Title: Diffusion
1Diffusion
- Structural Bases of Social Network Diffusion
- Dynamic limitations on diffusion
- Implications / Applications in the diffusion of
Innovations
2Diffusion
Two factors that affect network
diffusion Topology - the shape, or form, of the
network - simple example one actor cannot pass
information to another unless they are either
directly or indirectly connected Time - the
timing of contacts matters - simple example an
actor cannot pass information he has not yet
received.
3Diffusion Topology features
- Connectivity refers to how actors in one part of
the network are connected to actors in another
part of the network. - Reachability Is it possible for actor i to
reach actor j? This can only be true if there is
a chain of contact from one actor to another. - Distance Given they can be reached, how many
steps are they from each other? - Number of paths How many different paths
connect each pair?
4Network Toplogy
Consider the following (much simplified) scenario
- Probability that actor i infects actor j (pij)is
a constant over all relations 0.6 - S T are connected through the following
structure
S
T
- The probability that S infects T through either
path would be 0.09
5Why Sexual Networks Matter
Now consider the following (similar?) scenario
S
T
- Every actor but one has the exact same number of
partners - The category-to-category mixing is identical
- The distance from S to T is the same (7 steps)
- S and T have not changed their behavior
- Their partners partners have the same behavior
- But the probability of an infection moving from S
to T is - 0.148
- Different outcomes different potentials for
intervention
6Probability of infection over independent paths
- The probability that an infectious agent travels
from i to j is assumed constant at pij. - The probability that infection passes through
multiple links (i to j, and from j to k) is the
joint probability of each (link1 and link2 and
link k) pijd where d is the path distance. - To calculate the probability of infection passing
through multiple paths, use the compliment of it
not passing through any paths. The probability
of not passing through path l is 1-pijd, and thus
the probability of not passing through any path
is (1-pijd)k, where k is the number of paths - Thus, the probability of i infecting j given k
independent paths is
Why matter
Distance
7Probability of infection over non-independent
paths
- To get the probability that I infects j given
that paths intersect at 4, I calculate
Using the independent paths formula.
8Network Topology Ego Networks
Mixing Matters
- The most commonly collected network data are
ego-centered. While limited in the structural
features, these do provide useful information on
broad mixing patterns relationship timing. - Consider Laumann Youms (1998) treatment of
sexual mixing by race and activity level, using
data from the NHSLS, to explain the differences
in STD rates by race - They find that two factors can largely explain
the difference in STD rates - Intraracially, low activity African Americans are
much more likely to have sex with high activity
African Americans than are whites - Interracially, sexual networks tend to be
contained within race, slowing spread between
races
9Network Topology Ego Networks
- In addition to general category mixing,
ego-network data can provide important
information on - Local clustering (if there are relations among
egos partners. Not usually relevant in
heterosexual populations, though very relevant to
IDU populations) - Number of partners -- by far the simplest network
feature, but also very relevant at the high end - Relationship timing, duration and overlap
- By asking about partners behavior, you can get
some information on the relative risk of each
relation. For example, whether a respondents
partner has many other partners (though data
quality is often at issue).
10Network Topology Ego Networks
Clustering matters because it re-links people to
each other, lowering the efficiency of the
transmission network.
Clustering also creates pockets where goods can
circulate.
11Network Topology Partial and Complete Networks
Once we move beyond the ego-network, we can start
to identify how the pattern of connection changes
the disease risk for actors. Two features of the
networks shape are known to be important
Connectivity and Centrality.
- Connectivity refers to how actors in one part of
the network are connected to actors in another
part of the network. - Reachability Is it possible for actor i to
infect actor j? This can only be true if there
is an unbroken (and properly time ordered) chain
of contact from one actor to another. - Given reachability, three other properties are
important - Distance
- Number of paths
- Distribution of paths through actors
(independence of paths)
12Reachability example All romantic contacts
reported ongoing in the last 6 months in a
moderate sized high school (AddHealth)
63
(From Bearman, Moody and Stovel, 2004.)
13Network Topology Distance number of paths
- Given that ego can reach alter, distance
determines the likelihood of an infection passing
from one end of the chain to another. - Diffusion is never certain, so the probability of
transmission decreases over distance. - Diffusion increases with each alternative path
connecting pairs of people in the network.
14Probability of Diffusion
by distance and number of paths, assume a
constant pij of 0.6
1.2
1
10 paths
0.8
5 paths
probability
0.6
2 paths
0.4
1 path
0.2
0
2
3
4
5
6
Path distance
15Probability of Diffusion
by distance and number of paths, assume a
constant pij of 0.3
0.7
0.6
0.5
0.4
probability
0.3
0.2
0.1
0
2
3
4
5
6
Path distance
16Return to our first example
2 paths
4 paths
17Reachability in Colorado Springs (Sexual contact
only)
- High-risk actors over 4 years
- 695 people represented
- Longest path is 17 steps
- Average distance is about 5 steps
- Average person is within 3 steps of 75 other
people - 137 people connected through 2 independent paths,
core of 30 people connected through 4 independent
paths
(Node size log of degree)
18Network Topology Centrality and Centralization
- Centrality refers to (one dimension of) where an
actor resides in a sexual network. - Local compare actors who are at the edge of the
network to actors at the center - Global compare networks that are dominated by a
few central actors to those with relative
involvement equality
19Centrality example Add Health
Node size proportional to betweenness centrality
Graph is 45 centralized
20Centrality example Colorado Springs
Node size proportional to betweenness centrality
Graph is 27 centralized
21Network Topology Effect of Structure
22Network Topology Effect of Structure
Simulated diffusion curves for the observed
network.
23Network Topology Effect of Structure
The effect of the observed structure can be seen
in how diffusion differs from a random network
with the same volume
24Network Topology Effect of Structure
25Network Topology Effect of Structure
Mean number of independent paths
26Network Topology Effect of Structure
Clustering Coefficient
27Network Topology Effect of Structure
Mean Distance
28Network Topology Effect of Structure
29Network Topology Effect of Structure
30Timing Sexual Networks
A focus on contact structure often slights the
importance of network dynamics. Time affects
networks in two important ways 1) The structure
itself goes through phases that are correlated
with disease spread Wasserheit and Aral, 1996.
The dynamic topology of Sexually Transmitted
Disease Epidemics The Journal of Infectious
Diseases 74S201-13 Rothenberg, et al. 1997
Using Social Network and Ethnographic Tools to
Evaluate Syphilis Transmission Sexually
Transmitted Diseases 25 154-160 2) Relationship
timing constrains disease flow a) by spending
more or less time in-host b) by changing the
potential direction of disease flow
31Sexual Relations among A syphilis outbreak
Changes in Network Structure
Rothenberg et al map the pattern of sexual
contact among youth involved in a Syphilis
outbreak in Atlanta over a one year period.
(Syphilis cases in red)
Jan - June, 1995
32Sexual Relations among A syphilis outbreak
July-Dec, 1995
33Sexual Relations among A syphilis outbreak
July-Dec, 1995
34Data on drug users in Colorado Springs, over 5
years
35Data on drug users in Colorado Springs, over 5
years
36Data on drug users in Colorado Springs, over 5
years
37Data on drug users in Colorado Springs, over 5
years
38Data on drug users in Colorado Springs, over 5
years
39What impact does this kind of timing have on
diffusion?
The most dramatic effect occurs with the
distinction between concurrent and serial
relations. Relations are concurrent whenever
an actor has more than one sex partner during the
same time interval. Concurrency is dangerous for
disease spread because a) compared to serially
monogamous couples, and STDis not trapped inside
a single dyad b) the std can travel in two
directions - through ego - to either of his/her
partners at the same time
40Concurrency and Epidemic Size Morris
Kretzschmar (1995)
1200
800
400
0
0
1
2
3
4
5
6
7
Monogamy
Disassortative
Assortative
Random
Population size is 2000, simulation ran over 3
years
41Concurrency and disease spread
42A hypothetical Sexual Contact Network
8 - 9
C
E
3 - 7
2 - 5
B
A
0 - 1
3 - 5
D
F
43The path graph for a hypothetical contact network
E
C
B
A
D
F
44Direct Contact Network of 8 people in a ring
45Implied Contact Network of 8 people in a ring All
relations Concurrent
46Implied Contact Network of 8 people in a
ring Mixed Concurrent
2
3
2
1
1
2
2
3
47Implied Contact Network of 8 people in a
ring Serial Monogamy (1)
1
8
2
7
3
6
5
4
48Implied Contact Network of 8 people in a
ring Serial Monogamy (2)
1
8
2
7
3
6
1
4
49Implied Contact Network of 8 people in a
ring Serial Monogamy (3)
1
2
2
1
1
2
1
2
50Timing Sexual Networks
- Network dynamics can have a significant impact on
the level of disease flow and each actors risk
exposure
This work suggests that a) Disease outbreaks
correlate with phase-shifts in the connectivity
level b) Interventions focused on relationship
timing, especially concurrency, could have a
significant effect on disease spread c) Measure
and models linking network topography to disease
flow should account for the timing of romantic
relationships
51Timing Sexual Networks
52Degree or Connectivity?
Large-scale network model implications
Scale-Free Networks
Many large networks are characterized by a highly
skewed distribution of the number of partners
(degree)
53Degree or Connectivity?
Large-scale network model implications
Scale-Free Networks
Many large networks are characterized by a highly
skewed distribution of the number of partners
(degree)
54Degree or Connectivity?
Large-scale network model implications
Scale-Free Networks
The scale-free model focuses on the
distance-reducing capacity of high-degree nodes
55Degree or Connectivity?
Large-scale network model implications
Scale-Free Networks
The scale-free model focuses on the
distance-reducing capacity of high-degree nodes
- Which implies
- a thin cohesive blocking structure and a fragile
global topography - Scale free models work primarily on through
distance, as hubs create shortcuts in the graph,
not through core-group dynamics.
56Degree or Connectivity?
Empirical Evidence
Project 90, Sex-only network (n695)
3-Component (n58)
57Degree or Connectivity?
Empirical EvidenceProject 90, Drug sharing
network
Connected Bicomponents
N616 Diameter 13 L 5.28 Transitivity
16 Reach 3 128 Largest BC 247 K gt 4 318 Max
k 12
58Degree or Connectivity?
Empirical EvidenceProject 90, Drug sharing
network
Multiple 4-components
59Degree or Connectivity?
Building on recent work on conditional random
graphs, we examine (analytically) the expected
size of the largest component for graphs with a
given degree distribution, and simulate networks
to measure the size of the largest bicomponent.
For these simulations, the degree distribution
shifts from having a mode of 1 to a mode of
3. We estimate these values on populations of
10,000 nodes, and draw 100 networks for each
degree distribution. Newman, Strogatz,
Watts 2001 Molloy Reed 1998
60Degree or Connectivity?
61Degree or Connectivity?
62Degree or Connectivity?
Very small changes in degree generate a quick
cascade to large connected components. While not
quite as rapid, STD cores follow a similar
pattern, emerging rapidly and rising steadily
with small changes in the degree
distribution. This suggests that, even in the
very short run (days or weeks, in some
populations) large connected cores can emerge
covering the majority of the interacting
population, which can sustain disease.
63Empirical Models for Diffusion
Macro-level models Typically model diffusion as
a growth rate process over some population.
Recent models include more parameters to get
better fits
Y is the proportion of adopters, bo a rate
parameter for innovation and b1 a rate parameter
for imitation. This is the Bass Model, after
Bass 1969.
These models really only work on the rate of
change, and assume random mixing.
64Empirical Models for Diffusion
Add peer effects
Were w is a weight matrix for contact between
actors.
65Empirical Models for Diffusion
66Empirical Models for Diffusion
67Empirical Models for Diffusion