STRATEGIC COMPETITIONS OVER NETWORKS. A DISSERTATION SUBMITTED TO THE DEPARTMENT OF ECONOMICS AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY Carlos R. Lever May 2010
105
Embed
STRATEGIC COMPETITIONS OVER NETWORKS. A …dg223bd6609/dissertation-augmented.pdfFrom Doug I also learned to be extremely focused on delivering the point of the paper. Extremely, extremely,
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
STRATEGIC COMPETITIONS OVER NETWORKS.
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ECONOMICS
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Carlos R. Lever
May 2010
http://creativecommons.org/licenses/by-nc/3.0/us/
This dissertation is online at: http://purl.stanford.edu/dg223bd6609
I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.
Matthew Jackson, Primary Adviser
I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.
Manuel Amador
I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.
B. Bernheim
Approved for the Stanford University Committee on Graduate Studies.
Patricia J. Gumport, Vice Provost Graduate Education
This signature page was generated electronically upon submission of this dissertation in electronic format. An original signed hard copy of the signature page is on file inUniversity Archives.
iii
Abstract
I present three applications of network theory to economic applications. The first chap-
ter studies strategic spending in voting competitions with social networks. It finds that
equilibrium spending targets voters whose position in the network has a high eigenvector
influence. The second chapter studies how eigenvector influence changes when disconnected
components of a network begin interacting. The result has implications for inequality in in-
vestment with social spill-overs and for consumption decisions with social influence. The final
chapter studies pricing competitions in infrastructure networks. It finds how the structure
of the network influences pricing behavior and market surplus.
iv
Preface.
This dissertation studies three applications of network models in economic settings. Network
models in economics seek to incorporate insights and techniques from social network analysis
and graph theory into economics. Simultaneously, economists seek to improve these models
by using our models of decision-making and strategic reasoning to explain how individuals
chose to form their network of relationships.
Chapter 1 analizes the effect of social networks on voting competitions. When political
parties carry out a campaign they know that people are influenced by the opinion of their
social neighbors. With the growth of social networking sites we are collecting more detailed
information than ever before on who talks with whom, but currently we don’t have a theory
to understand how this will alter political campaigns. To address this I propose a model of
strategic spending where two parties compete by targeting resources to convince voters to
chose them. I assume that parties can use the information on the structure of social rela-
tionships to target their resources more effectively and I study the implications for spending
strategies.
The same model can be applied to study of lobbying. Competing interest groups in
Congress have to decide how to allocate their time and resources over different legislators. On
the other hand, legislators develop relationships with each other over repeated interactions
across different legislatures. Understanding the role of each legislator in the structure of
relationships is valuable to decide whom to lobby. With this interpretation in mind I test
the predictions of the model using data on campaign contributions by interest groups in
the US Congress. Consistent with the model, I find that the influence of the position of a
legislator in the network is a significant predictor of the campaign contributions by lobbies
even after controlling for possible confounds. In contrast, the prediction of previous models
of strategic spending is not supported by the data.
In joint work with Benjamin Golub, Chapter 2 develops the theory of networks to un-
derstand how the influence of each member changes when the network changes. Through
our previous research, both Ben and I had became interested in self-referential measures of
v
influence. We were interested in understanding these measures changed when you changed
the structure of the network. As we studied the problem we found that there was a gap in
the literature. Previous papers by the markov chain literature had expressions that worked
as long as the network remained connected, but disconnected groups feature prominently in
the study of trade networks and strategic network formation games, so we decided to expand
the set of tools.
The main contribution of the chapter is that we derived a closed-form solution for how
influence changes when disconnected groups start interacting. Ben deserves the credit for
finding the main insight for the proof and writing the main proposition. We then jointly
worked in correcting the proofs and in finding applications of the result to economic settings.
We came up with two important applications: investment decisions with strong social spill-
overs and consumptions decisions with strong social influence.
Finally chapter 3 analyzes pricing competitions in infrastructure networks. In the short-
run, infrastructure networks severely limit which buyers can interact with which sellers,
inducing an imperfect competition. This chapter seeks to understand how the structure of
the network influences pricing behavior and welfare. It provides an important lesson for
network models that are not explicit on how agent are competing through the network: if
competition is very aggressive, having access to more people through the network can be
harmful: the increased competition offsets any possible benefit. Therefore rational players
refrain from getting as much connections as they can.
vi
Acknowledgements
I have benefitted greatly from my advisors.
Doug Bernheim provided very high quality advice in key moments. His ability to frame
an issue in clear and concise terms is nothing short of awe-inspiring. A typical experience
with him would involve me pitching an idea that I had been working on for months, only to
have him describe it back to me in a way that completely transformed the way I had been
thinking about it, and yet immediately seemed as the absolutely obvious way to view things.
From Doug I also learned to be extremely focused on delivering the point of the paper.
Extremely, extremely, focused. I tend to resist any change in my work, but Doug’s sharp
comments stick with you for a long time. Sometimes I would mull over them for months
before I was ready to detach myself from work and accept I had to drop the non-essential
features. In the end it was always for the best.
After having two senior advisors on my committee, I needed someone as approachable as
Manuel Amador to go over the myriad concerns that happen int he day-to-day of research.
Finding him was a stroke of luck, as he was not in my main fields.1 But Manuel is so
sharp that there wasn’t a noticeable language-barrier. Besides giving sharp advice he was
always available, sometimes on a daily basis (albeit he wasn’t necessarily punctual). He also
provided encouragement and motivation in key moments.
My primary advisor, Matt Jackson, deserves a special mention. He played an immense
role in forming me as a researcher. Matt is extremely generous with his time, patient with
ideas in early stages and receptive to methodologies that are not conventional. He always
pushed me to judge my ideas by their economic implications, not by the complexity of the
methods behind them. I learned from him in a very wide range of settings: by taking courses
with him, by being his TA, through his reading group and through interacting with him at
conferences.
Matt has a unique style of advising. There are several episodes worth mentioning. In the
1It’s a shame that the Stanford theory group does not currently have young professors to perform such arole.
vii
beginning of our relationship he supervised my second year paper. He made me go through
so many rounds of revisions that I ended up submitting it six months late. In hindsight
this was one of the best investments I have made. Irrespective of the content of my paper,
investing to write clearly in my first work has had a big return.
Matt also did an incredible job in putting together a reading group where students
actively give feedback to each other and there is a lot of horizontal interaction. Matt has
managed to establish an informal environment where ideas under development can be freely
discussed. To achieve this, Matt seemed to actively hold back in giving feedback to force
the students to participate. At the beginning this generated after a couple of sessions of
uncomfortable silence, until the students in the group became more and more proactive in
giving feedback. Eventually this lead to a self-sustaining path where the students share the
weight of the work in the group.
Through his example, Matt also taught me how to defend my ideas without taking
criticism personally. As with all new fields in research, especially fields that abandon the
established methodological paradigm, network models attract a passionate backlash. Both in
his networks course and in his seminar presentations, I saw Matt field visceral questions that
went beyond questioning the substantive question and sometimes got downright offensive.
Matt always reacted with a relaxed demeanor, separating his work from his person. He
would accept the comments that improved the ideas, ignore any implications unrelated to
the work and relentlessly push back on comments that were incorrect. Over these years I
have seen more than one of his students adopt this relaxed but firm approach to presenting.
Finally I must mention that behind every great man there is a great woman, or in Matt’s
case, three: his wife and two daughters. In this area Matt also teaches a powerful lesson by
example. Matt is always open about the role of his family in his life, and makes a strong
impression by keeping a balance between work and family. Grad school is full of subtle and
not-so-subtle messages that pressure people into being very narrow-minded about work. In
discussions among students we would always point to Matt as a guiding counterexample.
Beside my advisors I spent a valuable time learning from my professors: Jon Levin, Andy
Skrzypacz, Paul Milgrom, Ilya Segal, Monika Piazzesi, Matt Harding, Giacomo deGiorgi,
Brian Knutson and Antonio Rangel.
I was incredibly lucky to have a strong cohort of students at Stanford. I learned as much
from them as from the faculty. In the theory field, I had endless conversations on all sort
of topics with Aaron Bodoh-Creed, Alex Hirsch, Juuso Toikka and Marcello Miccoli. Their
tough questions in the gradlloquium have been the harshest I have ever faced, but it was
all “creative destruction”. Beyond the theorists, I benefitted from other extremely smart
viii
people in my cohort: Albie Bollard, Alessandra Voena, Neale Mahoney and Max Floetotto.
Matt’s research group allowed me to interact with a large quantity of students interested
in networks and political economy, both from Stanford or visiting. There are too many to
mention all, but Matt Elliot, Ben Golub and Jeanne Hagenbach deserve special mention.
Working on a paper with Ben Golub was always a pleasurable if challenging experience, as
I had to keep catching up with his physics intuitions.
I was also lucky to have one of my closest friends, Gaby Calderon, with me at Stanford.2
She kept me in balance by providing a connection to home while being in a unique position
to understand my present-day dilemmas. The other members of the Mexican community
did a great job in welcoming and mentoring me: Quique Seira, Alex Ponce, Rodrigo Barros,
Alejandrina Salcedo, Luis Fernando Perez, Carlos Mery, Santiago Ocejo, and Carlos Gamez.
I have tried to repay them by greeting the new generations of Mexicans at Stanford.
Most important of all, this long and consuming project would not have been possible
without the support of my family: my parents, Delia and Carlos, and my three amazing
sisters, Mariana, Paloma and Victoria. All foreign student bear a substantial cost from being
away from his family and society. Those that stay behind share the cost while receiving a
negligible part of the benefits.
Finally I want to mention that my responsibility to my country has never been far from
my mind. I come from a place where half the population lives in an unacceptable poverty.
Were it not from the extreme inequality in Mexico, everybody could have the opportunity to
seek the good life in the pursuit of happiness. This is the reason I decided to enter economics
and I haven’t given up on it. I think the best way to pay what I owe to my family, friends
and advisors, is through being a productive member of society, using my gifts to give back
to the dispossessed.
2I like to think I played some part in that luck by lobbying her during the flyout.
This paper studies political competitions when voters influence each other’s opinion. When
people are deciding how to vote or which product to buy, they discuss their decision with
people in their social environment. Studying the pattern of social relationships is important
in understanding how individuals are influenced directly and indirectly by the opinion of
others. Currently we do not have a model of competition that takes these effects into
account.
Using techniques from social network analysis, I propose a model where persuaders strate-
gically assign resources across voters based on thier position on a social network. My model
allows a rich structure of influence between individuals. For example, I allow for influence
to be asymmetric between individuals and put no restriction on the number of people they
talk to.
My main finding is that there is a unique equilibrium in pure-strategies. In that equi-
librium, persuaders spend on each voter in proportion to the influence of his position in the
network as measured by the eigenvector centrality of the position. These measures had been
found in the sociology literature.1 Persuaders also adjust to spend less on voters who in the
margin are harder to persuade.
Previous papers on strategic spending in political campaigns and lobbying have found
that resources should be targeted toward voters who have a higher probability of casting a
1See Wasserman and Faust (1994a); Bonacich (1987); Bonacich and Lloyd (2001).
1
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 2
pivotal vote.2 In contrast, pivotal voters are not targeted at all in my model, only influential
voters are. This yields new predictions on campaign spending. For example, my model
predicts that resources will be spent on voters who have an influential position on the social
network even if they are very unlikely to swing their vote.
Adding the network influence switches spending from pivotal voters to influential voters
because the network spillovers undermine targeting. Pivotal voters are important for elec-
tions because they have the maximum impact in the outcome of the election. This is true
even with network spillovers, but persuaders can no longer effectively persuade individual
voters because their opinions mix with the opinions of their neighbors.
To test the model I match data on campaign contributions by lobby groups with data
on cosponsorships networks in the US Congress. I identify the effect of network influence by
analyzing how the contributions of each legislator vary from one electoral year to the next.
I find that changes in network influence are a significant predictors of changes in campaign
contributions for the House of Representatives, while pivot probabilities are not. After
controlling for several confounds I find that increasing network influence by one standard
deviation increases the campaign contributions by 44,834 US dollars (10.5% of the average
contributions received by a Representative).
In the Senate the result is reversed, network influence is not a significant predictor but
being pivotal to break the filibuster is. These finding have to be taken with caution because
the Senate generates very few observations on political campaigns due to the low number of
legislators and their long office tenures. For example, we cannot statistically reject that the
DeGroot weight in the Senate has the same magnitude that in the House.
My paper brings together two literatures. On the social networks side there has been
much work on identifying the influential members in networks but almost no work has been
done on how this information would be used in competitions. There are a vast number of
measures of network influence.3 But my model predicts that eigenvector measures are the
correct measure.
On the strategic persuasion side, there is a literature on counter-active lobbying4 and
strategic spending in presidential elections,5 but these papers do not allow for voters to
influence one other.
The only previous works on political competitions with network effects is the work in
Galeotti and Mattozzi (2008), which build a model of information disclosure revealed when
2See Shubik and Weber (1981); Snyder (1989).3See Jackson (2008a); Wasserman and Faust (1994a).4Austen-Smith and Wright (1994, 1996).5Merolla et al. (2005).
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 3
voters inform themselves through a social network. Their work focuses on the amount of
information revealed when political parties have an incentive to hide their platforms and on
the types of candidates selected to run. Their work puts much less emphasis on the structure
of the network.
Social networks will be increasingly more important for future political campaigns because
growth in social networking sites is generating more data than ever before on the structure
of social networks. This will allow a level of targeting that would have been inconceivable
a decade ago. Furthermore younger are receiving a larger amount of information through
social networking sites. In a survey by the Pew Research Center on the 2008 presidential
election, 27% of people under 30 reported getting information on the campaign through
social networking sites. The number rose to 37% if you only consider those between 18 and
24 years. This drastically differed from the 4% of people in their 30s and the less than 1%
of people above 40 who reported using these sites.
This chapter is structured as follows: Section 3.2 sets-up the model. Section 1.3 solves
the model. Section 1.4 extends the model for competitions in proportional representation
systems. For most of the chapter I assume persuaders have a fixed amount of resources,
Section 1.5 solves the model when persuaders have to raise their resources at a cost. Section
1.6 tests the model with data on legislative cosponsorship networks and data on campaign
contributions in the US Congress. Section 1.7 concludes. All tables are included at the end
of the chapter.
1.2 The Model
1.2.1 The persuaders
Two persuaders, A and B, have to decide how to spend resources over a group of voters
who will choose between them. The persuaders A, B can be thought of as political parties
or competing lobbies. Interpreting A, B as political parties is straightforward: the parties
have to convince voters to chose them and whoever gets a majority wins.6
To interpret the model as a lobbying competitions, A and B have to be taken as two
opposing lobbies who are fighting over a bill in Congress. One lobby wants the bill to pass
and the other wants it to fail. The voters the must convince are the legislators that vote on
the bill. Lobbies must then target their resources over different legislators to convince them
to vote in their preferred direction. Without loss I will assume lobby A wants the bill to
6Except in the US presidential system, where parties need a majority of the electoral college.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 4
pass while lobby B prefers the status quo be sustained.
I will assume that A and B only care about winning the election. They do not care by
how many votes. In the model there will be uncertainty on the votes, so A and B will seek to
maximize their probability of winning. Since bills need a qualified majority of votes to pass,
I will also solve the model for supermajority rules. In these cases, I will assume without loss
that A needs a qualified majority and that B wins whenever A fails to obtain it.
Section 1.4 solves an alternative model where A and B wish to maximize the percentage
of votes they receive. This is particularly important for political systems with proportional
representation where the number of seats in congress depends on the share of the vote. The
results are qualitatively similar, but competitions with a majoritarian rule illustrate better
the difference of adding network influence to a model of strategic spending.
1.2.2 The voters
There is a finite number N of voters that select between A and B. A subscript i denotes
voter i. All voters have to chose A or B, so turn-out is not an issue.
Each voter will have an opinion vi of the relative value of A relative to B. A larger vi will
be more favorable to A. These opinions are a summary statistic of the relevant information
required to chose between A and B. For example, vi could capture the perception on which
candidate is more able to deal with a financial crisis, which candidate is more determined to
carry out difficult reforms or which candidate has more charisma.
Games of strategic spending frequently only have strategies in complicated mixed strate-
gies and characterizing all of them is hard.7 Solving for these along with network influence
would be intractable. To avoid it I will assume voters chose probabilistically. Increasing the
opinion vi will only increase the probability that voter i choses A over B.
I will assume that increasing the relative value vi smoothly increases the probability of
voting for A. The easiest way to model this is to reparametrize opinions so that vi correspond
exactly to the probability i choses A. To be concrete, I assume that votes are cast to maximize
the following utility function.
U(voting for A)− U(voting for B) = vi − ηi
Where vi ∈ (0, 1) and each ηi is distributed Uniform[0, 1] and drawn independently across
agents. Voter i choses A if vi is greater than ηi which occurs with probability vi. Voter i
choses B with probability (1− vi). For technical reasons I assume everybody has a positive
7These are called Colonel Blotto games by the literature. See Roberson (2006) for a great reference.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 5
probability of choosing both candidates, although the probability a voter swings his vote can
be arbitrarily small.
The stochastic element ηi represents uncertainty about the elements that determine a
vote. This shock need not be random from the point of view of the voter, it only matters
that it’s unknown by the persuaders. There can be all sorts of elements that make voters
have a change of heart when they make their decision. For example, a voter might decide to
change his vote because his happened to shake hands with a candidate at a rally; it has even
been reported that that bad weather can change the outcome of an election by influencing
turnout differently for Democrats than Republicans.Gomez et al. (2008) From the persuaders
perspective, these elements are hard to forecast or control.
To model persuasion and network influence I will assume that persuaders can spend
money to change opinions and that a voter’s opinion is influenced by the opinions of his
social neighbors on the network. I proceed to explain how this happens next.
1.2.3 The timing of the game
The game is divided in several stages which are qualitatively different. Inside these stages
are periods which repeat similar actions.
Let vti represent the opinion of voter i at period t.
• The initial stage: (Round 0) Voters begin with an opinion v0i .
• The persuasion stage: (Round 1) Persuaders simultaneously spend resources to
influence the decision of the voters. Section 1.2.4.
• The network stage: (Rounds 2 through T ) After persuaders spend all their budget,
voters update their opinion parameter through the social network. Section 1.2.5.
• Final stage. After updating their opinion several times through the network, the ηi’s
are realized. Voters pick A with probability vTi .
1.2.4 The persuasion stage
During the persuasion stage, persuaders simultaneously spend resources to influence opin-
ions. Every persuader has a fixed amount RA, RB of resources to spend. I solve the model
when persuaders have to raise resources at a cost in Section 1.5.
Let (ai, bi) be, respectively, the percentage of resources persuader A and persuader B
spend on voter i, so (aiRA, biRB) are the amounts in units of resources.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 6
I assume persuaders can change opinions through the following contest success function.
v1i : R2
+ → [0, 1]
v1i (0, 0) = v0
i
v1i (aiRA, biRB) =
v0i (aiRA)γ
v0i (aiRA)γ + (1− v0
i )(biRB)γ; γ > 0
I chose this functional form for tractability. Below are it’s main characteristics.
• It takes values in [0, 1] and varies smoothly with the amount of resources each persuader
spends.
• A persuader can only completely convince a voter by spending and infinite amount of
resources. Opinion v1i tends to 1 as aiRA →∞ and v1
i tends to 0 as biRB →∞.
• If both persuaders spend the same amount of resources, aiRA = biRB, then the opinion
of voter i doesn’t change: v1i = v0
i . This is a symmetry assumption.
• If both persuaders scale the amount they are spending on voter i by any positive factor,
the opinion v1i is left unaffected.
v1i (λx, λy) = v1
i (x, y);∀λ > 0
This occurs because the contest-success function only depends on the ratio of resources
spent on each voter, (aiRA)/(biRB).
• The marginal persuadability decreases when A, B scale-up their resources.
∂v1i
∂x(λx, λy) <
∂v1i
∂x(x, y);∀λ > 1
This is crucial to get equilibria in pure-strategies.
Contest-success functions have been used in the economics literature to study strategic
spending in tournaments, arms races and competitions.8 Skaperdas provides axiomatizations
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 7
Shubik & Weber use the contest-success function above to solve a smooth Colonel Blotto
game.10 Snyder uses a slightly more general function that is does not depend on the ratio
of resources but has all the other characteristics above. The results are qualitatively similar
but the ratio formula gives convenient analytical solutions that depend on percentage of
resources spent. My model is different in that it allows influence through the network. 11.
The parameter γ measures determines the impact of resources on opinions. For a large
γ, a small difference in the level of spending between A and B dramatically swings opinions
in one direction or the other. As γ tends to infinity, the game becomes a standard Colonel
Blotto game.
1.2.5 The network stage
After persuaders have spent all their budget, voters update their opinion by taking a weighted
average of the opinion their neighbors on a social network. The network is exogenous and
common-knowledge by the persuaders.
Every round of updating, the opinion vti evolves according to
vt+1i =
N�
j=1
Mijvtj; with
N�
j=1
Mij = 1 and Mi,j � 0
The parameter Mij represents the weight voter i puts on voter j’s opinion and Mii
represents how much voter i keeps his previous opinion. Each voter has a unit of attention
his divides between the opinions of his neighbors and his own. The weights are non-negative.
A network is a row-stochastic matrix M with non-negative entries whose rows sum-up
to 1. It summarizes all the structure of who listens to whom and how voters influence each
other. Voters can have asymmetric weights on each other’s opinion; Mij can be different
than Mji. It can even be that voter i influences j but the converse is not true. For example,
bloggers can influence the opinion of their readers without reciprocating by following their
readers’ twits.
The biggest challenge of models with network influence is keeping track of the evolution
of opinions when the structure of influence is complex. This is even more complicated for
strategic spending in majoritarian competitions because it’s hard to calculate how opinions
determine the probability of winning.
10Shubik and Weber (1981)11Snyder (1989)
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 8
By assuming the process above we can set-up the evolution of opinions as a linear-
transition system. That allows us to apply powerful tools from linear-algebra and markov-
chain theory to study the problem.
Let vt be the vector of opinions at time t. This vector evolves according to:
vt+1 = Mv
t = M tv
1
There are two ways of interpreting this influence. It can be interpreted as a model of
information processed in a bounded rational way or as a model of social preferences.
In a an informational world, there is a common-value v that captures the true difference
in value between A and B, but voters do not know it. Instead they have disaggregated
information, or opinions, and they try to update their assessment of v through the opin-
ion of their neighbors. As people update their opinion, the information gets disseminated
through society. Voters update their opinion many times using their neighbors’ opinions to
incorporate the new information that propagates through the network.
Updating opinions through a linear-process with the same weights is not the optimal
bayesian way of processing information. For one, linear-updating assumes that the marginal
influence of voter j’s opinion over voter i is independent of the value of v1j and of the
initial opinion v0j . In the actual political campaigns world, it’s perfectly plausible that
seeing a republican express support for a democratic candidate is more informative than
seeing a moderate democrat express a strong democratic opinion, although both carry some
information.
Linear-updating would be optimal in a world with a normal prior on the true v and with
signals vi that are normally distributed. But even in this world, voters should change the
Mij weights every period to adjust for redundancies in information.12
In spite of these drawbacks, the model can be justified because in general the optimal
bayesian can be quite cumbersome to solve. Therefore boundedly-rational agents might
restrain themselves using simple linear-updating rule. This interpretation was put forward
by DeGroot in the statistics literature and DeMarzo, Vayanos and Zwiebel in the economics
literature. 13
Furthermore, under appropriate conditions on the network structure, this myopic linear-
updating process still processes information in a reasonable way. The work in Golub and
Jackson (2010) shows that myopic linear-updating provides a consistent estimate of the true
v for a large network, as long as the influence of any individual and of any finite group of
12See DeMarzo et al. (2003); Golub and Jackson (2010); Acemoglu et al. (2010) for more on this issues.13DeGroot (1974); DeMarzo et al. (2003)
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 9
individuals is not bounded away from zero.
In the second interpretation of network influence, there is no true parameter v. Instead
voters have a preference for choosing A or B and vi measures the intensity of their preference.
Furthermore, voters have a social preference to vote like their neighbors.14 The problem is
that different voters want to imitate different people, so every voters have to continuously
update their preferences to match their neighbors. The linear-process assumes agents are
myopic in doing so.
For now I will assume that as the network stage advances, voters are willing to completely
change their initial preferences to imitate those around them, but in section 1.3.5 I will extend
the model so that voters chose based on two criteria: which candidate is better according to
an unchangeable ideological dimension; and according to which candidate is better along a
persuadable social dimension. The second dimension corresponds to vi.
1.2.6 Some network definitions
I will refer to the voters as the nodes of the network and will say there is a link from i to
j if Mij > 0. A network is directed if there can be a link from node i to node j without a
link from j to i. A directed network is path-connected if for every pair of nodes i, j there is
a directed path from i to j and a directed path back. That is, either i is connected to j or
there exists a series of nodes {k1, . . . , kn} such that {Mi,k1 , Mk1,k2 , . . . ,Mkn−1,kn , Mkn,j} > 0.
I will also need to assume that my network is aperiodic. Aperiodicity is a technical
condition that is verified if at least one voter places a positive weight on his previous opinion.
I will assume this throughout. See Jackson (2008a) for more details on the definitions.
For any network M that is path-connected and aperiodic, we can find it’s associated
DeGroot weights.
Definition 1 (The DeGroot Weights). Let M be a weighted-directed network which is
row-stochastic. Suppose the network is path-connected and aperiodic. Define the DeGroot
weights of network influence, or simply the DeGroot weights, as the unique left-
eigenvector of matrix M that corresponds to the eigenvalue 1 and whose entries have been
normalized to one. I denote it by s.
14You could also call it an altruistic preference in the sense that the utility of a voter is a weighted averageof his utility vi and the utility of his neighbors.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 10
In math, s is the unique vector such that
sM = s with�
si = 1
This completes the set-up of the model. I now proceed to solve for the equilibria.
1.3 Solving for equilibria.
The following result by DeGroot (1974) is important to solve for equilibria. Under mild con-
ditions on the network, in the long-run all opinions converge to a consensus. This consensus
is a weighted-average of the initial opinion of every voter. The weight each voter receives is
given by the DeGroot weight of the voter.
Theorem 2 (DeGroot 1974). Let M be a path-connected, aperiodic network which is row
stochastic. For any initial vector of opinions v1 ∈ RN we have:
limt→∞
M tv = v∗
1...
1
Where v∗ is
v∗ =�
siv1i
The DeGroot weight of a voter only depends on his position in the network. It does not
depend on his actual opinion nor on the opinion of his neighbors. Therefore we can identify
the influential voters without having to solve for equilibrium spending.
I now present the main result of the paper. If there is a large-enough number of rounds of
network-updating, there is at most one unique pure-strategy equilibrium. In this equilibrium
A and B spend the same percentage any given voter i and spend in proportion to the DeGroot
weight of that voter.
This is stated formally in Proposition (3). Proposition (4) shows that the equilibrium
exists and is the unique as long as the opinion of voters is not too responsive to campaign
spending. Section 1.3.3 explains the example through a stylized example.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 11
It is hard to believe that voters reach a consensus voting probability during political
campaigns. This is not essential to solve for equilibria, it’s just easier to solve the model. In
Section 1.3.4 and 1.3.5, I solve two extensions of the model without consensus voting.
In Sections 1.3.4, I assume the network is not path-connected but composed of disjoint
groups that are path-connected. The DeGroot consensus result holds inside each group, but
not for society as a whole. In Section 1.3.5, I allow individuals to have an unchangeable
“ideology”.
1.3.1 Solving the model with consensus.
Remark 1. Suppose T → ∞. A strategy profile (a,b) constitutes pure-strategy nash
equilibrium if and only if (a,b) solves
max(a1,...,an)
�siv
1i (aiRA, biRB)
s.t.�
ai = 1
and
min(b1,...,bn)
�siv
1i (aiRA, biRB)
s.t.�
bi = 1
To be precise, for a large enough T , the maximizer of the DeGroot consensus is an
epsilon-optimum of the probability A wins. This follows from the uniform convergence of
opinions to the DeGroot consensus and from the fact that the probability of winning is
uniformly continuous in (a1, . . . , aN). I will not dwell in this point. The details for proving
this are well-established but cumbersome. I will directly assume that persuaders maximize
a monotone transformation of the DeGroot weights.
Proposition 3 (On the structure of equilibria). Let M be a path-connected, aperiodic
network. Suppose T = ∞. Then the unique pure-strategy nash equilibrium in spending, if
there exists such an equilibrium, is:
(a∗i , b∗i ) =
siv1i (RA, RB)
�1− v1
i (RA, RB)�
�sjv1
j (RA, RB)�1− v1
j (RA, RB)�
Proof. This proof is an adaptation of the Shubik & Weber proof to my environment.
I first prove that a pure-strategy equilibrium must be in the interior by the contrapositive.
Suppose that bi = 0. A can spend an arbitrarily small quantity on i to obtain v1i = 1. Since
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 12
v1j (·, bj) is continuous for aj > 0, persuader A can increase v1
i by a discrete amount while
decreasing v1j for some j by an arbitrarily small amount. This increases the DeGroot consen-
sus by a discrete amount. Since this is true for an arbitrarily small change in expenditure,
persuader A has no best-response and the strategies cannot constitute an equilibrium.
Next I show that persuaders spend the same percentage on each voter. From the First
Order Conditions (FOCs) I obtain:
si∂v1
i
∂ai= sj
∂v1j
∂aj
si∂v1
i
∂bi= sj
∂v1j
∂bj
⇒ ∂v1i /∂ai
∂v1i /∂bi
=∂v1
j /∂aj
∂vji /∂bj
;∀i, j
By homogeneity of v1, I apply Euler’s law to get
ai∂v1
i
∂ai+ bi
∂v1i
∂bi= 0
−∂v1i /∂ai
∂v1i /∂bi
=bi
ai
From the FOCs we know that the left-hand side must be the constant across i. Therefore
ai/bi must be constant for all voters. This means both A and B must be spending the same
fraction of their resources on each voter: a∗i = b∗i .
We now know both persuaders spend the same percentage on a given voter, but we don’t
know what this percentage is. To find out I use the FOCs.
∂v1
∂a(a∗i RA, b∗i RB) =
∂v1i
∂a(b∗i RA, b∗i RB) =
1
b∗i
∂v1i
∂a(RA, RB) =
γ
b∗iv1
i (RA, RB)�1− v1
i (RA, RB)�
Where the second equality comes from the fact that the partial derivative of v1 is ho-
mogenous of degree -1. I now substitute this in the first order condition for A.
siγ
b∗iv1
i (RA, RB)�1− v1
i (RA, RB)�
= sjγ
b∗jv1
j (RA, RB)�1− v1
j (RA, RB)�
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 13
⇒b∗jb∗i
=sjv1
j (RA, RB)�1− v1
j (RA, RB)�
siv1i (RA, RB)
�1− v1
i (RA, RB)�
Since this is true for any two voters and�
ai =�
bi = 1, I conclude that
a∗i = b∗i =siv1
i (RA, RB)�1− v1
i (RA, RB)�
�sjv1
j (RA, RB)�1− v1
j (RA, RB)�
Proposition (3) does not prove a pure-strategy equilibrium exists. It only shows that con-
ditional on one existing, it must have the stated strategies. To complement this Proposition
(4) shows such an equilibrium exists and is the unique equilibrium of the game is voters are
not too responsive to persuasion. Specifically, if γ < 1 the persuader’s objective function is
strictly quasi-concave and FOCs are necessary and sufficient to find an a best-response.
Uniqueness follow because equilibria in zero-sum games are interchangeable. Since the
best-responses are unique, there can be no other strategy profile that constitutes an equilib-
rium.
If γ > 1 the stated strategies might still be an equilibrium, but there might be other
equilibria in mixed strategies. In those situations all equilibria would be payoff equivalent
because this is a zero-sum game.15 We also know that as γ → ∞, the stated strategies
cannot be an equilibrium, because the game approaches a standard Colonel Blotto game
that has no pure strategy equilibria. (And the equilibrium correspondence as γ → ∞ is
upper-hemicontinuous).
Proposition 4 (Existence and uniqueness of an equilibrium). Take the same assumptions
as in Proposition (3). If γ < 1, the stated strategies are the unique equilibrium of the game.
Proof. Because the probability of winning is a monotone transformation of v∗, they share
the same maximizers. Taking the derivative of v∗ I get
∂2v∗
∂2a= si
�γ
a
�2v1
i (a, b∗)�1− v1
i (a, b∗)��
1− 2v1i (a, b∗)− 1
γ
�
Which is strictly negative whenever γ < 1. Therefore a∗ is the unique best-response to
b∗. Mutato mutandis, b∗ is the unique best-response to a∗. This proves existence.
Uniqueness follows because equilibria for zero-sum games are interchangeable. Take any
equilibrium of the game: (σa, σb). (These are potentially mixed-strategies.) It must be
15See the minimax theorem in Mas-Colell et al. (1995).
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 14
that (σa, b∗) and (a∗, σb) are also equilibria. Since a∗ is the unique best-response to b∗, and
viceversa, I conclude that (σa, σb) = (a∗, b∗).
1.3.2 Discussion: Where are the pivotal voters?
Previous models of strategic spending had found that persuaders should target their resources
on pivotal voters. 16 A voter is pivotal for the election if, conditional on the votes of the
others, his choice changes the outcome. Let N be the number of votes A needs to win. The
the probability voter i is pivotal under v1(RA, RB), represented by qi, is
qi =�
S⊂N\{i}|S|=N−1
�
j∈S
v1j (RA, RB)
�
j� /∈Sj� �=i
�1− v1
j�(RA, RB)�
Pivotal voters are important because persuaders only care about winning, which means
pivotal voters have the highest expected marginal benefit. Spending money to change a vote
that is not pivotal is a waste of resources. This reasoning carries through to elections with
networks, yet pivotal voters play no role in equilibrium spending. Why is this so?
The shift from pivotal to influential voters happens because the network prevents re-
sources from being targeted. When T = ∞ persuaders cannot change the opinion of a voter
in isolation, their opinions mix with the opinion of their neighbors. Persuaders respond by
working to convince society as a group by focusing on the DeGroot consensus. The most
effective way to change this is to target the influential voters.
Even though society only reaches a consensus in the limit, the network blunts targeting
starting on the the first round of updating. A similar shift in resources must occur even
without a consensus. For example, in Section 1.3.5 voters have an ideology parameter that
prevents a consensus in the voting probabilities, but persuaders still target the influential
voters instead of the pivotal voters.
It’s likely that the equilibrium spending with a finite T would involve some combination
pivot probabilities and network influence. Unfortunately, solving for this can be quite cum-
bersome. Section 1.3.4 shows one way of approaching the problem by assuming that society
is composed of disconnected groups. As T tends to ∞, each group reaches a consensus, but
disagreements can persist across groups. Persuaders respond by spending more on pivotal
groups but targeting the influential voters inside the group.
Is there a systematic relationship between pivotal and influential voters? Theoretically
no, these two concepts are orthogonal. One can always construct a network where pivotal
16See Shubik and Weber (1981); Snyder (1989).
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 15
voters are the same as influential voters, or construct a network where influential voters are
completely different than pivotal voters. This is a consequence of linear-updating. Under
linear-updating influence is independent of opinions, but the probability of being pivotal
crucially depends on them. This is shown in the following example.
1.3.3 A Parent-Child example.
Two voters, a parent and a child, have to decide between two (almost) identical products: A
and B. The main difference is that product A is sponsored by a popular cartoon character.
The child is very much convinced that A is better than B so v0child = p ≈ 1. The parent is
of the opposite state of mind. For symmetry, assume v0parent = 1− p.
To decide which product they want, the parent and the child are going to take a vote.
Product B is the status quo object, both the parent and the child have to vote for A to buy
it. Suppose the persuaders, firms A and B, have the same amount of resources to spend on
advertising.
Because of the unanimity rule, a voter is pivotal only if the other voter choses for A. The
parent will be pivotal with probability p and the child with probability 1−p. It’s much more
likely that the parent’s vote will be decisive for the election. Both firms react rationally by
heavily targeting the parent. In equilibrium, both firms spend a fraction p of their budget
on persuading the parent and a fraction 1− p on persuading the child.
Suppose instead that before taking the decision the parent and the child will deliberate
about the decision. The parent feels it’s important to give an equal weight in the decision
to his child’s opinion. The child, being a childish, pays very little attention to the parent.
She places ξ/2 ≈ 0 weight on the parent’s opinion and 1− ξ/2 on her own.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 16
The matrix representation of the network is
M =
�Mparent,parent Mparent,child
Mchild,parent Mchild,child
�=
�1/2 1/2
ξ/2 1− ξ/2
�
The corresponding DeGroot weights are
s =
�sparent
schild
�=
�ξ
1+ξ1
1+ξ
�
Given this, if the parent and the child talk for long enough, the opinion of the child will
almost completely prevail. Knowing this, the firms would spend a large fraction of their
resources on the child, ξ/(1+ξ).
Which is the right model? Different products might have different levels of communica-
tion. The parent might not be willing to discuss with the child what is the right type clothes
for playing in the snow. On the other hand, the car drive from San Francisco to LA will give
the child ample time to convince the parent they should go to Disneyland instead of the LA
Museum of Contemporary Art.
1.3.4 Solving the model without consensus: targeting disconnected
groups.
The previous analysis focused on networks that were path-connected. This section will
extending the analysis to disconnected groups to shows how pivot probabilities and network
influence interact with each other.
Assume the voters can be partitioned into m disjoint groups such that each group is
path-connected and aperiodic. Label them {I1, . . . , Im, . . . , Im}.Theorem (2) implies each group will reach a “consensus” the long-run, but different
groups might end up with different opinions. The DeGroot weights can be constructed for
each group. Let s be the eigenvector of stacked DeGroot weights for each group. (The entries
of the vector belonging to the same group must sum-up to 1.)
Let Nm be the group size of Im and let qm be the average pivot probability in Im.
qm =1
Nm
�
i∈Im
qi
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 17
The unique pure-strategy equilibrium of the game is
a∗i = b∗i =Nmqmsiv1
i (RA, RB)�1− v1
i (RA, RB)�
�m� Nm� qm�
�j∈Im� sjv1
j (RA, RB)�1− v1
j (RA, RB)� ; i ∈ Im
From this I can rewrite the relative spending across voters and across groups as
a∗ia∗j
=siv1
i (RA, RB)�1− v1
i (RA, RB)�
sjv1j (RA, RB)
�1− v1
j (RA, RB)� ; if i, j ∈ Im.
�i∈Im
a∗i�j∈Im� a
∗j
=qmNm
�i∈Im
siv1i (RA, RB)
�1− v1
i (RA, RB)�
qm�Nm��
j∈Im� sjv1j (RA, RB)
�1− v1
j (RA, RB)�
Spending across groups is proportional to the average pivot probability, the size of the
group and a network-average of the marginal persuadability. Spending across voters inside
each group is proportional to the DeGroot weights inside the group.
Even if the real network is path-connected, reaching a consensus could take an arbitrary
long-period of time. This is especially true of a network with high homophily. Homophily
refers to the tendency of individuals to interact with individuals who are similar to them.17
The most persistent disagreements in societies with homophily are disagreements across
different groups. Homophily decreases the speed of convergence of opinions across groups,
but increases the convergence within groups, as shown in Golub and Jackson (2009). For
these societies the disconnected network might be a better approximation to model campaign
spending.
1.3.5 Solving the model without consensus: adding ideology to
the model.
To show that convergence in voting probabilities is not crucial for network effects to dominate
pivot probabilities, I will add ideology to the model. Let θi ∈ [0, 1] be an ideology parameter
for voter i. The closer θi is to 1 the more inclined the voter is to support A. These parameters
are common-knowledge between A and B.
17See McPherson et al. (2001) for many examples of groups that exhibit homophily.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 18
Voters maximize the following utility function.
U(voting for A) = u(vi; θi)− ηi
U(voting for B) = 0
Where u : [0, 1]2 → (0, 1) is continuous function and strictly increasing in vi and in θi.
With this parametrization the probability that voter i choses A is u(vi, θi).
For example, u could be
u(vi; θi) =vi + θi
2
Everything else remains as before. Persuaders can spend to change vi through the contest
success function and voters update vi through the network. The ideology is fixed cannot be
changed by spending nor by the opinions of other people.
In the information interpretation of the model, the parameter θi represents preferences,
while vi represents information. The underlying assumption is that when voters interact
through the network they are able to separate information from ideologies.
In the preference interpretation of the model, the θi represent the private aspects of choice,
those that cannot be influenced by other people, while vi represents the social dimensions of
choice, those aspects that voters wish to match with their social neighbors.
In a pure-strategy equilibria with T = ∞, equilibria of this model are identical to a
model without ideology. Since persuaders cannot change ideologies they focus instead on
influencing the consensus. This is spelled out in detail in Remark 2.
Remark 2. In a pure-strategy equilibrium with T → ∞, persuader A maximizes
v∗ while persauder B minimizes v∗. To see this note that if all probabilities are between
zero and one, the probability persuader A will win the election is a strictly increasing in vi
for any i. Now let v1, v2 be two possible values for v∗ such that v1 > v2. Since the probability
voter i choses A is strictly-increasing in vi, the distribution of votes under a consensus of v1
first-order stochastically dominates the distribution under v2. Therefore the probability A
wins is just a monotone transformation of v∗.18
Therefore the unique pure-strategy equilibrium is as stated in Proposition 3.
18This argument is only true for pure-strategy equilibria.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 19
1.4 Competitions in proportional representation sys-
tems
In this section I solve for equilibria when persuaders want to maximize the share of voters
who select them. This model can also be applied to electoral systems with proportional
representation, where parties get seats in parliament in proportion to the share of votes they
get in the election.
The main result is qualitatively the same as before: persuaders spend over voters in
proportion to another eigenvector based measure of network influence: Bonacich influence.19
Pivotal voters do not matter in proportional representation systems because there is no
threshold to win the election.20
Solving for equilibria in proportional competitions requires less assumptions. For ma-
joritarian competitions, I could solve for equilibria with T = ∞. Proportional competitions
allows me to solve for competitions with a finite T . Instead I will assume T is random
and follows a geometric distribution. This allows me to relate my result with the Bonacich
measure, an important influence measure in the sociology literature.21
The random variable T follows a geometric distribution if the probability the game moving
to t + 1 conditional on reaching round t is constant for all t.
Prob(T � t + 1|T � t) = δ; for δ ∈ (0, 1)
When voters stop deliberating they vote for A with probability vTi .
Conditional on B’s strategy, persuader A solves
max(a1,...,aN )
ET
� �vT
i
�= max
(a1,...,aN )(1− δ)
∞�
t=1
δt−1�
i
vti
As δ → 1 the game approaches the game with T = ∞, so this game is more general.
I can get a stronger result because in proportional competition the objective function of
the persuaders is linear, while in majoritarian competition it was highly non-linear near the
threshold of votes required to win. The non-linearity doesn’t matter in the limit, but it’s
hard to analyze for any finite time-horizon.
19This measure is known as Bonacich centrality in the sociology literature, but to be consistent with myapplication I call it influence.
20Indeed, there is no such thing as a pivotal voter.21See Ballester et al. (2006); Bramoulle et al. (2009) for the relationship between Bonacich influence and
Nash equilibria in games with linear-best responses.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 20
Definition 5. Fix δ ∈ (0, 1). The vector s of Bonacich influence weights for a matrix
M is
s = (1− δ)(1/N, . . . , 1/N)[I − δM ]−1
Proposition 6. Suppose each persuader wants to maximize the percentage of voters that
selects him. Then the unique pure-strategy nash equilibrium, if it exists, is:
a∗i = b∗i =siv1
i (RA, RB)�1− v1
i (RA, RB)�
�sjv1
j (RA, RB)(1− v1j (RA, RB))
If γ < 1, this is the unique equilibrium of the game.
Proof. Take (a,b) ∈ (0, 1)n. I simply show that the objective function of each persuader is
equal to�
siv1i . The rest of the proof follows the logic in Proposition (3) and Propostion
(4) . Setting-up the persuader A’s maximization problem we have
maxa1,...,aN
(1− δ)∞�
t=1
δt−1�
i
vti ∼ max
a1,...,aN
(1− δ)(1, . . . , 1)∞�
t=1
δt−1M t−1v
1
= maxa1,...,aN
(1− δ)(1, . . . , 1)[I − δM ]−1v
1
∼ maxa1,...,aN
s · v1
1.5 Competition with fundraising
Until now I have assumed that the amount of resources was exogenous. In this section I
analyze the possibility that persuaders have to raise resources at a cost.
I find that in equilibrium the ratio of resources raised is independent of the network
influence, of the specific campaign rules and of the initial distribution of opinions. The ratio
of resources only depends on the relative costs each persuader has for raising resources. The
absolute level of resources raised does depend on the rules and the distribution of opinions,
but in ways that are hard to characterize.
I assume each persuader has to pay an cost cj(Rj)k to raise resources with k > 1. The
parameter cj determines the marginal cost of raising resources.
In the first stage of the game, persuaders simultaneously collect resources and the amounts
they raise become common-knowledge. In the second stage, persuaders decide where spend
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 21
it. By backward induction the spending patterns in the second stage have to be the same as
in Proposition (3).
The second stage pay-offs only depend on the ratio of resources collected. Let r = RA/RB
be such ratio and let π(r) be the second-stage pay off for persuader A. I can write each
persuader’s maximization problem as one that only depends on r.
maxRA
π(r)− cARkA = max
rπ(r)− cA(rRB)k
maxRB
�1− π(r)
�− cBRk
B = maxr
(1− π(r))− cB(RA/r)k
To solve for the equilibrium r∗ I again appeal to the ‘hidden symmetry’ of the game. The
FOCs for the problem are
dπ
dr− kcArk−1Rk
B = 0
−dπ
dr+ kcBr−k−1Rk
A = 0
Solving this yields a solution that is independent of π.
r∗ =�cB
cA
�1/k
If cA = cB both persuaders will raise the same amount of resources and their probability
of winning will not change from that determined by the initial opinion of voters plus the
network updating.
Since marginal benefit only depends on r∗ we can find the absolute level of resources by
equate the marginal cost to the marginal benefit in the FOCs above. From this I can derive
two easy comparative statics.
• Constants everything else, if voters are less persuadable, γ decreases, the total amount
of resources raised by each persuader decreases.
• Suppose the marginal cost parameters increase proportionally. That is, (cA, cB) changes
to (λcA, λcB) with λ > 1. Then the total amount of resources raised by each persuader
decreases.
For majoritarian elections the marginal benefit of resources increases with the probability
the election will be decided by a pivotal vote. Persuaders spend more money on elections
that are likely to be close.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 22
The network has an ambiguous effect on campaign spending because it can make the
election more or less close. For example, if everybody is very likely to choose for A except
for one very influential voter, the competition with the network will be more close than
without it. On the contrary, one very influential voter can tilt a large number of undecided
voters, making the competition less close.
1.6 Testing the model with data on lobbying.
My model puts stringent conditions on the behavior of voters and requires a large number
of rounds of network-updating to solve for equilibria. To show if my model is still a useful
approximation to the real world I will test it using data on lobbying expenditures in the U.S.
Congress. My main aim is to test if lobbyists spend more on legislators who have a larger
network influence.
The estimation proceeds in three steps. First I measure the bilateral influence between
each pair of legislators: the weights of the links. Next I calculate the global influence of
each legislator by calculating his DeGroot weights. Finally I regress campaign contributions
on network influence and on a measure of the pivotality of each legislator to see which is
a better predictor. In my data I will observe the same legislator several times, so I will be
able to use the time-variation of campaign contributions and network influence to control
for confounds.
To build the network I use data on the cosponsorship structure of bills in the U.S.
Congress. Every time a bill is proposed in Congress it must have a sponsoring legislator.
Other legislators can sign up as cosponsors of the bill. I will use the frequency of cosponsor-
ship as a proxy for bilateral influence. Every time legislator j cosponsors one of legislator’s
i’s bills, I interpret that i has some influence over j. This data is very convenient because
there is a direction of influence (from cosponsor to sponsor) and because legislators cospon-
sor together many times in the same congress and across different congresses. This allows
me to build a weight for each link.
The cosponsorship data I use has all the bills, resolutions and amendments between 1972
and 2006, from the 93rd to the 109th Congress. The data was collected from the library of
Congress by Fowler (2006a,b).
To construct the network for electoral year t, I take each pair of legislators i, j who
served in year t and construct the entry Mji by counting the number of times that legislator
j cosponsored a bill sponsored by i in any congress where they both served together before
year t. I do the corresponding thing to measure Mji.
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 23
After I counted all the cosponsorships, I divide the cosponsorships from j to i by the
total number of times j cosponsored with anybody else. Therefore the influence from i to
j is measured by the frequency with which j cosponsors i’s bills relative to how often j
cosponsors with anybody else. This makes the network matrix row-stochastic.
Links in the network accumulate over time for legislators that remain in Congress. Since
this strongly biases the DeGroot weights in favor of more senior legislators, I control for
seniority in the regression.
A problem with the data is that some bills are cosponsored by a almost everybody in
congress. This probably has more to do with the content of those bill rather than the influence
of the sponsoring legislator. The distribution of cosponsors per bill decreases exponentially
but spikes up when the number of cosponsors approaches the half the chamber. These peaks
hint that these cosponsored bills involve position signaling by the majority party instead of
influence by the sponsoring legislator.
To deal with this I do two things: I drop the bills that have more legislators than the
threshold where the distribution peaks up (215 legislators for the House and 49 for the
Senate). I then weigh down the links between cosponsors and sponsors by the number of
cosponsors in each bill. If j cosponsored i’s bill along with 9 other legislators, I assign a
weight of 1/10 from j to i. Running the regression without these adjustments yields similar
coefficients but higher standard errors.
The model assumes legislators also place weight on their own opinion, but I do not observe
self-links in the data. To identify the DeGroot weights I assume that all individuals put the
same weight on their own opinion. As long as this weight is positive, the DeGroot weights
will be the same regardless of what value one choses.22
To measure lobbying expenditures I use the campaign contributions by Political Action
Committees (PACs) using data from the Federal Elections Committee from 1990 to 2006.
The data is made available by the Center for Responsive Politics.23
It’s generally considered that PACs donate to get access to legislators and to influence
their vote. But PACs also donate to help elect legislators who are affine to their positions.
Therefore these expenditure do not correspond exactly to the lobbying expenditures in my
model. But this is just an extraneous source of variation in my dependent variable. It
should not bias my estimate for network influence, but will increase the standard errors in
my regression. As such it only makes it harder to test if network influence significantly
22To see this let α > 0 be the weight each legislator puts on himself and M be the network matrix whosemain diagonal is zero and whose rows sum to one. The true network would be αI +(1−α)M but the largestleft-eigenvector of αI + (1− α)M is also the largest left-eigenvector of M .
23http://www.opensecrets.org
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 24
predicts campaigns contributions.
PAC contributions are not the only way interest groups spend resources on legislators.
Interest groups can also hire full time lobbyists. The predictions of the model refer to the
total resources lobbies spend on a legislator, both through PAC contribution and through
lobbyists. As long as these two expenditures are positively correlated, PAC contributions
works as a proxy for lobbying spending.
In each electoral year, many bills are presented and many different lobbies compete over
separate issues. I interpret each bill as an independent realization of my model, with an
interest group on each side of the issue that spends according to network influence or pivot
probabilities. Even if there are many lobbies influencing a single bill, once the content of the
bill is fixed there are only two sides to the issue: those who are in favor and and those who
are not. Groups of lobbies spending in a coordinated matter should spend as in my model. If
on each bill lobbies spend as my model predicts, the total contributions a legislator receives
in a year should also be proportional to the influence of each legislator.
Do lobbies target the same legislators? Unfortunately I cannot test this directly, because
I cannot match PAC contributions to specific bills.
Are lobbies strategic in their campaign contributions? Do they target legislators with
opposing views? There is evidence that some PACs are very strategic in their contributions.
In the 2006 electoral cycle, the top contributing PAC was the National Association of Realtors
which gave to 49% to Democratic candidates and 51% to Republican.
A strand of papers on counteractive lobbying have found that for the number of affine
lobby groups engaged in persuading a particular legislator is positively correlated with the
number of rival lobby groups who try to persuade him.24
At same time many interests groups that have a PAC do not have a lobbyist and the
majority of groups that have a lobbyist do not have a PAC. Nevertheless, as reported in
Tripathi et al. (2002), the groups that have both account for a large share of the total
contributions. “[Although] groups that have both a lobbyist and a PAC account for only
one-fifth of all groups in our sample, these groups account for fully 70% of all interest group
expenditures and 86% of all PAC contributions.”
24Austen-Smith and Wright (1996, 1994)
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 25
1.6.1 The House Specification
To test out hypothesis I run two specifications, ordinary least-squares (OLS) and OLS with
fixed-effects. The specifications for the House are
1 Solomon Gerald 105 1.792 Skeen Joe 102 0.27222 Stark F. Pete 103 1.734 McHenry Patrick 109 0.27203 Gilman Benjamin 106 1.693 Taylor Gene 106 0.27194 Stark F. Pete 104 1.657 LaHood Ray 105 0.27185 Gilman Benjamin 105 1.583 Barcia James 104 0.27166 Crane Philip M. 105 1.556 Davis Danny 105 0.27147 Stark F. Pete 105 1.550 Blunt Roy 105 0.27138 Crane Philip M. 106 1.450 Foxx Virginia 109 0.27139 Gilman Benjamin 107 1.398 Clement Bob 106 0.271110 Crane Philip M. 104 1.371 Ganske Greg 105 0.271111 Solomon Gerald 104 1.354 Matheson Jim 109 0.270912 Stark F. Pete 102 1.348 Miller Dan 105 0.270713 Stark F. Pete 106 1.338 Brown Jr. Henry 109 0.270614 Gilman Benjamin 104 1.317 Pappas Michael 105 0.270615 Shaw E. Clay Jr. 106 1.311 Nussle Jim 103 0.2705
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 33
Summary Statistics for the SenateVariable Mean Std. Dev. Min MaxCampaign contributions* 1,255 811 -6 3,910DeGroot Weights 1.08 0.72 0.09 4.30Relative Majority Pivot Prob 1.00 0.12 0.24 1.18Relative Filibuster Pivot Prob 1.00 0.08 0.31 1.09Seniority 14.1 8.7 2 48Number of Cosponsors 84.3 45.4 9 234Bills Sponsored 88.7 7.2 34 100
*In thousands of 2006 dollars
Most Influential Senators Most Pivotal SenatorsDeGroot Relative
Senator Con- weights Senator Con- filibustergress (sum=100) gress pivot prob
1 Dole Robert 103 4.63 Wyden Ron 108 1.0922 Dole Robert 102 4.30 Bingaman Jeff 108 1.0913 Dole Robert 101 4.28 Reid Harry 108 1.0904 Bingaman Jeff 108 3.70 Durbin Richard 108 1.0895 Bingaman Jeff 109 3.67 Mikulski Barbara 108 1.0876 Kennedy Ted 103 3.48 Cantwell Maria 108 1.0867 McCain John 107 3.44 Feingold Russell 108 1.0868 Kennedy Ted 104 3.36 Domenici Pete 103 1.0869 Kennedy Ted 108 3.30 Lautenberg Frank 108 1.08510 Kennedy Ted 107 3.28 Levin Carl 108 1.08511 Hatch Orrin 106 3.24 Coats Daniel 102 1.08412 Kennedy Ted 109 3.18 Stabenow Debbie 108 1.08413 McCain John 106 3.17 Schumer Charles 108 1.08314 Kennedy Ted 102 3.17 Byrd Robert C. 108 1.08315 Kennedy Ted 105 3.16 Murray Patty 108 1.081
CHAPTER 1. STRATEGIC SPENDING IN VOTING COMPETITIONS 34
Observations 3822 3822R2 0.115 0.116Number of Representatives 928 928
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
Table 1.5: The table shows thew regression with the decomposition of the DeGroot weightsinto direct and indirect influence. In column 1, the direct influence is measured by thenumber of legislators who cosponsored with each legislator. In column 2 the direct effectweighs each cosponsor by the number of bills they cosponsored. In both cases the indirecteffect is the DeGroot weights minus the direct effect.
Chapter 2
The leverage of weak ties:
how linking groups affects inequality.
2.1 Introduction
Why do some groups in society invest so much in certain activities — such as education or
crime — compared with other groups? How does the distribution of investment depend on
social structure? In particular, how does it depend on the nature of the interactions between
mostly separate, weakly interlinked groups in society?
Several studies — most prominently Ballester, Calvo-Armengol, and Zenou (2006) and
Calvo-Armengol, Patacchini, and Zenou (2009) — have found that network centrality relates
to investment levels in social games with linear externalities over a network structure, both
in theory and in the data. Thus, we address the above questions and others by studying
how network centrality is affected when previously separate groups are first connected.
Our results show that, within the Calvo-Armengol, Patacchini, and Zenou model of
investment choices in the presence of social network complementarities, asymmetries in the
weak ties that link homophilous groups strongly affect group-level asymmetries in investment.
The details of the slight intergroup connections matter in subtle and extremely stark ways for
the distribution of investment; the weak ties have a lot of leverage. In particular, even when
the intrinsic productive capacities of all individuals are identical, the shares of investment
undertaken by different equally-sized groups can be arbitrarily asymmetric as a result of
arbitrarily weak intergroup interactions.
We identify two main effects. Suppose we begin with two previously non-interacting
groups X and Y and create an intergroup connection between a member x of group X and
39
CHAPTER 2. THE LEVERAGE OF WEAK TIES 40
a member y of group Y . If x was making a small investment relative to his own group and
y was making a large one relative to her group, then group X will, as a result, have a larger
share of the overall equilibrium investment in the newly connected society. Indeed, the ratio
of the overall effort level of group X to that of Y will be approximately proportional to the
ratio of the pre-connection effort level of y to to that of x. A second, different effect is that if
x benefits more from the connection to y than vice versa, then group X will, as a result, have
a proportionally larger share of the overall investment. Both effects remain strong no matter
how weak the link between groups is; only the asymmetry in the relationship matters.
The two effects complement each other, leading to even greater inequality when both
types of asymmetry are present; we fully characterize the impact of both phenomena on
investment shares with a simple formula. We also fully characterize how each individual’s
investment changes and find that the agents with the intergroup connections increase their
shares of investment relative to their groups, while the ratios of investments of other agents
within a given group do not change.
There are two ways to interpret our results: as a warning about how inequality can depend
crucially on the structure of social relationships on inequality across sparsely connected
groups; or as a warning against our modeling approach for sparsely connected groups.
From the modeling perspective, network models with strong linear spill-overs have become
very common in the study of social networks.1 These models are popular because they give
a tractable framework with very little assumptions on the size or structure of the network.
Currently the biggest challenge of the theoretical models of economic networks is finding
tractable models that allow for complicated patterns of influence across large networks.
Our results give a dire warning for any attempt to bring these models to the data. The
relative centrality across groups that are sparsely connected is very sensitive to measurement
error in the strength of the links across groups. Unless the researcher is very confident of
these measures, the measured centralities must be treated with extreme care.
Our finding isn’t completely negative. Our results also show that the relative centrality
inside groups that are well connected is robust to perturbations that connect the group to
other parts of the network and to perturbations in the link structure inside the group.
From the inequality perspective, an important characteristic we observe in many social
networks is that people tend to form relationships with people who are similar to them
by race, gender, age, class, geographical location and other factors.2 This effect is called
homophily in the sociological literature and while it is widely believed that it matters for
1Ballester et al. (2006); Calvo-Armengol et al. (2009); DeMarzo et al. (2003); Bramoulle et al. (2009);DeGroot (1974); Golub and Jackson (2010, 2009) and see Chapter 1.
2See McPherson et al. (2001) for a comprehensive survey.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 41
social outcomes, there is not much of a theoretical understanding of exactly how it matters.
Our results suggests that the details of the links bridging homophilous groups can be very
important to determine relative outcomes of groups in societies. These details include the
identity of the individuals that have relationships across groups and the position they have
in the network of their own group, as well as the relative direct influence between bridge
agents of different groups.
Inequality across different social groups in society has been widely documented. For in-
stance, in data from the National Longitudinal Survey of Youth (1971–91 waves), 52 percent
of white American males have completed high school at age 24, but only 38 percent of Latin
Americans males have done so.3 As such, we cannot rule out that the structure of social
relationships is influencing the outcomes.
On the normative side, there are many social interventions and programs whose explicit
purpose is to link separate social groups. This includes cultural exchange between coun-
tries as well as charities such as the Big Brothers and Big Sisters of America, which pairs
vulnerable youth with mentors from different social groups. In an equilibrium model, such
interventions have indirect effects on many agents beyond the ones they directly involve, and
ultimately on inequality of investment within a society. Understanding these externalities
is important to designing such of interventions, and our work takes a step toward such an
understanding.
On the way to our results, we make two general contributions that are of independent
interest for the study of centrality measures in networks.
First, we show that eigenvector centrality is the limit of Bonacich centrality in an ap-
propriate sense. This extends a result proved in Bonacich (1991) to a much larger class
of interaction matrices. It also establishes a broad connection between two of the most
important centrality measures used in practice: Bonacich centrality, on the one hand, and
eigenvector centrality, on the other.4
Second, we derive a characterization of the eigenvector centrality of a network obtained
by connecting two previously disconnected networks. Beyond the application to a model of
investment with complementarities, this characterization applies to eigenvector centralities
in other settings such as the Google PageRank measure, the theory of Markov chains (where
the eigenvector centrality vector is the stationary distribution) and numerous applications in
3See (Cameron and Heckman, 2001) for this and more statistics on racial gaps in school achievement andan overview of some of the literature.
4Bonacich centrality was introduced in Bonacich (1987). Eigenvector centrality is also known as Katzprestige (Katz, 1953). It measures influence in an important model of belief evolution (French, 1956; DeGroot,1974), and is also used in the computation of the Google PageRank measure (Langville and Meyer, 2006).
CHAPTER 2. THE LEVERAGE OF WEAK TIES 42
sociology. In these settings, too, arbitrarily weak intergroup links can have arbitrarily large
effects on the distribution of centrality across groups. We comment on some of the broader
implications of this in Section 2.4.
The intergroup connections that we study have been a focus of investigation across several
fields. Granovetter’s (1973) seminal analysis of how social networks mediate learning about
jobs emphasized the informational importance of weak ties connecting otherwise separate
communities. Burt (1992) developed an influential theory of structural holes, exploring the
advantaged positions enjoyed by the agents incident to a bridge — a feature that is also
present, for different reasons, in our analysis. Both works have spawned large literatures in
sociology. In a recent paper on the spread of misinformation in social networks by Acemoglu,
Ozdaglar, and ParandehGheibi (2010), bridges are important in characterizing when agents
can successfully spread their potentially biased beliefs. Bridges in graphs are also well-
studied objects in graph theory (see, for example Diestel (2005)). Our work links the notion
of bridges with the study of centrality which, as mentioned above, has been an important
area in sociology and network analysis more generally (Wasserman and Faust, 1994b).
The paper is organized as follows. Section 2.2 sets up the model, presents the main
definitions on networks and provides our main result: a formula for how connecting groups
changes eigenvector centrality. Section 2.3 then uses the result to derive new predictions
for two economic applications, investment decisions with positive spillovers through a so-
cial network (Section 2.3.1) and consumption decisions with social influence (Section 2.3.2).
In Section 2.4, we summarize the main insights and comment on some implications and
extensions of the results.
2.2 How linking groups affects centralities.
2.2.1 Setup
Let there be n agents (also called nodes) N = {1, . . . , n}. A weighted social network is a
matrix A with non-negative entries aij. An entry aij represents the weight of the link from i
to j. We will assume that A is a column-stochastic matrix,�
i aij = 1 for all j. (See Section
2.3.1 and Section 2.3.2 for an economic justification.) We assume that aii = 0 for all i, which
means we don’t allow self-links.
Below are some common definitions for networks.5
A walk in A is a sequence of nodes i1, i2, . . . , iK , not necessarily distinct, such that
5See Jackson (2008b) for these and other definitions relating to networks.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 43
aikik+1> 0 for each k ∈ {1, . . . , K − 1}. We say that a walk i1, i2, . . . , iK goes from i1 to iK .
The length of the walk is defined to be K − 1. The weight of the walk is defined to be the
product�K−1
k=1 aikik+1. A walk is closed if it starts and ends at the same node, i.e. i1 = iK .
A path is a walk whose nodes are distinct. A network A is path-connected (also called
irreducible) if for every pair of agents i, j, there is a directed path from i to j and back.
We are interested in studying changes in eigenvector centrality, as defined below.
Definition 7 (Eigenvector centrality). Given a nonnegative, path-connected matrix A, the
eigenvector centrality e(A) is the right-hand (column) eigenvector of A with nonnegative
entries. This eigenvector is unique up to scale; we will always assume its entries sum to 1.6
We may also speak of the eigenvector centrality of a group X ⊂ N , written eX(A), which
is just the sum of the eigenvector centralities of its members.
In many economic applications, we are interested in studying how the centrality of an
agent changes when we change the weight of a link or when we connect disconnected groups.
When a network is path-connected, there are formulas for the derivative of the eigenvector
centralities with respect to the weight of any link; we discuss some of these in Section 2.5.
Unfortunately these formulas do not apply for disconnected networks. Since linking discon-
nected groups features prominently in studies of trade, peer effects and strategic network
formation, it’s important to have tools to address this scenario.
Our main result below bridges this gap and provides a closed-form solution for the change
in eigenvector centrality when we link disconnected groups. To get closed-form solutions we
focus on a single link connecting two segments of society. This two-way bridge is defined
formally below.
Definition 8 (A two-way bridge). Let X, Y be a partition of the agents. We say a path-
connected matrix A has a two-way bridge if there exist x ∈ X and y ∈ Y so that
axy > 0, ayx > 0
and
aij = 0;∀(i, j) ∈ (X × Y ) \ {(x, y)}
We call x and y the bridge nodes.
6All vectors in the paper are column vectors unless otherwise noted. The existence and uniqueness of theeigenvector centrality are a consequence of the Perron-Frobenius theory of nonnegative matrices, and arediscussed in (Meyer, 2000, Section 8.3).
CHAPTER 2. THE LEVERAGE OF WEAK TIES 44
Figure 2.1: The nodes (x, y) form a two-waybridge between groups X and Y .
Figure 2.2: This is not a two-way bridge, be-cause the node that connects X to Y is notthe same that connects Y to X.
In words, A has a two-way bridge if it can be divided into two path-connected groups
X and Y such that there is a single pair of agents (x, y) who have influence across groups.
Having a single link that connects the groups will be crucial to get a clean closed-form
solution, but we believe most of our qualitative results hold if the groups have many bridge
links but are sparsely connected. We discuss more on these matters in Section 2.5.
It will be important for our purposes to talk about the restriction of the network to a
subgroup. Given a matrix A with a two-way bridge between X and Y and bridge nodes
x ∈ X and y ∈ Y , let AX denote the matrix with index set X so that
aXij =
aij if j �= x
aij/(1− ayx) if j = x.
The entries are normalized to make AX column-stochastic while preserving the relative
influence of x over the other members of X. We define AY analogously. These restrictions
of A capture the interactions that would exist if there were no bridge between X and Y .
Intuitively, we can think that in the beginning X and Y belong to separate networks
each with column-stochastic interaction matrices AX and A
Y . Then we connect them by
creating the link between x and y, giving rise to a new combined interaction matrix A.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 45
2.2.2 The main result
We now present our main result. The following theorem characterizes the eigenvector cen-
tralities of all agents in the connected network based on their eigenvector centralities in the
restricted networks.
Theorem 9. Assume A is path-connected and has a two-way bridge between X and Y with
bridge nodes x ∈ X and y ∈ Y . Let
cx = 1− ayx(1− ex(AX)) and cy = 1− axy(1− ey(A
Y ))
and
r =axy
ayx· ey(AY )
ex(AX)· cx
cy. (2.1)
Then
ei(A) =1
1 + r·
ei(AX) · rc−1x if i = x
ei(AX) · rc−1x (1− ayx) if i ∈ X \ {x}
ei(AY ) · c−1y if i = y
ei(AY ) · c−1y (1− axy) if i ∈ Y \ {y}.
The formulas are simple to interpret. The centrality of an agent in the connected net-
work is proportional to the his centrality in the disconnected network multiplied by certain
correction factors that depend on the weights of the two-way bridge. The agents in in X
have their centralities multiplied by a factor of r, and then adjusted by the correction factor
c−1x . The agents in in Y have an analogous correction term. The fraction (1 + r)−1 on the
outside is merely a normalization.
Notice that the ratios of centralities between members of the same group remain the same
except when a bridge node is involved. Relative to the other agents in his group, the bridge
agent x has his centrality increased by a factor of (1 − ayx)−1. This generates a perverse
incentive: when agent x increases the benefits he gives to y in the other group, all members
of his group lose centrality, but x gains a higher relative position inside of X.
The logic for the proof is straightforward. Since we know the eigenvector is unique we
guess it has the form above and verify it. The real challenge was coming up with the candidate
guess. We present the intuition below. Reviewing this will clarify why our approach requires
connecting X and Y through a single two-way bridge.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 46
Extending our approach to multiple links is difficult, but once the groups have become
connected by single-link, we can apply derivative formulas for path-connected networks. See
Section 2.5.
2.2.3 The intuition for the formula
The key to guess the formula for eigenvector centrality is to use the fact that the eigenvector
centrality of A is the same as the stationary distribution of a Markov chain with a transition
matrix which is the transpose of A. This Markov chain is one in which a particle randomly
hops around the nodes N with a probability aji of moving from node i to node j. While the
intuitions of this process are rather far from the economic applications we present below, the
advantage of the Markov chain approach is that we can use probabilistic results to study our
problem. This provides natural intuitions that motivate the expression.
Suppose at time n = 0 the particle starts a random walk from node x. Let Wn be the
position of the particle at time n, which is a random variable. Analogously define a Markov
chain starting at x corresponding to AX and denote the position of that particle at time n
in that process by WXn . This process corresponds to a particle that walks only in X, with
the transition probabilities prescribed by AX .
Define the random variable
Ti = inf{n ≥ 1 : Wn = i}
This is the time of the first visit to i after time 0. A well-known formula (Durrett, 2005,
Chapter 5, (4.3)) states that for any M ⊆ N
eM(A) =Ex
��Tx−1n=0 1{Wn∈M}
�
Ex(Tx). (2.2)
Here, the subscripts on the expectations remind us that the particle starts at x and 1{Wn∈M}
is the indicator random variable which takes value 1 if and only if the particle is in M at
time n.
This tells us, for example, that:
Claim 1. For i ∈ X such that i �= x
ei(A)
ex(A)= (1− ayx) ·
ei(AX)
ex(AX).
CHAPTER 2. THE LEVERAGE OF WEAK TIES 47
This is because (2.2) applied to M = {x} yields:
ex(A) =1
Ex(Tx). (2.3)
When the formula is applied to M = {i}, we reason as follows. After leaving x, the particle
does one of two things. (1) It may go to y with probability ayx, in which case it cannot
possibly hit i before its first return to x (recall that there is a bridge from x to y which,
by definition, is the only way to get between the two groups). (2) Or it may go somewhere
within X with probability (1− ayx), and conditional on this event its travels throughout X
before its next return to x have the same probability distribution as they did without the
bridge. Thus we may write
ei(A) =(1− ayx)EX
x
��Tx−1n=0 1{W X
n =i}
�
Ex(Tx). (2.4)
The superscript on the expectation indicates that it refers to the chain conditional on staying
in X, whose transition matrix is given by the transpose of AX . Dividing (2.4) by (2.3) and
using (2.2) for the chain corresponding to AX yields the claim. An analogous claim holds
for Y .
This pins down the new eigenvector centralities inside each group. To finish we only
have to determine the relative influence across groups. For this it suffices to consider the
eigenvector condition
ex(A) =�
j
axjej(A).
2.3 Motivating examples
We know use our formula to provide new results for two economic problems. In both ex-
amples, eigenvector centrality will characterize the equilibrium decisions in the limit with
strong social influence.
2.3.1 Investment decisions with strong social spillovers
We work with a special case of the model of Ballester, Calvo-Armengol, and Zenou (2006)
and define the following investment game. Each agents i selects an effort level zi ≥ 0, which
can be interpreted as human capital investment. Agent i selects zi to maximize:
CHAPTER 2. THE LEVERAGE OF WEAK TIES 48
ui(z1, · · · , zn) = zi −1
2cz2
i + γ�
j �=i
aijzizj. (2.5)
Here, c > 0 is a marginal cost parameter; γ > 0 is a social spillover parameter and aij is an
entry of the column-stochastic matrix A.
The interpretation is that each person’s utility is a combination of three things. First,
there is a linear own-effort effect which we normalize to have a unity coefficient. Second, there
is a convex cost in own effort introduced by the quadratic second term and parametrized
by c. We assume everybody has the same marginal cost of effort. Finally, there are social
complementarities. An agent j spends an amount of time aij “teaching” agent i. The
teaching has complementarities in the level of knowledge of the teacher and of the student.
The benefit i receives from j is increasing in zi and in zj. The total benefit i receives from
other people’s “lessons” is given by γ�
j aijzizj. Agent i’s marginal benefit of investing in
education is increasing in the investment level of the people connected to her through the
social network.
We assume that each agent j spends the same amount of time teaching other agents.7
With an appropriate choice of γ we normalize this so that�
j aij = 1 and the matrix A
is column-stochastic. This amounts to assuming that no agent has more potential to teach
than others, and corresponds to our focus on the network creating inequality as opposed
to ex-ante differences between agents. Note, however, that this is not a strong symmetry
assumption. The weight from i to j might be different from the weight from j to i. The
returning link might even not exist at all.
In order to focus on the consequences of changing the network, we will treat the values
for aij as exogenous.
Theorem 11 below characterizes the equilibria of the game. It is almost follows from
Theorem 1 in Ballester, Calvo-Armengol, and Zenou (2006). Before the statement, we need
one definition.
Definition 10. Given a nonnegative matrix M, provided that [I − αM]−1 is well-defined
and nonnegative, the vector of Bonacich centralities of M with parameter α is
b(M, α) = [I− aM]−11 =
∞�
k=0
αkM
k1, (2.6)
where 1 is the n× 1 vector of ones. Let B(M, α) be the sum of entries of b(M, α).
7A model where every agent spends the same fixed-amount to “learn” rather than “teach” would bestrategically equivalent to out second motivating example. See Section 2.3.2.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 49
Theorem 11. The investment game has a Nash equilibrium if and only if γc < 1. When it
exists, the Nash equilibrium is unique and is given by:
z∗ =
1
cb
�A,
γ
c
�.
All proofs appear in the Proof Sections.
The intuition for this result is that investment is proportional to network position as
measured by Bonacich centrality. The agents who invest the most are the ones who most
benefit from feedback loops of social complementarities. This can be seen in equation (2.6)
which shows that the Bonacich centrality of agent i is a weighted sum of the walks in A that
start at i. This is just a multiplier effect. The size of the multiplier is determined by the
structure of the network and each agent’s position determines how much he or she benefits.8
Bonacich and eigenvector centrality are closely related: as the network feedback-loops
become large, Bonacich centrality converges to eigenvector centrality. This is stated generally
in Theorem 12, which extends a theorem in Bonacich (1991). Bonacich shows the result for
symmetric matrices using a diagonalization argument. Our proof is quite different and does
not require symmetry or even diagonalizability.
To state the result precisely and in full generality, we need one more technical notion —
that of aperiodicity, which is standard in the study of Markov chains. It is a mild technical
condition that holds generically and requires that the greatest common divisor of the lengths
of all closed walks be 1. For example, this condition holds if there is at least one closed walk
with two agents and at least one closed walk with three agents.
Theorem 12. Given a nonnegative path-connected aperiodic matrix M whose largest eigen-
value in magnitude is µ, we have
limα↑µ−1
b(M, α)
B(M, α)= e(M).
The relationship between Bonacich centrality and eigenvector centrality can be under-
stood in the following way: the Bonacich centrality of i is computed by starting with a
baseline centrality of 1 (which corresponds to the linear own-effort term in our investment
game) and sums all walks starting at i, with walks of length k getting weight αk.
8When γc � 1, the social spillovers are so big that there is no equilibrium, because agents would always
want to invest more. One way to see this is to note that this is a supermodular game, so the best-responsemapping converges to the lowest Nash equilibrium when the mapping starts from the lowest action. Whenγc � 1, this dynamic is explosive, so no equilibrium can exist.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 50
Eigenvector centrality measures influence by giving equal weight to all walks starting at
i. The number of such paths is infinite, but by taking appropriate limits, we can still make
comparisons between nodes. Eigenvector centrality pins down these centralities. Adding a
normalization determines the levels.
The higher α is, the greater is the importance of long walks for Bonacich centrality. In the
limit, the baseline effect and the short-distance walks are completely insignificant. Therefore
as the network feedback become large, the ratio of Bonacich centralities converges to the
ratio of eigenvector centralities.
This allows us to use our result on eigenvector centrality to study the investment game
when there are strong social spillovers, focusing on the distribution of investment across
society. For a given interaction matrix A and parameters γ, c, let the equilibrium investment
share of a set X of agents be defined by
sX(A, γ/c) =
�i∈X z∗i�i∈N z∗i
,
where z∗ is the equilibrium as characterized in Theorem 11. Note that even though z
∗
itself depends on the levels of both γ and c, the equilibrium shares as defined above depend
only on the ratio γ/c, which explains the notation. We will also write si(A, γ/c) instead of
s{i}(A, γ/c), and use the boldface notation s(A, γ/c) for the vector with si(A, γ/c) in the
ith position. We will drop the arguments when they are clear from context.
A corollary of Theorem 12 is that when social spillovers are big, these shares are well
approximated by eigenvector centrality.
Corollary 13. Assume A is path-connected and aperiodic. As γ/c approaches 1 from below,
the investment share of a group X approaches the eigenvector centrality of X with respect
to A. That is:
limγc ↑1
s(A, γ/c) = e(A).
As a consequence of Theorem 9, we can characterize the relative investment shares of the
two groups.
Corollary 14. Assume A is path-connected and has a two-way bridge between X and Y
with bridge nodes x ∈ X and y ∈ Y . Then
eX(A)
eY (A)=
axy
ayx· ey(AY )
ex(AX)· 1− ayx(1− ex(AX))
1− axy(1− ey(AY )).
CHAPTER 2. THE LEVERAGE OF WEAK TIES 51
In view of Corollary 13, this means that the ratio of investment shares is given by the
above equation. A simpler formula can be obtained by taking a limit as the link between X
and Y becomes weak.
Corollary 15. Let A be as above and axy = kayx for some constant k > 0. Then
limayx→0
eX(A)
eY (A)= k · ey(AY )
ex(AX).
This simplified formula contains the main insights about how the bridge affects each
group’s centrality.
First, the ratio of the investment of group X to the investment of group Y is directly
proportional to the ratio of the weights between the bridge-nodes: (axy/ayx). If agent x gets
more benefit from the complementarity with y than vice versa, then group X ends up doing
most of the investment in the society after the connection is made.
Second, the ratio of the investment of group X to the investment of group Y is directly
proportional to the ratio ey(AY )/ex(AX). This is the ratio of the bridge agents’ investment
shares in their own groups before the bridge is created. In other words, connecting a relatively
high-investing member of group Y to a relatively low-investing member of group X is good
for the investment share of group X in the combined network.
Third, and perhaps most surprisingly, arbitrarily weak links can have arbitrarily large
effects on the investment shares in the combined network. Even in the limit as the level of
interaction between the two groups tends to zero, the shares of investment can remain quite
unbalanced. Indeed, the formula in Corollary 15 shows that there is a discontinuity as a link
is added; without the link, two groups may invest the same amounts, but after an arbitrarily
weak link is introduced, they may become very unequal.
Fourth, the investment ratio of the groups X and Y depends only on the link between
them and on the within-group investment shares of the bridge nodes. The relative size of
each group is irrelevant as well as other traditional metrics on network. It does not matter
which group is more densely connected or which group has a larger network diameter.
This inequality in investment does not go away if the link between both groups is strong.
The next theorem shows that the comparative statics for strong links go in the same direction:
the investment share of X relative to Y increases if y increases the amount of help she gives
to x; it is also increasing in the influence of y within his own group Y .
Theorem 16. The ratio of investments eX(A)eY (A) is strictly increasing in axy and ey(AY ). It is
strictly decreasing in ayx and ex(AX).
CHAPTER 2. THE LEVERAGE OF WEAK TIES 52
2.3.2 Consumption with strong social influence
We now study a model of consumption with social influence. Every agent makes a consump-
tion decision zi ∈ R and is influenced by the choice of his neighbors. Agents simultaneously
make their consumption decision to maximize:
ui(zi, z−i) = −1
2
�zi − θi
�2 − β
2
�zi −
�
j
aijzj
�2
The parameter θi represents agent i’s ideal choice in a world without social influence.
It is fixed and common knowledge. We call it the autarky ideal point of i. The weight β
represents the agent’s preference for acting like his neighbors. Each parameter aij is an entry
of a matrix of influence A and represents the influence of j over i. We assume the matrix
of influence A is row-stochastic, so every agent is trying to match a weighted average of the
decisions of his neighbors.
Agents face a trade-off between being closer to their autarkic ideal point and being close
to their neighbors’ decision. Because being away from either point has an the increasing
marginal cost, agents chose something in-between.
The theorem below characterizes the equilibria of the game:
Theorem 17. The unique Nash equilibrium of the game is
z∗ =
1
1 + β
�I− β
1 + βM
�−1�θ1, . . . , θN
��.
Corollary 18. In the limit when social preferences dominate (β →∞), all agents make the
same consumption choice which is a weighted-sum of the autarkic ideal points:
limβ→∞
z∗i =�
j
ej(A)θj;∀i.
The equilibrium of the consumption game is very similar to that of the investment game.
The strategies in Theorem 17 involve a modified version Bonacich centrality. There is one im-
portant difference between the games: in the investment game with strong social spillovers,
each agent invests different amounts in proportion his eigenvector centrality. In the con-
sumption game with strong social spillovers, all agents invest the same amount, which is a
weighted sum of the autarkic ideal points. The weights are the eigenvector centralities.
Using Theorem 14, we see that when group X and Y start interacting (become linked)
the new equilibrium can be very biased toward the ideal-points of one of the groups. In
CHAPTER 2. THE LEVERAGE OF WEAK TIES 53
particular if ayx � axy and ey(AY ) � ex(AX) with one strict inequality, the ideal-points of
group X will have a larger weight in the final consumption decision.9
In the limit when the bridge-links are close to zero as in Corollary 15, the new consump-
tion is
z∗ =r
1 + r
�
i∈X
ei(AX)θi +
1
1 + r
�
j∈Y
ej(AY )θj
where r is the ratio
r =ayxey(AY )
axyex(AX)
We can take the analysis one step further by assuming that consumption preferences
evolve over time. Suppose that the consumption game is played repeatedly many times,
but that the ideal points θi change randomly every period. After the θi’s are realized they
become common-knowledge and all agents play the static Nash equilibrium.
Suppose all agents in the same group draw θi from the same distribution but that the
realizations of θi are independent across agents. Let σ2x, σ
2y be the variances of the distribution
for groups X and Y . Suppose that half the members of society belong to group X and half
to group Y . Before the groups are connected the average of agents’ period-to-period variance
in consumption is
σ2z∗ = 2
��1
2
�2σ2
x
�
i∈X
�ei(A
X)�2
+�1
2
�2σ2
y
�
j∈Y
�ej(A
Y )�2
�
Assume that the variance in the equilibrium-consumption in group X is larger than in
group Y .
σ2x
�
i∈X
�ei(A
X)�2
> σ2y
�
j∈Y
�ej(A
Y )�2
This can happen for two (non-exclusive) reasons: the variance of tastes per agent in X
is larger than in Y : σ2x > σ2
y ; or the eigenvector centralities in Y group are more evenly
balanced than in X:�
j∈Y
�ej(AY )
�2<
�i∈X
�ei(AX)
�2. This second effect occurs because
social influence between members of the same network “averages out” the shocks in θi. But
if the eigenvector centrality is distributed very unevenly there is little “averaging-out”. In
the extreme case, all agents listen to a single member of the group and the group variance
9Here and throughout the section ayx and axy are switched with respect to the formula in Corollary 14because here the matrix A is row-stochastic, so we have to transpose it to apply Theorem 9.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 54
in consumption is identical to the variance of that individual’s θi.
Now assume groups X and Y become connected. The new the variance in consumption
becomes
σ2z∗ = σ2
x
�
i∈X
�ei(A)
�2+ σ2
y
�
j∈Y
�ej(A)
�2,
which in the limit as axy, ayx → 0 becomes
σ2z∗ =
� r
1 + r
�2σ2
x
�
i∈X
�ei(A
X)�2
+� 1
1 + r
�2σ2
y
�
j∈Y
�ej(A
Y )�2
Notice that there is a 2 that dropped out when we connect the two groups. This represents
the drop in variance because society it now twice as large, so it has more members “averages
out” the shocks in θi. This potential reduction is diminished depending on how we connect
the groups. When r → 1 the variance in consumption is just the variance in group y and
there is no “averaging-out” effect across groups, so the period-to-period variance increases.
Therefore connecting the groups makes the consumption decisions more homogenous
inside each period and might increase the average variance in consumption from period-to-
period. This last effect occurs because the tastes of the volatile group has a disproportionate
influence in the equilibrium consumption decisions.
2.4 Empirical implications:
the sensitivity of centrality measures
Our results also point out that the global properties of centrality measures can be very
sensitive to small, local perturbations when there are islands bridged by weak ties, as there
often are in real social networks. Imbalances in these weak links can lead to large global
imbalances of centrality, so that small absolute changes in some network parameters can lead
to large absolute changes in the centralities.
This has implications for the empirical analysis of networks.10 When there is noise in
the measurement of the network and the network has an “island-like” structure, our analysis
suggests that the relative centrality across groups should be treated with caution, since they
may be very sensitive to measurement error in the links across groups.
Suppose, for example, that there is one link between agents x and y bridging two groups,
10We thank Matt Jackson for emphasizing the implications of this to us.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 55
and the estimate of the complementarity axy that agent x gets from agent y is twice the
estimate of ayx; however, in reality, the two are equal. This is quite plausible if both quantities
axy and ayx are small, so that the error is very small in absolute terms but leads to a large
error in the estimate of the ratio axy/ayx. In this case, the measurement error causes all the
computed centralities in the group of agent x to be off by a factor of two! On the other
hand, Theorem 9 shows that the ratios of centralities within a group will be fairly accurate,
because they essentially do not depend on the weak intergroup links.
More generally, this seems to suggest that, in the presence of measurement error, eigen-
vector centrality is not a robust tool for comparing centralities of nodes located in parts
of a network that are only weakly connected to each other. At the same time, it is quite
reasonable for comparing centralities of agents in a thick region of the network. Making this
speculation more precise could be a fruitful direction for further work.
2.5 Several links between groups
Our analysis characterizes how investment and consumptions decisions — or, more generally,
network centralities — change when previously separate groups start interacting. A natural
question is what happens when further links are added beyond the first, so that we are
looking at two islands connected by a few links as opposed to just one bridge. This question
is answered in a beautiful paper of Conlisk (1985). Start with a path-connected column-
stochastic A and consider perturbating a single link: Aji → Aji + � and Aki → Aki − �. All
other entries of A are held fixed. This corresponds to agent i increasing his complementarity
or interaction with j (the “favored node”) at the expense of k (the “disfavored node”). The
decrease is necessary to satisfy the assumption that the total complementarity shared out
by any agent is fixed, so that increasing the benefit to one neighbor requires decreasing the
benefit to another.
Conlisk shows that, under this perturbation, for any �, we have
∂e�
∂�= ei(A) · (wj� − wk�),
where wj� is the mean first passage time from j to � in the Markov chain whose transition
matrix is given by the transpose of A. (See our intuition in Section 2.2.3.) This is a measure
of the network distance between j and �. Thus, the centrality of node � changes in proportion
to the difference wj�−wk�. If � is closer to j (the favored node) she will gain centrality while
if she is closer to the disfavored node k she will lose it. The magnitude of the change is also
CHAPTER 2. THE LEVERAGE OF WEAK TIES 56
proportional to the centrality of the node i, the agent doing the redistributing.
The intuitions of the formula cohere with those of our main result. Adding directed links,
all else equal, helps those who get more weight (and the agents near them) and hurts those
who lose weight (and the agents near them). This effect is stronger when the originating
node is more influential.
Indeed, in combination with Conlisk’s results, our treatment of bridges completes the
picture on the comparative statics of the centralities of normalized network matrices (equiv-
alently, the stationary distributions of Markov chains). For any change one might like to
consider, the results working in tandem can completely characterize the effects of the pertur-
bation in an intuitive way — though an exact computation of the quantities would require
solving a system of differential equations.11 Of course, since the relevant centralities are solu-
tions to a well-studied fixed-point problem, one could always simply compute them explicitly
with a program like MATLAB. The virtue of the comparative statics formulas is that they
explain exactly what matters and how, which sheds more light on positive and normative
economic questions than a computational approach. Indeed, these formulas remove some of
the mystery of network centralities. Not only do we know that they satisfy some desirable
fixed-point property, but we also know, in a reasonably explicit way, how they change when
interesting changes happen in the network.
2.6 Conclusion
We presented a new theoretical result that provides a simple, closed-form solution for how
eigenvector centrality changes when disconnected groups become linked. Previous work had
studied only the derivative of eigenvector centralities in connected networks.
The result is particularly useful for comparing group centralities. We find that depending
on the way the groups are connected, the influence of one group can dominate the influence
of the other. In particular, if a member with a central position in group Y becomes linked
to a member with a noncentral position in group X, the resulting eigenvector centrality of
11One change which may at first seem tricky to treat is that of introducing a connection between twoseparate groups which is not a two-way bridge, as in Figure 2.2. For example, it may be that there isa directed link from x ∈ X to y ∈ Y , but the only link in the reverse direction comes back from somey� �= y in Y to x� ∈ X, which may or may not be the same as x. Using techniques similar to those usedto prove our main result, one can characterize the post-perturbation centralities exactly. The formulas arenot particularly elegant. But one can also characterize them by combining Conlisk’s result with ours. Onewould add a two-way bridge at first (say between x and y), and then use the Conlisk formulas to characterizewhat happens when one of its directions is gradually replaced with the correct return link from y� to x�. Wethank Tomas Rodriguez-Barraquer for pointing out that such a change can be substantively important, forexample when academic disciplines are first linked via two separate one-way bridges.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 57
group X will be proportionally larger than that of group Y . This occurs regardless of the
relative size of the groups and other network metrics: their clustering, the density of their
connections, or their diameter.
We also showed how the result is useful through two economic applications. In one,
individuals invest in human capital and generate positive externalities by teaching others
through a social network. In equilibrium, agents invest in proportion to their eigenvector
centrality, which means that one group may end up with a disproportionate share of the
total investment after the groups become connected.
In the second example, every period agents locate themselves in a one-dimensional con-
sumption space by balancing their private tastes with their desire to match the decisions
of their neighbors. In equilibrium, the group with a higher eigenvector centrality has a dis-
proportionate weight in determining the final location of consumption decisions in society.
Furthermore, connecting two groups, somewhat counterintuitively, increases the variance
in consumption from one period to the next. This happens because a small group has a
disproportionate influence in the final decision.
Our results caution that network models with linear-spill over should be used with caution
when dealing with sparsely connected groups. Empirical measures of the relative centrality
across groups are extremely sensitive to measurement error in the links across groups. From
the modeling perspective, the stark conclusions for investment and consumption decisions
across groups also raise questions about the modeling assumptions. The usual justification
for linear spill-overs is that it is good local approximation to any continuos influence function.
Our results point that this approximation is too sensitive to be of use for groups that are
sparsely connected.
At the same time, our results highlight that centrality inside groups that are well con-
nected can be robustly measured as it does not change with small perturbations across
groups.
One of the most interesting directions for further work is to endogenize the network.
Our approach views the network is exogenous, which is reasonable when constraints like
languages, occupations, and geography determine much of the interaction that goes on in
the short and medium run. In that case, the formation of the new links across groups arises
exogenously because these parameters are varied by external forces. Nevertheless, over longer
time scales agents do have choices in their interactions, though these are constrained and
costly. The main challenge is to formulate a model which can encompass a fairly broad range
of constraints on the network formation problem (such as different costs of linking between
different pairs of agents) and produce tractable and interesting network formation dynamics.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 58
Proof sections for the chapter.
Proof of Theorem 9
The proof amounts to checking that the claimed specification — call it �e(A) — satisfies the
eigenvector condition
�ei(A) =�
j
aij�ej(A)
for every i. To do this, we use the facts that e(AX) is an eigenvector centrality (i.e. an
eigenvector with eigenvalue 1) of AX , and e(AY ) is an eigenvector centrality of A
Y . For
example, to check the condition for i = x, we compute:
�
j
axj�ej(A) = axy�ey(A) + axx�ex(A) +�
i∈X\{x}
axi�ei(A)
= axy · 1
1 + r· 1
1− axy(1− ey(AY ))· ey(A
Y )
+ axx · r
1 + r· 1
1− ayx(1− ex(AX))· ex(A
X)
+�
i∈X\{x}
axi ·r
1 + r· 1− ayx
1− ayx(1− ex(AX))· ei(A
X)
= axy · 1
1 + r· 1
1− axy(1− ey(AY ))· ey(A
Y ) (Using the definition of AX)
+ aXxx(1− ayx) ·
r
1 + r· 1
1− ayx(1− ex(AX))· ex(A
X)
+�
i∈X\{x}
aXxi ·
r
1 + r· 1− ayx
1− ayx(1− ex(AX))· ei(A
X)
= axy · 1
1 + r· 1
1− axy(1− ey(AY ))· ey(A
Y )
+r
1 + r· 1− ayx
1− ayx(1− ex(AX))
�
i∈X
aXxiei(A
X)
= axy · 1
1 + r· 1
1− axy(1− ey(AY ))· ey(A
Y ) (Using the definition of e(AX))
+r
1 + r· 1− ayx
1− ayx(1− ex(AX))· ex(A
X)
CHAPTER 2. THE LEVERAGE OF WEAK TIES 59
If we now plug in r from (2.1) and simplify, this is equal to
r
1 + r· 1
1− ayx(1− ex(AX))· ex(A
X) = �ex(A).
This verifies the eigenvector condition for i = x. The calculations for the other indices are
equally straightforward.
Proof of Theorem 11
Part 1. If γc < 1 a Nash equilibrium exists and is unique.
Both of these statements follow from the same reasoning as Theorem 1 in Ballester,
Calvo-Armengol, and Zenou (2006). The proof works by solving the first order conditions,
which are a linear system of equations. The condition γc < 1 guarantees the linear system
has a solution, given by z∗ below, which proves existence.
z∗ =
1
cb
�A,
γ
c
�.
To show uniqueness, note that payoffs are linear in the actions of the other players,
so when opponents play mixed strategies, only the expectations of their choices matter.
Additionally there is a unique best response to any mixed strategy because the cost of effort
is convex. Therefore all equilibria in this regime are in pure strategies. When γc < 1, only
z∗ solves the first order conditions for a pure-strategy Nash equilibrium.
Part 2. If γc ≥ 1, there is no pure-strategy Nash equilibrium.
Suppose, toward a contradiction, that α := γ/c ≥ 1 and there is an equilibrium z∗. The
first-order conditions imply that
z∗ = αAz
∗ +1
c1,
By recursively substituting the entire right-hand side for z∗ we get, for every natural
number K:
z∗ =
1
c
�K−1�
k=0
αkA
k
�1 + αK
AKz.
Since A is column-stochastic, so is Ak for every k. In particular, column i of
�Kk=0 αk
Ak
sums to�K
k=0 αk ≥ K, and thus some entry of that column exceeds K/n. Choosing K so
that K/(cn) exceeds the maximum entry of z∗ yields a contradiction.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 60
Part 3. If γc ≥ 1, there is no Nash equilibrium.
Take a mixed-strategy Nash equilibrium F, which is a vector of cumulative distribution
functions. Create a pure strategy profile z where each j sets zj to the expectation of his
random investment under Fi. Using linearity of expectation as well as the fact that the
payoff of i is linear in each zj, we find that the function zi �→ ui(zi; z−i) is the same as
zi �→ ui(zi;F−i), and so what was a best response remains a best response. This gives
a pure-strategy Nash equilibrium. Thus, by the previous part, there cannot be a mixed-
strategy Nash equilibrium in the regime γc ≥ 1.
Proof of Theorem 12
Let � · � be the supremum norm on Rn and let � · � also denote the induced matrix norm
when the argument is a matrix. Define T = µ−1M. Aperiodicity of M implies that T
k has
a positive diagonal entry for some high enough k (Durrett, 2005, p. 310). That, in turn,
implies that all eigenvalues of A are smaller than µ (Meyer, 2000, Section 8.3). From this it
follows that G = limk→∞Tk is well-defined and that
G =e(A)e(A�)�
e(A�)�e(A),
where e(A) is as in Definition 7 (Meyer, 2000, Section 8.3). The prime notation denotes
transposition.
Claim 1. lima↑1(1− a)(I− aT)−1 exists and is equal to G.
Proof. Fix δ > 0. Choose K so large that for k > K we have��T
k −G�� < δ/2 and then
choose a < 1 so that����K
k=0(1− a)ak��� < δ/4. The Neumann series
(1− a)(I− aT)−1 =∞�
k=0
(1− a)akT
k
CHAPTER 2. THE LEVERAGE OF WEAK TIES 61
converges (Meyer, 2000, Section 7.10). Now,
�����
K�
k=0
(1− a)akT
k −G
����� =
�����
∞�
k=0
(1− a)akT
k −∞�
k=0
(1− a)akG
�����
≤K�
k=0
(1− a)ak�Tk −G�+∞�
k=K+1
(1− a)ak�Tk −G�
≤ 2K�
k=0
(1− a)ak +∞�
k=K+1
(1− a)ak�Tk −G�
≤ 2 · δ
4+
δ
2
∞�
k=K+1
(1− a)ak
≤ δ.
Here we have used the triangle inequality and the fact that Tk and G are both stochastic,
so have matrix norm at most 1.
Recall that that by Definition 10 we have
(1− a)b(T, a) = (1− a)(I− aT)−11.
Thus, for any � > 0, there is a δ > 0 so that the statement
�(1− a)(I− aT)−1 −G� < δ
implies �����bi(T, a)
B(T, a)−
�j Gij�
j,k Gjk
����� =
�����(1− a)bi(T, a)
(1− a)B(T, a)−
�j Gij�
j,k Gjk
����� < � for all i.
Recalling that �j Gij�
j,k Gjk= ei(T)
and putting everything together with Claim 1 shows that for every � > 0, all high enough
a < 1 satisfy ����bi(T, a)
B(T, a)− e(T)
���� < � for all i.
Since b(T, a) = b(M, aµ−1) and eigenvectors are invariant to scale, this completes the proof.
CHAPTER 2. THE LEVERAGE OF WEAK TIES 62
Proof of Theorem 16
We will show that the derivative of the logarithm of the ratio is positive.
First for axy:
∂ log�
eX(A)eY (A)
�
∂axy=
∂
∂axy
�log(axy)− log(ayx) + log
� ey(AY )
ex(AX)
�+
log�1− ayx(1− ex(A
X))�− log
�1− axy(1− ey(A
Y ))��
=1
axy+
1− ey(AY )
1− axy(1− ey(AY ))> 0
Likewise for ey(AY ) we have
∂ log�
eX(A)eY (A)
�
∂ey(AY )=
1
ey(AY ))− axy
1− axy(1− ey(AY ))
=1− axy
ey(AY ))�1− axy(1− ey(AY ))
� > 0
The results for ayx are ex(AX) follow by symmetry.
Proof of Theorem 17
Taking the first-order conditions for each agent’s consumption decision we obtain:
z∗ =
β
1 + βAz
∗ +1
1 + βθ
Which is identical to the FOCs of the investment game with γ = β and c = 1 + β. The
rest of the proof follows from Theorem 11.
Proof of Corollary 18
This follows from Claim 1 of Theorem 12.
Chapter 3
Price competition on a buyer-seller
network.
3.1 Introduction
In standard models of competition, firms can sell to all consumers. Nevertheless, in many
interesting markets, firms have a restricted access to consumers and must compete with other
firms under asymmetric potential consumers. This restriction can come from technological
constraints, for example when consumers have to invest in a technological platform in order
to buy from firms; from geographical constraints, as in land or maritime trade routes; or in-
frastructure constraints, as in the gas, electricity or water markets, where the transportation
infrastructure is fixed in the short run.1
This paper proposes using a network to model the possible transactions between sellers
and buyer by relating the graph-theoretical properties on the network with strategic pricing
and welfare. The network allows for a rich structure of limited interaction between sellers
and buyers. Therefore, we address the age-old question of what lies between the extreme
cases of a perfect monopoly and perfect competition. Networks model are not the first to
allow for local interaction. The closest class of model are the spatial location models, also
called Hotelling models.2 We view network models as providing a tractable alternative for
1In the US, natural gas pipelines are privately owned and offer a a bundled service of purchase anddelivery to to local distribution companies and large industrial buyers. The industry has traditionally beenheavily regulated but in 1996 the Federal Energy Regulatory Commission (FERC) decided to allow market-based transportation rates to allow for a greater supply-demand responsiveness. A necessary condition toderegularize a pipeline was showing it did not have significant market power. For a more detailed analysissee McAfee and Reny (2007).
2For a reference see Tirole (1988).
63
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 64
different environments.
Our main findings are that under price competition, aggregate surplus decreases mono-
tonically as the network becomes less connected but the payoffs for individual firms are
not monotonic. Aggregate surplus is maximizes by a fully connected network which gener-
ates perfect competition. The maximum inefficiency occurs in disconnected networks where
each firm is a local monopolist. We find that stable networks are midway between the two
extremes.
We find that firms do not always want to have access to more consumers because it could
increase the aggressiveness of competition. Stable networks are those where firms balance the
trade-off between more potential consumers and more aggressive competition. We also find
that firms do not want to have the ability to price-discriminate because it makes competition
more aggressive.
Two important previous papers on the matter are Kranton and Minehart (2001) and
Corominas-Bosch (2004). Both papers deal with buyer-seller networks. The main difference
between these papers and ours is the market protocol used to determined prices and alloca-
tions. We model the market through price (Bertrand) competition. Kranton & Minehart use
an ascending-bid auction that simultaneously determines prices for all sellers. Corominas-
Bosch analyzes an alternating-offers bargaining protocol. More recent paper by Blume et al.
(2007) solves a model where traders balance supply and demand by setting a spread be-
tween the price for buyers and for consumers; while Elliot (2010) finds the core allocations
for trades with heterogenous values.
In contrast to our model, all of the cited papers find that the allocation is efficient
conditional on the feasibility constraint imposed the network.3 Even degenerate networks
where firms are monopolists have full efficiency! Therefore their choice of the underlying
economic environment does not allow us to see how the network structure moves welfare
with through it’s influence on the market power of each agent. The only thing we can be
sure of is that the network structure does not add any interesting inefficiency of it’s own.
In our model we allow both full efficiency and for inefficiencies dor the networks that
correspond to the standard cases studied in economics: perfect competition and monopoly
pricing. We then ask how the network mediates the inefficiency as we move get networks
that are more similar to one or the other.
To analyze which networks are likely to form we take two approaches. First we study
the properties of pairwise-stable networks. Pair-wise stable networks are between networks
with local monopolists and the network with perfect competition. Next we study a couple of
3Corominas-Bosch does find inefficiencies in the way the network is formed.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 65
entry games where before the pricing game firms chose which links they want to form with
consumers in the other firm’s niche. It turns out that the equilibrium of the entry games are
always pairwise-stable.
The paper is structured as follows: Section 3.2 sets up the basic duopoly model; Section
3.2.1 solves an example of competition between tollway stations on a road network that
shows our main findings and contains most of the intuition for solving the model; Section
3.2.2 solves for the unique Nash equilibrium of the duopoly model. Section 3.2.3 does some
comparative static exercises for welfare. Section 3.2.4analyze what would change if we allow
for price-discrimination. Section 3.3 analyzes which networks are likely to form. Finally,
Section 3.4.1 extends the results to a case of oligopolistic competition where consumers can
either buy from one firm or all of them and closes with a discussion on the challenges to
solve more general oligopoly models.
3.2 The duopoly model
F1 F2
C1
��
��
C1,2��
��
��
��
C2
��
��
Figure 3.1: A number C1 of consumers can only buy from Firm 1, a number C2 from F2;and a number C1,2 can buy from any firm.
Our model has a finite number consumers and that there are two firms, labeled F1 and
F2. Firms and consumers are represented by nodes on a network. Let C1 denote the number
of consumers who can only buy from Firm 1, C2 the number who can only buy from F2, and
C1,2 the number who can buy from both. Consumers must be linked to a firm to be able to
buy it’s product or service. In Figure 3.1, for example, C1 can only buy from F1 while C1,2
can buy from both F1 and F2.
The network structure is exogenously determined before any transactions are carried out
and is common knowledge between the players. We will discuss how the network is formed
in Section 3.3.
We refer to the C1 and C2-type consumers as locked-in consumers, to the C1,2-type as
mobile consumers and label the firms such that Firm 1 is the one with weakly more
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 66
locked-in consumers: C1 � C2.
Consumers can only buy one product. They has a value for the product that is private
information. Products from different firms are perfect substitutes. Values drawn indepen-
dently from a common distribution function. Let Q(p) be the probability that a consumer
has a value higher or equal to p. That is, Q is defined over [0,∞) → [0, 1] and is decreasing.
We assume Q is differentiable.
For tractability we assume that Q has a decreasing marginal revenue in prices.
Q(p)− p∂Q
∂pis decreasing in p
This assumption to gives tractable predictions for firms who only focus on the revenue
from their locked-in consumers. In our model, firms face a trade-off between the gains
from acting more monopolistically and acting more competitively. The decreasing marginal
revenue property ensures there is a unique solution for a firm who acts like a monopolist.
Also, for the price range below this unique solution, any increases in prices lead to increases
in the revenue from “locked-in” consumers. This monotonicity of revenues will be useful in
solving for equilibrium. We discuss this in more detail in Section 3.2.2.
The nodes Cj can also be thought as different locations and we can interpret Q as the
total demand at each location, where we normalized the maximum possible demand at each
location has been normalized to Cj. In this case a sufficient condition to have the decreasing
marginal revenue party is that the demand has a weakly decreasing elasticity, which can be
checked by verifying the demand is log-concave.
Firms engage in Bertrand competition. Each firm simultaneously announces a price it will
charge to any consumer for buying her product. Firms cannot price discriminate. We show
the consequences of allowing price discrimination in Section 3.2.4. Consumers first observe
all prices and then chose to buy from the cheapest firm available to them with probability
Q(p). They randomize equally across firms in case of a tie.
Firms maximize their expected profits. We assume firms have a constant marginal cost
for selling their products which we normalize to zero. Firms have no fixed costs. Therefore
profits are simply the quoted price times the quantity sold, which is determined by the prices
of both firms as above. Let πj(p1, p2) denote the profits for Firm j when firms quote (p1, p2).
A pure strategy for firm j is a simply a price pj she wishes to quote. Firms can quote
any non-negative number.
pj ∈ [0,∞); j ∈ {1, 2}
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 67
In general to find equilibria we will need to allow firms to randomly select with a price.
Definition 19. A mixed strategy pricing scheme for Firm j if it is a distribution function
σj : [0,∞) → [0, 1].
Abusing notation, we use πj(σ1, σ2) for the expected profit for firm j when firms are
randomizing according to (σ1, σ2). A subscript −j denotes the firm who is not j. In the
oligopoly section it denotes all firms that are not j.
To solve the game we look for Nash equilibria. Because the behavior of consumers is
completely captured by Q(p), we focus on the strategies for the firms.
Definition 20. A Nash Equilibrium of the pricing game is a strategy profile, (σ1, σ2), such
that no player j can deviate to an alternative mixed strategy σ�j and get a higher payoff:
πj(σj, σ−j) � πj(σ�j, σ−j);∀σ�j �= σj,∀j ∈ {1, 2}
We denote by a star-supercript strategies that constitute an equilibrium, (σ∗1, σ∗2). Let
(π∗1, π∗2) denote the corresponding expected profits.
Next we develop an example to preview how to solve the game and expose some of the
results of the paper. Those interested in going directly to the details of solving the model
can skip to Section (3.2.2).
3.2.1 An Example
Two consumers, A and B, wish to cross a river to go to an Irish pub. To do so, they must
cross through one of two available bridges. Each bridge is independently operated by a firm
that charges a toll to any consumer who wishes to cross. Consumers are willing to pay up to
1 to cross the river. The only cost they incur is that of the toll. In addition, an exogenous
road network limits the choice of bridges for each consumer. For example, in Figure (3.4),
Consumer A can only cross through the bridge operated by F1 while Consumer B can choose
to cross through any bridge. Firms simultaneously post their tolls (prices). After observing
them consumers choose which bridge to cross, if any.
In the Monopolistic Network, Figure (3.2), each firm knows she has a monopoly and
quotes prices to extract all surplus:
p∗1 = p∗2 = 1
π∗1 = π∗2 = 1
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 68
The Irish Pub��
������������ ��
������������
F1 F2
CA
��
CB
��
Figure 3.2: (The Monopolistic Network) Each Firm is a monopoly in their local market
The Irish Pub��
������������ ��
������������
F1 F2
CA
������������������������������� CB
�������������������������������
Figure 3.3: (The Competitive Network) A network with perfect competition
The Irish Pub�� ��
F1 F2
CA
��
CB
�� ��
Figure 3.4: (The Mixed Case Network) An intermediate case
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 69
The Competitive Network, Figure (3.3), is a standard Bertrand competition where firms
undercut each other until they vanish their profits.
p∗1 = p∗2 = 0
π∗1 = π∗2 = 0
In the Mixed Case Network, Figure (3.4), F1 has a locked-in consumer (C1) from where
she can extract all her valuation by charging p1 = 1. Alternatively she can try to compete
against F2 for the mobile consumer, CB, at the cost of reducing the price charged to her
locked-in consumer, CA. No pure strategy Nash equilibrium exists in this game: If any firms
charges a positive price, at least one of the two firms would want to charge slightly below.4 If
both Firms charged zero, F1 could profitably deviate to charging 1, extracting all the surplus
of her locked-in Consumer. This is the same logic we use to rule out pure strategy equilibria
in the duopoly model.
There is unique a mixed strategy equilibrium. This is described below with it’s corre-
sponding expected profits.
F1 : Post p = 1 with probability 12 . With probability 1
2 chose p < 1 and then price according
to:
σ∗1(p|p < 1) = 2− 1
p;∀p ∈ [
1
2, 1)
F2 : Price according to:
σ∗2(p) = 2− 1
p;∀p ∈ [
1
2, 1)
π∗1 = 1; π∗2 =1
2
From the example we can already see that firms are not always better off by having access
to more consumers. In the Mixed Case Network, Firm 2 receives a payoff strictly larger than
in the Competitive Network because the more aggressive competition dissipates profits.
Another result that generalizes to the duopoly model is that in equilibrium Firm 1 receives
a payoff exactly equal to the monopoly rents from her locked-in consumers. It is fairly direct
to deduce her payoff could not be smaller than this. The fact that it’s exactly the same
comes from Firm 1’s has a higher opportunity cost for charging low prices which places it at
a disadvantage to compete for the mobile consumers. This does not mean Firm 1 only sells
4Such a firm would in fact not have a best response because of an openness problem.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 70
to it’s locked-in consumers. As we see in the example she sells to all potential consumers
with some probability.
Equilibrium pricing strategies for our model are also qualitatively similar to those in the
example. Firm 1 quotes the price the maximizes her monopoly payoffs (1 in the example) with
a discrete probability and both firms mix continuously over prices below. In the example,
both firms can be sure to sell to both consumers if they set a price equal to 1/2. This is the
minimum price Firm 1 is willing to quote. Any price lower must forcefully yield an expected
profit lower than what she can assures herself by charging the monopoly price. Since quoting
lower prices is strictly dominated for Firm 1 they are also ruled out for Firm 2.
3.2.2 Solving the duopoly model
Let’s return to analyzing an arbitrary network with two firms and any number of consumers.
To solve the model it’s useful to derive two useful properties from the decreasing marginal
revenue property of the distribution of consumer values, Q(v). First, there is a unique price
pM that maximizes revenue from locked-in consumers. It turns this price is the same price
for all firms because the number of ”locked-in” consumers only change the optimization by
rescaling it. Second, over the price range below below pM any increase in prices yields higher
revenues from the locked-in consumers.
Proposition 21. For any firm with Cj > 0 consider the maximization over the revenue from
only her locked-in consumers:
maxp∈[0,∞)
pQ(p)Cj
If Q(·) has the decreasing marginal revenue property there is a unique price, labeled
pM , that solves the monopolist’s problem. This price is independent of Cj. Furthermore, for
any prices p, p� such that p < p� � pM , p� yields a higher revenue than p from the locked-in
consumers.
Proof. First we show existence. Note that the limit of the objective function as p goes to
infinity is zero. For low values of p it is strictly positive. Therefore we can restrain attention
to a compact subset of the price space. The theorem of the maximum guarantees existence.
Now we show uniqueness. Take the derivative of the objective function:
Cj
�Q(p) + p
∂Q
∂p
�
By the decreasing marginal revenue property derivative of the objective function crosses
zero only once and from above. Call pM the value where it crosses zero. This is the unique
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 71
maximum and does not depend on Cj. Furthermore, for values below pM , the derivative is
positive, so any infinitesimal increase in the price increases revenues. By the fundamental
theorem of calculus, the change in revenue between any two prices below pM is equal to
the integral of this derivative, therefore revenue increases monotonically until it reaches it’s
maximum.
A direct implication of Proposition 21 is that any price above pM is strictly dominated:
decreasing the price to pM would strictly increase the revenue from the locked-in consumers,
while weakly increasing the probability of winning the mobile consumers.
Proposition (22) shows that there are no pure strategy Nash equilibria except for two
extreme cases. One is the standard Bertrand competition which corresponds to C1 = C2 = 0;
the other is when each firm is a separate monopolist, C1,2 = 0.
Proposition 22. There is no pure strategy equilibria for the model with C1 > 0 and C1,2 > 0.
Proof. An pure strategy equilibrium with with a positive price can be ruled out by the
standard Bertrand argument as follows. Take a strategy profile (pj, p−j) such that p−j <
pj � pM , then player −j would wish to deviate to pricing arbitrarily closer to pj. Now
assume pj = p−j > 0, then any player could deviate to an arbitrarily smaller price obtaining
a discrete gain from avoiding the probability of a tie while charging essentially the same
price. Finally, assume (pj, p−j) = (0, 0), then F1 could deviate to extracting a positive profit
by quoting pMand selling to her locked-in consumers.
We therefore look for mixed strategy nash equilibria. It will turn out that the equilibrium
is unique and involves mixing over a continuous interval. Proposition 23 solves for the
equilibrium payoffs. To obtain them the proof first solves for some standard properties
of continuous mixed strategies and Bertrand competition: ties cannot occur with positive
probability; strategies have no atoms below the upper bound on prices, pM ; strategies must
have a common support for the two firms; strategies cannot have gaps in their support and
must not have gaps below the upper bound on prices, pM . These are enough to pin down the
equilibrium payoffs. Corollary 24 solves for the unique equilibrium strategies by plugging
the equilibrium payoffs into the mixed strategy indifference condition.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 72
Proposition 23. The unique equilibrium payoffs for the duopoly network pricing game are:
F1:
π∗1 = pMQMC1
F2:
π∗2 = π∗1C2 + C1,2
C1 + C1,2
Proof. See Proofs Section (3.4.2).
Corollary 24. The unique mixed strategy equilibrium for the duopoly network pricing game
are:
For F1: “Stay out” with probability�1− C2+C1,2
C1+C1,2
�by charging the monopoly price;
For F2: Always “go in”: charge below pM with probability 1.
For Both: Conditional on going-in mix according to:
σj(p|p < pM) = 1− C1
C1,2
�1− QMpM
Q(p)p
�
Proof. See Proofs Section (3.4.2).
3.2.3 Welfare
We can use the payoffs in Proposition (23) to do comparative statics on welfare by changing
the number of consumers.
Firm 1’s payoff does not depend on the size of the mobile market, C1,2, nor on the size
of Firm 2’s locked-in market, C2. She gets a profit which is exactly the same as if extracted
the monopoly surplus from her locked-in consumers. This does not mean Firm 1 doesn’t sell
to the mobile consumers. She does with positive probability. Nevertheless, in equilibrium
Firm 1 must be made indifferent between quoting the monopoly price and quoting below.
As such the mixing must be such that any increase in expected sales by lowering prices must
be exactly offset the decrease in price.
Firm 2 has a profit that is strictly increasing the number of her own locked-in consumers.
Her profits are also strictly increasing in the number of mobile consumers, C1,2. Surprisingly
her payoffs are also strictly increasing in the number of locked-in consumers Firm 1 has.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 73
The driving force is the effect on competition. A larger number of C1 makes Firm 1 less
aggressive, which is beneficial for Firm 2. Firm 2’s profits are smaller than those of Firm 1
but larger than the rents from the monopoly rents out of her locked-in consumers.
π∗1 � π∗2 � C2pMQM
The inequalities are strict whenever, respectively, C1 > C2 and C1,2 > 0.
Every consumer is weakly better off by any increase in the number of mobile consumers,
C1,2, because the distribution of prices shifts downwards. The old distribution first-order
stochastically dominates the new distribution.
Aggregate surplus is maximized by equating demand to marginal cost. Because of we
normalized marginal cost to zero this is equivalent to producing to fulfill all demand. Higher
prices create deadweight loss.
3.2.4 Price discrimination
Firms would prefer not to be able to price-discriminate between mobile and locked-in con-
sumers. If both firms could discriminate in prices, profits from the mobile consumers would
completely dissipate. Firm 1 would be indifferent between being able to price discriminate
or not, she doesn’t receive any profit from mobile consumers anyway. Nevertheless, if there
were an arbitrarily low cost from being able to price-discriminate, she would be deterred
from trying. Firm 2 would strictly prefer price discrimination were not possible as long as
C1 > C2.
Each firm would wish to be able to discriminate if the other firm could not. If a firm were
the only one with the ability to discriminate, she could compete for the mobile consumers
without sacrificing any rents from her locked-in consumers. The discriminating firm would
get the same profit from the mobile consumers as Firm 2 in our duopoly model with C2 = 0.
The firm’s total profit would be that plus the monopoly rents from it’s locked-in consumers.
As long as the other firm had some locked-in consumers, the discriminating firm could get a
positive gain from this.
The table below summarizes this. All payoffs have to multiplied by pMQM . This table
is not symmetric for the case where both firms cannot price-discriminate because Firm 2 is
able to use her lower opportunity cost for lowering prices to compete more agressively.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 74
F2 can discriminate F2 cannot discriminate
F1 can discriminate C1,C2 C1 + C2C1,2
C2+C1,2, C2
F1 cannot discriminate C1, C2 + C1C1,2
C1+C1,2C1,C1
C2+C1,2
C1+C1,2
We are not proposing the table above as a game that is played by the firms. Whether
price discrimination is possible or not, and how costly it would be to implement, depends on
the application at hand. For example, in the bridge example it might not be cost-effective to
identify the geographical origin when a large number of consumers is crossing at rush hour.
3.3 Which networks are likely to form?
Our analysis up to now has taken the network as given. Given that the network structure
can have a big influence on welfare it is important to know which networks are more likely
to arise. To do so we will proceed in two different ways.
We will first ignore the precise protocol of how links are formed and and look for a
solution from cooperative game theory. We will look for networks that are pairwise-stable
with respect to the payoffs induced by the pricing game. We find that stable networks are
always between the two extreme cases of networks with local monopolists and networks with
perfect competition.
We will next analyze two “entry” games where we precisely specify how and when firms
can form links. These games will consist of two stages, an entry stage and a pricing stage. The
network will be determined at the entry stage after which the pricing stage will start treating
the network as given. The pricing stage will correspond exactly to our pricing model. It will
turn out that the equilibria of the entry game are a subset of the pairwise-stable networks.
Entry games are more than a device to understand which networks are likely to form.
They are interesting in their own right and have a long tradition in game-theoretic models
of industrial organization.5 Their aim is to understand under what circumstances and by
what actions, if any, can an incumbent firm credibly deter competitors from entering their
market.
The entry games considered here are different from the traditional ones in that entry
can be partial: we allow firms to establish links with some consumers without having allow
access to all consumers in the incumbent’s niche.5Tirole (1988) is the required reference.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 75
3.3.1 Pairwise-stable networks
Fix the total number of consumers,6 for a given network we ask if a firm has an incentive to
add or delete links to the existing consumers. By adding and deleting links, firms change the
network. The payoff for a firm from a network will be her expected payoff of the pricing game
that would be carried out under such a network. When we say firms have incentives to add
or delete links, we are doing a comparative statics across the payoffs from different networks
in our pricing model. There is no cost to forming links. Unlinked firms and consumers get
a payoff of zero. We want to characterize all the networks that are pairwise-stable.
Definition 25 (pairwise-stable networks). A network is pairwise-stable if :
• No unlinked firm-consumer pair can become linked and increase the payoff of both
members.
• No firm or consumer can delete one of their links and increase their individual payoff.
The Competitive Network of the tollway example we analyzed in Section 3.2.1 was not
pairwise-stable. Any of the two firm could increase their profit by deleting a link. Both
the Monopolistic Network and the Mixed Network were pairwise stable, even though Firm
2 received a strictly smaller payoff in the later. This because Firm 1 has no strict incentive
to add or remove the link that connect it to Consumer B.
Only the firms’ incentives determine if a network is pairwise-stable because, as we saw
in the welfare section (Section 3.2.3), adding a link between any firm and any consumer
increases the expected payoffs of all consumers.
Firms would never want to delete links to their locked-in consumers and would always
want to add a link to consumers that are not linked to the other firm.
Firm 1 is always indifferent between adding links or not to consumers that are linked to
Firm 2. These would only increase C1,2 and decrease C2 by one, but this does not affect her
equilibrium payoffs.
Firm 2 would have an incentive to add a link to a consumer of type C1 if and only if:
C2 + C1,2 + 1 < C1
That is, only if her potential consumer once the link is added is not larger than Firm 1’s
locked-in market.6This is one of the only parts of the paper where the fact that the number of consumers are integers
matters. Multiplying the number of all consumers by a constant does not change the the strategic incentives.Thus allowing the number of consumers to be a continuous variable can be thought as an approximation tolarge numbers of consumers with some adequate rescaling.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 76
Except for integer constraints, Firm 2 would want to increase links at low levels of C1,2
but would want to decrease them at high levels. (As long as C2 < C1) In this sense, pairwise-
stable networks are bounded away from the networks with local monopolists and from those
with perfect competition.
3.3.2 Entry game 1
Suppose Firm 1 is the incumbent and initially has links with all potential consumers. Firm
2 enters by forming links with the consumers. She can decide to form any number of links.
After Firm 2 makes her decision, the network is fixed and the pricing game is played out.
How many links would Firm 2 wish to form? What would be the final network formed?
Firm 2 solves:
max0�E�C1
(C1 − E)pMQM E
(C1 − E) + E
Where E is the level of entry measured by the number of links formed to enter.
Using the results in Section 3.3.1 we know that Firm 2 would continue to add links while
the following condition holds:
E + 1 < C1
Therefore Firm 2 would stop forming links much before she takes away all of Firm 1’s
locked-in market, to keep Firm 1 from being too aggressive in the subsequent pricing game.
Roughly speaking she would only enter half of the potential market.
A similar effect has been previously shown in the entry games literature. It was labeled
by Fudenberg and Tirole (1984) as the “fat cat effect”. The “fat cat effect” refers to a
situation of strategic over-investment in capital by the incumbent firm to act less agressive
conditional on an entry she knows she cannot deter. The effect is also present in the multi-
market oligopoly model of Bulow et al. (1985).
Our focus on the entry stage is different than the previous models. There the incumbent
chose to be a “fat cat” by over-investing in capital. In our model it’s the entrant who choses
to keep the incumbent as a “fat cat” by only entering partially.
Would Firm 1, given a chance, wish to delete links to her initial consumers to accomodate
Firm 2 in better terms? Could she perhaps even deter entry by sustaining a “lean and
hungry” stance described by Fudenberg and Tirole (1984)? The answer is negative. If Firm
preemptively broke some links, Firm 2 would form links to any unlinked consumer and then
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 77
continue to form links while the condition holds:
C2 + C1,2 + 1 < C1
The final amount of C1 would remain unchanged. 7 Since Firm 1’s payoff only depends
on this amount, she would never wish to break links to accomodate Firm 2.
3.3.3 Entry game 2
Suppose now that initially Firm 1 and Firm 2 have two separate monopolies. That is, at
the initial stage there are only locked-in consumers. We keep the convention that Firm 1
starts out as the dominant firm: C1 > C2. The game starts with the the entry stage where
firms are allowed to simultaneously form links with consumers in the other firms locked-in
market. Once all link are formed the network is fixed and the pricing game is carried out.
This could be the case of two separate regional monopolies held in place by a restriction
to sell across borders. The entry stage occurs after a free trade agreement lifts restrictions
across borders, but firms still have to decide in which locations across the border they want
to set-up a point of sale. It could also represent two similar products that in the pre-entry
stage could not reach the same consumers because of regulation or technological restrictions.
This happened in the phone and cable service markets where technological advances allowed
cable providers to supply phone services and viceversa, although companies still had the
ability to decide in which regions they would operate.
In the entry stage firms simultaneously announce the number of links they will form with
the other firm’s locked-in consumers. This announcement is their level of entry. We seek to
find how many links would each firm wish to form in a Nash equilibrium.
Proposition 26. The set of Nash Equilibria of the Entry Game 2 is:
E∗1 ∈ Integer
�[0, C2]
�; E∗
2 = Integer�(C1−C2
2 − 1, C1−C22 ]
�
Where the Integer function maps intervals to the set of integers inside that interval.
Proof. Their profits from a strategy profile (E1, E2) are π1(E1, E2) as described below.
π1(E1, E2)
pMQM=
�(C1 − E2); if C2 − E1 � C1 − E2
(C2 − E1)(C1+E1)(C2+E2) ; if C2 − E1 > C1 − E2
7Unless Firm 1 broke links with more than half her locked-in consumers in which case she would have aneven smaller payoff!
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 78
π2(E1, E2)
pMQM=
�(C1 − E2)
(C2+E2)(C1+E1) ; if C2 − E1 � C1 − E2
(C2 − E1) if C2 − E1 > C1 − E2
First we rule out any strategy profile that makes Firm 2 the dominant firm after entry.
Take any strategy profile such that:
C2 − E1 > C1 − E2
Firm 1’s best response would be to set her entry level as low as possible: E1 = 0. For this
strategy Firm 2’s best response is Integer[C1−C22 ]. This level of entry is not enough to make
Firm 2 the dominant firm after entry. Therefore there are no equilibria where this happens.
If after entry Firm 1 is still going to be the dominant firm, then Firm 2’s unique best
response is:
E∗2 =
�(C1 − C2
2− 1,
C1 − C2
2]�
For this profile Firm 1’s best response set is:
E∗1 ∈ Integer
�[0, C2]
�
In equilibrium Firm 2 always enters partially. The level of entry is always low enough
such that Firm 1 still remains the dominant firm after entry. Firm 1 has multiple best
responses but has no strict incentive to enter Firm 2’s locked-in market. Doing so does not
alter her payoff nor alter Firm 2’s preferences over entry levels. Adding an arbitrarily small
cost to forming links would reduce the set of equilibria to a unique strategy profile where
only Firm 2 enters as described above.8
Consumer surplus and aggregate surplus strictly increase with the partial entry.
In equilibrium the Firms have no incentive to change they’re links. Because only firms’
incentives matter to determine which networks are pairwise stable, equilibrium networks of
the entry games must necessarily be pairwise stable.
8Except for the case where C1−C22 is an integer. In that case adding an arbitrarily small cost would
diminish Firm 2’s entry level by one link.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 79
3.4 Extending the results for oligopolistic competition
For applications it would be useful to extend the model to allow for an arbitrary number
of firms. Unfortunately it quickly becomes intractable. The potential types of consumers
include all possible subsets of firms. These grow exponentially.
Nevertheless, we can solve a special case of oligopolistic competition where consumers
come in two types: locked-in consumers who can only buy from one firm and perfectly mobile
consumers who can buy from all firms. We will find that most of the pricing behavior and
payoffs extend to this model.
There are several applications that fit this type of environment. For example, a situ-
ation where some people are already locked into a specific technology while other people
are waiting for the prices to be determined before they decide which technology to adopt.
Small fixed-costs for adopting technology could be easily included in the model. Another
pertinent environment is one where some consumers buy online and can see the all the price
information, while others just go to their local store and cannot react to prices from other
firms. A third application would treat the network as a model of brand loyalty. Firms face
a trade-off between tendering to completely loyal consumers or competing for extreme price
seeking consumers.
In this model the pricing action will happen between the most agressive firms, those
with the lowest opportunity cost for lowering their prices. All other firms will give up on
capturing the mobile market and stay out by pricing the monopoly price.
We solve this model in Section 3.4.1. We will discuss what we know and what we don’t
for the general oligopolistic model in Section 3.4.2.
3.4.1 An arbitrary number of firms competing in a single market
The model we are considering has a finite number J of firms each with some of locked-in
consumers.9 We label firms such that:
C1 � C2 � . . . � CJ > 0
In addition to their locked-in consumers, all firms can compete to sell in a large global
market that has a number CG of consumers. Consumer values and the pricing game are as
before.9Our main results for strategies and payoffs are still true when some firms might not have any locked-in
consumers. We rule this out for ease of exposition.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 80
There is always an equilibrium where Firms J and J−1 play as described in the duopoly
model while all other firms price pM . We refer to this pricing behavior as “staying out”
because the equilibrium probability that they will capture the global market is zero. Propo-
sition (27) shows proves that this is an equilibrium.
The intuition behind the proof is that firms face a trade-off between extracting all the
surplus from their locked-in consumers or lowering their price to try to capture to the global
consumers. There are two effects that keep firms with more locked-in consumers out. They
face a higher opportunity cost from lowering prices; and they to face an fiercer competition
if they enter because to win the global market they have to quote the minimum price of ALL
firms going in. This two forces move in the right direction to sustain the equilibrium.
These strategies turns out to be the unique equilibrium when we assume that only two
firms have the have the smallest number of consumers. That is, when CJ−2 > CJ−1. This
is proven in Proposition (28). Uniqueness follows from the fact the is a price interval where
firms J and J − 1 must necessarily mix as in the duopoly model. But higher firms cannot
be made indifferent between quoting the monopoly price and a lower one. This is proven in
Proposition 28.
Proposition 27. The following strategies always constitute an equilibrium of the oligopoly
model:
σJ(p) = 1− C1
C1,2
�1− pMQM
pQ(p)
�
σJ−1(p) =
�CJ+CG
CJ−1+CG
�1− C1
C1,2
�1− pMQM
pQ(p)
��for p < pM
1 for p = pM
The strategies for j > J − 1 are: Quote pj = pM with probability 1.
Firms J − 1 and J play as in the duopoly model, all other firms just quote the monopoly
price.
Proof. See Proof Section (3.4.2).
Proposition 28. The strategies in Proposition (27) constitute the unique equilibrium strat-
egy profile if CJ−2 > CJ−1.
Proof. See Proof Section (3.4.2).
The additional assumption in Proposition (28) is necessary for uniqueness. For example,
if it fails we can construct symmetric equilibria where all the firms tied at the bottom
enter. Nevertheless all equilibria are payoff equivalent and only firms with the two lowest Cj
parameter can actively mix.
CHAPTER 3. PRICE COMPETITION ON A BUYER-SELLER NETWORK. 81
3.4.2 Remarks on the general oligopoly model
What can we say for networks with any number of consumers who can arbitrarily buy from
any subset of the firms?
Existence of equilibrium is guaranteed by the main the existence theorem for discontin-
uous games in Dasgupta and Maskin (1986).10 Broadly, for the theorem to hold we require
the following:
• The set of pure strategies be a compact set of Rm. We fulfill this by restricting
attention to the game with prices between zero and the monopolist price, because prices
above are weakly dominated for all firms. Any equilibrium of the restricted game will
be an equilibrium of the unrestricted game.
• Payoff functions must be continuous except for subset of a continuous man-
ifold of dimension smaller than the strategy space. This continuous manifold
can be defined in our model as follows. For each j ∈ 1, . . . , J define: