1

Network Methods for Behavior Change

Thomas W. Valente, PhD

Professor

Preventive Medicine, Keck School of Medicine

University of Southern California

tvalente@usc.edu

22

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

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 Interventions1. Identify opinion leaders or key players to act as change agents 2. Create network-based groups/positions3. Identify leaders within groups or match leaders to groups4. Snowballing / Contact tracing / Respondent Driven Sampling5. Rewire Networks

1. More/less cohesive 2. More/less centralized3. More/less dense4. Change core-peripheriness5. Etc.

6. Other: 1. Triadic Structures2. Identify low threshold adopters3. Reaching critical mass4. Reporting back to group/dialogue5. 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

0

20

40

60

80

100

1 2 3 4 5 6 7 8 9 10

Time

Per

cen

t A

do

pte

rs

Opinion Leaders

Random

Marginals

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 (N=255, 52 networks, retention>90%)

Kelly, Amirkhanian, Kabakchieva et al., BMJ, 2006

Unprotected Intercourse with Casual Partners

010203040506070

baseline,n.s.

3-monthfollowup,p=0.02

12-monthfollowup,p=0.009

Assessment Points

% p

ast

3 m

on

ths

Experimental

Control

9

Young MSM Egocentric Network HIV Prevention Trial, Bulgaria/Russia (n=276, 52 networks, retention

>84%) Amirkhanian, Kelly, Kabakchieva et al., AIDS, 2005

Any Unprotected Intercourse

01020304050607080

baseline,n.s.

3-monthfollowup,p=0.0001

12-monthfollowup,p=0.06*

Assessment Points

% p

ast

3 m

on

ths

Experimental

Control

* 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

1. Use different types of opinion leaders at different stages:

– In-degree– Betweenness – Closeness

2. Use different types of opinion leaders for different groups:

– In-degree in highly cohesive groups– Betweenness in fractured groups

13

1

10

11

12

13

14

15

16

17

18

19

2

20

21

22

23

24

25

26

27

28

29

3

30

31

32

33

34

35

36

37

4

56

7

8

9

14

FREEMAN'S DEGREE CENTRALITY MEASURES----------------------------------------------------------------------

Diagonal valid? NOModel: 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

Method Technique

1. Celebrities Program recruits well-known people to promote behavior.

2. Self-selection Staff requests volunteers in-person or via mass media and those who volunteer are selected.

3. Self-identification Surveys are administered to the sample, and questions measuring leadership are included. Those scoring highest on leadership scales are selected.

4. Staff selected Program implementers select leaders from those whom they know.

5. Positional Approach Persons who occupy leadership positions such as clergy, elected officials, media and business elites, and so on are selected.

6. Judge’s Ratings Persons who are knowledgeable identify leaders to be selected.

7. Expert Identification Trained ethnographers study communities to select leaders.

8. Snowball method Index cases provide nominations of leaders or are in turn interviewed until no new leaders are identified.

9. Sample Sociometric Randomly selected respondents nominate leaders and those receiving frequent nominations are selected.

10. Sociometric All (or most) respondents are interviewed and those receiving frequent nominations are selected.

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

1

10

11

12

13

14

15

16

17

18

19

2

20

21

22

23

24

25

26

27

28

29

3

30

31

32

33

34

35

36

37

4

56

7

8

9

20

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

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

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

Borgatti’s Key Player Program

• Nodes of high in degree may overlap and so just selecting on in-degree may not be helpful

• Borgatti’s 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• Heterogeneity = 0.156

28

The Effects of a Social Network Method for Group Assignment Strategies on Peer Led Tobacco Prevention Programs in Schools

Thomas W. Valente, PhDBeth R. Hoffman, MPH

Anamara Ritt-Olson, MAKara Lichtman, MA

C. Anderson Johnson, PhD

Am. J. of Public HealthFunded by NCI/NIDA,

Transdisciplinary Tobacco Use Research Center

29

Opinion LeadersIndividuals Receive the Most Nominations

110

11

12

13

14

15

16

18

192

20

21

22

2324

25

26

27

28

29

3

30

31

32 33

37

38

39

4

40

41

42

43

44

5

51

53

56

57

58

6

7

8

9

Data from Coleman et al. 1966

30

110

11

12

13

14

15

16

18

192

20

21

22

2324

25

26

27

28

29

3

30

31

32 33

37

38

39

4

40

41

42

43

44

5

51

53

56

57

58

6

7

8

9

1

2

3

4

5

6

78

9

10

11

12

13

14

15

16

17

18

19

20

21

2223

24

25

26

27

28

29

30

31

32

33

Networked ConditionSociogram 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

Condition Description

Opinion Leader &

Random

Leaders chosen by students and randomly assigned to groups

Teacher Leaders and their groups are defined by the teacher

NetworkedLeaders chosen by students and assigned to groups of students that chose them

34

Study Design

Chips Flavor Total

Schools 8 8 16

Opinion Leader

Teacher Network Opinion Leader

Teacher Network

Classes 15 12 13 16 16 15 87

Students 359 281 310 363 349 298 1960

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

Pre-test

March, 2001

Curric.

March-May

Post-test

May, 2001

1 Year Post

May, 2002

•Attitudes–General Attitude (.87)–Self Efficacy (.68)–Social Consqnces (.54)

•Susceptibility to smoke

•Smoking Attitudes–General Attitude (.90)–Self Efficacy (.71)–Social Consqnces (.56)

•Susceptibility to smoke

•Program appeal–Like Curriculum (.89)–Friend Support (.79)–Like Program (.81)

•Susceptibility•Smoking Initiation

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 can’t think of five names in this class, then leave the extra lines blank. You can name yourself if you want. 

  FIRST NAME LAST NAME ROSTER NUMBER  

1      

2      

3      

4      

5      

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)

(N=1961; k=84; Beta Coefficients)

Program Friends Curriculum

Female 0.15** 0.18** 0.13**

Smoking Prevalence 0.10** 0.14** -0.06

FLAVOR -0.01 0.02 0.0

Teacher Condition -0.08* -0.03 -0.07

Network Condition -0.14** -0.09** -0.04

Network*FLAVOR 0.05 0.05 0.0

R2 4 5 2

*p<.01; **p<.001

41

Regression Results on Post Program Attitudes (Lower Scores Better, Beta Coefficients)

Outcomes

Smoking Attitude

Self

Efficacy

Social Consequences

Baseline attitude 0.60** 0.52** 0.55*

Female -0.03 0.02 0.03

Smoking Prevalence 0.08* 0.12* -0.02

FLAVOR 0.01 0.01 0.07**

Teacher Condition -0.04 0.01 0.0

Network Condition -0.07* -0.09** -0.01

Network*FLAVOR 0.06 0.04 0.01

R2 39 29 31

*p<.01; **p<.001

42

Susceptibility to Smoke Adjusted

Odds Ratio

Susceptibility at baseline 7.30***

Female 0.87

Smoking Prevalence 1.08*

FLAVOR 0.95

Teacher Condition 0.95

Networked 0.44***

Network*FLAVOR 2.17*

N=1960; *p<.05; **p<.01; ***p<.001

43

Classroom Level Analysis(N=k=84; Beta Coefficients)

Smoking Attitude

Self

Efficacy

Social Consequences

Percent

Intention

Baseline attitude 0.70** 0.45** 0.72** 0.54**

Smoking Prevalence 0.12 0.34 -0.06 0.20

FLAVOR 0.12 0.05 0.21 0.03

Teacher Condition -0.07 0.01 0.01 -0.01

Network Condition -0.16** -0.24* -0.11 -0.38***

Network*FLAVOR 0.12* 0.07 0.09 0.28*

R2 62 38 51 48

*p<.01; **p<.001

44

1-Year Change in Smoking by Curricula & Implementation Condition

0

2

4

6

8

10

12

Total Random Network Teacher

Implementation Condition

Cha

nge

in S

mok

ing

(%)

Chips

Flavor

45

0

1

2

Curriculum & Implementation Conditions

Ad

jus

ted

Od

ds

Ra

tio

sAORs 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

28 Control 25 TND Networked22 TND Regular

14 Continuation High SchoolsRecruited for the Study

Baseline Survey Administered (N=938)

Pre-test Surveys

75 Classes Randomized

1 Year Surveys (N=541)

Study Design

49

50

Associations (β Coefficients) for Study Conditions on Current Substance Use

Tobacco Alcohol Marijuana Drugs Total

TND 0.07 0.21 0.09 0.03 0.06

Network -0.40 -0.23 -0.64** -0.37** -0.37**

Network*

Peer Use

0.16 0.25 0.34* 0.28* 0.19**

*p<0.05; **p<0.01

Regression controls for baseline level, age, grade, gender, ethnicity, # friends, # friends in school, ties sent, ties received, social support, # friends who engage in behavior.

50

51UNIVERSITY OF SOUTHERN CALIFORNIA INSTITUTE FOR

PREVENTION RESEARCH

TND Network Increased Substance Use for Students with Peer Users

-1.5

-1-.

50

.5C

hange in S

ubsta

nce U

se

0 2 4 6 8Peer Use

Control TND

TND Netw ork

Network

TND

Control

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

59 MCA Index Patients

307 Social Network Members Named(306 Individualsa)

31 Named Alters Enrolled

31 Alters Enrolled – Not Previously Named

357 Social Network Members Named(339 Individualsb)

291 New Social Network Members(283 Individualsc)

30 Previously Named by Index Patients (25 Individualse)

17 Enrolled Alters(17 Individuals)

19 Index Patients(14 Individualsd)

NOTE

aOf the 306 unique individuals: 305 were named once and 1 was named twice.bOf the 339 unique individuals: 322 were named once, 16 were named twice and 1 was named three times.cOf the 291 unique individuals: 275 were named once and 8 were named twice.dOf the 14 unique individuals: 10 were named once, 3 were named twice, and 1 was named three times.eOf the 25 unique individuals: 20 were named once and 5 were named twice.

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 Numbers=Cohesion Increase When Added

Negative Numbers=Cohesion Decrease When Deleted

1 2 3 4 5 6 7 8…1 0 0.0011 0.0006 0.0005 0.0039 0.0005 -0.0012 -0.00042 0.0011 0 -0.0012 0.0008 0.0037 0.0015 -0.0034 0.00093 0.0006 -0.0012 0 0.0004 0.0037 0.0016 0.0011 0.00054 0.0005 0.0008 0.0004 0 0.0039 0.0009 -0.0011 0.00045 0.0039 0.0037 0.0037 0.0039 0 0.0043 0.0059 0.00426 0.0005 0.0015 0.0016 0.0009 0.0043 0 -0.0015 0.00087 -0.0012 -0.0034 0.0011 -0.0011 0.0059 -0.0015 0 -0.0018 -0.0004 0.0009 0.0005 0.0004 0.0042 0.0008 -0.001 0

…

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 -0.008937.0000 2.0000 1.0000 0.0000 -0.008929.0000 27.0000 1.0000 0.0000 -0.007827.0000 29.0000 1.0000 0.0000 -0.00787.0000 2.0000 1.0000 0.0000 -0.00342.0000 7.0000 1.0000 0.0000 -0.003423.0000 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

Granovetter’s 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 changePotential 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

1. Triadic Changes

2. Identify low threshold adopters

3. Reaching critical mass

4. Simply collecting data may be important

5. Reporting back to the group and discussing networks may be important.

6. 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