Networks Economics

Francis Bloch, Margherita Comola, Gabrielle Demange,

.
today Gabrielle Demange

Paris School of Economics

September 18, 2017

Gabrielle Demange (PSE-EHESS) Networks 1 / 37

General Outline

Introduction, Background definitions on networks GD 18 Sept

Network formation GD 25 Sept
random: Erdos Renyi
strategic: pair-wise Stability
Intro to econometric methods on network formation, MC 3 Oct
Measuring influence of the nodes according to the network structure
GD 9 Oct
Risk sharing FB 16 Oct
Games on networks (1) -23 Oct GD
coordination games, games with complementarities

Gabrielle Demange (PSE-EHESS) Networks 1 / 37

 Communication-diffusion Nov 6 GD
Exchange of opinions : de Groot learning model and its developments
(non strategic), strategic models
Experimental models of diffusion MC Nov 14
IO FB Nov 20
Games on networks (2) - Nov 27 GD
Experiments MC 5 Dec
Targeting FB 11 Dec

Gabrielle Demange (PSE-EHESS) Networks 2 / 37

 Grades The grade will be composed of
(1) your homework and attendance (50 %)
(2) an exam (50 %)
Office hours: after the lecture.
All the information is on this Website.
When you are not present at a lecture, ask another student for
information.
Homeworks: Type the answers and print them. Do not send me pdf.
Answers are due on time. If you are not present at the due time, pass
your answers to another student.

Gabrielle Demange (PSE-EHESS) Networks 3 / 37

Outline

1 Introduction

2 Fundamentals notions
Graphs
Degrees, Walks, Components
Diameter, average length path

3 Statistical approach
Regular tree, Erdos-Renyi
Clustering/homophily
The degree distribution
Thresholds and phase transition

Gabrielle Demange (PSE-EHESS) Networks 4 / 37

 Introduction

What are networks?

In social and economics areas, networks represent interactions among
units (individuals, firms)
friendship, family
production networks, financial networks
Web
First analysis: in sociology
Introduces statistics on a network, measure of positions in the
networks, importance of the structure ( ’weak’ ties versus strong ties)
Biology, to analyse spead of epidemics
Later in statistical physics and computer science in particular to study
large networks, diffusion processes

Gabrielle Demange (PSE-EHESS) Networks 5 / 37

 Introduction

Questions in economics

Interactions though networks are pervasive

trade: most trades are not conducted in market places
risk sharing (in countries under development)
information sharing: opinions, jobs, rumors
trust, favors, boardrooms, crimes
How do networks form, how do they influence behavior?
Bilateral links are at the root of the interactions
External effects: what is the interplay between the ’local’ interactions
and the whole network structure

Gabrielle Demange (PSE-EHESS) Networks 6 / 37

 Introduction

Questions in economics

Economic networks often exhibit specific structures

Look at some basic statistics.
Try to explain link formation through simple mechanisms
e.g. game theory approach
N.B.: distinguish network formation and game on a network

Gabrielle Demange (PSE-EHESS) Networks 7 / 37

 Fundamentals notions Graphs

Description

(N, G )
N = {1, · · · , n} set of nodes
G =set of links (edges)
Link: a pair (i, j)
Matrix Representation
Incidence matrix g of a network:
gij = 0 if (i, j) is not a link
gij = 1 if (i, j) is a link
Undirected : g is symmetric
ex.: friendship
Directed
Influence, referral : the link (i, j) if i influences j (can choose the
opposite convention)
Web network: i points to j
Gabrielle Demange (PSE-EHESS) Networks 8 / 37
 Fundamentals notions Graphs

Weighted networks

In some contexts,
P deflate by the number of links from i:
divide gij by k gik (i’s ’Out-degree’ ):
g
P ij
k gik

ex.: time spent as a function of the number of friends, Web

More generally: valued network → matrix with ≥ 0 elements

Gabrielle Demange (PSE-EHESS) Networks 9 / 37

 Fundamentals notions Graphs

Example: Trade

Gabrielle Demange (PSE-EHESS) Networks 10 / 37

 Fundamentals notions Graphs

Graph/matrix

A graph is represented by a Matrix. When and why is it useful?

Useful if the links (friend or not) are binary and rare
Or specific structure: hierarchies, star, ’hubs’
A matrix is more general: can be weighted

Gabrielle Demange (PSE-EHESS) Networks 11 / 37

 Fundamentals notions Graphs

Graphical representation

Technically difficult for large-scale networks

Non-neutral
May sometimes include choices (threshold)

Gabrielle Demange (PSE-EHESS) Networks 12 / 37

 Fundamentals notions Degrees, Walks, Components

Degree

Degree: number of neighbors

P
i’s Out-degreeP: j gij
i’s In-degree j gji
Web network:
i’s In-Degree=number of links pointing to i
i’s Out-Degree=number of links pointing from i
Degree is a local measure: Does not account for the whole network,
i.e. for indirect relationships 6= ’prestige/power’ ( studied later on)

Gabrielle Demange (PSE-EHESS) Networks 13 / 37

 Fundamentals notions Degrees, Walks, Components

Degrees are not sufficient (see later)

Gabrielle Demange (PSE-EHESS) Networks 14 / 37

 Fundamentals notions Degrees, Walks, Components

Walk/path

Walk: sequence of successive nodes related by a link

i1 , i2 , · · · , ik+1 such that i`−1 , i` in G
Path: all nodes are distinct; length k
Cycle:a walk with i1 = ik
geodesic: shortest path between two nodes

Gabrielle Demange (PSE-EHESS) Networks 15 / 37

 Fundamentals notions Degrees, Walks, Components

Counting Walks

g 1 (i)= number of links from i= i’s out-degree

g k 1 (i)= number of (directed) walks from i of length k
Status, prestige, vote:
Better to influence a neighbor who influences many neighbors
→ network matters

Gabrielle Demange (PSE-EHESS) Networks 16 / 37

 Fundamentals notions Degrees, Walks, Components

Components

(N, G ) is connected if there is a path between any two nodes

(the matrix g is irreducible)
Component: maximal connected subgraph
Only components matter for the interactions

Gabrielle Demange (PSE-EHESS) Networks 17 / 37

 Fundamentals notions Degrees, Walks, Components

Example 1. Illustrate components

Gabrielle Demange (PSE-EHESS) Networks 18 / 37

 Fundamentals notions Diameter, average length path

Diameter, average length path

Diameter: largest geodesic

(infinite if g unconnected)
Average length path (in a component)
depends less on outliers

Gabrielle Demange (PSE-EHESS) Networks 19 / 37

 Fundamentals notions Diameter, average length path

Small world

When any two units are linked with a ’small’ number of links
Milgram experience (1967): contact ’unknown’ targets
median ’5.5’ for the successful
called ’six degrees of separation
On the Web: 11 Albert, Jeong, Barabási (1999)
Co-authorships
in Economics: 9,5 Goyal et al (2004)

Gabrielle Demange (PSE-EHESS) Networks 20 / 37

 Fundamentals notions Diameter, average length path

How to interpret these results?

Average length path and diameter are often difficult to compute

Compute them in two benchmark models:
1 Regular tree: same degree for each node except for the leaves (Caylee)
2 Erdos Renyi random model
links drawn at random, independently across pairs

Gabrielle Demange (PSE-EHESS) Networks 21 / 37

 Statistical approach Regular tree, Erdos-Renyi

Diameter in a regular tree

Degree d except for the leaves

Reach d(d − 1)k−1 nodes in k steps
Total in at most k steps : d + · · · + d(d − 1)k−1 ≈ (d − 1)k
To reach the whole set, need roughly ` with n = (d − 1)`

log (n)
`≈
logd
` increases rather fast with n:
Exemple: Compute d for ` = 6 and n = 6.7 million
no small world

Gabrielle Demange (PSE-EHESS) Networks 22 / 37

 Statistical approach Regular tree, Erdos-Renyi

Erdos Renyi model

p = probability of a link (i, j), independent across pairs

G (n, p)= family of graphs
The degree distribution is binomial
Average degree : d = (n − 1)p
Large n, small p, approximated by a Poisson distribution with
(n − 1)p = d
dk
proba of degree k → e −d
k!

Gabrielle Demange (PSE-EHESS) Networks 23 / 37

 Statistical approach Regular tree, Erdos-Renyi

Erdos Renyi. A draw

Gabrielle Demange (PSE-EHESS) Networks 24 / 37

 Statistical approach Regular tree, Erdos-Renyi

Erdos Renyi. A draw

Gabrielle Demange (PSE-EHESS) Networks 25 / 37

 Statistical approach Regular tree, Erdos-Renyi

Random graph

Cycles, triangles : no exact computation for the length

But the importance of cycles tends to vanish:
one can show
log (n)
`≈
logEd
n = 6.7 billion, d = 50
log (n)
`= ≈6
logd

Gabrielle Demange (PSE-EHESS) Networks 26 / 37

 Statistical approach Clustering/homophily

Clustering coefficient/triangles

i’s clustering coefficient: fraction of i’s friends who are friends

number of i’s friends who are friends

number of pairs of i’s friends
Average over i:
Related to number of triangles
Trees: no triangles

Gabrielle Demange (PSE-EHESS) Networks 27 / 37

 Statistical approach Clustering/homophily

Clustering-cd

Random Erdos Renyi network:

the fact that two individuals are i’s friends is independent of their
own friendship
→ The clustering coefficient becomes small with n (because p
becomes small)
Violated in most social networks
Homophily: the tendency of individuals to be linked to other
individuals with similar characteristics
→ introduces dependance in links’ formation
Increases the number of triangles (and more generally of clusters)

Gabrielle Demange (PSE-EHESS) Networks 28 / 37

 Statistical approach Clustering/homophily

Clustering/Homophily

Important for dissemination

Effects in different directions depending on the process of influence
Extreme homophily induces separate components
How to define a ’homophily’ index? see exercise

Gabrielle Demange (PSE-EHESS) Networks 29 / 37

 Statistical approach The degree distribution

The degree distribution

More informative than the average degree

Large networks, ex: Web.
Compare with a random model with the same expected degree

Gabrielle Demange (PSE-EHESS) Networks 30 / 37

 Statistical approach The degree distribution

Web

Albert, Jeong, Barabási (1999)

Gabrielle Demange (PSE-EHESS) Networks 31 / 37

 Statistical approach The degree distribution

Fat tails

More nodes with high degree than predicted by a fitted random model
True for many social networks: co-authorships, citations...
True in other contexts -city size, wealth- as first identified by Pareto
(1896)
Described by Power law/Scale free distributions

Gabrielle Demange (PSE-EHESS) Networks 32 / 37

 Statistical approach The degree distribution

Power law/Scale free

For high enough k

P(k)= proba of degree k = ck −α for positive α, c constant
→ Linearity in a log-log diagram (see www)
P(mk)
Scale free: ratio P(k) is independent of k
For a random model the ratio goes to zero with k: decay is
exponential

Gabrielle Demange (PSE-EHESS) Networks 33 / 37

 Statistical approach The degree distribution

Web

Gabrielle Demange (PSE-EHESS) Networks 34 / 37

 Statistical approach Thresholds and phase transition

Diffusion: Importance of the network

Disease spreads via connections in the network

Nodes are linked if one would ’infect” the other
How many people will it reach ?

Gabrielle Demange (PSE-EHESS) Networks 35 / 37

 Statistical approach Thresholds and phase transition

Thresholds and phase transition

Consider sequence of graphs with increasing number of nodes n.

If np < 1 , then a graph in G(n, p) will almost surely have no
connected components of size larger than O(log (n)) as n → ∞.
If np → d > 1, then a graph in G(n, p) will almost surely have a
unique giant component containing a positive fraction of the vertices
as n → ∞. No other component will contain more than O(log (n))
vertices.
A phase transition occurs at t(n) = 1/n

Gabrielle Demange (PSE-EHESS) Networks 36 / 37

 Statistical approach Thresholds and phase transition

Connectedness

ln n
n is a sharp threshold for the connectedness of G (n, p):
If p < (1−)
n
ln n
, then a graph in G (n, p) will almost surely contain
isolated vertices, and thus be disconnected.
If p > (1+)
n
ln n
, then a graph in G (n, p) will almost surely be
connected.

Gabrielle Demange (PSE-EHESS) Networks 37 / 37