Metodološki zvezki, Vol. 3, No.

2, 2006, 185-216

Comparing Social Networks: Size, Density, and

Local Structure
Katherine Faust1

Abstract
This paper demonstrates limitations in usefulness of the triad census for
studying similarities among local structural properties of social networks.
A triad census succinctly summarizes the local structure of a network using
the frequencies of sixteen isomorphism classes of triads (sub-graphs of
three nodes). The empirical base for this study is a collection of 51 social
networks measuring different relational contents (friendship, advice,
agonistic encounters, victories in fights, dominance relations, and so on)
among a variety of species (humans, chimpanzees, hyenas, monkeys, ponies,
cows, and a number of bird species). Results show that, in aggregate,
similarities among triad censuses of these empirical networks are largely
explained by nodal and dyadic properties – the density of the network and
distributions of mutual, asymmetric, and null dyads. These results remind
us that the range of possible network-level properties is highly constrained
by the size and density of the network and caution should be taken in
interpreting higher order structural properties when they are largely
explained by local network features.

1 Introduction
This paper addresses several issues concerning local structure in social
networks. Most generally, it continues the work of Skvoretz and Faust (Faust and
Skvoretz, 2002; Skvoretz and Faust, 2002) modeling similarities in the structural
features of diverse social networks. It also extends the idea of “structural
signatures” for these comparisons (Skvoretz and Faust, 2002). In addition to these
methodological contributions, the empirical example provides insights into the
local nature of the structures of a diverse collection of social networks and in
doing so challenges the basis for comparative modeling of higher order (macro)
structures in networks. In particular, this paper uses triad censuses for network

1
Department of Sociology and Institute for Mathematical Behavioral Sciences; University of
California, Irvine; Irvine, CA 92697 USA; [email protected]
186 Katherine Faust

comparison and points out limitations in their usefulness for that purpose by
showing that these triad distributions are largely explained by more local
properties – network density and the dyad distribution.
The general question addressed throughout this paper is: What accounts for
similarities among triad censuses from diverse social networks? The analytic
strategy is straightforward. Similarities among observed triad distributions for a
collection of social networks are represented using correspondence analysis and
the resulting dimensions of similarity are interpreted using local network
properties.

2 Comparing networks
The majority of social network studies are case studies of a single group or setting.
Relatively less attention has been paid to comparisons using networks from
multiple settings. Studies employing multiple networks focus on one of two
distinct general questions. The first asks whether networks of a specific relational
content, in aggregate, exhibit common structural tendencies. The second asks what
structural features distinguish among different kinds of social relations. In
approaching the first sort of question, some studies examine the same relation
measured in multiple settings. Empirical examples include friendships in schools
or classrooms (Bearman, Jones, and Udry, 1997; Hallinan, 1974b; Leinhardt, 1972;
Snijders and Baerveldt, 2003), social interactions in workplaces (Johnson, Boster
and Palinkas, 2003), and social and economic relations in communities or villages
(Laumann and Pappi, 1976; Rindfuss et al., 2004; Entwisle et al., n.d.) and so on.
Wasserman (1987) and Pattison and Wasserman (1999) describe methodology for
these comparisons. In addition, there are studies in which roughly similar
relations are compared across different settings. Bernard, Killworth, and Sailer’s
studies of informant accuracy using observations and verbal reports of interactions
are an example of such applications (Bernard et al., 1984) as is Freeman’s study of
group structure of social interactions in different settings (Freeman, 1992). One of
the most prolonged projects along these lines is the classic work by Davis,
Holland, and Leinhardt using the sociometric data bank, a collection of
sociometric measurements of positive interpersonal sentiments from different
settings to investigate the presence of structural balance, clustering, hierarchy, and
transitivity (Davis, 1970; Davis and Leinhardt, 1972; Holland and Leinhardt, 1971,
1973). Similarly, Butts (2001) investigated the degree of complexity in social
networks gathered using different data collection protocols.
In contrast, another line of research is concerned with distinctions among
diverse kinds of social relations in disparate groups. The work of Skvoretz and
Faust is a case in point (Faust and Skvoretz, 2002; Skvoretz and Faust, 2002).
Using exponential random graph models (Wasserman and Pattison, 1996), they
compared the direction and magnitudes of parameters characterizing local
Comparing Social Networks: Size, Density, and Local Structure 187

structure in graphs and calculated measures of dissimilarity between graphs for a

variety of social networks. Results showed differences in the “structural
signatures” of different kinds of relations, notably antagonistic relations such as
fighting and dominance on the one hand, and relations of affection (friendship,
liking) and affiliation on the other. Differences between species were apparent
only for the first kind of relation, where humans showed tendencies toward
mutuality and in-stars and away from transitivity whereas non-human primates
showed tendencies in the opposite direction on these properties (Skvoretz and
Faust, 2002).
The current work continues the line of inquiry initiated by Skvoretz and Faust.
In particular, it uses the triad census as a vehicle for comparisons to investigate
local structural similarities among a collection of 51 networks of different
relational contents and measured on different species.

3 Notation
Social networks consist of social relationships between pairs or sets of social
units, such as directed friendship choices between school children, victories in
antagonistic encounters between fighting deer, or advice seeking between
corporate managers. Formally, a social network for a directed dyadic relation
consists of a set of social units, referred to as actors, and a set of linkages between
pairs of actors, referred to as ties. Social networks are commonly represented by a
graph or directed graph. In a directed graph nodes represent the social units in the
network and arcs represent the directed ties between pairs of actors. A directed
graph with node set V and arc set E is denoted G(V,E), with n the number of nodes
in the graph. A social network or its associated directed graph can also be
presented in a sociomatrix with n rows and columns indexing actors (in identical
order). An entry in the sociomatrix codes the tie from the row actor to column
actor. When a tie is either present or absent, the relation is dichotomous, taking
on values of 0 or 1. In general self ties are undefined.

4 Local structure, isomorphism classes and subgraph

censuses
There has been considerable and enduring interest in local structure in networks
since the early years of social network studies. Local structure consists of
configurations and properties of small subgraphs of nodes and arcs, most notably
dyads and triads. A dyad is a subgraph of two nodes and the possible arcs between
them. In a directed graph there are three isomorphism classes of dyads: mutual
(M), asymmetric (A) ignoring the direction of the arc, and null (N). A triad is a
188 Katherine Faust

subgraph of three nodes and the arcs between these nodes. In a directed graph
each triad is isomorphic with one of sixteen isomorphism classes or triad types,
shown in Figure 1. Holland and Leinhardt (1970) first proposed the now standard
MAN notation for triads. This notation records the number of mutual (M),
asymmetric, (A) and null (N) dyads in each triad, along with further indication of
the direction of ties, when there is more than one triad with a given number of
mutual, asymmetric, and null dyads. In Figure 1 the triad types are arranged
vertically by the number of ties present.

003

012

102 021D 021U 021C

111D 111U 030T 030C

201 120D 120U 120C

210

300

Figure 1: Triad isomorphism classes with MAN labelling.

Comparing Social Networks: Size, Density, and Local Structure 189

A standard vehicle for studying local structure is a subgraph census. A

subgraph census records the frequency of each isomorphism class observed in a
directed graph or network. For example, a dyad census records the number of
mutual, asymmetric, and null dyads in a network. Similarly, a triad census records
the number of triads in each of the sixteen triad isomorphism classes.
The appeal of the triad census as a means for investigating structural patterns
in social networks lies in the fact that it succinctly summarizes a large amount of
information about a network while retaining information about theoretically
 n
important structures. In a directed graph of n nodes there are   triads. This
 3
quantity increases rapidly as the size of the graph increases, making summary into
sixteen isomorphism classes a substantial simplification (Wasserman and Faust,
1994). Nevertheless, this summary retains important information about local
features of the network and allows one to test hypotheses about the prevalence of
structural properties such as transitivity or intransitivity.

5 Macro structures and the triad census

Research employing triads and triad censuses has proved fruitful and long lived,
largely due to the important theoretical properties embodied in triads and the links
they afford between local (micro) structures and global (macro) structures (Davis,
1967, 1970; Davis and Leinhardt, 1972; Holland and Leinhardt, 1971; Johnsen,
1985, 1986, 1989, 1998; Friedkin, 1998). Whereas micro structures pertain to
small subgraphs and properties measured on them, macro structures characterize
the entire graph or network. Usefulness the triad census for testing theoretical
macro structures arises because some theoretical macro structures are contradicted
by specific configurations of triads. Support for the macro theory is evaluated by
examining empirical networks for the occurrence of triads inconsistent with the
theory. Theories that have been expressed in triadic terms include structural
balance, clusterability, ranked clusters, and transitivity.
Structural balance is one of the most straightforward theories expressed in
triadic terms. In early work in this area, Cartwright and Harary (1956) generalized
Heider’s (1946) cognitive balance notion to structural balance. As a macro
structure, a balanced signed graph has two subgroups where all ties within each
subgroup have positive signs and all ties between the two subgroups have negative
signs. Structural balance also can be examined using directed rather than signed
graphs (Johnsen, 1985, 1986, 1998), where mutual ties take the place of positive
ties and null ties take the place of negative ties. In a balanced directed graph all
mutual ties are within subgroups and all null ties are between subgroups. In a
balanced directed graph only two kinds of triads are permitted: {300 and 102}.
Other triads violate the theory.
190 Katherine Faust

The idea of structural balance was extended to the notion of clusterability by

Davis (1967) to allow more than two subgroups. In a clusterable signed graph all
positive ties are within subgroups and negative ties are between subgroups. In a
clusterable directed graph three triads are permitted: {300, 102, 003}. All other
triads violate the theory. As a substantive example, clusterability would
characterize a relation such as friendship if there were multiple cliques of mutual
friendships in a population, but no friendships between cliques.
Further generalizations of these ideas include ranked clusters (Davis, 1970;
Davis and Leinhardt, 1972) and transitivity (Holland and Leinhardt, 1971).
Ranked clusters extend the clustering model to allow directed ties between
clusters. The permitted triads for this model are {300, 102, 003, 120D, 120U,
030T, 021D 021U}. This macro model would represent a population with multiple
friendship cliques ranked in prestige or popularity, in which friendships between
cliques are directed from lower to higher status clique members. Transitivity holds
for a triple of distinct nodes if, whenever the i
j tie and the j
k ties are
present, then the i
k tie is present. The transitivity model includes one triad in
addition to those for ranked clusters: {300, 102, 003, 120D, 120U, 030T, 021D
021U, 012}. The macro structure for transitivity provides for separate systems of
ranked clusters within a population. As a substantive example, this macro model
would characterize friendships in a population where different categories of people
maintained separate systems of ranked cliques.
Since these structural theories imply different profiles of triads, we expect to
observe dissimilar triad distributions for social networks of diverse kinds of social
relations. For example, relations of dominance (Chase, 1974) generally exhibit
hierarchical patterns and would be expected to contain triads consistent with
transitivity. On the other hand, affectionate interpersonal relations, such as
friendship, would be expected to form subgroups of mutual ties, consistent with
balance or clusterability. Thus, triad distributions should be useful for studying
similarities and dissimilarities in the local structural properties of diverse social
networks. The following section describes methodology for these comparisons.

6 Data and analysis strategy

6.1 Data

The empirical base for this investigation is a collection 51 of social networks

representing a variety of types of relations and animal species. This collection
includes relations of dominance, friendship, advice seeking, grooming, fights,
social grazing, non-agonistic social acts, email communications, and confiding, to
name a few. Relations are measured among many different species, including
humans, baboons, colobus monkeys, cows, hyenas, ponies, red deer, sparrows,
Comparing Social Networks: Size, Density, and Local Structure 191

willow tits, and vervet monkeys. These networks are a sample of social networks
of the sort generally seen in the social network and ethology literatures and were
compiled from widely available sources (for example the standard data in
UCINET, journal articles, and other published sources). The Appendix lists and
describes the networks and their sources. All networks are coded to be
dichotomous and the ties are directed. The heterogeneity of relations in this
sample is an advantage for the current analysis since the goal is to examine
similarities and dissimilarities among the networks, rather than structural
tendencies in a single type of relation. The initial expectation is that networks
with similar social relations should exhibit similar structural tendencies and thus
have similar profiles of triad censuses. Data to investigate this expectation consist
of the triad census (expressed as relative frequencies) for each network in the set
of 51. This information is arrayed in a matrix with the 51 networks on the rows
and 16 triad types on the columns. The entries are the relative frequencies of each
triad type for each network.

6.2 Analysis strategy

The logic of the analysis is as follows. The first step represents similarities among
the networks based on their triad censuses and among triad types based on their
distributions across networks. Correspondence analysis is used to produce a low
dimensional representation of the similarities among networks and among types of
triads. The second step in the analysis seeks to interpret the spatial configurations
of networks and triads using local structural properties of the networks
Correspondence analysis (Blasius and Greenacre, 1998; Greenacre and Blasius,
1994; Weller & Romney, 1990) is a method for studying relationships in two-way
arrays, and results in a low dimensional representation of similarities in the data.
It is accomplished through decomposition of a matrix into its basic structure using
singular value decomposition (Clausen 1998; Weller and Romney, 1990; Digby
and Kempton, 1987). In practice, a “normalized” version of the matrix is
decomposed: entries in the original matrix are divided by the square root of the
product of the row and column marginal totals prior to singular value
decomposition. Let A be a rectangular matrix of positive entries with g rows and
1 1
− −
h columns (where g ≥ h). Two diagonal matrices R 2 2
and C have entries equal
to reciprocals of the row and column totals of A:

 1 
= diag  
1
−

 a 
2

 i+ 
R

(6.1)
192 Katherine Faust

 1 
= diag 
1
−

 a+ j 
C 2 (6.2)
 
Correspondence analysis is a singular value decomposition of the normalized
1 1
= XΛY′ where Λ
− −
matrix: R 2 AC 2
is a diagonal matrix of singular values, {λk } ,
and X and Y are the left and right singular. Graphic displays presented below use
principal coordinates, uik (for rows) and v jk (for columns) where:

a+ + (6.3)
uik = λk xik
ai +

a+ +
v jk = λk y jk (6.4)
a+ j

On each dimension these scores have weighted means equal to 0.0 and
weighted variances equal to the squared singular values:

∑ uik = ∑ u jk + j = 0
g
ai + h a (6.5)
i =1 a+ + i =1 a+ +

∑ uik2 = ∑ u 2jk + j = λ2k

g
ai + h a
a+ + i =1 a+ + (6.6)
i =1

Squared singular values express the amount of variation (chi square distance)
that is explained by each dimension in the model. The total amount of variation in
the data is referred to as inertia (Greenacre, 1984; Greenacre and Blasius, 1994;

∑λ
W
2
Clausen, 1998) and is equal to the sum of the squared singular values: k .
k =1

Table 1: Descriptive Statistics for 51 Networks.

Mean
Nodal Proportion Proportion Proportion Proportion
Size Degree Density Mutual Asymmetric Null Transitive
Mean 20.51 5.06 0.37 0.15 0.43 0.42 0.64
Std. Deviation 16.89 2.93 0.20 0.17 0.28 0.30 0.20
Minimum 4 0.55 0.02 0.00 0.01 0.00 0.21
Maximum 73 13.75 0.86 0.81 0.93 0.98 0.98
Percentiles 25 10 2.62 0.21 0.03 0.19 0.15 0.52
50 14 4.64 0.43 0.07 0.40 0.41 0.63
75 28 6.83 0.50 0.23 0.73 0.69 0.80
Comparing Social Networks: Size, Density, and Local Structure 193

7 Results
7.1 Descriptive statistics and triad censuses

Descriptive statistics for the networks are presented in Table 1 2. These results
show considerable variability among the networks in their size, density, dyadic
distributions, and tendencies toward transitivity. Networks range from 4 to 43
members, with densities ranging from 0.02 to 0.86.
The triad censuses for the 51 networks are presented in Table 2. Censuses
were calculated using an adapted version of the SAS program described by Moody
(1998). Glancing at these distributions shows some notable distinctions among the
networks. First, 003 (all null) triads are prevalent in friendships between
adolescent boys (cole1 and cole2), grazing preference between cows (cowg), social
licking between cows (cowl), and dominance between nursery school boys (kids2),
accounting for more than 50% of the triads in these distributions. Completely
mutual triads, 300, are rare across the networks, but reach almost 50% in the
network of grooming between chimpanzees (chimp3). The 030T transitive triad is
prevalent in agonistic bouts between baboons (baboon3), threats between Highland
ponies (ponies), fights between adult rhesus monkeys (rhesus1 and rhesus6),
dominance between sparrows (sparrow), and aggressive encounters between
juvenile vervet monkeys (vervet2a).

7.2 Correspondence analysis, network and triad spaces

Turning now to similarities among the 51 tirad distributions, scores for the first
four dimensions of the correspondence analysis of the network-by-triad census
array are presented in Table 3, for both networks and triad types. The first four
dimensions account for 33.8%, 24.9%, 10.7% and 8.3% of the inertia,
respectively (77.7% of the total inertia).

2
The density of a network is the proportion of possible ties that are present. The measure of
mutuality is the proportion of dyads that are mutual: M/(M+A+N). The proportion of dyads that
are asymmetric and null are computed similarly. The measure of transitivity is the number of
transitive triples divided by the number of triples that meet the condition for possibly being



transitive. Specifically, it is the proportion of i, j, k triples where the i j tie and the j k ties are


present in which the i k tie is also present.
194 Katherine Faust

Table 2: Triad distributions for 51 Networks, expressed as percentages.

Tryad type
Network 003 012 102 021D 021U 021C 111D 111U 030T 030C 201 120D 120U 120C 210 300
baboonf 12 35 02 10 09 07 05 01 14 01 00 03 00 01 01 00
baboonm1 00 00 00 00 00 00 00 00 15 00 00 00 50 00 15 20
baboonm2 00 05 00 10 10 00 00 10 40 00 00 00 20 00 05 00
baboonm3 02 18 02 02 00 08 01 02 51 00 00 03 10 00 01 00
banka 13 36 02 12 08 05 03 01 15 00 01 03 00 01 00 00
bankc 21 33 08 06 12 04 05 04 04 00 01 02 02 00 00 00
bankf 38 29 11 05 04 03 04 02 00 00 00 02 01 00 00 01
banks 07 13 01 19 01 04 00 13 09 00 02 01 14 03 10 03
cattle 12 31 02 13 07 10 02 02 17 00 00 02 01 01 00 00
chimp1 00 08 01 10 19 06 01 04 42 00 00 01 05 02 01 00
chimp2 00 00 00 04 00 00 02 08 06 00 02 12 19 08 27 11
chimp3 00 00 02 00 00 00 04 00 00 00 17 00 01 00 27 49
cole1 81 12 06 00 00 00 00 00 00 00 00 00 00 00 00 00
cole2 79 14 06 00 00 00 00 00 00 00 00 00 00 00 00 00
colobus1 00 00 00 25 00 00 00 25 00 00 00 00 25 00 25 00
colobus2 00 00 10 20 00 00 20 00 00 00 00 10 00 10 20 10
colobus3 02 08 19 00 00 02 10 13 00 00 21 02 02 02 08 08
cowg 93 02 04 00 00 00 00 00 00 00 00 00 00 00 00 00
cowl 87 12 01 00 00 00 00 00 00 00 00 00 00 00 00 00
eiesk1 49 23 15 01 02 01 02 03 00 00 01 00 01 00 01 00
eiesk2 39 23 17 02 03 01 05 04 01 00 02 01 01 00 01 01
eiesm 20 09 12 06 00 01 01 13 01 00 14 01 02 00 08 10
fifth 29 32 21 02 02 01 02 03 01 00 01 02 01 01 01 01
fourth 19 25 28 04 02 02 05 05 01 00 03 02 01 01 03 01
hyenaf 55 31 01 06 01 02 00 01 02 00 00 00 01 00 00 00
hyenam 28 34 02 15 04 05 00 04 07 00 00 00 01 00 00 00
kids1 06 11 13 05 01 04 08 14 03 00 06 03 04 03 12 07
kids2 60 28 04 02 02 02 01 01 01 00 00 00 00 00 00 00
macaca 09 13 21 03 02 02 09 09 00 00 10 02 02 01 10 07
nfponies 21 41 00 12 06 08 00 00 13 00 00 00 00 00 00 00
patasf 00 00 00 03 03 05 00 00 81 01 00 05 02 00 00 00
patasg 38 29 14 02 01 03 03 04 00 00 02 00 00 01 01 01
ponies 00 01 00 02 04 04 01 02 52 00 00 17 12 01 04 00
prison 82 12 05 00 00 00 00 00 00 00 00 00 00 00 00 00
reddeer 06 20 09 06 00 00 00 26 00 00 09 00 11 00 06 09
rhesus1 00 00 00 03 00 09 00 03 74 00 00 03 09 00 00 00
rhesus2 00 10 00 00 00 40 00 00 50 00 00 00 00 00 00 00
rhesus4 01 07 00 10 10 16 01 01 45 01 00 04 01 02 01 00
rhesus5 00 25 00 10 08 16 00 01 37 00 00 01 02 00 00 00
rhesus6 00 03 01 13 01 05 00 02 61 01 00 10 02 01 03 00
silver 00 00 03 06 01 01 02 06 19 01 00 21 09 03 17 12
sparrow 01 08 00 09 05 09 01 01 54 01 00 05 03 01 00 00
third 10 21 19 05 04 04 06 08 03 00 04 05 02 01 06 03
tits 00 00 02 09 00 04 00 05 21 00 00 14 16 04 20 05
vcbf 35 22 20 03 01 02 03 06 01 00 01 02 01 00 01 01
vcg 09 14 17 05 02 02 06 12 03 00 06 05 03 02 08 05
vcw 28 27 14 05 03 04 03 05 02 00 00 03 02 01 01 01
vervet1a 01 11 00 16 08 08 03 03 37 00 00 05 05 02 01 00
vervet1m 02 07 01 15 12 02 04 04 34 00 01 08 05 02 03 00
vervet2a 00 00 01 08 01 00 01 01 71 00 00 14 04 00 01 00
Comparing Social Networks: Size, Density, and Local Structure 195

Table 3: Correspondence analysis of triad censuses, scores for networks and triads.

Dimension
Network 1 2 3 4
baboonf 0.078 -0.360 -0.448 0.190
baboonm1 -0.980 1.141 1.395 0.307
baboonm2 -0.803 -0.187 0.378 0.473
baboonm3 -0.656 -0.573 0.110 -0.131
banka 0.133 -0.371 -0.420 0.191
bankc 0.402 -0.166 -0.422 0.283
bankf 0.756 -0.107 -0.175 0.117
banks -0.341 0.388 0.126 0.597
cattle 0.024 -0.411 -0.361 0.172
chimp1 -0.667 -0.563 -0.171 0.105
chimp2 -0.767 1.182 0.542 0.346
chimp3 -0.548 2.390 0.080 -1.675
cole1 1.358 -0.206 0.620 -0.222
cole2 1.336 -0.209 0.582 -0.205
colobus1 -0.641 0.906 0.516 1.320
colobus2 -0.406 1.013 -0.392 0.223
colobus3 -0.014 1.214 -0.841 -0.388
cowg 1.478 -0.225 0.869 -0.314
cowl 1.414 -0.267 0.772 -0.252
eiesk1 0.955 -0.026 -0.018 0.002
eiesk2 0.784 0.077 -0.213 0.037
eiesm 0.208 0.904 -0.289 -0.298
fifth 0.700 0.072 -0.388 0.124
fourth 0.517 0.274 -0.605 0.147
hyenaf 0.971 -0.275 0.242 0.036
hyenam 0.456 -0.286 -0.174 0.273
kids1 -0.102 0.782 -0.397 0.142
kids2 1.092 -0.229 0.250 -0.041
macaca 0.117 0.848 -0.602 -0.084
nfponies 0.304 -0.462 -0.293 0.153
patasf -1.081 -0.965 0.181 -0.570
patasg 0.797 0.011 -0.206 0.078
ponies -0.985 -0.427 0.354 -0.099
prison 1.369 -0.222 0.651 -0.225
reddeer80 0.768 -0.347 0.019 0.023
rhesus1 -1.071 -0.829 0.302 -0.393
rhesus2 -0.750 -1.008 -0.344 -0.644
rhesus4 -0.729 -0.721 -0.217 -0.172
rhesus5 -0.463 -0.695 -0.349 -0.060
rhesus6 -0.938 -0.675 0.094 -0.243
silver -0.799 0.642 0.341 -0.015
sparrow -0.783 -0.710 -0.058 -0.206
third 0.154 0.370 -0.550 0.188
tits -0.837 0.493 0.480 0.299
vcbf 0.725 0.112 -0.235 0.091
vcg 0.045 0.632 -0.461 0.122
vcw 0.539 0.003 -0.266 0.187
vervet1a -0.634 -0.477 -0.136 0.138
vervet1m -0.622 -0.291 -0.100 0.200
vervet2a -1.059 -0.739 0.265 -0.343
vervet2m -0.802 -0.426 -0.041 0.015
196 Katherine Faust

Triad 1 2 3 4
003 1.209 -0.179 0.452 -0.139
012 0.489 -0.207 -0.394 0.183
102 0.542 0.450 -0.657 0.053
021D -0.360 -0.066 -0.166 0.520
021U -0.243 -0.485 -0.399 0.258
021C -0.426 -0.693 -0.437 -0.295
111D -0.011 0.597 -0.738 0.066
111U -0.189 0.606 -0.220 0.623
030T -0.933 -0.760 0.126 -0.302
030C -0.681 -0.460 -0.195 -0.263
201 -0.011 1.538 -0.843 -0.940
120D -0.737 0.017 0.093 0.033
120U -0.859 0.589 0.999 0.732
120C -0.501 0.708 -0.199 0.390
210 -0.603 1.311 0.295 0.181
300 -0.543 1.955 0.256 -1.142

3.0

2.5 chimp3

2.0

1.5
chimp2 colobus3
baboonm1
Dimension 2

colobus2
1.0 colobus1 eiesm
macaca
kids1
silver vcg
tits Quintiles of Density
.5 banks third
fourth
vcbf
eiesk2
vcwfifth
patasg
eiesk1
5
0.0 baboonm2 bankc bankf kids2cole2
cole1
vervet1m hyenam hyenaf prison
cowg
cowl
baboonf
banka
cattle reddeer80 4
vervet2m
ponies vervet1a nfponies
chimp1
baboonm3
-.5 rhesus6
vervet2a rhesus5
sparrow
rhesus4
rhesus1 3
patasfrhesus2
-1.0 2

-1.5 1
-1.5 -1.0 -.5 0.0 .5 1.0 1.5 2.0

Dimension 1

Figure 2a: Correspondence analysis of triad censuses, network space.

Comparing Social Networks: Size, Density, and Local Structure 197

3.0

2.5

300
2.0
201
1.5 210 Number of Ties
Dimension 2

1.0 6
120C
120U 111U
111D
102 5
.5
120D 4
021D
0.0 012 003
3
030C 021U
-.5 021C
030T 2

-1.0 1

-1.5 0
-1.5 -1.0 -.5 0.0 .5 1.0 1.5 2.0

Dimension 1

Figure 2b: Correspondence analysis of triad censuses, triad space.

Graphic display of the scores for the first two dimensions are in Figures 2a and
2b, for network and triad spaces respectively. In each figure, points that are close
in space have similar profiles. In Figure 2a networks that are close to one another
have similar profiles of proportions in their triad censuses, whereas those that are
far apart have different triad census profiles. In this figure we see that triad
censuses for the networks of fights between yearling rhesus monkeys (rhesus2) and
fights between patas monkeys (patasf) are similar to each other and different from
networks of grooming between chimpanzees (chimp3). Triad distributions in these
three networks are different from the distributions for grazing preference between
cows (cowg), social licking between cows (cowl) and friendships in a prison
(prison). Figure 2b presents the similarity space for triads. In this figure triad
types that occur in similar proportions across the collection of networks are close
together. There is a clear diagonal pattern in the similarity space for triad types,
related to the number of ties in the triad. Triads with more ties (300) are toward
the upper left of the figure whereas triads with fewer ties (003) are toward the
lower right. Symbols for points in Figure 2b code the number of ties in the triad
(from 0 to 6) and in Figure 2a code the density of the network (in quintiles).
We can also take a joint view of the relationship between networks and triad
types, though, for ease of visualization the two configurations are presented in
separate plots in Figures 2a and 2b. Viewing the displays in Figures 2a and 2b
together, we see in the upper left of the plots that grooming between chimpanzees
(chimp3), agonistic encounters between chimpanzees (chimp2), non-agonistic
198 Katherine Faust

social acts between colobus monkeys (colobus3), and dominance between male
baboons (baboonm1) are associated with 300 triads. In the lower left of the
figures, fights between yearling rhesus monkeys (rhesus2) and fights between
female patas monkeys (patasf) are associated with the 030T, 030C, and 021C
triads. In the lower right, cows grazing and cows licking (cowg and cowl),
friendships between adolescents (cole1 and cole2) and friendships in a prison
(prison) are associated with the 003 triad.

7.3 Interpreting the correspondence analysis dimensions

What are the bases for resemblance among these triad censuses? To investigate
this question, let us look more closely at the dimensions of the correspondence
analysis solution, focusing first on the network space. Specifically, the following
correlations and scatterplots explore whether similarities among triad distributions
for these networks are largely patterned by nodal and dyadic properties.

2.0

1.5

1.0
Dimension 1

.5

0.0

-.5

-1.0

-1.5
0.0 .2 .4 .6 .8 1.0

Density

Figure 3: Scatterplot of correspondence analysis network space dimension 1 and the

network density, N = 51 networks.
Comparing Social Networks: Size, Density, and Local Structure 199

2.0

1.5

1.0
Dimension 1

.5

0.0

-.5

-1.0

-1.5
-.2 0.0 .2 .4 .6 .8 1.0

Proportion Null Dyads

Figure 4: Scatterplot of correspondence analysis network space dimension 1 and the

proportion of null dyads, N = 51 networks.

2.5

2.0

1.5

1.0
Dimension 2

.5

0.0

-.5

-1.0

-1.5
-.2 0.0 .2 .4 .6 .8 1.0

Proportion Mutual Dyads

Figure 5: Scatterplot of correspondence analysis network space dimension 2 and the

proportion of mutual dyads, N = 51 networks.
200 Katherine Faust

1.5

1.0
Dimension 3

.5

0.0

-.5

-1.0
-.2 0.0 .2 .4 .6 .8 1.0

Proportion Null Dyads

Figure 6: Scatterplot of correspondence analysis network space dimension 3 and the

proportion of null dyads, N = 51 networks Y = 0.67 − 4.9 X + 5.2 X 2 .

1.5

1.0

.5
Dimension 4

0.0

-.5

-1.0

-1.5

-2.0
-.2 0.0 .2 .4 .6 .8 1.0

Proportion Asymmetric Dyads

Figure 7: Scatterplot of correspondence analysis network space dimension 4 and the

proportion of asymmetric dyads, N = 51 networks Y = −0.59 + 3.82 X − 4.03 X 2 .
Comparing Social Networks: Size, Density, and Local Structure 201

As suggested by the patterning of triad space, the first dimension of the

correspondence analysis network space is associated with the density of the
network – the correlation between coordinates on the first dimension and network
density is r = -0.882. However, the correlation is even stronger with the
proportion of dyads that are null; r = 0.994. The second dimension is associated
with the level of mutuality in the network. This dimension correlates 0.941 with
the proportion of dyads that are mutual. The third dimension is quadratic function
relating to the proportion of dyads that are null; the best fitting quadratic equation
relating scores on dimension three to the proportion of null dyads is
Y = 0.67 − 4.9 X + 5.3 X 2 with r 2 = 0.792. The fourth dimension is related in a
quadratic form to the asymmetry in the network, albeit weakly
( Y = −0.59 + 3.82 X − 4.03 X 2 , r 2 = 0.410). Scatterplots of these relationships are
shown in Figures 3 through 7.
Turning to the correspondence analysis space for triads, the first dimension
correlates 0.962 with the proportion of null dyads in the triad; the second
dimension correlates 0.956 with the proportion of mutual dyads; the third
dimension is related in a quadratic function to the proportion of null dyads
( Y = 0.67 + 0.21 X − 3.25 X 2 , r 2 = 0.592); and the fourth dimension is a quadratic
function of the proportion of asymmetric dyads ( Y = −0.53 + 3.28 X − 3.06 X 2 , r 2 =
0.519).
These results demonstrate that similarities among the triad distributions for
these social networks are largely explained by very local properties, the nodal and
dyadic distributions at most. In other words, modeling resemblance among these
triad distributions does not require triadic level properties.
Canonical correlation analysis (Tatsuoka, 1971) provides a way to summarize
the overall degree of linear relationship between the dimensions of the
correspondence analysis and dyadic properties of the networks. The canonical
correlation between two sets of variables is the maximum correlation between
linear combinations of the variables in each set. For this analysis the first set
consists of the four dyadic measures (the proportion null dyads, the proportion
mutual dyads, the best fitting quadratic function of the proportion of asymmetric
dyads, and the best fitting quadratic function of the proportion of null dyads) and
the second set consists of the scores from the first four dimensions of
correspondence analysis. Canonical correlation analysis is repeated separately for
the network and the triad spaces.
202 Katherine Faust

Table 4: Canonical correlation loadings.

4a Loadings from canonical correlation analysis, network space

Canonical Variate

Dyadic Measures
1 2 3 4
Proportion null -0.996 0.060 0.065 0.005
Proportion mutual 0.357 -0.929 -0.095 0.010
Quadratic function of proportion null -0.042 0.127 -0.978 0.157
Quadratic function of proportion asymmetric 0.257 -0.306 0.503 -0.766

Correspondence Analysis Dimensions

Dimension 1 -0.998 -0.048 0.041 0.008

Dimension 2 0.044 -0.993 -0.054 -0.099
Dimension 3 -0.045 0.072 -0.973 -0.216
Dimension 4 -0.003 0.085 0.223 -0.971

4b Loadings from canonical correlation analysis, triad space.

Canonical Variate

Dyadic Measures
1 2 3 4
Proportion null 0.947 -0.042 0.316 0.027
Proportion mutual -0.426 -0.900 -0.009 0.086
Quadratic function of proportion null -0.076 -0.159 -0.970 -0.165
Quadratic function of proportion asymmetric -0.298 0.513 0.049 0.804

Correspondence Analysis Dimensions

Dimension 1 0.930 -0.235 0.281 0.040

Dimension 2 -0.308 -0.921 0.066 0.228
Dimension 3 -0.044 -0.059 -0.976 0.206
Dimension 4 0.075 0.512 -0.077 0.852
Comparing Social Networks: Size, Density, and Local Structure 203

For the network space, there are four significant canonical correlations (all
with p < .0001), equal to 1.0, 0.999, 0.951, and 0.754, respectively. Together the
first four canonical variates account for 86.7% of the variance in the first four
dimensions of the correspondence analysis of the network space and 90.4% of the
variance in the four dyadic network measures. Canonical loadings for the
variables in the two sets are reported in Table 4a (The canonical loading is the
correlation of the variable with the linear combination.). For the triad space there
are also four significant canonical correlations (the first two with p < .0001, and
the second two with p < .001), having values of 0.997, 0.979, 0.825, and 0.714
respectively.
In light of the fact that the first four dimensions of the correspondence analysis
network space explain 77.7% of the inertia in the triad census distributions for
these 51 networks, and dyadic level network properties account for 86.7% of this
space, one might argue that around 67% (0.867 x 0.777) of the similarity among
these triad distributions is accounted for by dyadic level network properties.
Similarly, around 62% (0.802 x 0.777) of the similarity among triad types is
accounted for by dyadic features of these triads.
In summary, these results demonstrate that, in aggregate, similarities in the
triad censuses for a wide range of different social networks can largely be
accounted for by dyadic level features of the networks. A reasonable estimate is
that around two thirds of the variance among the networks’ triad distributions is
accounted for by no more than nodal and dyadic features. How are we to interpret
these results, and what are their implications for our understanding of social
network structure? The following section addresses these questions.

8 Discussion and implications

Two related questions are raised by these results. First, what gives rise to the
finding that similarity among triadic distributions is largely accounted for by lower
order (nodal or dyadic) properties? Second, what are the implications of these
result for comparisons of triad distributions in social networks? Clues to the
answer these questions are found throughout the literature on triads and are
provided more directly in literature on effects of network size and density on graph
level indices. Four points are pertinent: effects of size and density on graph level
measures; comparative use of triad censuses; absence of social structure in many
social networks; and distinction between triad distributions and configurations of
local structural properties.
204 Katherine Faust

8.1 Size and density

Network size has been recognized an issue in studies of triads since their earliest
empirical use. In his 1970 paper on clustering and hierarchy, Davis (1970)
analyzed 742 sociomatrices and presented results of triad distributions and
statistics for triadic cycles separately for networks in five ranges of sizes.
However, he does not reveal his rationale for doing so. Davis does note that for
the 210 triad, which is not permitted under the ranked clusters model, “results are
catastrophic in the larger groups” (1970: 845-846). Davis attributes the surplus of
210 triads in large networks to the frequency of fixed choice data collection
designs that force otherwise mutual ties to be asymmetric. Johnsen (1985, 1986,
1998) pays considerably more attention to network size and its impact on possible
micro and macro structural models. Picking up on Davis’ (1970) results for
networks of different sizes, Johnsen (1985) observes that for network sizes 8 to 13,
data perfectly fit the ranked clusters of mutual cliques model. Friedkin (1998:
143) addresses the point more extensively in his discussion of macro models for
large networks. With respect to the 003 triad (all null dyads) he notes that “…
especially in large groups, the possibility of three N linked actors should not be
forbidden” (Friedkin, 1998: 143, emphasis in the original). 3 So, authors have
recognized the relationship between network size and triadic macro models, but
have not addressed the issue directly.
Importantly, network density and not network size, per se, is crucial for
understanding the distribution of triadic configurations. The density, d, of a
network is the number of arcs in the network divided by the possible number of
arcs. If the average number of arcs from each node is fixed, the total number of
arcs increases as a linear function of network size but the possible number of arcs
increases with the square of network size. Thus, if the average number of arcs
from each node is fixed, a reasonable assumption if actors can only maintain a
limited number of ties, then network density must decrease with network size.
Why is this important?
Given the density of the network, the range of possible triad distributions
is heavily constrained. Some triad types have extremely low probabilities, simply
because of the density of the network. Formulae for the probability of each of the
16 triad types, given network density, are presented in Skvoretz et al. (2004). To
illustrate, the probability of a 300 triad (all mutual ties) is equal to d 6 . In a
network with 10 actors and 3 ties per actor, the density is equal to .33, and the
probability of a 300 triad is 0.0014. If the size of the network is increased to 100,

3
Network size is also mentioned in discussions of its effect on statistics for testing structural
hypotheses, but largely in the context of its effect on the standard errors. It is easier to detect
departures from expected frequencies in larger groups than in smaller groups (Leinhardt 1972,
Hallinan 1974a, Holland and Leinhardt 1975).
Comparing Social Networks: Size, Density, and Local Structure 205

but the mean number of ties per actor remains 3, the density decreases to 0.03, and
the probability of a 300 triad is 0.0000000008.

Table 5: Triad census of the routing data.

Triad Count Percent

003 322,769,974,374,083 99.99252370409
012 23,955,959,979 0.00742143658
102 175,605,448 0.00005440169
021D 882,596 0.00000027342
021U 109,179 0.00000003382
021C 444,490 0.00000013770
111D 4917 0.00000000152
111U 15,508 0.00000000480
030T 17,107 0.00000000530
030C 111 0.00000000003
201 1002 0.00000000031
120D 899 0.00000000028
120U 1120 0.00000000035
120C 96 0.00000000003
210 121 0.00000000004
300 69 0.00000000002
Total 322,794,107,416,725
Adapted from Vladimir Batagelj and Andrej Mrvar (2001): A
subquadratic triad census algorithm for large sparse networks with
small maximum degree. Social Networks, 23, 237–243.

To put the effect of size and density in sharper perspective for comparing
empirical networks, consider the network of routing linkages on the internet
analyzed by Batagelj and Mrvar (2001). This network has 124,651 nodes and
207,214 edges. The mean number of ties per node is 3.3 and the density is
0.000027. The triad census for the routing network, expressed as a percentage
distribution, is in Table 5. There are 322,794,107,416,725 triads in this network,
of which 322,769,974,374,083, or 99.99%, are completely null (type 003). This is
slightly less than the percent expected, given the density of the network
(99.99999999%). In this network 69 triads, 0.00000000002%, are entirely mutual
(type 300), which is more than the expected percent of 8.6x10-63 %. For
comparative purposes, consider what happens when the routing triad census is
included in a correspondence analysis with the set of 51 networks analyzed above.
The extremely low density of this network severely constrains its possible
locations in the space of similarities among networks. Figure 8 presents the first
two dimensions of the correspondence analysis of 52 networks (the first four
dimensions account for 34.1%, 24.4%, 11.3% and 8.2% of the inertia). As
expected, Batagelj and Mrvar’s network is in the far lower right corner with the
other low density networks. Given its density, it is virtually impossible for it to
occupy any other region of the space.
206 Katherine Faust

3.0

2.5

2.0

1.5
Dimension 2

1.0

.5

0.0
Batagelj & Mrvar

-.5

-1.0

-1.5
-2.0 -1.5 -1.0 -.5 0.0 .5 1.0 1.5 2.0 2.5

Dimension 1

Figure 8: Correspondence analysis dimensions 1 and 2 for 52 triad censuses including

Batagelj and Mrvar’s internet routing network, network space.

In contrast, consider the network of non-agonistic social encounters in a group

of 5 colobus monkeys (labeled colobus2 in Figure 2). This network is toward the
upper left corner of both correspondence analysis figures for the first two
dimensions (Figures 2 and 8). In the colobus monkey network the mean number of
ties per monkey is 2.4, less than the mean for the routing network, and its density
is 0.6. Of its 10 triads, none is completely null (003), though based on the density
of the network 0.4% are expected to be so. This network has one triad that is
completely mutual (type 300); 10% compared to an expectation of 4.7%. Given
the network’s density it is almost impossible for it to occupy the lower right corner
of the correspondence analysis space and be associated with the 003 triad.
Clearly these two illustrative networks are extreme in terms of size and
density. However, regardless of other features they might share, because of their
vastly different densities their triad distributions can not be similar.
The profound effect of density on graph level indices (GLIs) more generally is
clearly demonstrated in Anderson Butts and Carley’s (1999) analyses. They
conclude: “As we have seen, both size and density have powerful – and complex –
interactions with other GLIs. These interactions stem from fundamental
constraints on the space of graphs, constraints that severely limit the combinations
of GLI values which can be realized on a given graph” (1999: 257). This is
undoubtedly demonstrated in the triad census results presented here.
Comparing Social Networks: Size, Density, and Local Structure 207

8.2 Comparing triad distributions

The fact that, in aggregate, similarities among observed triad censuses are largely
patterned by nodal and dyadic properties does not preclude higher order (i.e.
triadic) structure in individual networks. In fact, 80% of the networks in the
collection of 51, exhibit tendencies toward transitivity more than three standard
deviations greater than expected, given their dyadic distributions. (Methodology
for these tests is described in Holland and Leinhardt, 1976). Rather, the result
suggests that the space of possible triad distributions is so constrained by network
density and dyadic properties that comparing these distributions among networks
that differ in these tendencies is uninformative. Little higher-order variability
among the triad censuses remains to be explained.

8.3 Is there social structure in social networks?

As a third point addressing the questions posed above, it is important not to lose
sight of the possibility that many social networks have little or no social structure.
Holland and Leinhardt (1979) define social structure in network terms as the
presence of higher order properties that are not adequately explained by nodal or
dyadic tendencies. In other words, triads are the lowest level at which there is
interesting social structure. With respect to the detection of triadic tendencies in
the bank of 384 sociomatrices, Holland and Leinhardt (1979) found that when
observed triad distributions were compared to those expected under different
conditional distributions, a higher level of conditional distribution allowed fewer
sociomatrices to exhibit significant tendencies away from intransitivity. They
concluded that “… what was previously thought to be structure was spurious”
(1979: 77) and that about 40% of the networks had no social structure.
Indeed, there may be many social networks in which there is little or no social
structure in Holland and Leinhardt’s sense. This finding is reinforced more
recently in Butts’s (2001) examination of the complexity of social networks. If
networks exhibit patterns such as structural balance or classes of equivalent actors,
then they should relatively “simple” compared to random graphs. They should not
be algorithmically complex. However, evidence from networks collected by
various methods (observations, self reports, and reports of others’ ties) fails to
support this expectation, once network density is taken into account. The
observation that graph structure is largely explained by local properties leads Butts
to pose the conditional uniform graph distribution hypothesis: “…the aggregate
distribution of empirically realized social networks is isomorphic with a uniform
distribution over the space of all graphs, conditional on graph size and density”
(Butts, 2001: 67). Moreover, if this hypothesis is correct “… much of what will be
found—or not found—in any given graph will be driven heavily by density and
graph size.” (Butts, 2001: 69).
208 Katherine Faust

Reiterating these observations, once we have taken lower order graph

properties into account, there may be very little higher order structure in many
social networks.

8.4 Configurations rather than triad distributions

As a final piece of the puzzle, reconsider the theoretically important structural

information contained in triads. The discussion in section 5 above highlighted the
linkage between macro structural theories and the particular triads that are
consistent or inconsistent with them. These macro theories imply different triad
census distributions A complementary approach to network structures links local
properties and macro theories by focusing on configurations of relational ties
between collections of actors rather than on the presence or absence of specific
triads types. For example, transitivity is expressed for an ordered triple of distinct
actors (i, j, k) such that if the i
j tie is present and the j
k tie is present, then
the i
k tie is present. Transitivity is violated for an ordered triple of distinct
actors (i, j, k) if the i
j tie is present and the j
k tie is present, but the i
k is
absent. A network perfectly characterized by transitivity has transitive triple
configurations but not intransitive triple configurations (other triple configurations
are moot with respect to transitivity). The problem with using a triad census to
examine transitivity is that individual triads types contain multiple configurations
of ordered triples, some of which might be transitive and others intransitive. To
illustrate, consider the 210 triad which is forbidden for the balance, clustering,
ranked clusters, and transitivity macro models. This triad contains three ordered
triples that are transitive and one that is intransitive, so on the whole it is shows a
greater tendency toward rather than away from transitivity.
A slightly different implication of focusing on configurations of ties rather
than triad censuses is presented in Davis’s (1979) review of the Davis, Holland,
and Leinhardt work on triadic structure in networks. The persistent presence of
the 210 triad led their research toward the study of transitivity and a focus on i,j,k
triplets and other configurations rather than on triad distributions. As Davis notes,
the focus on transitivity and sentiment patterns in triplets represented, for him, a
retreat from a sociological perspective and “drifting back toward psychology”
(Davis, 1979: 58) and “a slide from global structure to microanalysis” (1979: 60).
The results presented in this paper need not herald a turn away from sociology
toward psychology. Rather, they remind us that system size and density, by
mathematical necessity, constrain possible patterns of social organization
(Mayhew and Levinger, 1976; Butts, 2001). Moreover, if lower-order properties
(density and dyad distributions) account for a substantial portion of the systematic
patterning of similarities among networks, parsimony and good scientific practice
require that we not exert effort “explaining” the remainder.
Comparing Social Networks: Size, Density, and Local Structure 209

Acknowledgements
I would like to thank Vladimir Batagelj, Carter Butts, Lin Freeman, Kim Romney,
John Skvoretz and participants in the UCI Social Network Research Group for
comments on this research. I am grateful to Anuška Ferligoj, Andrej Blejec, and
Janez Stare for the opportunity to present a version of this paper at the 2005
Applied Statistics Conference, Bled, Slovenia, September 2005.

References
[1] Anderson, B.S., Butts, C., and Carley, K. (1999): The interaction of size and
density with graph-level indices. Social Networks, 21, 239-267.
[2] Batagelj, V. and Mrvar, A. (2001): A subquadratic triad census algorithm for
large sparse networks with small maximum degree. Social Networks, 23, 237–
243
[3] Bearman, P.S., Jones, J., and Udry, J.R. (1997): National Longitudinal Study
of Adolescent Health: Research Design. Carolina Population Center.
Unpublished manuscript.
[4] Bernard, H.R., Killworth, P., Kronenfeld, D., and Sailer, L. (1984): The
problem of informant accuracy: The validity of retrospective data. Annual
Review of Anthropology, 13, 495-517.
[5] Blasius, J. and Greenacre, M. (1998): Visualization of Categorical Data. New
York: Academic Press.
[6] Butts, C.T. (2001): The complexity of social networks: theoretical and
empirical findings. Social Networks, 23, 31-71.
[7] Cartwright, D. and Harary, F. (1956): Structural balance: A generalization of
Heider’s theory. Psychological Review, 63, 277-293.
[8] Chase, I.D. (1974): Models of hierarchy formation in animal societies.
Behavioral Science, 19, 374-382.
[9] Claussen, S. (1998): Applied Correspondence Analysis. Thousand Oaks:
Sage.
[10] Davis, J.A. (1967): Clustering and structural balance in graphs. Human
Relations, 20, 181-187.
[11] Davis, J.A. (1970): Clustering and hierarchy in interpersonal relations:
Testing two graph theoretical models on 742 sociomatrices. American
Sociological Review, 35, 843-851.
[12] Davis, J.A. (1979): The Davis/Holland/Leinhardt studies: An overview. In
P.W. Holland and S. Leinhardt (Eds): Perspectives on Social Network
Research, 51-62. New York: Academic Press.
210 Katherine Faust

[13] Davis, J.A. and Leinhardt, S. (1972): The structure of positive interpersonal
relations in small groups. In J. Berger, M. Zeldith Jr. and B. Anderson (Eds):
Sociological Theories in Progress, 2, 218-251. Boston: Houghton Mifflin.
[14] Digby, P.G.N. and Kempton, R.A. (1987): Multivariate Analysis of
Ecological Communities. London: Chapman and Hall.
[15] Entwisle, B., Faust, K. Rindfuss, R., and Kaneda, T. (n.d.): Networks and
contexts: Variation in the structure of social ties.
[16] Faust, K. and Skvoretz, J. (2002): Comparing networks across space and time,
size and species. Sociological Methodology, 32, 267-299.
[17] Freeman, L.C. (1992): The sociological concept of `group': An empirical test
of two models. American Journal of Sociology, 98, 152-166.
[18] Friedkin, N. (1998): A Structural Theory of Social Influence. New York:
Cambridge University Press.
[19] Gifi, A. (1990): Nonlinear Multivariate Analysis. New York: John Wiley.
[20] Greenacre, M. (1984): Theory and Applications of Correspondence Analysis.
New York: Academic Press.
[21] Greenacre, M. and Blasius, J. (Eds) (1994): Correspondence Analysis in the
Social Sciences: Recent Developments and Applications. New York:
Academic Press.
[22] Hallinan, M.T. (1974a): The Structure of Positive Sentiment. New York:
Elsevier.
[23] Hallinan, M.T. (1974b): A structural model of sentiment relations. American
Journal of Sociology, 80, 364-378.
[24] Heider, F. (1946): Attitudes and cognitive organization. Journal of
Psychology, 21, 107-112.
[25] Holland, P.W. and Leinhardt, S. (1970): A method for detecting structure in
sociometric data. American Journal of Sociology, 76, 492-513.
[26] Holland, P.W. and Leinhardt, S. (1971): Transitivity in structural models of
small groups. Comparative Group Studies, 2, 107-124.
[27] Holland, P.W. and Leinhardt, S. (1972): Some evidence on the transitivity of
positive interpersonal sentiment. American Journal of Sociology, 77, 1205-
1209.
[28] Holland, P.W. and Leinhardt, S. (1976): Local structure in social networks.
Sociological Methodology, 7, 1-45.
[29] Holland, P.W. and Leinhardt, S. (1979): Structural Sociometry. In P.W.
Holland and S. Leinhardt (Eds): Perspectives on Social Network Research,
63-83. New York: Academic Press.
[30] Johnsen, E.C. (1985): Network macrostructure models for the Davis-
Leinhardt set of empirical sociomatrices. Social Networks, 7, 203-224.
Comparing Social Networks: Size, Density, and Local Structure 211

[31] Johnsen, E.C. (1986): Structure and process: Agreement models for
friendship formation. Social Networks, 8, 257-306.
[32] Johnsen, E.C. (1989): The micro-macro connection: Exact structure and
process. In F. Roberts (Ed): Applications of Combinatorics and Graph Theory
to the Biological and Social Sciences, 169-201. New York: Springer-Verlag.
[33] Johnsen, E.C. (1998): Structures and processes of solidarity: An initial
formalization. In P. Doreian and T. Fararo (Eds): The Problem of Solidarity:
Theories and Models, 263-302. Amsterdam: Gordon and Breach.
[34] Johnson, J.C., Boster, J.S., and Palinkas, L.A. (2003): Social roles and the
evolution of networks in extreme and isolated environments. Journal of
Mathematical Sociology, 27, 89-121,
[35] Laumann, E.O. and Pappi, P. (1976): Networks of Collective Action: A
Perspective on Community Influence Systems. New York: Academic Press.
[36] Leinhardt, S. (1972): Developmental change in the sentiment structure of
children’s groups. American Sociological Review, 37, 202-212.
[37] Mayhew, B.H. and Levinger, R.L. (1976): Size and density of interaction in
human aggregates. American Journal of Sociology, 82, 86-110
[38] Moody, J. (1998): Matrix methods for calculating the triad census. Social
Networks, 20, 291-299.
[39] Pattison, P. and Wasserman, S. (1999): Logit models and logistic regressions
for social networks: II. Multivariate relations. British Journal of
Mathematical and Statistical Psychology, 52, 169-193.
[40] Rindfuss, R.R., Jampaklay, A., Entwisle, B., Sawangdee, Y., Faust, K., and
Prasartkul, P. (2004): The Collection and Analysis of Social Network Data in
Nang Rong, Thailand. In M. Morris (Ed): Network Epidemiology: A
Handbook For Survey Design and Data Collection, 175-200. Oxford: Oxford
University Press.
[41] Skvoretz, J., Fararo, T.J., and Agneessens, F. (2004): Advances in biased net
theory: Definitions, derivations, and estimations. Social Networks, 26, 113-
139.
[42] Skvoretz, J. and Faust, K. (2002): Relations, species, and network structure.
Journal of Social Structure, 3.
http://zeeb.library.cmu.edu:7850/JoSS/skvoretz/index.html
[43] Snijders T.A.B. and Baerveldt, C. (2003): A Multilevel Network Study of the
Effects of Delinquent Behavior on Friendship Evolution. Journal of
Mathematical Sociology, 27, 123-151.
[44] Tatsuoka, M.M. (1971): Multivariate Analysis: Techniques for Educational
and Psychological Research. New York: John Wiley and Sons.
[45] Wasserman, S.S. (1977): Random directed graph distributions and the triad
census in social networks. Journal of Mathematical Sociology, 5, 61-86.
212 Katherine Faust

[46] Wasserman, S.S. and Faust, K. (1994): Social Network Analysis: Methods and
Applications. New York: Cambridge University Press.
[47] Wasserman, S.S. and Pattison, P. (1996): Logit models and logistic
regressions for social networks: I. An introduction to markov graphs and p*.
Psychometrika, 61, 401-425.
[48] Weller, Susan C. and A. Kimball Romney (1990): Metric Scaling:
Correspondence Analysis. Newbury Park, CA: Sage.

References for social network data

[1] Anderson, C.J., Wasserman, S., and Crouch, B. (1999): A p* primer: Logit
models for social networks. Social Networks, 21, 37-66.
[2] Appleby, M. (1983): The probability of linearity in hierarchies. Animal
Behaviour, 31, 600-608.
[3] Clutton-Brock, T.H., Greenwood, P.J., and Powell, R.P. (1976): Ranks and
relationships in highland ponies and highland cows. Zeitschrift fur
Tierpsychologie, 41, 202-216.
[4] Coleman, J. (1964): Introduction to Mathematical Sociology. New York:
Free Press.
[5] Dunbar, R.I.M. and Dunbar, E.P. (1976): Contrasts in social structure among
black-and-white colobus monkey groups. Animal Behaviour, 24, 84-92.
[6] Frank, L.G. (1986): Social organization of the spotted hyaena crocuta crocuta.
II: Dominance and reproduction. Animal Behaviour, 34, 1510-1527.
[7] Freeman, L.C., Freeman, S.C., and Romney, A.K. (1992): The implications of
social structure for dominance hierarchies in red deer. Animal Behaviour, 44,
239-245.
[8] Freeman, S.C. and Freeman, L.C. (1979): The Networkers Network: A Study
of the Impact of a New Communications Medium on Sociometric Structure.
Social Science Research Reports, 46. Irvine, CA: University of California.
[9] Hall, K.R.L. and DeVore, I. (1965): Baboon social behavior. In I. DeVore
(Ed): Primate Social Behavior: Field Studies of Monkeys and Apes, 53-110.
New York: Holt, Rinehart, and Winston.
[10] Hausfater, G. (1975): Dominance and Reproduction in Baboons (Papio
cynocephalus), a Quantitative Analysis. New York: S. Karger.
[11] Horrocks, J. and Hunte, W. (1983): Maternal rank and offspring rank in
vervet monkeys: An appraisal of the mechanisms of rank acquisition. Animal
Behaviour, 31, 772-782.
[12] Kaplan, J.R. and Zucker, E. (1980): Social organization in a group of free-
ranging patas monkeys. Folia Primatologica, 34, 196-213.
Comparing Social Networks: Size, Density, and Local Structure 213

[13] Kikkawa, J. (1980): Weight change in relation to social hierarchy in captive

flocks of silvereyes, zosterops lateralis. Behaviour, 74, 92-100.
[14] Lahti, K., Koivula, K., and Orell, M. (1994): Is the social hierarchy always
linear in tits. Journal of Avian Biology, 25, 347-348.
[15] MacRae, D. (1960): Direct factor analysis of sociometric data. Sociometry,
23, 360-371.
[16] McGrew, W.C. (1972): An Ethological Study of Children's Behavior. New
York: Academic Press.
[17] Nishida, T. and Kasuhiko H. (1996): Coalition strategies among male
chimpanzees of the Mahale Mountains, Tanzania. In W.C. McGrew, L.F.
Marchant, and T. Nishida (Eds): Great Ape Societies, 114-134. New York:
Cambridge University Press.
[18] Parker, J.G. and Asher, S. (1993): Friendship and friendship quality in middle
childhood: Links with per group acceptance and feelings of loneliness and
social dissatisfaction. Developmental Psychology, 29, 611-621.
[19] Pattison, P., Wasserman, S., Robins, G., and Kanfer, A.M. (2000): Statistical
evaluation of algebraic constraints for social networks. Journal of
Mathematical Psychology, 44, 536-568.
[20] Reinhardt, V. and Reinhardt, A. (1981): Cohesive relationships in a cattle
herd (bos indicus). Behaviour, 76, 121-151.
[21] Roberts, J.M.Jr. (1994): Fit of some models to red deer dominance data.
Journal of Quantitative Anthropology, 4, 249-258.
[22] Roberts, John M.Jr. and Browning, B.A. (1998): Proximity and threats in
highland ponies. Social Networks, 20, 227-238.
[23] Robins, G., Pattison, P., and Wasserman, S. (1999): Logit models and logistic
regressions for social networks: III. Valued relations. Psychometrika, 64, 371-
394.
[24] Sade, D.S. (1967): Determinants of dominance in a group of free-ranging
rhesus monkeys. In S.A. Altmann (Ed): Social Communication among
Primates, 99-114. Chicago: University of Chicago Press.
[25] Sade, D.S. (1989): Sociometrics of macaca mulatta III: N-path centrality in
grooming networks. Social Networks, 11, 273-292.
[26] Schein, M.W. and Fohrman, M.H. (1955): Social dominance relationships in a
herd of dairy cattle. The British Journal of Animal Behavior, 3, 45-55.
[27] Strayer, F.F. and Janet Strayer (1976): An ethological analysis of social
agonism and dominance relations among preschool children. Child
Development, 47, 980-989.
[28] Tufto, J., Solberg, E.J., and Ringgsby, T. (1998): Statistical models of
transitive and intransitive dominance structures. Animal Behaviour, 55, 1489-
1498.
214 Katherine Faust

[29] Tyler, S.J. (1972): The behaviour and social organization of the new forest
pony. Animal Behavior Monographs, Volume 5, Part 2.
[30] Vickers, M. and Chan, S. (1981): Representing Classroom Social Structure.
Melbourne: Victoria Institute of Secondary Education.
[31] Wasserman, S.S. and Faust, K. (1994): Social Network Analysis: Methods and
Applications. New York: Cambridge University Press.
[32] Watt, D.J. (1986): Relationship of plumage variability, size and sex to social
dominance. Animal Behaviour, 34, 16-27.

Appendix: List of social networks

The following list gives the label for each network along with a reference for the
source of the data.

• baboonf: dominance interactions between female and one adult male

baboons (Figure 3-8, page 69, Hall and DeVore, 1965)
• baboonm1 and baboonm2: dominance between male baboons (Table 3-2,
page 60, Hall and DeVore, 1965)
• baboonm3: outcomes of agonistic bouts between male baboons (Table XI,
page 39, Hausfater, 1975)
• banka: advice in a bank office (Table 5, page 558, Pattison et al., 2000)
• bankc: confiding in a bank office (Table 5, page 558, Pattison et al., 2000)
• bankf: close friends in a bank office (Table 5, page 558, Pattison et al.,
2000)
• banks: satisfying interaction in a bank office (Table 5, page 558, Pattison
et al., 2000)
• cattle: contests between dairy cattle (Figure 1, page 49, Schein and
Fohrman, 1955)
• chimp1, chimp2, and chimp3: Three relations between Chimpanzees. pant-
grunt calls (Table 9.3, page 119), initiation of dyadic agonistic
confrontations (Table 9.4, page 119), and initiation of grooming (Table
9.14a, page 126). Data from Nishida and Hosaka (1996).
• cole1 and cole2: friendship at two time points between adolescent boys
(Table 14.5 pages 450-451, Coleman, 1964)
• colobus1, colobus2, colobus3: non-agonistic social acts between colobus
monkeys in a small group (Table I, page 86, Dunbar and Dunbar, 1976)
• cowl, cowg: Two relations between cows, bos indicus. social licking
(Figure 7, page 130) and social grazing (Figure 4, page 126). Data from
Reinhardt and Reinhardt (1981).
Comparing Social Networks: Size, Density, and Local Structure 215

• eiesk1: EIES data, rating of acquaintanceship between social science

researchers at time 1. Recoded 3,4=1, <3=0. (Freeman and Freeman, 1979;
Table B.8, page 745, Wasserman and Faust, 1994)
• eiesk2: EIES data, rating of acquaintanceship between social science
researchers at time 2. Recoded 3,4=1, <3=0. (Freeman and Freeman, 1979;
Table B.9, page 746, Wasserman and Faust, 1994)
• eiesm: EIES data frequency of message sending between social science
researchers, Recoded “1” if any message was sent. (Freeman and Freeman,
1979; Table B.10, page 747,Wasserman and Faust, 1994)
• fifth: friendships between fifth graders (Table 3, page 44, Anderson et al.,
1999, data from Parker and Asher, 1993)
• fourth: friendships between fourth graders (Table 3, page 44, Anderson et
al., 1999, data from Parker and Asher, 1993)
• hyenaf, hyenam: Dominance, among females and among males Hyaena,
crocuta crocuta. Dominance among adult females (Table I, page 1513) and
dominance among males (Table V, page 1519). Data from Frank (1986).
• kids1: initiated agonism between children (Figure 2, page 986, Strayer and
Strayer, 1976)
• kids2: dominance among boys in a nursery school (Figure 5.5, page 125,
McGrew, 1972)
• macaca: Grooming between Macaca Mulatta. (Table 1, page 274). Data
from Sade (1989).
• nfponies: threats between ponies (Table XIV, page 122, Tyler, 1972)
• patasf and patasg: Two relations between Patas monkeys: fighting (Table
III, page 202) and grooming (Table V, page 205). Data from Kaplan and
Zucker (1980).
• ponies: Threats between Highland ponies. (Table 2, page 3). Data from
Roberts and Browning (1998), originally in Clutton-Brock, Greenwood,
and Powell (1976).
• prison: closest friendships in a prison (Table 1, page 363, MacRae, 1960)
• reddeer80: Winner and loser in encounters between Red deer stags, Cervus
elaphus L. (Figure 1a, page 601) Data from Appleby (1983). Data also in
Freeman, Freeman, and Romney (1992) and Roberts (1994).
• rhesus1: fights between adult female rhesus monkeys (Table 1, page 105,
Sade, 1967)
• rhesus2: fights between yearling rhesus monkeys (Table 2, page 107, Sade,
1967)
• rhesus4: fights between adult rhesus monkeys (Table 4, page 108, Sade,
1967)
• rhesus5: fights between adult rhesus monkeys (Table 7, page 110, Sade,
1967)
216 Katherine Faust

• rhesus6: fights between adult rhesus monkeys (Table 8, page 111, Sade,
1967)
• silver: Victories in encounters between Silvereyes, zosterops lateralis.
(Table 1, page 94). Data from Kikkawa (1980).
• sparrow: Dominance between Sparrows, zonotrichia querula, including
both attacks and avoidances (Figure 2, page 19). Data from Watt (1986).
• third: friendship between third graders (Table 1, page 42, Anderson et al.,
1999, data from Parker and Asher, 1993)
• tits: dominance between Willow tits, parus montanus. (Table 1, page
1492). Data from Tufto, Solberg, and Ringgsby (1998). Data originally in
Lahti, Koivula, and Orell (1994).
• vcbf: best friends between seventh graders (Table 3, page 385, Robins,
Pattison, and Wasserman, 1999, from Vickers and Chan, 1981)
• vcg: get on with between seventh graders (Table 11, page 422, Wasserman
and Pattison, 1996, from Vickers and Chan, 1981)
• vcw: work with between seventh graders (Table 12, page 423,
Wasserman and Pattison, 1996, from Vickers and Chan, 1981)
• vervet1a and vervet2a: Vervet monkeys (Cercopithecus aethiops sabaeus),
juveniles from two troops (1 and 2) dyadic aggressive / submissive
interactions, both mothers absent (Table II, page 776), Data from Horrocks
and Hunte (1983).
• vervet1m and vervet2m: Vervet monkeys (Cercopithecus aethiops sabaeus),
juveniles from two troops (1 and 2) dyadic aggressive/ submissive
interactions, both mothers present (Table I, page 775), Data from Horrocks
and Hunte (1983).