Network Methods for Behavior Change

1
Network Methods for Behavior Change

• Thomas W. Valente, PhD
• Professor
• Preventive Medicine, Keck School of Medicine
• University of Southern California
• tvalente_at_usc.edu

2
Interventions Definitions

• Using network data to change behaviors
• Change individual and community/organizational
  level
• Not exactly clear what constitutes a network
  intervention, for now
• Any change program that uses network data to
• Select change agents
• Define groups
• Affect network structure
• Assist Behavior Change program implementation

2
3
Theory Will Guide

• The type of change desired will be guided by
  theory
• In some cases want to increase cohesion in others
  increase fragmentation
• Increase/decrease centralization
• E.g., slowing spread of STDs requires different
  strategy than accelerating adoption of office
  automation

4
6 Types of Network Interventions

• Identify opinion leaders or key players to act as
  change agents
• Create network-based groups/positions
• Identify leaders within groups or match leaders
  to groups
• Snowballing / Contact tracing / Respondent Driven
  Sampling
• Rewire Networks
• More/less cohesive
• More/less centralized
• More/less dense
• Change core-peripheriness
• Etc.
• Other
• Triadic Structures
• Identify low threshold adopters
• Reaching critical mass
• Reporting back to group/dialogue
• Others?

5
1. Opinion Leaders

• The most typical network intervention
• Easy to measure
• Intuitively appealing
• Proven effectiveness
• Over 20 studies using network data to identify
  OLs and hundreds of others using other OL
  identification techniques

6
Diffusion Network Simulation w/ 3 Initial
Adopter Conditions
7
HIV Sexual Risk Reduction Social Network
Intervention Trials in Eastern Europe

• HIV sexual risk reduction behavior interventions
  within indigenous friendship-based social
  networks in Eastern Europe - J.A. Kelly, Ph.D.
  and Y.A. Amirkhanian, Ph.D. (CAIR).
• Social networks of Roma ethnic minority and of
  young MSM were identified, recruited, assessed to
  identify sociometric leader of each network, and
  then randomized into either immediate or delayed
  intervention condition.
• The leaders of the intervention networks attended
  9-session training program and carried out HIV
  prevention conversations with their own network
  members.
• Intervention outcomes were compared between
  experimental and control groups at Baseline, 3-
  and 12-months.

8
Roma Egocentric Network HIV Prevention Trial,
Sofia, Bulgaria (N255, 52 networks,
retentiongt90)Kelly, Amirkhanian, Kabakchieva et
al., BMJ, 2006
9
Young MSM Egocentric Network HIV Prevention
Trial, Bulgaria/Russia (n276, 52 networks,
retention gt84) Amirkhanian, Kelly, Kabakchieva
et al., AIDS, 2005
12-month p indicates significance in difference
between Bulgaria and Russia The long-term
effects remained strong in Bulgaria.
10
Achievable in UCINET

• Do not symmetrize data
• Compare degree scores with other centrality
  measures
• Compare degree scores with Key Player analysis

11
Other Centrality Measures

• In-degree preferable, easy
• Theoretical diffusion processes may suggest other
  centrality measures, closeness, betweenness
• Can use other measures as tie-breakers (i.e., 2
  nodes of same in-degree choose one with higher
  closeness)

12
Sophisticated World

• Use different types of opinion leaders at
  different stages
• In-degree
• Betweenness
• Closeness
• Use different types of opinion leaders for
  different groups
• In-degree in highly cohesive groups
• Betweenness in fractured groups

13
(No Transcript)
14
FREEMAN'S DEGREE CENTRALITY MEASURES -------------
--------------------------------------------------
------- Diagonal valid? NO Model
ASYMMETRIC Input dataset
C\MISC\DIFFNET\OL\com18
1 2 3 4
OutDegree InDegree NrmOutDeg
NrmInDeg ------------ ------------
------------ ------------ 19 26 5.000
2.000 13.889 5.556 20 27
5.000 7.000 13.889 19.444 3
11 5.000 5.000 13.889
13.889 4 12 5.000 6.000
13.889 16.667 5 13 5.000
6.000 13.889 16.667 6 14
5.000 7.000 13.889 19.444 25
31 5.000 7.000 13.889
19.444 8 16 5.000 6.000
13.889 16.667 9 17 5.000
8.000 13.889 22.222 10 18
5.000 1.000 13.889 2.778 11
19 5.000 3.000 13.889
8.333 12 2 5.000 2.000
13.889 5.556 13 20 5.000
1.000 13.889 2.778 14 21
5.000 11.000 13.889 30.556 15
22 5.000 4.000 13.889
11.111 34 6 5.000 5.000
13.889 13.889 17 24 5.000
6.000 13.889 16.667 . . .
15
10 Methods Used to Identify Peer Opinion Leaders
(Valente Pumpuang, 2007)
16
Implementation Issues

• Do you just turn leaders loose?
• Schedule 1-1 between leaders members
• Have leaders give formal presentations
• Have leaders call a meeting
• Allow leaders to decide how to promote change
• Continuum of Passive to Active OL Involvement

17
2. Network Based Groups

• Sets of people/nodes that are densely connected
• Groups can reinforce (or inhibit) the behavior
  change process
• Behavior change may be appropriate for groups
• Finding groups

18
Defining Groups

• Components
• Cliques/Kplexes/Cycles, etc.
• Newman-Girvan algorithm
• Provides mutually exclusive groups
• Provides measure of group fit
• Ken Frank at MSU has group programs
• Can also use positional analysis such as CONCOR
  to identify equivalent positions

19
(No Transcript)
20
Newman-Girvan 6 Groups
21
Implementation Issues

• Do groups need to be the same size?
• In school based programs, usually they do
• In organizations they can vary somewhat but then
  group size becomes an issue
• Does the socio-demographic composition of the
  group matter?
• Most cases groups will be homogenous
• Some cases may need to impose homogeneity on the
  group (sex education, e.g.)

22
Positions rather than Groups

• Positions may be more relevant than groups
• Hierarchical position may be relevant (e.g.,
  supervisors versus line staff)
• Positions may identify hierarchy and clustering
  at the same time
• Issues for group implementations are similar to
  those for positions

23
(No Transcript)
24
Do groups need leaders?

• May be sufficient to let groups determine
  leaders or leadership preference
• Behavior change issue is controversial
• Behavior change process is controversial
• May be preferred to impose some leadership
  structure
• Behavior change process is accepted
• Goals are well-defined

25
3. Match Leaders to Groups

• Rather than have leaders unattached, assign them
  to people who think they are leaders
• Leadership is local
• Emphasizes homophily between leaders and members
• Builds on naturally occurring networks
• Leaders can be more effective if assigned to
  those who nominate them

26
Borgattis Key Player Program

• Nodes of high in degree may overlap and so just
  selecting on in-degree may not be helpful
• Borgattis Key player program avoids this problem
  somewhat, but it does not directly (yet) indicate
  specifically who covers whom (who is connected to
  whom)

27
Key Player Results

• Baseline Fragmentation 0.000
• Baseline Heterogeneity 0.000
• Initial set (based on betweenness) is 20 12 22
• Fit of initial set 213.000
• Round 1, 2 iterations. Fit 783.000
• Round 2, 1 iterations. Fit 213.000
• Round 3, 1 iterations. Fit 213.000
• Round 4, 2 iterations. Fit 503.000
• Round 5, 3 iterations. Fit 783.000
• Round 6, 2 iterations. Fit 783.000
• Round 7, 2 iterations. Fit 783.000
• Round 8, 1 iterations. Fit 213.000
• Round 9, 1 iterations. Fit 213.000
• Round 10, 1 iterations. Fit 213.000
• Key players are
• 12. 2
• 16. 23
• 20. 27
• Fragmentation 0.158

28
The Effects of a Social Network Method for Group
Assignment Strategies on Peer Led Tobacco
Prevention Programs in Schools
Thomas W. Valente, PhD Beth R. Hoffman,
MPH Anamara Ritt-Olson, MA Kara Lichtman, MA C.
Anderson Johnson, PhD Am. J. of Public
Health Funded by NCI/NIDA, Transdisciplinary
Tobacco Use Research Center
29
Opinion Leaders Individuals Receive the Most
Nominations
Data from Coleman et al. 1966
30
Networked Condition
Sociogram based on ties
Optimal leader/learner matching
31
Tobacco Use Prevention Among Adolescents in
Culturally Diverse California
32
TTURC IRP Project

• Test of a Culturally Tailored Tobacco Prevention
  Curriculum
• Two curricula created and implemented in 16
  middle schools
• Compared against 8 control schools
• CHIPS standard social influences program
• FLAVOR culturally tailored
• Data collected in 6th, 7th and 8th grades

33
Comparison of 3 Conditions
34
Study Design
35
Objectives

• Evaluate the feasibility of a network method for
  identifying leaders and creating workgroups for
  school-based tobacco prevention curriculum.
• Nested within a study of FLAVOR, a culturally
  tailored program, being compared to CHIPS! a
  standard social influences curriculum.
• Determine whether more effective than random
  groups and teacher defined ones.

36
Data
37
  Who are the five BEST LEADERS in this class?
  Think about the five people in this class who
would make the best leaders for working on group
projects. Write up to 5 names on the lines below,
starting with the best leader on the first line.
After you write their name look at the list of
names on the roster that has been provided. Match
the name to the number and write the number in
the boxes. If you cant think of five names in
this class, then leave the extra lines blank. You
can name yourself if you want.  
Also asked who are your five best friends
38
A Network of Leader Nominations
39
Group Assignments for One Network Class
40
Regression Results on Post Program Appeal (Lower
Scores Better) (N1961 k84 Beta Coefficients)
41
Regression Results on Post Program Attitudes
(Lower Scores Better, Beta Coefficients)
42
Susceptibility to Smoke
43
Classroom Level Analysis (Nk84 Beta
Coefficients)
44
1-Year Change in Smoking by Curricula
Implementation Condition
45
AORs for Curricula Implementation Condition on
1-Year Change in Smoking
Regression controls for age, gender, ethnicity,
parent foreign born, parent education, SAS,
parental smoking, and scholastic achievement
46
Results Summary

• Network condition
• was most appealing
• reduced pro-tobacco attitudes
• reduced susceptibility
• Network effect was dependent on curriculum

47
TND Network

• How would a network condition compare to an
  existing evidence-based program?
• TND is a tobacco and drug use prevention
  curriculum tested in multiple setting.
• Created TND Network designed to be TND plus
  interactivity and network method for leader and
  group definitions.

48
TND Network

• Background
• TND evidence based program for reducing
  substance abuse among adolescents in school.
• TND Network modified TND to be more
  interactive, led by trained peer opinion leaders.
• Objectives
• Determine whether TND Network was effective at
  reducing current use
• Would it create deviancy training?

48
49
Study Design
14 Continuation High Schools Recruited for the
Study
Baseline Survey Administered (N938)
Pre-test Surveys
75 Classes Randomized
28 Control
25 TND Networked
22 TND Regular
1 Year Surveys (N541)
49
50
Associations (ß Coefficients) for Study
Conditions on Current Substance Use
50
51
TND Network Increased Substance Use for Students
with Peer Users
Network
TND
Control
UNIVERSITY OF SOUTHERN CALIFORNIA ? INSTITUTE FOR
PREVENTION RESEARCH
52
STEP Replication

• STEP trial method was ineffective
• Lacked personal data so could not test for
  mediators or interactions

53
Valente program or UCINET

• Valente program is helpful in that it can be
  adapted for specific needs
• Can be run on multiple networks
• Possible also to implement the idea in UCINET
• Also possible to implement on the fly with a
  show of hands, for example.

54
Network Method in UCINET (sort of)
55
Conclusions

• Network methods were effective at changing short
  term outcomes
• First turn-key network-based interventions
• Network implementation methods are sensitive to
  curriculum.

56
Implementation Issues

• Have assignments and information readily
  available, we had 1 week or less to collect
  network data and return the leaders and groups
• Concerns about confidentiality

57
Leaders v. Popular Students

• The strategy for working with peer leaders has
  some merit.
• Data also show that popular students were still
  more likely to become smokers at 1 year.
• However, we used peer leaders as those were
  nominated to be peer leaders, not those most
  frequently nominated as friends.
• Future interventions may want to use friendship
  networks, not leader networks for interventions.

58
4. Snowballs, Contact Tracing Respondent Driven
Sampling

• Epidemiologists have employed contact tracing for
  years often data are not published or publicly
  available
• Several studies using snowball methods to recruit
  a sample
• Some instances using snowball methods to recruit
  intervention group
• http//www.respondentdrivensampling.org/

59
Vaccine Preparedness Network Study
59
60
Network linkages among indexes and alters in
which at least one alter was enrolled (794 links,
59.2).
61
Latkin et al. (2009)

• Network based peer education program among IDUs
  in
• Chang Mai, Thailand Philadelphia PA
• 414 Indexes with 1,123 participants (2.71 per
  network)
• Intervention consisted of 6 2-hour small group
  sessions over 4 weeks (indexes got 2 booster
  sessions)

62
Latkin et al., Results
63
Snowball -Implementation Issues

• Ties are homophilous
• Need coupons as incentives for recruiters
• 10 works fine for most applications but it is
  probably going up to 20
• Need to ID coupons
• Challenge to keep ID numbers straight
• Might be able to use automated debit cards

64
4 Types of Data Collection Strategies
65
5. Rewiring Networks

• Make network(s)
• More/less cohesive
• More/less centralized
• More/less transitive
• Finding links that span structural holes (Burt)
• More generally, finding the link or links that
  can or should be changed

66
Experiences

• Valdis Krebs has considerable experience working
  in organizations
• Cross et al. (2003) paper showed network change
  after intervention
• Many studies may be proprietary

67
Rewire Calculations

• Calculate original metric
• Change link
• Delete existing link
• Add non-existence link
• Calculate new metric
• Put difference in new matrix

68
Change Matrix ScoresPositive NumbersCohesion
Increase When Added Negative NumbersCohesion
Decrease When Deleted
69
Dyadic List
Links Added 7.0000 31.0000 0.0000 1.0000
0.0078 31.0000 7.0000 0.0000 1.0000 0.0078
7.0000 24.0000 0.0000 1.0000 0.0071 24.0000
7.0000 0.0000 1.0000 0.0071 14.0000 7.0000
0.0000 1.0000 0.0071 7.0000 14.0000 0.0000
1.0000 0.0071 25.0000 7.0000 0.0000 1.0000
0.0066 7.0000 25.0000 0.0000 1.0000 0.0066
7.0000 10.0000 0.0000 1.0000 0.0065 10.0000
7.0000 0.0000 1.0000 0.0065 . . .
Links Deleted 2.0000 37.0000 1.0000 0.0000
-0.0089 37.0000 2.0000 1.0000 0.0000
-0.0089 29.0000 27.0000 1.0000 0.0000
-0.0078 27.0000 29.0000 1.0000 0.0000
-0.0078 7.0000 2.0000 1.0000 0.0000
-0.0034 2.0000 7.0000 1.0000 0.0000
-0.0034 23.0000 26.0000 1.0000 0.0000
-0.0026 . . .
70
Get Node-Level Measure

• Sum row and column scores (Vitality Index)
• Average row and column scores

71
Granovetters SWT Bridges
72
Links vs. Nodes

• Can aggregated change scores to the nodes so you
  know which nodes to target.
• It may be easier to work with nodes than with
  links.
• Or it may be preferable to work with links rather
  than nodes.

73
Bridges Potential Bridges

• Bridges
• Systematically delete each link
• Calculate change in APL
• Sort links by the degree of change
• Potential Bridges
• Systematically add each possible link
• Calculate change in APL
• Sort links by the degree of change

74
Implementation Issues

• Do we know which network structural metric to
  maximize (e.g., density example)?
• Is it a zero-sum (i.e., does one need to keep the
  of links constant)?
• How to control for naturally occuring network
  dynamics (people enter/leave the network, change
  affiliations, etc.)?

75
Implementation Issues (2)

• Simply re-wiring links will probably not work
• There are reasons networks are the way they are
  (e.g., people make bonds with those they like).
• Network dynamics people come and go and this
  will affect overall network properties

76
6. Other Methods

• Triadic Changes
• Identify low threshold adopters
• Reaching critical mass
• Simply collecting data may be important
• Reporting back to the group and discussing
  networks may be important.
• Others?

77
Response Rates

• Traditionally been a lot of concern having less
  than 100
• Evidence shows that network data still helpful
  with reasonable RR
• Costenbader Valente (2003)
• Degree is very robust to missing data
• Many centrality measures are robust to missing
  data
• Borgatti et al. (2005)

78
Advantages of Network Methods (1)

• Capitalize on existing interpersonal
  relationships
• Use community input
• Establishes a learning organization /community
• Build social capital
• Are Empowering

79
Advantages of Network Methods (2)

• Can be replicated
• Fidelity can be measured
• Expands array of intervention options
• Creates data!

80
Caution

• The effects of network interventions may be a
  product of soliciting community input and
  involvement and it is this empowerment which
  creates positive social change not the effects of
  using network data.

81
Segmentation

• Geographic
• Demographic
• Psychographic
• Sociometric
• The messenger is the message
• The messenger is as important as the message