04 Diffusion and Peer Influence (2016)

“you are who you associate with”
Time
Diffusion & Peer Influence

1. Diffusion
A. Compartmental Models
B. Network Diffusion
i. Topology
ii. Timing
iii. Structural Transmission
a. Complex contagion
2. Peer Influence

Coleman, Katz and Menzel,
“Diffusion of an innovation among
physicians” Sociometry (1957)
Our substantive interest in networks
is often in how things move through
them, from disease to ideas to
behavior.
Network Diffusion & Peer Influence
Basics

Figure 2. Binge Drinking
Predicted Probabilities by
Gender and Friends’ Prior
Drinking.
Derek A. Kreager, and Dana L. Haynie American
Sociological Review 2011;76:737-763
Romantic partnerships
in high school lead to
adoption of partners’
friends’ behaviors
Basics

Table 4. Proportion of Ties Created or Maintained Over Time by Network Process and
Depression Level.
David R. Schaefer et al. American Sociological Review
2011;76:764-785
Health effects on ties
Depressed students form ties through non-normative network processes

Basics
Classic (disease) diffusion makes use of compartmental models. Large N and homogenous
mixing allows one to express spread as generalized probability models.
Works very well for highly infectious bits in large populations…
SI(S) model – actors are in only two
states, susceptible or infectious.
See: https://wiki.eclipse.org/Introduction_to_Compartment_Models for general introduction.
SIIR(S) model – adds an “exposed” but
not infectious state and recovered.

Basics – might even help understand the zombie apocalypse
http://loe.org/images/content/091023/Zombie%20Publication.pdf

Basics
Network Models
Same basic SI(R,Z,etc) setup, but connectivity is not assumed random, rather it is structured by
the network contact pattern.
If pij is small or the network is very clustered, these two can yield very different diffusion
patterns.*
Real Random
*these conditions do matter. Compartmental models work surprisingly well if the network is large, dense or the bit highly
infectiousness…because most networks have a bit of randomness in them. We are focusing on the elements that are unique/different for
network as opposed to general diffusion.

Basics
If 0 < pij < 1

Basics
If 0 < pij < 1
0.01
0.06
0.11
0.26
0.46

In addition to* the dyadic probability that one actor passes something to another
(pij), two factors affect flow through a network:
Topology
- the shape, or form, of the network
- Example: one actor cannot pass information to another unless they are
either directly or indirectly connected
Time
- the timing of contact matters
- Example: an actor cannot pass information he has not receive yet
*This is a big conditional! – lots of work on how the dyadic transmission rate may differ across
populations.
Key Question: What features of a network contribute most to diffusion potential?
Network diffusion features
Use simulation tools to explore the relative effects of structural
connectivity features

• A network has to be connected for a bit to pass over it
• If transmission is uncertain, the longer the distance the lower the likelihood of spread.
0
0.2
0.4
2 3 4 5 6
Path distance
probability
Distance and diffusion (p(transfer)=pij
dist
Here pij of 0.6
We need:
(1) reachability
(2) distance
(3) local clustering
(4) multiple routes
(5) star spreaders

• Local clustering turns flow “in” on a potential transmission tree
Arcs: 11
Largest component: 12,
Clustering: 0
Arcs: 11
Largest component: 8,
Clustering: 0.205
We need:
(1) reachability
(2) distance
(4) multiple routes
(5) star spreaders

• The more alternate routes one has for transmission, the more likely flow should be.
• Operationalize alternate routes with structural cohesion
We need:
(1) reachability
(2) distance
(4) multiple routes
(5) star spreaders

Probability of transfer
by distance and number of non-overlapping paths, assume a constant pij of 0.6
0
0.2
0.4
0.6
0.8
1
1.2
2 3 4 5 6
Path distance
probability
1 path
C
P
X Y
10 paths
5 paths
2 paths
Cohesion  Redundancy Diffusion

0 1 2 3
Node Connectivity
As number of node-independent paths
C
P
X Y
Structural Cohesion:
A network’s structural cohesion is equal to the minimum number of
actors who, if removed from the network, would disconnect it.

STD Transmission danger: sex or drugs?
Structural core more realistic than nominal core
C
P
X Y
Data from “Project 90,” of a high-risk population in Colorado Springs

• Much of the work on “core groups” or “at risk” populations focus on high-degree nodes. The
assumption is that high-degree nodes are likely to contact lots of people.
We need:
(1) reachability
(2) distance
(4) multiple routes
(5) star spreaders

Assortative mixing:
A more traditional way to think about “star” effects.

• Simulation study: How do these different features compare over a collection of observed nets?
• For each network trial:
• Fix dyadic transmission probability
• Randomly select a node as seed
• Trace the diffusion path across the network
• Measure speed & extent of spread
• Model extent of spread by structural characteristics
• First run: Add Health:
• simple diffusion process
• dyadic probability set to 0.08
• 500 trials in each network.
• Then expand to :
• Different assumptions of dyadic transmission probability
• Different data set (Facebook)
• Complex diffusion models
Network diffusion features: simulation test

Define as a general measure of the “diffusion susceptibility” of a graph’s structure as the
ratio of the area under the observed curve to the area under the curve for a matching
random network. As this gets smaller than 1.0, you get effectively slower median
transmission.

Table 2. OLS Regression of Relative Diffusion Ratio on Network Structure
Variable Model 1 Model 2 Model 3 Model 4 Model 5
Intercept 1.62***
1.90***
1.02***
1.81***
1.71***
Connectivity
Distance -0.207***
-0.179***
-0.171***
Independent Paths -0.077***
-0.056***
-0.052***
Distance x Paths 0.023***
0.015***
0.016***
Clustering
Clustering Coefficient -0.692***
-0.653***
-0.454***
Grade Homophily -0.026**
-0.007 -0.009*
Peer Group Strength -0.868***
-0.141 -0.146
Degree Distribution
Degree Skew -0.023 -0.007 -0.002
Assortative Mixing -0.189*
-0.059 -0.071
Control Variables
Network Size/100 0.005***
-0.005***
-.005***
0.004*
0.002**
Proportion Isolated -0.007 -1.106***
-.984***
-0.300*
0.058
Non-Complete -0.006 -0.052*
-.078**
-0.006 0.018
Adj- R2
0.85 0.76 0.60 0.90 0.93
N 124 124 124 124 121

Figure 4. Relative Diffusion Ratio
By Distance and Number of Independent Paths
0.4
0.6
0.8
1
1.2
2.3 2.8 3.3 3.8 4.3 4.8 5.3 5.8 6.3
Average Path Length
Observed/Random
k=2
k=4
k=6
k=8
Figure 4. Relative Diffusion Ratio
By Distance and Number of Independent Paths
0.4
0.6
0.8
1
1.2
2.3 2.8 3.3 3.8 4.3 4.8 5.3 5.8 6.3
Average Path Length
Observed/Random
k=2
k=4
k=6
k=8

Traditional “core group” models have a local-vision understanding of risk: those
with lots of ties (high degree) are the focus for intervention and actions.
•In the short time-windows necessary for STD transfer, low-degree networks are
the relevant features for transmission. What sorts of networks emerge when
average degree (in the short run) is held to small numbers?
•How does the shape of the degree distribution matter? If activity is homogeneous
do we get fundamentally different networks than if it is very heterogeneous?
A closer look at emerging connectivity

Partner
Distribution
Component
Size/Shape
Emergent Connectivity in low-degree networks

In both distributions, a giant
component & reconnected core
emerges as density increases, but
at very different speeds and
ultimate extent.

What distinguishes
these two distributions?
Shape

What distinguishes these two distributions?
Shape:
The scale-free network’s signature is the long-tail
So what effect does changes in the shape have on connectiv

 Volume 
DispersionxSkewness

Search Procedure:
1) Identify all valid degree distributions with
the given mean degree and a maximum of
6 w. brute force search.
2) Map them to this space
3) Simulate networks each degree distribution
4) Measure size of components &
Bicomponents

Based on work supported by R21-HD072810 (NICHD, Moody PI), R01 HD068523-01 (NICHD, Moody PI), R01 DA012831-05 (NIDA Morris, Martina PI),

C:45%, B: 8.5%

C:83%, B: 36%

C:99%, B: 86%

Largest Component
(at least 1 path)
Largest Bicomponent
(at least 2 paths)

In addition to* the dyadic probability that one actor passes something to another
(pij), two factors affect flow through a network:
Topology
- the shape, or form, of the network
- Example: one actor cannot pass information to another unless they are
either directly or indirectly connected
Time
- the timing of contact matters
- Example: an actor cannot pass information he has not receive yet
*This is a big conditional! – lots of work on how the dyadic transmission rate may differ across
populations.
Key Question: What features of a network contribute most to diffusion potential?
Relational Dynamics
Use simulation tools to explore the relative effects of structural
connectivity features

Three relevant networks
Discussions of network effects on STD spread often speak loosely of “the network.”
There are three relevant networks that are often conflated:
1) The contact network. The set of pairs of people connected by sexual contact. G(V,E).
2) The exposure network. A subset of the edges in the contact network where timing
makes it possible for one person to pass infection to another.
3) The transmission network. The subset of the exposure network where disease is actually
passed. In most cases this is a tree layered on (2) and rooted on a source/seed node.
Relational Dynamics

Contact network: Everyone, it is a connected component
Who can “A” reach?
Relational Dynamics

Exposure network: here, node “A” could reach up to 8 others
Relational Dynamics

Transmission network: upper limit is 8 through the exposure links
(dark blue). Transmission is path dependent: if no transmission to
B, then also none to {K,L,O,J,M}
Exposable Link (from A’s p.o.v.)
Contact
Relational Dynamics

The mapping between the contact network and the exposure network is based on
relational timing. In a dynamic network, edge timing determines if something can flow
down a path because things can only be passed forward in time.
Definitions:
Two edges are adjacent if they share a node.
A path is a sequence of adjacent edges (E1, E2, …Ed).
A time-ordered path is a sequence of adjacent edges where, for each pair of
edges in the sequence, the start time Si is less than or equal to Ej S(E1) <
E(E2)
Adjacent edges are concurrent if they share a node and have start and end
dates that overlap. This occurs if:
S(E2) < E(E1)
Concurrency
Relational Dynamics

A B C
D
time 1 2 3 4 5 6 7 8 9 10
AB
BC
CE
E
CD
2 - 71 - 3
S(ab) E(ab)
S(bc)
E(bc)
S(ce) E(ce)
The mapping between the contact network and the exposure network is based on
relational timing. In a dynamic network, edge timing determines if something can flow
down a path because things can only be passed forward in time.
Concurrency
Relational Dynamics

The constraints of time-ordered paths change our understanding of the system structure
of the network. Paths make a network a system: linking actors together through indirect
connections. Relational timing changes how paths cumulate in networks.
Indirect connectivity is no longer transitive:
A B C D1 - 2 3 - 4 1 - 2
Here A can reach C, and C and reach D. But A cannot reach D (nor D A). Why?
Because any infection A passes to C would have happened after the relation between C
and D ended.
A B C D1 - 2 3 - 4 1 - 2
Relational Dynamics

Edge time structures are characterized by sequence, duration and overlap.
Paths between i and j, have length and duration, but these need not be
symmetric even if the constituent edges are symmetric.
Relational Dynamics

1 2 2 2 2 2 1
1 1 2 2 2 2 2
2 1 1 2 2 2 2
2 2 1 1 2 2 2
2 2 2 1 1 2 2
2 2 2 2 1 1 2
2 2 2 2 2 1 1
1 2 2 2 2 2 1
Implied Contact Network of 8 people in a ring
All relations Concurrent
Reachability = 1.0
Relational Dynamics

1 1 1 1 1 1 1
1 1 1 1 1 1 1
1 1 1 1 1 1 1
1 1 1 1 1 1
1 1 1 1 1
1 1 1 1
1 1 1
1 1
Serial Monogamy (1)
1
2
3
7
6
5
8
4
Reachability = 0.71
Relational Dynamics

1 1 1
1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1
1 1 1
1 1 1 1 1
1 1 1 1 1
Mixed Concurrent
2
2
1
1
2
2
3
3
Reachability = 0.57
Relational Dynamics

Serial Monogamy (3)
1 1 1
1 1 1
1 1 1
1 1 1
1 1 1
1 1 1
1 1 1
1 1 1
1
2
1
1
2
1
2
2
Reachability = 0.43
Relational Dynamics

1
2
1
1
2
1
2
2
Timing alone can change mean reachability from
1.0 when all ties are concurrent to 0.42.
In general, ignoring time order is equivalent to
assuming all relations occur simultaneously –
assumes perfect concurrency across all relations.
Relational Dynamics

A B C D E F
A 0 1 2 2 4 1
B 1 0 1 2 3 2
C 0 1 0 1 2 2
D 0 0 1 0 1 1
E 0 0 0 1 0 2
F 1 0 0 1 0 0
While a is 2 steps from d, and d is 1 step from e, a and e
are 4 steps apart.
This is because the shorter path from a to e emerges after
the path from d to e ended.
4
2
1
Path distances no longer simply add
Relational Dynamics

A
B C
D
3 - 4
E
The geodesic from A to D is AE, ED and is two steps long.
But the fastest path would be AB, BC, CD, which while 3 steps long
could get there by time 5 compared to time 7.
I ignore this feature for the remainder…
Aside: “shortest” can now be replaced with “fastest”
Relational Dynamics

Concurrency affects exposure by making paths symmetric, which increases exposure down
multiple “branches” of a contact sequence.
Consider a simplified example:
All edges concurrent
a b c d e f g h i j k l m
a . 1 1 1 1 1 1 1 1 1 1 1 1
b 1 . 1 1 1 1 1 1 1 1 1 1 1
c 1 1 . 1 1 1 1 1 1 1 1 1 1
d 1 1 1 . 1 1 1 1 1 1 1 1 1
e 1 1 1 1 . 1 1 1 1 1 1 1 1
f 1 1 1 1 1 . 1 1 1 1 1 1 1
g 1 1 1 1 1 1 . 1 1 1 1 1 1
h 1 1 1 1 1 1 1 . 1 1 1 1 1
i 1 1 1 1 1 1 1 1 . 1 1 1 1
j 1 1 1 1 1 1 1 1 1 . 1 1 1
k 1 1 1 1 1 1 1 1 1 1 . 1 1
l 1 1 1 1 1 1 1 1 1 1 1 . 1
m 1 1 1 1 1 1 1 1 1 1 1 1 .
Relational Dynamics

All edges except gd concurrent
a . 1 1 1 1 1 1 1 1 1 1 1 1
b 1 . 1 1 1 1 1 1 1 1 1 1 1
c 1 1 . 1 1 1 1 1 1 1 1 1 1
d 1 1 1 . 1 1 1 1 1 1 1 1 1
e 1 1 1 1 . 1 1 1 1 1 1 1 1
f 1 1 1 1 1 . 1 1 1 1 1 1 1
g 0 0 0 1 0 0 . 1 1 1 1 1 1
h 0 0 0 1 0 0 1 . 1 1 1 1 1
i 0 0 0 1 0 0 1 1 . 1 1 1 1
j 0 0 0 1 0 0 1 1 1 . 1 1 1
k 0 0 0 1 0 0 1 1 1 1 . 1 1
l 0 0 0 1 0 0 1 1 1 1 1 . 1
m 0 0 0 1 0 0 1 1 1 1 1 1 .
Concurrency affects exposure by making paths symmetric, which increases exposure down
multiple “branches” of a contact sequence.
Relational Dynamics

Concurrency affects exposure by making paths symmetric, which increases
exposure down multiple “branches” of a contact sequence.
All edges gd after all others:
a . 1 1 1 1 1 1 0 0 0 0 0 0
b 1 . 1 1 1 1 1 0 0 0 0 0 0
c 1 1 . 1 1 1 1 0 0 0 0 0 0
d 1 1 1 . 1 1 1 0 0 0 0 0 0
e 1 1 1 1 . 1 1 0 0 0 0 0 0
f 1 1 1 1 1 . 1 0 0 0 0 0 0
g 0 0 0 1 0 0 . 1 1 1 1 1 1
h 0 0 0 0 0 0 1 . 1 1 1 1 1
i 0 0 0 0 0 0 1 1 . 1 1 1 1
j 0 0 0 0 0 0 1 1 1 . 1 1 1
k 0 0 0 0 0 0 1 1 1 1 . 1 1
l 0 0 0 0 0 0 1 1 1 1 1 . 1
m 0 0 0 0 0 0 1 1 1 1 1 1 .
The concurrency status of {dg}
determines which “side” of the graph
is exposed. Note this effect happens
at the system level – the correlation
between exposure and node-level
timing is essentially zero (d/g
excepted)
Relational Dynamics

Resulting infection trace from a simulation (Morris et al, AJPH 2010).
Observed infection paths from 10 seeds in
an STD simulation, edges coded for
concurrency status.
Relational Dynamics

Resulting infection trace from a simulation (Morris et al, AJPH 2010).
Relational Dynamics
Observed infection paths from 10 seeds in
an STD simulation, edges coded for
concurrency status.

Timing constrains potential diffusion paths in networks, since bits can flow through edges that have
ended.
This means that:
• Structural paths are not equivalent to the diffusion-relevant path set.
• Network distances don’t build on each other.
• Weakly connected components overlap without diffusion reaching across sets.
• Small changes in edge timing can have dramatic effects on overall diffusion
• Diffusion potential is maximized when edges are concurrent and minimized when they are
“inter-woven” to limit reachability.
Combined, this means that many of our standard path-based network measures will be incorrect on
dynamic graphs.
Relational Dynamics

The distribution of paths is important for many of the measures we typically construct on
networks, and these will be change if timing is taken into consideration:
Centrality:
Closeness centrality
Path Centrality
Information Centrality
Betweenness centrality
Network Topography
Clustering
Path Distance
Groups & Roles:
Correspondence between degree-based position and reach-based position
Structural Cohesion & Embeddedness
Opportunities for Time-based block-models (similar reachability profiles)
In general, any measures that take the systems nature of the graph into account will differ.
Structural measurement implications
Relational Dynamics – other implication of dynamic nets (briefly)

New versions of classic reachability measures:
1) Temporal reach: The ij cell = 1 if i can reach j through time.
2) Temporal geodesic: The ij cell equals the number of steps in the shortest path linking i to j
over time.
3) Temporal paths: The ij cell equals the number of time-ordered paths linking i to j.
These will only equal the standard versions when all ties are concurrent.
Duration explicit measures
4) Quickest path: The ij cell equals the shortest time within which i could reach j.
5) Earliest path: The ij cell equals the real-clock time when i could first reach j.
6) Latest path: The ij cell equals the real-clock time when i could last reach j.
7) Exposure duration: The ij cell equals the longest (shortest) interval of time over which i
could transfer a good to j.
Each of these also imply different types of “betweenness” roles for nodes or edges, such as a
“limiting time” edge, which would be the edge whose comparatively short duration places
the greatest limits on other paths.
Structural measurement implications
Relational Dynamics – other implication of dynamic nets (briefly)

Topology & Time interact: How relational sequencing affects diffusion is
conditioned by the structural patterns of relations.
Examples:
- Time limitations mean star nodes can’t interact with everyone at each
time tick; effects of high degree are thus limited by schedule/availabiltiy
- If within-cluster ties are also more frequent than between cluster ties,
then the effects of communities will be magnified.
-Multiple connectivity should provide routes around breaks built by
temporal sequence.
Structural Moderators of Timing Effects

Measures
Dependent variable: Reachability in the exposure graph. This is the proportion of
pairs in the network that are reachable in time.
Exposure Graph Density

Measures: Independent variables
Features of the topology. Of key interest is the level of structural cohesion.
1 2 3 4 5 6 7 8 9
1 -- 1 1 1 1 1 1 1 1
2 1 -- 2 2 2 1 1 1 1
3 1 2 -- 2 2 1 1 1 1
4 1 2 2 -- 2 1 1 1 1
5 1 2 2 2 -- 1 1 1 1
6 1 1 1 1 1 -- 2 2 1
7 1 1 1 1 1 2 -- 2 1
8 1 1 1 1 1 2 2 -- 1
9 1 1 1 1 1 1 1 1 --
Cell Value = highest k-connected
component pair belongs to.
Average = 1.25

1 2 3 4 5 6 7 8 9
1 -- 1 1 1 1 1 1 1 1
2 1 -- 3 3 3 2 2 2 1
3 1 3 -- 3 3 2 2 2 1
4 1 3 3 -- 3 2 2 2 1
5 1 3 3 3 -- 2 2 2 1
6 1 2 2 2 2 -- 2 2 1
7 1 2 2 2 2 2 -- 2 1
8 1 2 2 2 2 2 2 -- 1
9 1 1 1 1 1 1 1 1 --
Cell Value = highest k-connected
component pair belongs to.
Average = 1.6

Average connectivity

Volume Distance Connectivity
Nodes: 148
Mean Deg: 6.16
Density: 0.042
Centralization: 0.187
Nodes: 80
Mean Deg: 5.27
Density: 0.067
Nodes: 154
Mean Deg: 3.71
Density: 0.025
Nodes: 128
Mean Deg: 3.39
Density: 0.027
Mean: 3.59
Diameter: 5
Mean: 3.02
Diameter: 5
Mean: 4.99
Diameter: 8
Mean: 4.55
Diameter: 6
Largest BC: 0.51
Pairwise K: 1.57
Largest BC: 0.33
Pairwise K: 1.34
Largest BC: 0.08
Pairwise K: 1.07
Largest BC:
Pairwise K: 1.06
Exemplar independent variables
“HighCohesive”“LowCohesive”Network Diffusion & Peer Influence

“Low Cohesive”

Proportion of relations concurrent
DensityoftheExposureNetwork

Colors=different
nets
Panels=cohesion
level
Proportion of relations concurrent
DensityoftheExposureNetwork

1. Concurrency has a necessarily positive effect on potential diffusion
exposure
1. This implies that we should see greater transmission given greater
concurrency
2. This works by creating “multiple routes” in the exposure path
structure
2. Structural cohesion captures multiple routes in the contact graph
1. Higher levels of cohesion increase exposure by directly increasing
the underlying transmission substrate
3. There is a negative interaction between cohesion and concurrency: as
cohesion increases, the relative returns to concurrency decrease.
1. But this comes at the cost of a higher base-level of exposure.

Structural Transmission Dynamics: beyond disease diffusion
Complex Contagion
Thus far we have focused on a “simple” dyadic diffusion parameter, pij, where the
probability of passing/receiving the bit is purely dependent on discordant status of
the dyad, sometimes called the “independent cascade model” (), which suggests a
monotonic relation between the number of times you are exposed through peers.
High exposure could be due to repeated interaction with one person or weak
interaction with many, effectively equating:
Alternative models exist. Under “complex contagion” for example, the likelihood
that I accept the bit that flows through the network depends on the proportion of my
peers that have the bit.

1
1
2
3
Complex Contagion
Assume adoption requires k neighbors having adopted, then transmission can only
occur within dense clusters:

Complex Contagion
Assume pij=1, k=2, starting nodes in yellow

Complex Contagion
For this network under weak complex diffusion (k=2), the maximum risk size is 8.

Complex Contagion
For this network under weak complex diffusion (k=2), the maximum risk size is reaches 98%.
One of the Prosper
schools:
Start

Complex Contagion
Can lead to widely varying sizes of potential diffusion cascades. Here’s the
distribution across all PROPSPER schools:
Distribution is largely bimodal (even with a connected pair start)

Complex Contagion
Can lead to widely varying sizes of potential diffusion cascades. Here’s the
distribution across all PROPSPER schools:
The governing factors are (a) curved effect of local redundancy and (b) structural cohesion
Network Average Proportion Reached
k=2 complex contagion
MeanCascadeSize
Coh=0.3
Coh=1.2
Coh=2.2
Coh=3.2
Coh=4.1

Complex Contagion
Does get used for real health
work:
Here , authors assume a CC
process, seeded with observed
depressive cases, turn that into
a Markov model and ask what
parameters would maximize fit
from simulated to observed.

Complex diffusion is just the most well studied of the options that combine
transmission with some pairwise positional feature. This is a wide-open
area for future research.
The basic idea is that transmission is increased/decreased if there is some
third structural property that the susceptible & infected pair share.
This leads us into the general problem of peer influence models…when do
peers change each other’s behaviors?

Background:
• Long standing research interest in how our relations shape our attitudes
and behaviors.
• Most often assumed mechanism is that people (through conversation or
similar) change each others beliefs/opinions, which changes behavior.
This implies that position in a communication network should be related
to attitudes.
• Alternatives:
• Modeling behavior: ego copies behavior of alter to gain respect,
esteem, etc.
• Distinction: Ego tries to be different from (some) alter to gain
respect, esteem, etc.
• Access: Ego wants to do Y, but can only do so because alter provides
access (say, being old enough to buy cigarettes).
Peer Influence Dynamics

Background:
• Early work was ego-centric – people informed on their peers
•Seems to have inflated PI effects by ~50% or so…either through projection of
ego behavior onto peers or selective interaction (what alters do with ego may be
different than what alter does all the time).
•Then to cross sectional associations based on alter self-reports
•Better, but still likely conflates selection with influence
•Next to dynamic models:
•Ego Behavior(t) ~ f(ego behavior(t-1) + alter behavior (t-1) + controls
•Much better; still debate on (a) correct estimation functions, (b) unobserved
selection features that confound causal inference.
•Development of Actor-oriented models (SIENA)

Background:
•Finally: Experimental manipulation of peer exposure
•“Gold standard” for isolation of peer effects
•Likely strongly underestimates effects (as measure intent to treat, not take-
up of treatment, since people may not care about relations that can be
manipulated).
b(Peer(y)): Ego Inform < Alter Inform < Cross Sectional < Dynamic < Experimental. Still often
find peer effects, but my sense is that we’ve (strongly) over-corrected at this point.

Freidkin’s Structural Theory of Social Influence :
Two-part model:
Beliefs are a function of two sources:
a) Individual characteristics
•Gender, Age, Race, Education, Etc. Standard sociology
b) Interpersonal influences
•Actors negotiate with others

XBY )1( (1)
)1()1()(
)1( YWYY αα Tt
 
(2)
Y(1) = an N x M matrix of initial opinions on M issues for N
actors
X = an N x K matrix of K exogenous variable that affect Y
B = a K x M matrix of coefficients relating X to Y
a = a weight of the strength of endogenous interpersonal
influences
W = an N x N matrix of interpersonal influences

XBY )1( (1)
This is the standard sociology model for explaining anything: the General Linear Model.
It says that a dependent variable (Y) is some function (B) of a set of independent
variables (X). At the individual level, the model says that:

k
kiki BXY
Usually, one of the X variables is e, the model error term.

)1()1()(
)1( YWYY αα Tt
 
(2)
This part of the model taps social influence. It says that each person’s final opinion is
a weighted average of their own initial opinions
)1(
)1( Yα
And the opinions of those they communicate with (which can include their own current
opinions)
)1( T
αWY

The key to the peer influence part of the model is W, a matrix of
interpersonal weights. W is a function of the communication structure of the
network, and is usually a transformation of the adjacency matrix. In general:
 

j
ij
ij
w
w
1
10
Various specifications of the model change the value of wii, the extent to which
one weighs their own current opinion and the relative weight of alters.

1 2
3
4
1 2 3 4
1 1 1 1 0
2 1 1 1 0
3 1 1 1 1
4 0 0 1 1
1 2 3 4
1 .33 .33 .33 0
2 .33 .33 .33 0
3 .25 .25 .25 .25
4 0 0 .50 .50
1 2 3 4
1 .50 .25 .25 0
2 .25 .50 .25 0
3 .20 .20 .40 .20
4 0 0 .33 .67
Even
2*self
1 2 3 4
1 .50 .25 .25 0
2 .25 .50 .25 0
3 .17 .17 .50 .17
4 0 0 .50 .50
degree
Self weight:
1 2 3 4
1 2 1 1 0
2 1 2 1 0
3 1 1 2 1
4 0 0 1 2
1 2 3 4
1 2 1 1 0
2 1 2 1 0
3 1 1 3 1
4 0 0 1 1

)1()1()(
)1( YWYY αα Tt
 
Formal Properties of the model
When interpersonal influence is complete, model reduces to:
)1(
)1()1()(
01




T
Tt
WY
YWYY
When interpersonal influence is absent, model reduces to:
)1(
)1()1()(
0
Y
YWYY

 Tt
(2)

Formal Properties of the model
The model is directly related to spatial econometric models:
If we allow the model to run over t and W remains constant:
XBWYY )1()()(
αα  
eb  
XWYY
~)()(
α
Where the two coefficients (a and b) are estimated directly (See Doreian,
1982, SMR).
This is the linear network auto correlation model, best bet with cross-sectional
data (and randomization trick to estimate se)

Simple example
1 2
3
4
1 2 3 4
1 .33 .33 .33 0
2 .33 .33 .33 0
3 .25 .25 .25 .25
4 0 0 .50 .50
Y
1
3
5
7
a = .8
T: 0 1 2 3 4 5 6 7
1.00 2.60 2.81 2.93 2.98 3.00 3.01 3.01
3.00 3.00 3.21 3.33 3.38 3.40 3.41 3.41
5.00 4.20 4.20 4.16 4.14 4.14 4.13 4.13
7.00 6.20 5.56 5.30 5.18 5.13 5.11 5.10
By t=7, still variability in Y

1 2
3
4
1 2 3 4
1 .33 .33 .33 0
2 .33 .33 .33 0
3 .25 .25 .25 .25
4 0 0 .50 .50
Y
1
3
5
7
a = 1.0
1.00 3.00 3.33 3.56 3.68 3.74 3.78 3.81
3.00 3.00 3.33 3.56 3.68 3.74 3.78 3.81
5.00 4.00 4.00 3.92 3.88 3.86 3.85 3.84
7.00 6.00 5.00 4.50 4.21 4.05 3.95 3.90
By t=7, almost no variability in Y
T: 0 1 2 3 4 5 6 7
Simple example

Extended example: building intuition
Consider a network with three cohesive groups, and an initially random distribution of
opinions:

Simulated Peer Influence:
75 actors, 2 initially random opinions, Alpha = .8, 7 iterations

Extended example: building intuition
Consider a network with three cohesive groups, and an initially random distribution of opinions:
Now weight in-group ties higher than between group ties

Simulated Peer Influence:
75 actors, 2 initially random opinions, Alpha = .8, 7 iterations, in-group tie: 2

04 Diffusion and Peer Influence (2016)

Consider the implications for populations of different structures. For example, we might
have two groups, a large orthodox population and a small heterodox population. We can
imagine the groups mixing in various levels:
Little Mixing Moderate Mixing Heavy Mixing
.95 .05
.05 .02
.95 .008
.008 .02
.95 .001
.001 .02
Heterodox: 10 people
Orthodox: 100 People

In an unbalanced situation (small group vs large group) the extent of contact
can easily overwhelm the small group. Applications of this idea are evident
in:
•Missionary work (Must be certain to send missionaries out into the
world with strong in-group contacts)
•Overcoming deviant culture (I.e. youth gangs vs. adults)
•This is also the mechanism behind why most youth peer influence
is a *good* thing – most youth are well behavior and civic
minded…so are exerting positive influences on their peers.

Friedkin (1998) generalizes the model so that alpha varies across people.
(1) simply changing a to a vector (A), which then changes each person’s
opinion directly
(2) by linking the self weight (wii) to alpha.
)1()1()(
)( YAIAWYY  Tt
Were A is a diagonal matrix of endogenous weights, with 0 < aii < 1. A further
restriction on the model sets wii = 1-aii
This leads to a great deal more flexibility in the theory, and some interesting
insights. Consider the case of group opinion leaders with unchanging opinions
(I.e. many people have high aii, while a few have low):

Group 1
Leaders
Group 2
Leaders
Group 3
Leaders
Peer Opinion Leaders

Further extensions of the model might:
• Time dependent a: people likely value other’s opinions more early than later in a
decision context
• Interact a with XB: people’s self weights are a function of their behaviors &
attributes
• Make W dependent on structure of the network (weight transitive ties greater
than intransitive ties, for example)
• Time dependent W: The network of contacts does not remain constant, but is
dynamic, meaning that influence likely moves unevenly through the network
• And others likely abound….

There are two common ways to test for peer associations through networks.
The first estimates the parameters (a and b) of the network autocorrelation model directly,
the second transforms the network into a dyadic model, predicting similarity among
actors.
eb  
XWYY
~)()(
α
Peer influence model:
This is the linear network autocorrelation model, and as specified, the
model makes strong assumptions about equilibrium opinion and static
relations.
 Some variants on this also expand e to include alternative
autocorrelation in the error structure.

There are two common ways to test for peer associations through networks.
The first estimates the parameters (a and b) of the network autocorrelation model directly,
the second transforms the network into a dyadic model, predicting similarity among
actors.
eb  
XWYY
~)()(
α
Peer influence model:
Note that since WY is a a simple vector -- weighted mean of friends Y -- which can be
constructed and added to your GLM model. That is, multiple Y by a W matrix, and run the
regression with WY as a new variable, and the regression coefficient is an estimate of a. This is
what Doriean calls the QAD estimate of peer influence.
It’s wrong, a will be biased, but it’s often not terribly wrong if most obvious selection factors are
built int0 X

An obvious problem with this specification is that cases are, by definition, not
independent, hence “network autocorrelation” terminology.
In practice, the QAD approach (perhaps combined with a GLS estimator)
results in empirical estimates that are “virtually indistinguishable” from MLE
(Doreian et al, 1984)
The proper way to estimate the peer equation is to use maximum likelihood
estimates, and Doreian gives the formulas for this in his paper, and Carter Butts
has implemented in in R with the LNAM procedure.
An alternative is to use non-parametric approaches, such as the Quadratic
Assignment Procedure, to estimate the effects.

Peer influence through Dyad Models
Another way to get at peer influence is not through the level of Y, but by assessing the
similarity of connected peers. Recall the simulated example: peer influence is reflected in
how close points are to each other.

Peer influence through Dyad Models
The model is now expressed at the dyad level as:
ij
k
kkijij eXbAbbY  10
Where Y is a matrix of similarities, A is an adjacency matrix, and Xk is a matrix of similarities
on attributes
Advantages include ease of specifying relation-specific similarity functions. You can add
different features of a relation by adjusting/adding “Aij” variables.
Disadvantage is that now in addition to network autocorrelation, you have repeated cases (on
both sides).
But these can be dealt with using non-parametric modeling & testing techniques (QAP, for
example). (which we will go over this afternoon)

Peer Influence & Health: Current Lit & Controversies

Used the friend/relative
tracking data from a
larger heart-health study
to identify network
contacts, including
friends.

The network shows significant evidence of weight-homophily
to identify network
contacts, including
friends.

Effects of peer obesity on ego, by peer type
Edge-wise regressions of the form:
ControlsEgoAltAltEgo previouspreviousCurrentCurrent  )()()( 321 bbb
Ego is repeated for all alters; models include random effects on ego id
to identify network
contacts, including
friends.

This modeling strategy
pools observations on
edges and estimates a
global effect net of
change in ego/alter as a
control. Here color is a
single ego, number is
wave (only 2 egos and 3
waves represented).
ControlsEgoAltAltEgo previouspreviousCurrentCurrent  )()()( 321 bbb
1
Ego-Current
Alter Current
1 1
2 2 2
3 3 3
1 1 1
2 2 2
3 3 3
Peer Effect

Alterative specifications
include using change-
change models and
allowing for a random
effect of peers. This
allows for greater
variability in peer effects,
and the potential to model
differences.
ee
currentpreviouseCurrentprevious ControlstAltAlEgoEgo
bb
b

 )()(
1
Ego-Current
Alter Current
1 1
2 2 2
3 3 3
1 1 1
2 2 2
3 3 3
b
be1
be1
ee
previouspreviousCurrenteCurrent ControlsEgoAltAltEgo
bb
bbb

 )()()( 32
Or difference
models:

Critiques of C&F
The C&F studies – of obesity, but also other work on the FHS data – turn on the validity of the
causal association.
All turn on some issue of model miss-specification, typically:
• Can’t truly distinguish a network effect from other sources of common influence
• “Selection” (“homophily”) or “Common influence” (“Shared environment”)
• The most strident work in this area (Salizi
• Statistical errors
• Misinterpretation of confidence intervals
• Poorly specified/estimated models
C&H do a nice job of laying out their responses here:
http://jhfowler.ucsd.edu/examining_dynamic_social_networks.pdf and here:
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2597062/

Critiques of C&F
Cohen-Cole, E. and Fletcher, J. M. (2008). Detecting implausible social network effects
in acne, height, and headaches: longitudinal analysis. British Medical Journal 337
a2533.
Use the same models as C&F
on Add Health to show that
things which are theoretically
unlikely to be contagious
appear to be in this form of
model.
Note these coefficients are
substantially smaller than C&F
and only significant at the 0.1
level; and not robust to any
sensitivity analysis.

Critiques of C&F
Lyons, 2011.
1) C&F claim that differences in directional effects support a PI story:
• C& F: While mutual friends and
egoalter friends are > 0,
alterego is not, means ego is
emulating alter.
• Lyons notes these CIs
overlap too much to make
any claim about
distinguishing them from
each other.

Critiques of C&F Lyons, 2011.
2) Insufficient controls for Homophily
• C& F: Use of alter’s lagged Y to control for homophily. Logic is that any feature
that selected us to be friends at t-1 would have had it’s effect then.
• Lyons notes that current and lagged have opposite signs, which seems
suspect, and anyway is an insufficient control. He’s likely right here…
3) Directionality cannot distinguish the source of association
• C& F: the ordering: mutual, egoalter, alterego suggests an “esteem” model,
where ego copies the behavior of alter.
• Lyons argues that we would expect the same logic from a simple “foci” of
similarity. I don’t find this argument convincing.
4) Random permutation tests cannot establish 3-degree rule
• C& F: Association between alters at 1, 2, 3 degrees of separation are higher than
we’d expect by chance, based on a permutation test.
• Lyons invalid if the data are incomplete, which they certainly are. I don’t
find this argument convincing…data are always incomplete…

5) The models are statistically inconsistent (if not incoherent)
• C& F: Use separate models for each type of tie, with random effects on ego.
• Lyons notes that these really should be treated as simultaneous equations,
with shared error structures and so forth. Doing so (a) leads to unidentified
models that must force the estimation of the peer effect to 0. That observed
^0 indicates something amiss.
• Strikes me as a bit down in the weeds and I’m not convinced here that he’s
critiquing them for what they are really doing (argues there are more
equations than data, which is patently not true).

My sense is that the strategy C&F took was not fundamentally misguided, but the model
specification is probably thin; certainly in the obesity paper – less so in some of the later papers
appearing after these debates.
Ideally you’d have a much better direct model for selection – perhaps even a separate two-stage
model (see the Siena module), but here there are very limited observational controls, which would
have been easy to add. In later specifications, they do add fixed effects for ego and still find similar
results.
Commenting on the debate on SocNET – and a related conclusion that only experiments could
provide valid inference – Tom Snijders says:
“The logical consequence of this is that we are stuck with imperfect methods. Lyons argues as
though only perfect methods are acceptable, and while applauding such lofty ideals I still believe
that we should accept imperfection, in life as in science. Progress is made by discussion and
improvement of imperfections, not by their eradication.”
For a full general discussion, see : https://www.lists.ufl.edu/cgi-bin/wa?A2=ind1106&L=SOCNET&P=R11428

Shalizi & Thomas: PI is *generally* confounded
So long as there is an unobserved X that causes both ties and behavior, the
effect of peers is unidentified.

Only route out is to make X fully informed (or informing) by an observable Z….but
realistically there are few things that (a) cause behavior exclusively without any
selection pressure (a) or cause ties exclusively without any influence pressure (b)
(though note b is what experimental assignments do)
(X causes Z, not Y directly) (X causes A, not Y directly)

Should be noted that this is true for *any* effect – there’s always the
potential that an unobserved latent variable is creating a spurious effect;
This sort of work argues that the only solution is to use experimental (or,
sometimes, propensity score style models)…but that’s simply not always
feasible practically.
We need to beware of making the best the enemy of the good enough…lest
we make no progress at all…

Willard Van Quine, professor of philosophy and mathematics emeritus from
Harvard University who is regarded as one of the most famous philosophers
in the world, wrote his doctoral thesis on a 1927 Remington typewriter,
which he still uses. However, he "had an operation on it" to change a few
keys to accommodate special symbols. "I found I could do without the
second period, the second comma -- and the question mark.”
"You don't miss the question mark?”
"Well, you see, I deal in certainties."
Selection or Influence?

Is it all selection
•What do we know about how friendships form?
•Opportunity / focal factors
- Being members of the same group
- In the same class
- On the same team
- Members of the same church
•Structural Relationship factors
- Reciprocity
- Social Balance
•Behavior Homophily
- Smoking
- Drinking

-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Network Model Coefficients, In school Networks

How to correct this problem?
•Essentially, this is an omitted variable problem, and my “solution” has
been to identify as many potentially relevant alternative variables as I can
find.
•The strongest possible correction is to use fixed-effects* models that
control for all non-varying individual covariates. These have their own
problems…
•Dual model for influence & selection.
•Two-stage model “Heckman” sorts of models
•Dynamic SAOM models
*“Adding fixed effects to dynamic panel models with many subjects and few repeat observations creates severe bias towards zero coefficients. This has
been demonstrated both analytically (Nickell 1981) and through simulations (Nerlove 1971) for OLS and other regression models and has been well-known
by social scientists, including economists, for a very long time. In fact, CCF even note that they do not add fixed effects to their logit regression model for
this reason, but they strangely assert that fixed effects are necessary in the OLS model.” Estimating Peer Effects on Health in Social Networks : A
Response to Cohen-Cole and Fletcher; Trogdon, Nonnemaker, Pais J.H. Fowler, PhD and N.A. Christakis, MD, PhD

• Causal status of such similarity is hard to know,
• Identification strategies are stringent
• My sense is we’re over-correcting on this front; let’s figure out what’s
there first.
Selection
Y
X1 X2
Weak instruments bias us toward null effects
Y
X1 X2
I

Possible solutions:
• Theory: Given what we know about how friendships form, is it reasonable
to assume a bi-directional cause? That is, work through the meeting,
socializing, etc. process and ask whether it makes sense that Y is a cause of
W. This will not convince a skeptical reader, but you should do it anyway.
• Models:
- Time Order. Necessary but not sufficient. We are on somewhat firmer
ground if W precedes Y in time, but the Shalizi & Thomas problem of
an as-yet-earlier joint confounder is still there.
- Simultaneous Models. Model both the friendship pattern and the
outcome of interest simultaneously. Best bet for direct estimation
•Sensitivity Analysis:
I think the most reasonable solution…take error potential seriously,
attempt to evaluate how big a problem it really is.

Table 4. Selected SIENA Parameter Estimates: Parental Knowledge, Parental Discipline, and Drinking
a
Model 2
b SE t SD
Selection parameters
Alter effects: Who is more often named as a friend?
Parental knowledge -0.002 0.004 -0.47 0.003
Parental discipline -0.004 0.002 -1.55 0.001
Drinking 0.083 0.010 8.69 *** 0.007
Ego effects: Who names more friends?
Parental knowledge 0.044 0.007 5.85 *** 0.039
Parental discipline -0.002 0.005 -0.36 0.023
Drinking -0.011 0.021 -0.53 0.089
Similarity effects: Choosing friends similar to oneself
Parental knowledge 0.169 0.025 6.70 *** 0.101
Parental discipline 0.151 0.017 8.86 *** 0.035
Drinking 0.276 0.021 13.45 *** 0.006
Behavioral parameters: Influence on Drinking
Friends' attributes
Mean Parental knowledge -0.230 0.065 -3.56 ** 0.014
Mean Parental discipline -0.051 0.043 -1.18 0.011
Drinking mean similarity 1.162 0.110 10.56 *** 0.023
Control variables (individual level)
Parental knowledge -0.122 0.014 -8.93 *** 0.004
Parental discipline -0.043 0.009 -4.54 *** 0.002
***p < .001. **p < .01. *p < .05. †p < .10.
a
Models also include rate and shape parameters, structural parameters, and the full set of alter,
ego, similarity, and individual-level control parameters
SIENA model of drinking
Daniel T. Ragan, D. Wayne Osgood

Possible solutions:
I think the most
reasonable solution…take
error potential seriously,
attempt to evaluate how
big a problem it really is.

Possible solutions:
I think the most reasonable solution…take error potential seriously, attempt to evaluate
how big a problem it really is.
Sociological Methods & Research 2000

04 Diffusion and Peer Influence (2016)

More Related Content

What's hot (20)

Similar to 04 Diffusion and Peer Influence (2016) (20)

More from Duke Network Analysis Center (20)

Recently uploaded (20)

04 Diffusion and Peer Influence (2016)