10. Network Models
Network theory is a whole branch of science, but it’s relatively new in terms
of the last 20 or 30 years. We haven’t had a chance to take all that theory
out of the universities and apply it to ask: “What kinds of networks should
we build, and for what purposes?”
—Anne-Marie Slaughter
In this chapter, we cover models of networks. A comprehensive study of
networks would require multiple books. We have more modest goals. We
want to understand the basics of networks, to be able to name their parts,
and to ask why they matter for modeling. The answer we arrive at will be
that networks almost always matter. Any model we construct, be it of a
market, the spread of a disease, or the transmission of information, can be
enriched by embedding the actors in a network.1
Networks are ubiquitous. People talk of trade networks, terrorist
networks, and networks of volunteers. Species organize into food webs, a
form of network. Firms build supply chain networks. As already noted, the
financial system is usefully thought of as a network of promises to pay.
Networks have always been important to understanding social relations.
During much of human history, social networks were constrained by
geography and difficult to map. Due to technological advances, many social
interactions and economic transactions now take place over virtual
networks which can be analyzed using models.
The organization of this chapter follows the same structure-logicfunction format we applied to distributions. We first characterize the
structure of networks using statistical measures of degree, path length,
clustering coefficient, and community structure. Then we discuss common

classes of networks: random networks, hub-and-spoke networks,
geographic networks, small-world networks, and power-law networks.
After that we turn to the logic of how networks form. We construct microlevel processes that produce the network structures we see. Last, we take up
function, the question of why network structure matters. Here we focus on
five implications. We begin with the friendship paradox, and then describe
the six degrees of separation phenomenon and the strength of weak ties
property. Last, we take up the robustness of networks to node or edge
failure and the aggregation of information over networks. The chapter
concludes with a discussion of how networks influence model outcomes.

Network Structure
A network consists of nodes and edges that connect them. We refer to nodes
connected by an edge as neighbors and to a network as connected if it is
possible to get from any one node to any other along edges. Networks can
be represented as graphs, as lists of edges, or as matrices of zeros and ones,
where a one in row A and column B denotes an edge between node A and
node B. Though people prefer graphical representations of networks, lists
and matrices are better for representations for calculating network statistics.
The edges in a network can be directed—that is, pointing from one node
to another. In an information network, a directed edge denotes that one
person gets information from another. In an ecosystem network, a directed
edge from a red-tailed hawk to a gray squirrel represents that the hawk eats
the squirrel. Edges can also be undirected. Edges that connect friends are
drawn in this way. In an undirected network, the degree of a node equals the
number of edges that connect to it. Networks are characterized by a set of
network statistics. For each statistic, we can compute the network average
and the distribution across all nodes. The average degree of a friendship
network tells us, on average, how many friends each person has. The degree
distribution tells us if some nodes are more connected than others. Social
networks have more equal distributions than the World Wide Web, the
internet, and citation networks, all of which have long tails.

Network Statistics
Degree: The number of neighbors (also the number of edges) of a
node.
Path length: The minimum number of edges that must be traversed to
get from one node to another.
Betweenness: The number of paths of minimal length connecting two
other nodes that pass through a node.
Clustering coefficient: The percentage of a node’s pairs of neighbors
that are also connected by a edge.

Path length, the minimal distance between two nodes, varies inversely
with degree. As we add edges, we shorten the average length between
nodes. In an airline flight network, path length corresponds to the number
of flights, on average, a person needs to take to get from one city in the
network to another. Given a choice between two airlines, all else equal
(namely, prices), a traveler would prefer the one with lower average path
length. Average path length also correlates with information loss.
Information that passes through several people is more likely to suffer
distortion than information passed between only two people. The nodes on
minimal paths play critical roles in networks. If information takes the
shortest route, then it goes through the nodes on a minimal path. A node’s
betweenness score equals the percentage of minimal paths that go through a
node. In a social network, people with high betweenness scores know more
information and wield more power.
The final statistic, the clustering coefficient, equals the proportion of a
node’s pairs of neighbors who are also neighbors of one another. For
example, a person with 10 friends has 45 pairs of friends. If 15 of those 45
pairs are themselves friends, then the person’s clustering coefficient equals
. If all 45 friendships existed, then the person’s clustering coefficient

would equal 1, the maximal possible value. The clustering coefficient for
the entire network equals the average of the clustering coefficients of the
individual nodes.

image

Figure 10.1: A Hub-and-Spoke Network and a Geographic Network

Figure 10.1 shows two networks with thirteen nodes: a hub-and-spoke
network and a geographic network. In the hub-and-spoke network, the hub
has degree 12 and all other nodes have degree 1, for an average degree of
less than 2. The degree distribution is unequal. The hub has a distance of 1
to every node. All other nodes have a distance of 1 to the hub and a distance
of 2 to the other nodes. It follows that average path length is also less than
2. The hub, which lies on every minimal path between any two other nodes,
has a betweenness score of 1. The spoke nodes do not lie on any minimal
paths connecting other nodes, so they have a betweenness of 0. Finally, in
the hub-and-spoke network, no nodes connected to a node are connected to
one another. Therefore, the network has a clustering coefficient of zero.
In the geographic network, each node is connected to the two nodes to
its right and its left, so the average degree equals 4. Each node is distance 1
from four nodes, distance 2 from four nodes, and distance 3 from four
nodes. So the average distance equals exactly 2. The degree and distance
distributions for this graph are degenerate—every node has the same degree
and the same average distance. It can be shown that the betweenness of
each node equals .2 Each node has four neighbors, creating six pairs. Of
those six pairs, exactly three are connected: the two nodes to the immediate
left and right are each connected to an outer node and to each other.
Therefore, the clustering coefficient equals .
An alternative method for capturing clustering is to partition the nodes
into communities. In a junior high friendship network, the communities
might correspond to teenagers interested in the arts, athletics, or science. Or
they could be defined by race and gender. A network of political alliances
might partition into regional or ideological allies. Multiple methods exist
for determining communities. One approach sequentially removes edges
with the highest betweenness, as edges with high betweenness are more
likely to connect distinct clusters. Other approaches take the number of
communities as given and seek an optimal partitioning given an objective
function such as minimizing the number of edges between the communities
or maximizing the proportion of edges within communities.3
We can use community detection algorithms to ask questions of network

data. Studies show that people may reside in online bubbles. That is, we
may belong to communities of people who get their news from similar
sources. If so, that has implications for social cohesion. Prior to the creation
of the internet, that may have been true as well, but demonstrating it with
data would have been hard. Now data scientists can scrape the web to
identify the news sources that people frequent and tell us that, yes, in fact
we do live in bubbles to an extent. Models provide the formal definitions of
communities. Data tells us the strength of those communities. Using
judgment we can make wise inferences based on what the data say.

Common Network Structures
In analyzing networks, we encounter a problem of variety. A handful of
network statistics are incapable of pinning down the specific network
structure: one can construct billions of distinct networks with ten nodes and
an average degree of 2. An alternative approach to characterizing a network
is to test whether its statistical measures differ significantly from those of a
common network structure. For example, a scholar might gather data on
judicial citations and put it in network form by drawing an edge when one
judge cites another judge’s opinions. The graph of that network may appear
to possess interesting structures and clusters. We can test whether a network
is random by comparing the network’s statistics to those of a random
network that has the same number of nodes and edges. A random network
clustering coefficient equals the probability of a random edge because two
neighbors of a node are no more likely to contain an edge than any other
randomly chosen node.

Monte Carlo Method for Random Networks
To test whether a network with N nodes and E edges is random, we
create a large number of random networks with N nodes and E edges
and calculate distributions for degree, path length, clustering
coefficient, and betweenness. We then perform standard statistical
tests to accept or reject the hypothesis that the network’s statistics
could have been drawn from the simulated distributions.4

Theoretical models often assume a particular network structure. Many
assume random networks, while others assume regular geographic networks
such as when the nodes are arranged in a circle and each is connected to the
nearest nodes in each direction. Other geographic networks arrange the
nodes on a checkerboard and connect each node to its neighbors to the
north, south, east, and west. Most of the common geographic networks have
low degree—they connect to only the local neighbors—and relatively high
average path length. On geographic networks, betweenness and clustering
coefficient have no variation.
A third common type of network, a power-law network, has a power-law
degree distribution. A handful of nodes has many connections, but most
nodes have very few networks. A fourth type of network, a small-world
network, combines features of geographic and random networks.5 To
construct a small-world network, we begin with a geographic network and
then “rewire” it by randomly selecting an edge and replacing one of the
nodes it connects with a random node. If the rewiring probability equals
zero, we have a geographic network. If it equals 1, we have a random
network. In between, we have a small-world network, distinguished by
small clusters from the geographic network connected by random links to
other clusters. Social networks look similar to small worlds. Each person
has a cluster of friends as well as random friends.

image

Figure 10.2: Random, Geographic, Power-Law, and Small-World Networks

Network Formation: Logic
We now briefly describe models of network formation. These models
provide logic to explain network structures. Most of the network structures
that we encounter emerge from choices of individual actors to make
connections. That is true of friendship networks, the World Wide Web, and
power grids. These networks are not planned. Other networks, such as
supply chain networks, do result from planning. We would expect planned
networks to be robust to the failure of nodes. The fact that emergent
network structures are robust is more of a puzzle.
We have already discussed how to create random networks and smallworld networks. We create the former by randomly creating a set of nodes
and then drawing edges connecting random pairs of nodes. We create a
small-world network by first constructing a regular geographic network
(often by arranging nodes in a circle and connecting k neighbors in each
direction) and then randomly “rewiring” a proportion of the edges.
Models of the formation of the power grid rely on economic and
engineering principles. The network must deliver power to homes,
businesses, and the government. Whether the producers are for-profit
companies or public utilities, they have little incentive to create high
clustering, as it would be inefficient. This lack of clusters reduces the
robustness of the network. Economic and engineering considerations also
rule out long leaps: connections that reach far across the network. Power
companies do not build direct connections from Chicago to Dallas. People
and businesses, however, do. A Chicagoan might strike up a friendship with
someone from Dallas. A firm in Singapore might trade with a firm in
Detroit. As we see in the next section, these long leaps contribute to
network robustness.
To create a network with a long-tailed distribution, we can apply a
version of the preferential attachment model. We create nodes randomly
and then draw edges from new nodes to existing nodes. If we let the
probability of connecting to a node be proportional to its degree, we
produce a power-law degree distribution. In that model, early-arriving
nodes will be far more likely to be of high degree. A shortcoming of the

model is that it does not allow for any difference in node quality. Higher
quality nodes should have higher degree. The quality and degree network
formation model corrects that omission while also producing a long-tailed
distribution.

Quality and Degree Network Formation Model
Create d disconnected nodes. In each period t create a new node with
quality Qt drawn from a distribution F. Connect that node to d other
nodes based on the degree of those nodes. If Dit denotes the degree of
node i at time t, the probability of choosing node i given N nodes
equals:

image

If the quality of new nodes has a low mean and low variance, the model
resembles the standard preferential attachment model. If the quality
distribution has a long tail, then new nodes of very high quality can grow to
have large degree.6

Why Networks Matter: Function
In Chapter 1, we mentioned the friendship paradox, the fact that on any
network, on average, people cannot have more friends than their friends do.
The logic for why this holds can be shown using the hub-and-spoke
network. In that network, twelve people have one friend and one person has
twelve friends. The twelve people with one friend are all connected to the
hub, and the hub has twelve friends. That feature—the fact that high-degree
people are connected to more people—drives the result. On the hub
network, people, on average, have fewer than two friends. Yet, on average,
each person’s friends have more than eleven friends.
The friendship paradox holds for any network: academic citation
networks, email networks, sexual contact networks, banking networks, and
international trade networks. On average, the references cited by an
academic article receive more citations than the article itself; a country’s
trading partners, on average, trade with more countries than the country
itself; and the multiple species connected to a single species in a food
network have, on average more connections than the single species itself.
The disparity between the number of friends and the number of friends of
friends becomes more pronounced on networks with degree distributions
that are more dispersed. One analysis of friendships on Facebook found that
the average person has around two hundred friends and their friends, on
average, have more than six hundred friends.7

The Friendship Paradox
If any two nodes in a network differ in their degree, on average a node
has lower degree than its neighbors. In other words, on average,
people’s friends are more popular than they are.8

The logic of the friendship paradox extends to any attribute that
correlates with the number of friends. If active, happy, intelligent, wealthy,
and kind people have, on average, more friends, then a person’s friends will
be, on average, more active, happier, more intelligent, wealthier, and more
beautiful.9 Imagine a network in which 90% of unhappy people have four
friends and 10% have ten friends. Reverse the proportions for happy
people: 10% have four friends and 90% have ten friends. People’s friends
will disproportionately consist of people with ten friends. A large majority
of those people will be happy, so most people’s friends will be happier than
they are.
We now show the Six Degrees of Separation phenomenon, the claim that
any two people on the earth can be connected through six friends or fewer.
While the friendship paradox holds for any network, six degrees of
separation only holds for some types of networks. The phenomenon’s name
derives from an experiment carried out by Stanley Milgram in the 1960s.
Milgram sent packets to 296 individuals in Omaha, Nebraska, and Wichita,
Kansas, that were to be forwarded to an individual in Boston,
Massachusetts. The recipients had to follow the same rules. These
participants were only allowed to send the packets via mail to people that
they knew personally and whom they believed might have a greater chance
of knowing the target person in Boston, with instructions to do the same.
Individuals signed a roster to record the path and mailed postcards to the
researchers so the researchers could track breaks in the chain. Sixty-four of
the letters arrived in Boston. Of those that did, the average path length was
slightly less than six, hence the phrase “six degrees of separation.”
A second experiment run fifty years later on a much grander scale using

email created eighteen global targets and sent them to more than 20,000
people. The median path length of email chains was between five and
seven, depending on the geographic distance between the source and the
target. The length of paths found does not equal the minimal path length
between participants. The evidence therefore suggests that most people are
linked by fewer than six degrees.10
We construct a simplified version of the small-world network to give
intuition about the six degrees of separation phenomenon. Our version
assumes that individuals have a small cluster of clique friends, who all
know one another, and that they also have friends outside of those cliques,
whom we call random friends.11 Figure 10.3 shows an individual (denoted
by the black circle) with five clique friends and two random friends. It also
shows a selection of friends of the node’s friends (light gray circles).

image

Figure 10.3: A Node’s Clique Friends (C) and Random Friends (R)

These random friends might also be thought of as weak ties—people
who connect you to other communities of people. Our weak ties, the
random friends in our network, play an important informational role by
connecting communities with diverse interests and information. Hence,
sociologists speak of the strength of weak ties.12
This construction allows us to calculate the number of neighbors of
degree two (the friends of friends), by adding up all of the friends of
random friends but only adding the random friends of the clique friends. We
do not count the clique friends of the clique friends, as they are members of
the node’s clique. We calculate the number of friends of friends of friends
similarly. We add in all of the clique friends’ random friends’ friends, but
we do not add in a random friend’s clique friends’ clique friends, as they
have already been counted as neighbors of degree two. To produce the six
degrees of separation phenomenon, we apply the same logic to a network
with 100 clique friends and 20 random friends.

Six Degrees of Separation
Assume each node has 100 clique friends (C), all of whom are friends
with one another, and 20 random friends (R), who have no friends in
common with the node.
Degree one: C + R = 120
Degree two: CR + RC + RR = 2000 + 2000 + 400 = 4400
Degree three: CRC + CRR + RCR + RRC + RRR = 328,000
Degree four: 17, 360,000 13
Degree five: > 1 billion
Degree six: > 20 billion

By assuming no overlap in the friends of the random friends, the model
implicitly assumes an infinite population. An actual social network will
have overlap in friends as the degree increases. In a network that includes
overlap and other realistic features such as heterogeneity in the number of
friends, the values will differ from those calculated above. The relative
magnitudes of the number of neighbors of each degree will remain similar.
A person will have many more neighbors of degree three (friends of friends
of friends) than of degree two (friends of friends).
The large number of friends of degree three, over a quarter million in
our example, can be consequential. Unlike a person’s clique friends, a
person’s friends of degree three tend to live in different cities, attend
different schools, and have different information. They are more diverse.
They are also near enough for trust to be established: a friend of a friend of
a friend could be your roommate’s mother’s coworker, or your sister’s
boyfriend’s aunt. The number of friends of degree three, their diversity, and

their relative proximity make them an important asset. They can provide
new information and job opportunities. These are the people most likely to
help a person find a job, facilitate a move to a new city, or become a life or
business partner.

Network Robustness
Our last implication of network structure evaluates the robustness of
network properties, or how close the network is to node (or edge) failure.
The most essential property of a network is whether it remains connected.
We can use models to calculate the probability that a network remains
connected as a function of the number of nodes removed. We could also ask
what happens to average path length as nodes are removed. Applied to an
airline network, an analysis of path length robustness would tell us how
many extra flights would be needed if an airport were to shut down due to
weather or a power failure.
Here, we consider the question of how the size of the largest connected
component of the network, the giant component, changes as nodes
randomly fail. Figure 10.4 shows the size of the giant component for a large
random network and a large small-world network. In the random network,
the size of the giant component falls linearly at first. At a critical value
where the probability of an edge equals 1 divided by the number of nodes,
the size of the largest component falls to an arbitrarily small proportion of
the original network size. The small-world network shows no such abrupt
change. A majority of connections exist within the geographic clusters.
Each cluster can withstand the failure of multiple nodes. This feature
combined with the random links prevents the entire network from
collapsing.

image

Figure 10.4: Size of the Giant Component (G) as a Function of Node Failure

From figure 10.4 we can infer that sparse networks that lack local
clustering are susceptible to failure. We can apply that insight to the power
grid. It lacks the long leaps and the tight clusters that make the small-world
network robust. In a power grid, the failure of a node or a link cannot be
overcome by other links in a cluster or by a long connection to a working
node far away. Local failures can cascade through the network.14 In
contrast, the internet, which has a long-tailed degree distribution, is robust
to random node failure. The degree distribution implies that the vast
majority of nodes have few connections. Even if they fail, the network
remains connected.
Up to now, we have assumed random node failure. We can also consider
strategic node removal. Networks with long tails, like the internet, now
become non-robust. Strategic removal of the nodes of highest degree
destroys the network. The logic can be seen by considering the hub-andspoke network. When nodes are removed randomly, the network remains
connected unless the hub node is removed, a low-probability event.
Strategic removal, wiping out the hub, disconnects the network in one step.
With some networks, such as terrorist networks and drug supply
networks, we might want to disconnect a network. If those networks are
sparse, like the power grid, or have a long-tailed degree distribution, they
can be disconnected through strategic node removal. For the terrorist
network, this would entail arresting the most connected members. If those
networks resemble small worlds, they will be robust, even to the strategic
removal of nodes. Attempts to cut off any geographic segment of the
network will fail because of the random reconnections that connect the
segment to the rest of the network.

Summary
When we build network models of people, we often do so to capture social
influences, where the success, behavior, information, or beliefs of a person
in the network influence the success, behavior, information, or beliefs of
their friends. Behavior can be contextual or intrinsic; so too can a person’s
value or contribution to a collective enterprise. A person’s value or
contribution could be due to properties of that person, such as her brilliance,
her effort level, or her good fortune. A person’s success could also be due to
his network of friends and colleagues. This is an age-old question: Does
success depend on what you know or whom you know?
Imagine a group of scientists working together in a research lab. They
share advice, ideas, and knowledge. The number of academic papers,
patents, or scientific breakthroughs produced by a scientist depends on what
she knows, but it can also be influenced by whom she knows, on her
interactions with other scientists. By thinking in terms of contextual
features (friendship networks) as well as intrinsic attributes (individual
abilities), we can determine how much of a scientist’s success to attribute to
each.
Investment firms that hire away superstar fund managers based on the
belief that investment success depends mostly on talent have not had very
promising results. Empirical evidence shows that top investors also depend
on networks of colleagues who provide them with specific types of
information.15 That specific finding can be viewed within the lens of a
much larger literature (some of which is model-based) showing how a
person’s position in an organization influences success.
Success still correlates with ability. A business idea that makes its
investors millions was probably a good one. A scientist who publishes
hundreds of papers and receives numerous awards has high ability. At the
same time, those best positioned in the network make the largest
contributions. We can measure a person’s position using betweenness and
other measures of centrality. The people who occupy high-betweenness
positions in a network fill what Ron Burt calls structural holes between

communities, which we can identify using algorithms.16 Access to
information and ideas from multiple communities gives people who fill
structural holes power and influence. Filling a structural hole requires
certain talents and abilities. A person cannot just jump in and fill any hole.
She must build trust and understanding within each community. And she
must be conversant in the knowledge base of each community.
We can apply nearly identical logic to assess the value of firms and
assign power to countries. We can see a firm’s value as intrinsic and take a
balance sheet perspective by looking at assets and liabilities. We can also
look at the context in which the firm operates, such as its position in the
supply chain. Similarly, the power of a country depends on its resources and
its alliances. For both firms and countries, intrinsic attributes and
connectedness correlate. Those who occupy powerful positions in the
network also possess important attributes.
Our analysis as well as most of the literature considers the node as the
unit of analysis. The edges can be critical as well. Taking an even broader
perspective, the network itself may be an appropriate unit of analysis. For
example, teacher networks that allow ideas and information to flow
between classrooms can improve educational outcomes, and so a wellconnected administrator can effectively coordinate curriculum reform.
Similarly, a second-grade teacher knows a lot of information about the
students from his class who are going on to third grade; that information
could help the third-grade teacher. A mathematics teacher knows what
concepts students have yet to grasp; that information could help the science
teacher structure her lessons. Good schools, therefore, have strong faculty
networks. This is just one example of how network models can improve our
thinking.17

Myerson Value and Burt’s Structural Holes
People who fill structural holes connect communities in a network and
have more influence. A variety of statistical measures of a network,
such as betweenness, should correlate with occupying a structural
hole. An alternative measure of influence in a network, Myerson
value, relies on the logic of Shapley values. To compute Myerson
values, we construct a cooperative game on a network but only allow
coalitions that include connected components.
Consider three individuals arranged in a line. Assume that their
locations represent political ideologies with person B in the center, as
shown below. If we restrict coalitions to immediate neighbors, then A,
the left-most person, cannot connect to C, the right-most person,
unless person B also belongs to the coalition. To compute each
player’s Myerson value, we first assign added values to all feasible
coalitions. We then compute Shapley values for each possible
coalition, treating each as a distinct game. Last, we add up the Shapley
values for each coalition game to obtain Myerson values.

image

As an example, suppose that any coalition of two players produces
an output of value 10 and all three players together produce output of
value 14, we obtain the following: Players 1 and 3 have Myerson
values of 3, and Player 2 has a Myerson value of 8.18
Centrality measures such as betweenness are based only on the
network. Myerson values depend on a value function. By having both
measures we can disentangle the dependence of power on a person’s
position in the network and on the functions she performs. In our
example, the Myerson values for the three players, (3, 8, 3), correlate
perfectly with their betweenness scores (0, 1, 0). That will not always
be the case, particularly for more complicated networks and value
functions.