also Barrett et al. 2002).
This raises the question as to whether modern, postindustrial societies
also exhibit a similar pattern, with a discernible grouping of about 150 in-
dividuals embedded into the somewhat diffuse and dispersed social sys-
tems in which most of us now live. Recent approaches for estimating
personal network sizes in contemporary societies have asked respondents
to estimate the number of people they know in specific subpopulations of
54 Human Nature, Vol. 14, No. 1, 2003
known size (e.g., diabetics) to generate estimates of maximum network
size (McCarty et al. 2001). Such methods, while producing reliable esti-
mates of maximum network sizes that can have a number of applied im-
plications (e.g., Killworth et al. 1998), provide little information on which
relationships are valued within networks or the way in which networks
are maintained. Furthermore, they also estimate the maximum number of
individuals known, rather than identifying those people an individual
considers important and whose relationship they value. As a consequence,
they tell us little about the way in which humans may actively maintain
contact with a network of specific individuals.
In Western societies at least, the exchange of Christmas cards represents
the one time of year when individuals make an effort to contact all those
individuals within their social network whose relationships they value. As
a consequence, this activity provides a unique insight not only into the size
of individuals’ social networks, but also into the way in which these net-
works are structured, both in terms of demographic composition and the
frequencies with which individual relationships are serviced.
METHODS
A questionnaire was designed to be completed as individuals were
sending out their Christmas cards. Such timing should ensure that self-
reporting errors are minimized (a significant problem in previous studies:
Milardo 1988), since responses did not rely upon the memory of those
completing the questionnaire. In recent years, questionnaire design has
been the focus of considerable discussion (Milardo, ed. 1988; McCarty et
al. 1997); the main conclusion from this has been that questionnaires which
take more than a few minutes to complete or are too complex in their de-
sign tend to result in loss of concentration and poor levels of completion
(Dunbar and Spoors 1995). Because the information we required was both
detailed and lengthy, we preferred a design in which a small number of in-
dividuals distributed questionnaires to personal acquaintances. Although
this drastically reduces the number of individuals who receive question-
naires, the sense of obligation that a respondent owes to the distributor
from whom he/she received the questionnaire greatly increases the return
rate (proportion of all questionnaires completed and returned), especially
when that questionnaire is long and complex. Only one questionnaire was
completed per household, but there remains a potential for overlap be-
tween households that belong to the same extended social network. How-
ever, everyday experience suggests that even close friends in modern
urban society do not share all their friends and acquaintances; moreover,
there is no reason to suspect that the size of any one person’s social net-
Human Social Network Size 55
56 Human Nature, Vol. 14, No. 1, 2003
Table 1. Information Requested on Questionnaire To Be Listed for Each
Christmas Card
Category Definition
Distance Approximate distance to recipient in miles (overseas
individuals were listed by country—mean distance to
that country was later estimated for analysis)
Relationship Relationship of respondent to contact: relative (stating
degree of relatedness, e.g., brother, sister-in-law, niece),
work colleague, friend, neighbor
Social Status Social structure of individuals contacted: single individ-
ual, couple, or family (indicating structure of family,
e.g., husband, wife, 3 children) as well as which indi-
viduals within household were contacted directly (e.g.,
wife only)
Last Contact An estimate to the nearest month, or week if contact is
within last month, of when the respondent was last in
contact with their acquaintance, or to indicate that pre-
vious contact was by last Christmas card only. If a let-
ter or long message was included in the Christmas
card this was also noted.
Emotional Closeness A rating of how emotionally close the respondent felt to
the principal contact where 10 is very emotionally
close and 0 is not emotionally close at all
work dictates in any way the size of anyone else’s when they do not belong
to the same household. To minimize possible cultural effects (e.g., Kim and
McKenry 1998), all respondents were white British.
In completing the questionnaire, respondents were initially asked for
their age and sex and the number and identity of individuals living in their
household. Respondents were also asked to list those individuals to whom
they regularly send cards but were not doing so this year because they ex-
pected to see them at Christmas. These individuals, as well as household
members, were included in the analyses of network size since they are
clearly integral to the respondents’ social networks even though they
might not be included in the Christmas card network. For each Christmas
card, respondents were asked to provide a number of details about the in-
dividual (or individuals) to whom the card was being sent (see Table 1).
For certain analyses, responses were recoded to produce data in a more
quantifiable format. In terms of social status, the recipient household was
classified as one of three categories: individual, couple, or family. Simi-
larly, an individual’s relationship to a given recipient was coded as one of
four categories: relative, friend, neighbor, or work colleague. In order to
account for potential differences between genetic relatives and relatives by
marriage, we created two measures of relatedness; genetic relatedness (the
Human Social Network Size 57
coefficient of relatedness) and affinal relatedness (an index that mirrored
exactly the equivalent genetic relationship—i.e., a brother-in-law is con-
sidered to have the same coefficient of relatedness as a full biological
brother, namely 0.5). Although it is possible that not all affinal relatives
were identified as such on the questionnaires (few respondents, for exam-
ple, distinguished between biological and affinal nephews and nieces),
such that they were taken to be genetic kin, the two categories at least
allow a preliminary exploration of the extent to which distinctions are
made between biological and social relatives in managing social networks.
Because respondents were completing questionnaires on behalf of a
household (most Christmas cards are typically sent from all members of
the household, or at least the adults), we make no attempt to determine the
importance of gender in determining network size. Although sex differ-
ences in network size have been reported in previous studies (Dickens and
Perlman 1981; for empathy groups and support cliques only: Dunbar
and Spoors 1995), the sampling design we have used makes it inappropri-
ate to explore this issue in the present case.
All continuous variables were tested for normality, and where they were
found to deviate significantly from the normal distribution (Kolmogorov
Smirnov, p< 0.05), the data were log-transformed for certain analyses. All
tests are two-tailed.
RESULTS
Forty-three questionnaires were returned, between them involving a total
of 2,984 Christmas cards. The number of individuals contacted via each
card ranged from 1 to 9. The mean number of Christmas cards sent was
68.19 (range 11–149). Since many cards were sent to couples and families,
this results in a mean network size of 153.5 (84.5) (Figure 1). However,
if we consider only those individuals whom respondents stated they ac-
tively contacted (as opposed to everyone actually residing in the house-
hold to which the card was sent), then mean network size is 124.9 (68.0)
for the 22 questionnaires for which this distinction was made.
There is a near-significant quadratic relationship between maximum
network size and age of the respondent (Figure 2: r2= 0.132, F2,40 = 3.04,
p= 0.059). While there is considerable variation in maximum network size
throughout the entire age range, it is clear that large network sizes are not
especially characteristic of either young or old individuals. These results
remain broadly similar if the number of individuals actually contacted is
considered.
The recipients of cards can be differentiated by whether they live alone,
as a couple, or as a family. Significant differences exist in the proportion
58 Human Nature, Vol. 14, No. 1, 2003
Figure 1. Frequency distribution of total number of individuals in respondents’
social networks.
of the maximum network that is made up of these three categories of
recipients, and these compositions are also influenced by age (Figure 3: r2
= 0.559; F14,129 = 11.00, p< 0.001; household type factor: F2,129 = 30.4, p<
0.001; household by age interaction: F8,129 = 7.6, p< 0.001). Almost identical
relationships are observed if the analysis is conducted for those network
members actually contacted or by the number of Christmas cards sent.
Previous studies have often sought to examine network composition in
terms of kinship and other relationship categories. Within this sample,
there are significant differences in the proportion of maximum network
size that is made up by five main relationship types (genetic relative, affi-
nal relative, friend, neighbor, and work colleague), although these pro-
portions are not influenced by age (Figure 4: r2= 0.857; F24,215 = 57.45, p<
0.001; relationship type factor: F4,215 = 229.6, p< 0.001; relationship by age
interaction: F16,215 = 1.10, p> 0.30). The mean proportion of each relation-
ship type within the typical maximum Christmas card network is 0.21 for
genetic relatives, 0.04 for non-genetic relatives (thus 0.25 for all relatives
combined), 0.63 for friends, 0.04 for neighbors, and 0.08 for work col-
Human Social Network Size 59
Figure 2. Relationship between maximum network size and age.
leagues. Again the results are virtually identical if the analysis is repeated
by number of Christmas cards sent or only for those network members
contacted directly.
Stepwise least-squares regression was used to determine the factors that
best explain the patterns of contact within an individual’s network. The
best-fitting model, given in Table 2, incorporates seven variables in the final
model (r2= 0.394, F7,2909 = 269.66, p< 0.0001), with only the variable for
whether an individual is a spouse/partner excluded from the list of avail-
able independent variables. Time since last contact increases as distance to
the individual increases, decreases as emotional closeness increases, de-
creases if the individual is a work colleague, decreases if the contact is over-
seas, decreases as the coefficient of relatedness increases for both genetic
and affinal relatives, and increases with age. The sign of the relationship for
overseas contacts has changed relative to a simple bivariate correlation (r=
0.192, N= 2984, p< 0.0001), suggesting that people overseas are contacted
more frequently once distance, emotional closeness, and whether the indi-
vidual is a work colleague have been controlled for.
60 Human Nature, Vol. 14, No. 1, 2003
Figure 3. Mean proportion of total network made up of different household types
for five age categories.
In cases where recipients were contacted only once a year, respondents
sometimes included a letter or extended message in their card. This prac-
tice might indicate that these recipients were held in higher regard. For
this subset of the social network, distance to contact, emotional closeness,
and relatedness for genetic relatives all form significant components of a
logistic regression model determining whether or not a letter is included
with the card (Table 3). The probability of a letter being included with a
Christmas card increases with distance to contact and emotional closeness,
but decreases with genetic relatedness. The coefficient of relatedness for
affinal relatives does not form a significant component of this model.
However, since all affinal relatives may not have been identified as such on
the questionnaires (and thus classified as genetic relatives), there may still
be a partial effect for affinal relatives. This is suggested by the fact that
treating the coefficient of relatedness for all relatives combined as a single
independent variable results in a slight improvement on the overall model
(r2= 0.192, 2LL = 783.15, df = 3, p< 0.0001; relatedness term: r2= 0.014,
Human Social Network Size 61
Figure 4. Mean proportion of total network made up of kin, non-genetic relatives,
friends, neighbors, and work colleagues for five age categories.
Table 2. Stepwise Regression Analysis of Factors Determining Time to Most
Recent Contact (An independent variable not incorporated in the final
model is Spouse/Partner [YES/NO].)
r2F7,2902 p
Final Model 0.393 268.855 <0.0001
Variable r2B p
Constant — 0.292 <0.01
Log (Distance to Contact) 0.221 0.243 <0.0001
Log (Emotional Closeness) 0.099 0.610 <0.0001
Work Colleague (YES/NO) 0.051 0.385 <0.0001
Overseas (YES/NO) 0.012 0.245 <0.0001
Log (Relatedness—Genetic) 0.004 0.738 <0.0001
Log (Relatedness—Affinal) 0.004 0.922 <0.0001
Log (Age of Respondent) 0.002 0.217 <0.0001
62 Human Nature, Vol. 14, No. 1, 2003
Table 3. Logistic Regression Analysis of Factors Determining Whether a Letter Is
Included with the Christmas Card (for those individuals in the network
contacted by Christmas card only). Independent variables in initial model
are those included in the linear regression analysis.
r2
2 Log L χ2df p
Final Model 0.187 783.149 111.617 3 < 0.0001
Variable r2BWald df p
Log (Distance to Contact) 0.153 0.845 63.96 1 < 0.0001
Log (Emotional Closeness) 0.026 2.034 17.73 1 < 0.0001
Log (Genetic Relatedness) 0.008 6.961 4.75 1 < 0.03
Constant —4.479 121.80 1 < 0.0001
Wald = 7.78, df = 1, p= 0.005). Nevertheless, the possibility exists that the
respondent may only make the effort of writing a letter if the recipient is
their own genetic kin.
Recent studies have suggested that human social networks have a hier-
archical structure, with frequency of contact being used to differentiate be-
tween levels within the social network (Dunbar 1993; Dunbar and Spoors
1995). However, since frequency of contact is dependent on stochastic fac-
tors it is not surprising that there is considerable variation in the reported
hierarchical group size estimates (e.g., published sympathy group sizes
range from 7 to 20: Dunbar and Spoors 1995; Hays and Oxley 1986; Mc-
Cannell 1988; Rands 1988). The preceding analysis indicated that three
“passive” factors might influence frequency of contact between network
members independent of any intentional contact: distance to individual
and whether the contact is a work colleague or lives overseas. If the com-
bined influence of these factors is controlled for, we should get a clearer
picture of whether social networks are differentiated hierarchically on the
basis of frequency of contact.
Figure 5 displays cumulative network size against residual frequency of
contact, controlling for distance and whether the contact is a work col-
league or overseas. Reference lines are displayed to indicate the approxi-
mate hierarchical grouping levels identified by Dunbar (1998). Clear
inflection points can be observed at 7, 20, and 35 individuals, respectively.
Inspection of the pattern in Figure 5 suggests that further grouping levels
might also be identified at about 70 and 100 individuals.
Figure 6 suggests that, once stochastic factors have been accounted for,
emotional closeness is likely to be the key parameter underlying the fre-
quency of contact between individuals: time since last contact declines as
the level of emotional closeness increases. The apparently high contact
Human Social Network Size 63
Figure 5. Cumulative network size on the basis of frequency of contact, control-
ling for distance to contact and whether the contact is a work colleague or
overseas, for maximum network reached and those actively contacted. Hori-
zontal lines indicate approximate hierarchical grouping levels reported in the
literature (Dunbar 1998: support cliques 5; sympathy groups 12; bands
35; cognitive group size 150).
frequencies for individuals ranked as zero emotional closeness are almost
certainly a sampling artifact. Only four questionnaires utilized the zero
emotional closeness value, and two-thirds of these data were from a single
questionnaire. Thus the low degree of emotional differentiation employed
by this respondent, coupled with a small sample size for this emotional
closeness category, may be the explanation for the apparently anomalous
downturn on the left-hand side of Figure 6. Despite this, there is a clear
trend for contact latency to decline as emotional closeness increases, such
that those of the highest emotional closeness are invariably contacted on
a weekly basis. In turn, this suggests that emotional closeness may be
the key parameter underlying the hierarchical differentiation of social
networks.
64 Human Nature, Vol. 14, No. 1, 2003
Figure 6. Mean (and standard error) number of months since last contact between
individuals based on their degree of emotional closeness.
DISCUSSION
In contemporary Western societies, we regularly interact with a wide array
of other individuals: social networks of up to 5000 individuals have, for ex-
ample, been reported (Pool and Kochen 1978; Killworth et al. 1990). How-
ever, many of the individuals we meet in this way do not form a part of our
intimate social network, and we have no formal relationships with them.
In contrast, the relationship between neocortex size and group size across
primates suggests that humans live in groups (or, rather, social networks)
of approximately 150 individuals (Dunbar 1993). While there is some evi-
dence to support this claim in the ethnographic literature (Dunbar 1993),
no concerted attempt has so far been made to test this prediction in con-
temporary society.
We proposed that Christmas represents the one time of year when indi-
viduals in Western cultures make a concerted effort to make contact with
their entire social network, or at least with those individuals whose rela-
Human Social Network Size 65
tionships they value and consider important. In addition to personal face-
to-face contact, the sending of Christmas cards allows us to reach those in-
dividuals for whom physical distance or time prevent us from meeting in
person. Estimated on this basis, the mean network size for the individuals
in this sample was 153.5 individuals, with a slightly lower figure of 124.9
if only those individuals that respondents actually intended to contact are
considered (i.e., excluding certain other members of the recipient’s house-
hold). Both of these values fall well within the confidence limits (100 to
231) for human group sizes predicted on the basis of the relationship be-
tween neocortex size and group size across primates (Dunbar 1993), the
observed maximum network size being remarkably close to the value of
147.8 (45.8) predicted for humans. Furthermore, the network size for
those individuals actively contacted is close to the value of 134 obtained by
Killworth et al. (1984) in their attempt to estimate the number of individu-
als that subjects felt they could ask a favor of in small world experiments.
This could be interpreted as suggesting that social networks contain a
(small) number of individuals who might be considered peripheral to the
core network, but who are nonetheless included because they form part of
the immediate household or family of a core network member. Such indi-
viduals might not be considered appropriate people to seek favors from
and might not themselves be granted support so readily. However, the ex-
tent to which these results can be generalized to all individuals cannot re-
ally be addressed.
Previous studies have shown that there can be considerable variation in
social network size between individuals, with factors such as age (Dickens
and Perlman 1981), marital status (Rands 1988), gender (Dunbar and
Spoors 1995), physical attractiveness (Reis et al. 1982), personality (Wilson
1995), and levels of education, occupation, and income (Belle 1982) being
major influences on network size and structure. These factors almost cer-
tainly account for some of the variation observed in the present study. Al-
though our methodology precludes us from commenting on gender
differences in network size, and we did not request details of respondents’
social or personal circumstances, we are able to comment on the effects of
age. In our study, significant changes were observed in network structure
with age, most notably with respect to the demographic composition.
From about 30 years of age onward, couples and families make up a higher
a proportion of the total network, and this trend persists until late in life.
One likely explanation for this is that changes in network composition are
influenced by the presence of dependent children. However, despite the
changes in general network composition, there is little change in the pro-
portion of kin in the network, suggesting that these relationships remain
relatively constant throughout life.
Kinship is known to play an important role in human social relationships
66 Human Nature, Vol. 14, No. 1, 2003
(Dunbar and Spoors 1995; Hughes 1988; Keesing 1975), even though kin
do not necessarily account for a large proportion of the network. In Kill-
worth et al.’s (1984) study, less than 10% of the total network was identi-
fied as kin, with the vast majority (86%) considered friends. Dunbar and
Spoors (1995) reported a figure of 37.5% for the proportion of kin, although
this study focused on sympathy group sizes and so the figure is likely to
be an overestimate of kin in the total network because kin are more likely
to appear in the more proximal segments of the network. The proportion
of kin in the total network reported here (approximately 21%) lies com-
fortably between these two values. The fact that only 4% of the network
was identified as non-genetic (affinal) relatives could be explained by
one of two possibilities. First, it may be that not all affinal relatives were so
identified on the questionnaire, but may instead have been listed as bio-
logical relatives. This suggestion is given credence by the fact that the inci-
dence of single individuals and divorcees in our sample is unlikely to have
been such as to produce such a significant level of bias. The second possi-
bility is that affinal kin receive much less attention than biological kin. Ei-
ther way, the results of this study suggest that, even though most of an
individual’s social network is composed of unrelated individuals, kin (and
principally biological kin) are included with disproportionately high fre-
quencies given their actual representation within both the national and the
local populations.
Frequency of contact with members of the social network is determined
by two general groups of factors: extrinsic factors (that impact on the like-
lihood of face-to-face meetings) and intrinsic factors (that determine the
frequency with which individuals are intentionally contacted). Distance
and whether the individual is a work colleague are the primary extrinsic
factors influencing frequency of contact, although the former is supple-
mented by whether or not the individual concerned lives overseas. Note,
however, that the sign for this factor is reversed in the multivariate analy-
sis: overseas individuals are contacted more frequently than would be ex-
pected on the basis of distance, suggesting that special efforts may be
made to keep contact with those core network members that live so far
away that direct personal contact is precluded. This suggestion is given
some support by the fact that, in such cases, a letter is often also included
with the Christmas card (i.e., a greater effort is made to keep recipients
abreast of news about the sender’s family).
As far as intrinsic factors are concerned, emotional closeness and the co-
efficient of relatedness, for both genetic and affinal relatives, are the key
parameters influencing frequency of contact. In both cases, individuals are
contacted more frequently as closeness increases, reflecting the impor-
tance of these relationships. (Note that, although we can expect these vari-
Human Social Network Size 67
ables to be correlated with each other, the regression equation given in
Table 2 reports their independent effects.) With regard to relatedness, how-
ever, extrinsic factors could also be involved, since family gatherings
could result in a higher frequency of contact that is independent of direct
personal contact between individuals. Nevertheless, kin networks are also
highly interconnected and frequency of contact might be “policed,” en-
suring that high rates of interchange are maintained. Hames (1979), for ex-
ample, found that Ye’kwana villagers of Venezuela interact more often
with close relatives.
Previous studies have suggested that social networks may be hierar-
chically differentiated, with larger numbers of progressively less intense
relationships maintained at higher levels. Dunbar (1998) suggested that
clusterings of relationships tended to occur at 5 (support cliques), 12–15
(sympathy groups), and 35 (bands) individuals, with further higher-level
groupings at 500 and 1500–2000 (equating in the ethnographic literature to
mega-bands and tribes, respectively). Support cliques (defined as all those
individuals from whom one would seek advice, support, or help in times
of severe emotional or financial distress) averaged 4.72 (2.95) individu-
als in one UK sample (Dunbar and Spoor 1995) and 3.01 (1.77) in a US
sample (Marsden 1987), while an estimate of women’s “hair care” net-
works among the !Kung San yielded a mean of 3.8 (Sugawara 1984). The
value of 7 suggested by Figure 5 lies within the upper limits for the mar-
gin of error for these estimates. Published sympathy group sizes typically
lie within the range of 10–15 individuals (mean of 10.9 6.8: Buys and Lar-
son 1979; mean of 11.6 5.64: Dunbar and Spoor 1995), although values as
low as 7 (Hays and Oxley 1986) and as high as 15–20 individuals (McCan-
nell 1988; Rands 1988) have been reported. Much of this variation can
probably be explained by methodological differences between the studies:
Buys and Larson (1979), for example, considered those individuals whose
death would be personally devastating, whereas other studies have
tended to use frequencies of contact over various time periods. The value
of 21 suggested by Figure 3 lies at the upper limit of these estimates. How-
ever, two points are worth noting. First, frequency of contact displays a
considerable degree of stochastic variation and such methods are also
prone to self-reporting errors (Milardo 1988). Second, the value obtained
in this study may be an overestimate because all household members were
ascribed the same contact frequency but almost certainly do not share the
same level of intimacy with the respondent. Afigure in the region of 12–15
may thus not be unrealistic. Finally, in Dunbar’s (1993) analysis of the
group sizes characteristic of modern hunter-gatherer societies, a level of
social organization of 30 to 50 individuals (mean = 37.7 16.8) was ob-
served; these groupings were often described as bands or overnight camps
68 Human Nature, Vol. 14, No. 1, 2003
Table 4. Probability of the Four Grouping Levels Identified in Figure 5 Being an
Estimate of the Four Types of Grouping Hitherto Identified in the Literature.
In each case, the probability of identity is derived from the standard normal
deviate separating the observed (row) and putative (column) value, given
the variance in the latter. Values in bold indicate value is not significantly
different from column value.
Estimated Sizes of Grouping Types Identified in Literature
This Study 3.8 2.29 11.3 6.19 37.7 16.8 147.8 45.8
7z= 1.4
(p= 0.090)
z= 0.7
(p= 0.244)
z= 1.8
(p= 0.034) z= 3.1
(p= 0.001)
21 z= 7.5
(p«0.001) z= 1.6
(p= 0.058)
z= 1.0
(p= 0.162)
z= 2.8
(p= 0.003)
35 z= 13.6
(p«0.001) z= 3.8
(p< 0.001) z= 0.2
(p= 0.436)
z= 2.5
(p= 0.007)
153.5 z= 65.4
(p«0.001) z= 23.0
(p«0.001) z= 6.9
(p< 0.001) z= 0.1
(p= 0.450)
in the ethnographic literature. The grouping level at 35 individuals ob-
served in the present study is sufficiently close to this value to suggest that
its origins in hunter-gatherer societies may not be entirely ecological (as
has previously been assumed). We attempt no explanations for the appar-
ent grouping levels at 70 and 100 individuals that are tentatively sug-
gested by Figure 5. It is possible that these may in part be simple artifacts
of our methodology, which required that all individuals must be contacted
at least once a year if they were to be included in the study.
If we compare the four main grouping levels tentatively identified in
Figure 5 (7, 21, 35, and 153.5) with the mean values obtained for the vari-
ous groupings identified in the literature (means of 3.8 2.29, 11.3 6.19,
37.7 16.8, and 147.8 45.8), it is clear that the values observed in Figure
5 are better estimates of their nominally corresponding values from the lit-
erature than of any other group type (Table 4). In each case, the value from
Figure 5 is significantly different (as reflected in the number of standard
deviations that separate the two values) from all other values except the
one to which it nominally corresponds (or, in some cases, the adjacent
value). The only exceptions are the values of 7 and 21, which do not differ
significantly from the values of the nominal grouping or the next higher
one. However, this is principally a consequence of the fact that sympathy
groups have a relatively larger standard deviation than do either support
cliques or bands. This obscures the fact that the value of 21, for example,
actually lies closer to the mean value for sympathy groups than it does to
that for bands. Taken as a whole, Table 4 suggests that the deviations are
Human Social Network Size 69
least significant along the main diagonal, suggesting that there is an asso-
ciation between the grouping levels identified within this one sample and
those identified in previous studies that focused on more specific types of
relationship.
Note that these grouping levels are strictly cognitively defined: in effect,
they are cognitive constraints on the number of individuals that can be
maintained at a given intensity of relationship (presumably involving fa-
miliarity, trust, etc). The social characteristics attributed to these groupings
or the way they are used within the social system and the labels attached
to them are wholly open to negotiation: each society may make quite dif-
ferent use of these groupings. The only constraint is that if the social func-
tion requires or depends on a particular intensity of relationship, then that
function may be restricted to a particular size of group. While cultural dif-
ferences will exist in network structure (Kim and McKenry 1998), and net-
works within cultures may change through time (Ruan et al. 1997), the
general grouping levels defined by cognitive constraints should remain
consistent.
In summary, it seems that Christmas card networks provide useful in-
sights into human social networks and support the idea that they are uti-
lized to make (at least annual) contact with all those individuals whose
relationships are considered important. Total network sizes estimated from
Christmas card lists are remarkably close to the value of 150 predicted for
human social group size based on the relationship between group size and
brain size across primates. Furthermore, the hierarchical structure of
human social groups hinted at by other studies seems to have some basis
in reality, at least insofar as it is observed in the Christmas card networks of
this UK sample. Thus, even in contemporary western societies, where in-
dividuals are operating egocentric networks within a virtually infinite
array of social possibilities, social network size and differentiation reflect
the sociocentric networks observed in traditional societies, suggesting that
the cognitive constraints on network size may apply universally to all mod-
ern humans.
This project was funded by a grant from Hewlett Packard Research Laboratories
(Bristol) and the Economic and Social Research Council (ESRC). The support of the
ESRC is gratefully acknowledged. This work was part of the programme of the
ESRC Research Centre for Economic Learning and Social Evolution (ELSE). We are
grateful to all those who distributed and completed questionnaires, in particular
J. M. Hill and R. Grainger.
Russell Hill (B.Sc., M.Phil, Ph.D.) is an Addison Wheeler Research Fellow at
the University of Durham. His main research interests are in the evolution of
70 Human Nature, Vol. 14, No. 1, 2003
mammalian social systems. Robin Dunbar (B.A., Ph.D.) is a professor of evolu-
tionary psychology at the University of Liverpool. His research interests span
mammalian behavioral ecology, including humans, cognitive mechanisms, and
Darwinian psychology.
REFERENCES
Barrett, L., R. I. M. Dunbar, and J. E. Lycett
2002 Human Evolutionary Psychology. Basingstoke: Palgrave/Macmillan, and
Princeton: Princeton University Press.
Barton, R. A., and R. I. M. Dunbar
1997 Evolution of the Social Brain. In Machiavellian Intelligence II, A. Whiten
and R. Byrne, eds. Pp. 240–263. Cambridge: Cambridge University Press.
Belle, D. E.
1982 The Impact of Poverty on Social Networks and Supports. Marriage and
Family Review 5:89–103.
Buys, C. J., and K. L. Larson
1979 Human Sympathy Groups. Psychology Reports 45:547–553.
Dickens, W. J., and D. Perlman
1981 Friendship over the Life Cycle. In Personal Relationships 2: Developing Per-
sonal Relationships, S. Duck, and R. Gilmour, eds. Pp. 91–122. London: Aca-
demic Press.
Dunbar, R. I. M.
1991 Functional Significance of Social Grooming in Primates. Folia Primatolo-
gica 57:121–131.
1992 Neocortex Size as a Constraint on Group Size in Primates. Journal of
Human Evolution 22:469–493.
1993 Coevolution of Neocortical Size, Group Size and Language in Humans.
Behavioural and Brain Sciences 16:681–735.
1998 The Social Brain Hypothesis. Evolutionary Anthropology 6:178–190.
Dunbar, R. I. M., and J. Bever
1998 Neocortex Size Predicts Group Size in Some Insectivores. Ethology 104:
695–708.
Dunbar, R. I. M., and M. Spoors
1995 Social Networks, Support Cliques, and Kinship. Human Nature 6:273–
290.
Hames, R. D.
1979 Relatedness and Interaction among the Ye’kwana: A Preliminary Analy-
sis. In Evolutionary Biology and Human Social Behavior, N. A. Chagnon and
W. Irons, eds. Pp. 238–249. North Scituate: Duxbury Press.
Hays, R. B., and D. Oxley
1986 Social Network Development and Functioning during a Lifetime Transi-
tion. Journal of Personality and Social Psychology 50:305–313.
Hughes, A.
1988 Evolution and Human Kinship. Oxford: Oxford University Press.
Human Social Network Size 71
Johnson, E., H. R. Bernard, P. D. Killworth, G. A. Shelley, and C. McCarty
1995 A Social Network Approach to Corroborating the Number of AIDS/
HIV+ Victims in the U.S. Social Networks 17:169–187.
Keesing, R. M.
1975 Kin Groups and Social Structure. New York: Holt, Rinehart, Winston.
Killworth, P. D., H. R. Bernard, and C. McCarty
1984 Measuring Patterns of Acquaintanceship. Current Anthropology 25:391–
397.
Killworth, P. D., E. Johnson, H. R. Bernard, G. A. Shelley, and C. McCarty
1990 Estimating the Size of Personal Networks. Social Networks 12:289–312.
Killworth, P. D., C. McCarty, H. R. Bernard, G. A. Shelley, and E. C. Johnson
1998 Estimation of Seroprevalence, Rape and Homelessness in the United
States Using a Social Network Approach. Evaluation Review 22:289–308.
Kim, H. K., and P. C. McKenry
1998 Social Networks and Support: A Ccomparison of African Americans,
Asian Americans, Caucasians, and Hispanics. Journal of Comparative Family
Studies 29:313–334.
Kudo, H., and R. I. M. Dunbar
2001 Neocortex Size and Social Network Size in Humans. Animal Behaviour
62:711–722.
McCannell, K.
1988 Social Networks and the Transition to Motherhood. In Families and Social
Networks, R. M. Milardo, ed. Pp. 83–106. Newbury Park: Sage.
McCarty, C., H. R. Bernard, P. D. Killworth, G. A. Shelley, and E. C. Johnson
1997 Eliciting Representative Samples of Personal Networks. Social Networks
19:303–323.
McCarty, C., P. D. Killworth, H. R. Bernard, E. C. Johnson, and G. A. Shelley
2001 Comparing Two Methods for Estimating Network Size. Human Organiza-
tion 60:28–39.
Marino, L.
1996 What Dolphins Can Tell Us about Primate Evolution. Evolutionary An-
thropology 5:81–86.
Marsden, P. V.
1987 Core Discussion Networks of Americans. American Sociological Review
52:122–131.
Milardo, R. M.
1988 Families and Social Networks: An Overview of Theory and Methodology.
In Families and Social Networks, R. M. Milardo, ed. Pp. 13–47. Newbury Park:
Sage.
Milardo, R. M., ed.
1988 Families and Social Networks. Newbury Park: Sage.
Pool, I. S., and M. Kochen
1978 Contacts and Influence. Social Networks 1:5–51.
Rands, M.
1988 Changes in Social Networks Following Marital Separation and Divorce.
In Families and Social Networks, R. M. Milardo, ed. Pp. 127–146. Newbury Park:
Sage.
72 Human Nature, Vol. 14, No. 1, 2003
Reis, H., L. Wheeler, N. Spiegel, M. Kernis, J. Nezlek, and M. Perri
1982 Physical Attractiveness and Social Interaction II: Why Does Appearance
Affect Social Experience? Journal of Personality and Social Psychology 43:979–
996.
Ruan, D., L. C. Freeman, X. Dai, Y. Pan, and W. Zhang
1997 On the Changing Structure of Social Networks in Urban China. Social
Networks 19:75–89.
Sugawara, K.
1984 Spatial Proximity and Bodily Contact among the Central Kalahari San.
African Study Monograph (Supplement) 3:1–43.
Tschudin, A.
1997 Relative Neocortex Size and Its Correlates in Dolphins: Comparisons with Hu-
mans and Implications for Mental Evolution. Ph.D. Thesis, University of Natal.
Wilson, R.
1995 An Investigation into the Effects of Individual Differences upon the Size
and Composition of Individuals Social Networks. B.Sc. Dissertation, Univer-
sity of Liverpool.
