Social Network Analysis Thomas W' Valente

1
Social Network Analysis Thomas W. Valente

• Introduction (Mon.)
• 1) Introduction to Major Concepts Terminology
  History (Chapters 1 2)
• 2) Data Collection, Management Basic Measures
  (Chapter 3)
• 3) Personal/Ego-centric Methods (Chapter 4)
• Measures Methods (Tues. Wed.)
• 4) Centrality Centralization (Chapter 5)
• 5) Relational Measures Models (Chapter 6)
• 6) Structural Measures Models (Chapter 7)
• 7) Network Level (Chapter 8)
• Applications (Thurs.)
• 8) ERGM Agent Based Modeling (Chapter 9)
• 9) Diffusion of Innovations/Behavior Change
  (Chapter 10)
• 10) Review Summary (Chapter 12)

1
2
Network Analysis - Introduction and
HistoryRelations vs. Attributes

• Relations
• Who you know
• Who you talk to
• What you talk about
• What are your networks attributes
• How do their attributes affect you
• How does your network constrain or enhance your
  opportunities

• Attributes
• income
• education
• residential location
• gender
• ethnicity
• occupation

2
3
Rationale for Network Analysis

• Dissatisfaction with attribute theories of
  behavior.
• Desire to model human interaction.
• Fit methodology to known theories of contagion
  (disease spreads through contact so model
  contact).
• Opportunity to explore mathematically rigorous
  theories of human behavior.
• Develop targeted interventions.

3
4
Social Networks

• INSNA
• www.insna.com
• Socnet list service
• UCINET list service
• Social Networks
• J. of Social Structure
• Connections
• Sunbelt (2009 San Diego March 15-20 2010 Riva
  del Garda, (near Trento) Italy)

4
5
Primary Applications in PH

• HIV/STDs transmitted via sexual contact networks.
• FP RH programs to understand how information
  and persuasion flow through social networks.
• Community-based promotional programs.
• Respondent driven (snowball) sampling.
• Inter-organizational collaboration, coordination,
  and cooperation.
• Improve interventions such as school-based health
  promotion programs.

5
6
Major Research Questions

• Small world networks
• Scale free networks
• Social capital
• Homophily
• Diffusion/contagion/social influence
• Centrality (popularity, leadership)
• Individual v. network measures
• Network dynamics and stability
• Efficient network forms
• Agent based models (ABM)
• Interventions
• Algorithms Measures

6
7
1. Favorite Small World Story

• Coach Jordan
• Different types of small world phenomenon
• 2 strangers find common person
• 2 acquaintances find common person not before
  realized
• People who know one another meet in an unexpected
  place
• When we discover these connections we are
  surprised and reassured

7
8
Small World Networks

• Characterized by shorter path lengths than would
  be expected.
• Also have local clustering
• Small world research first conducted by Stanley
  Milgram
• Dormant for decades then revitalized by Watts
  Strogatz

8
9
2. Scale Free Networks

• Barabasi studied how networks grow
• Preferential attachment New nodes have
  preferences for how they join a network.
• Nodes prefer to attach to central nodes
• Gives rise to very central nodes
• Distribution of degree ( of connections a node
  has) is highly skewed
• Internet classic example because if you want
  people to visit your site you link to a high
  traffic site

9
10
Graph of degree distribution power law
10
11
3. Social Capital

• People have human (abilities), material
  (resources) and social capital (social
  connections).
• SC is the resources available in ones network-
  useful for getting a job, access to health care,
  information, etc.
• SC also exists at the community level
• SC can be negative

11
12
Moores Study of SC in PH
12
13
4. Homophily

• SW, SF, SC move from macro to micro, but what
  drives all of these are network choices at the
  individual (micro) level.
• People tend to affiliate with others like
  themselves.
• Ties are often among those of the same sex,
  ethnicity, social class, and so on

13
14
Friendship Network 6th Graders
14
15
5. Diffusion/contagion

• Ideas and practices spread through networks.
• As friends do something, you are more likely to
  do something
• Contagion is the special case of diffusion in
  which only contact is required for
  adoption/spread
• Diffusion is really the special case of social
  influence models
• Series of classic studies on diffusion and
  thresholds

15
16
Diffusion When Adopters Persuade Non-adopters at
a Rate of One Percent (Homogenous Mixing)
16
17
Diffusion for Random Mixing
17
18
6. Centrality/Popularity

• One diffusion driver is the behavior of opinion
  leaders
• OLs both reflect and drive the diffusion process
• Identify OLs by those who receive many
  nominations (typically top 10-15)
• Other measures of centrality- who is located at
  the center of the network

18
19
Centrality Measures

• Degree
• Betweenness
• Closeness (Integration, radiality)
• Flow
• Eigenvector
• Power
• Others (Complement)
• Many measured with different algorithms

19
20
7. Individual v. Network Level

• Individual behaviors are not independent of the
  network within which the behaviors occur.
• Individual network position is not independent of
  the network structure being central in a
  centralized network is different than being
  central in a decentralized network.

20
21
8. Network Dynamics Actor Co-Evolution Models

• SIENA Simulation Investigation for Empirical
  Network Analysis
• Simultaneously model individual and network
  changes over time.
• Compare influence versus selection
• Control network structural influences on
  behaviors and non-independence of data.

21
22
9. Efficient network forms

• What is best network?
• Inter-organizational collaboration or
  organizational behavior researchers want to know
  Is this structure optimal?
• Naïve view is that more density is better than
  less- not so.
• How centralized should a network be? Some cases
  centralization is good, others, not so good.
• Level of reciprocity, transitivity, clustering,
  overall path length, etc.

22
23
Association between Adoption and Change in
Density is Negative
Outcome Change
Change in Density
23
24
Efficiency?

• Finding efficient forms may completely depend on
  behavior being studied?
• What is tradeoff between individual satisfaction
  and network-level performance?
• Can optimal forms be created or are the most
  optimal ones those that exist?

24
25
10. Agent based models/simulations

• Often useful to simulate conditions to get a
  sense of what happens.
• One can create hypothetical networks based on
  some parameters (e.g., ave. of links per
  person, centralization, reciprocity)
• Can also simulate what happens to this network
  over time
• And what happens to behaviors within this network.

25
26
Simulated Diffusion in a Real Random
Network(Valente, T. W (in press). Models and
methods for innovation diffusion. In P.
Carrington, J. Scott S. Wasserman (Eds.)
Models and Methods in Social Network Analysis.
New York Cambridge University Press.
26
27
11. Interventions

• Network data may help improve health promotion
  programs
• The Messenger is the Message
• Sociometric segmentation
• Identify opinion leaders (over 15 studies)
• Identify leaders and match them to nominees
• Identify groups and find leaders within groups
• Snowball/Respondent driven sampling for
  recruitment

27
28
Opinion Leaders Individuals Receive the Most
Nominations
Sociogram based on ties
Optimal leader/learner matching
1(a)
1(b)
28
29
12. Algorithms Measures

• New algorithms to measure well known concepts are
  continuously being created
• E.g., Newman Girvan method for identifying
  groups
• Developing new algorithms also leads to new
  insights on networks, how they are organized and
  function
• E.g., complement centrality (the complement of a
  graph is its reverse, ties for non-ties and
  non-ties for ties) makes us think about the
  absence of ties

29
30
Measures

• Also, how do we know a network measures what we
  think it measures, validity and reliability are
  issues are to contend with
• Do different network methods return the same
  results.
• How are missing or incomplete data handled and
  are network measures robust?

30
31
Average Correlation between Centrality Measure
and a Random Subsample
31
32
Network Affects on Behavior

• This rich tapestry of approaches, models, and
  ideas rarely addresses behavior explicitly
  (except diffusion of course)
• This course will focus on the implications of
  each of these for behavior
• I will try to be brief in my description and
  explanation of the model or concept reserving
  time for how it affects behavior

32
33
SW, SF, SC

• Small world networks accelerates the spread of
  ideas, practices, disease
• Bridges, which create small world networks, make
  it possible for diseases to spread between
  continents and for information to travel quite
  widely
• The small world phenomenon accelerates the trust
  building process, and it can accelerate the sense
  of community.

33
34
Scale Free

• Indicates that network structures are likely to
  reinforce themselves
• Once at the top, you can stay at the top.
• It can be difficult to re-arrange networks.
• SF means many networks will be very centralized,
  have ties directed to one or a few people
• Centralization is likely to create authoritative
  structures.

34
35
Social Capital

• Social capital provides resources that people can
  use to advance their economic or social well
  being.
• We all know its who you know, not what you know.
  SC provides an attempt to theorize about social
  resources
• SC can be negative
• Do communities have SC?

35
36
Homophily/Diffusion

• Because people associate with others like
  themselves, their networks are often closed to
  outside information
• My friends think and do the same things I think
  and do
• Even it they dont, I project on to them
• Makes diffusion a difficult and slow process
• Yet, once someone I know and am close to adopts
  something, I am likely to find it relevant for me
• Strong ties important for adoption, weak ties
  for information and network level diffusion.

36
37
Centrality

• This is why central members are so important.
• They are seen as the same yet different than
  everyone else in the network
• Little hope to get diffusion going unless central
  members embrace the idea
• Yet, most ideas start on the periphery of the
  network because they are freed from social norms.
• Bridges

37
38
First Contact Diffusion (Rumor)/Random Seeds
38
39
History Development

• Network analysis often thought to have begun with
  Moreno (1934) but Freeman notes that many early
  sociologists studied networks
• Freeman has criteria for what constitutes a
  network study
• Structure
• Positions
• Graph

39
40
History (from Knoke Kuklinski)

• Gestalt psychoanalytic tradition
• Kurt Lewin - Conducted group dynamics studies and
  founded the Group Dynamics Center at U. of
  Michigan Field Theory
• Moreno 1934 Who Shall Survive chronicled effects
  of networks
• on problem children and attempted to use
  sociograms to understand child development.
• Heider - Balance theory
• Festinger - Cognitive Dissonance - social
  influences on perceptions
• Cartwright - Graph Theory
• Harary - Graph Theory
• Bavelas - Centrality studies
• Homans - Bank Wiring Room - Hawthorne study
• Hawthorne Yankee City studies
• Today U. Michigan Institute for Systems Science

40
41
Scotts History (Chapter 2)

• Manchester Anthropologists - Barnes, Mitchell,
  Bott, Boissevain Studies in African communities,
  structuration
• Harvard Group - Harrison White --gt Networks
  decomposed into roles and positions based on
  equivalence of relations with the rest of the
  network. White trained or influenced
  Granovetter, Bonacich, Breiger, Carly, Wellman,
  Burt?, and many others

41
42
Modern History

• Hawaii meetings in the mid-1970s
• INSNA created by Wellman, Connections founded,
  Sunbelt conferences started in 1981 (San Diego,
  Santa Barbara, Tampa Bay, New Orleans,
  Charleston, Europe).
• Annual Meeting rotates (West Coast, East Coast,
  Europe)
• Social Networks initiated in 1981 by Lin Freeman
  Freeman Brieger for 5 years now Snijders
  Doreian.
• Connections published items of interest to INSNA,
  gossip, abstracts, peer reviewed articles
  currently edited by Valente Coronges
• J. of Social Structure initiated 2001 electronic
  only (Carnegie Mellon U.)

42
43
Knoke Kuklinski (pages 7-21)

• 1. Two insights
• a. structure as roles positions
• b. attributes versus relations
• 2. Sampling units
• a. individuals
• b. groups
• c. organizations
• d. nations
• 3. Relational form
• a. ties
• b. joint involvement

43
44
Knoke Kuklinski (pages 7-21)

• 4. Relational content
• a. communication
• b. transactions
• c. kinship
• d. etc.
• 5. Relational measurement
• a. binary vs. valued
• b. symmetric vs.asymmetric (reciprocated)

44
45
Knoke Kuklinski (pages 7-21)

• 6. Measures
• a. connectedness
• b. density
• c. centrality
• d. integrativeness
• 7. Data Collection

45
46
Knoke Kuklinski 22-35

• Sampling and Measurement are issues in all
  research. In network research, boundary
  specification is an additional problem which is
  explicit.
• Realist approach researchers take on the
  subjective perceptions of actors. In such cases
  mutual relevance is important.
• Nominalist approach researchers use their own
  conceptual framework. Examples include
  laboratories (such as MIT's), small groups,
  organizations, villages, etc.
• Problems w/ SNA
• 1. restricted to small groups
• 2. doesn't solve boundary problem
• 3. Independence of observations is violated

46
47
Knoke Kuklinski 22-35

• Sampling in Large Populations
• 1. Granovetter - draw a random sample and list
  these names and see if respondents can name
  others. Does not produce network data, rather
  provides parameter estimates for the population.
• 2. Subgroup approach (Burt, Beniger) - develop a
  typology of attributes relevant to the object of
  study. For example, we might look at the
  distribution of education, age, occupation, race,
  etc. for an individual's ego-centered network.
  Most importantly what is the distribution of the
  object of study in ego's network, how many use
  FP, immunization, etc.
• Measures (1) Direct observation, (2) archival
  records, (3) self-report surveys. Criticism of
  Bernard Killworth, but remember their studies
  were not on representative samples.
• To reduce error (1) Precise and specific, and
  (2) avoid interviewer and interviewee fatigue.

47
48
History Unchained

• There had been social scientists who were
  comfortable with matrix manipulation methods.
• Developed tools and ideas which gained traction
  in the 1950s.
• Sociometry developed (Now J. of Soc. Psych.)

48
49
Unchained (2)

• SAS-issification (or SPSS) of social sciences
• Medical model of RCTs
• Pollsters desire to make population estimates
• Science of social science demanded that social
  scientists use random sampling methods and make
  population inferences

49
50
Unchained (3)

• Convergence of diffusion scholars discovered
  importance of social networks
• There was a crisis confronting the diffusion
  paradigm but no one saw it
• Study of diffusion and social context abandoned
  for political reasons (see Valente Rogers)

50
51
Unchained (4)

• Thus, networks, for the most part, was dormant
• Freeman and others kept it alive thru 70s and got
  INSNA going.
• Somewhat moribund late 1980s and early 1990s
• HIV/AIDS money and researchers gave it new life

51
52
Unchained (5)

• Computer Networks and WWW increased relevance
• Invasion of the Physicists (Bonacich, 2005)
  created publicity and a new threat of success
• Social Networking become important
• Today field is robust and growing and the idea of
  connectedness seems to have been accepted
  overnight.

52
53
Unchained (6)

• Hierarchical linear models
• Group randomized designs
• Clustering and survey design issues
• Random selection is a myth

53
54
What does this mean?

• Any study should consider social or network
  context
• Can be always be measured
• Studies that do not recognize, acknowledge and
  contend with collinearity, clustering and
  contextual affects are suspect

54
55
Behavioral Science

• Homophily Selection
• Social Learning Theory
• Selection
• Norms
• Perceived Social Consequences
• Lifespan Approaches

56
Public Health Applications

• Social Support
• AIDS/STDs Reproductive Health/FP
• Community Health
• Inter-organizational Relations
• Provider Performance

57
Week 2 Measures Methods

• There are a variety of ways to collect network
  data.
• Variety of methods creates confusion over what
  constitutes network analysis
• Type of method dictates type of analysis possible
• Tradeoff between depth and breadth

57
58
How do You Measure Context?Network Analysis Data
Types
58
59
1.A-B General Survey

• Have you been to a MD?
• Have you seen Dr. Jones?

59
60
2.A Roles

• How often do you turn to _____ for social
  support
• Mom
• Dad
• Sister
• Brother
• Friend
• Friend
• Etc.

60
61
2.B Egocentric

• Please tell me, within the last six months, the
  first names or initials of up to 5 people you
  talk to most often about important matters?

61
62
Ego-centric Network Data are Node-basedAnalyzed
completely in SPSS/SAS
Same or Higher Econ. Status
Primary Schooling
Older
Live Outside of Village
Confidants Friends
Kin (sister- in-law, brother)
62
63
Dyadic

• One way to analyze ego-centric data is to convert
  the data to dyadic
• In dyadic data, each relationship reported by the
  respondent is a case in the new dataset.
• Dyadic data
• Increase sample size
• Facilitate analysis
• Reflect theoretical processes of study (e.g., is
  behavior associated with homophily)

63
64
3. Network Sampling Strategy

25 Indexes 250 1st degree alters identified 100
alters are interviewed 2,000 2nd degree Alters
identified none interviewed in pilot study

Red nodes are interviewed alters Blue node are
not Interviewed
64
65
4.A Nominations Method
  Who are the five BEST FRIENDS in this class?
  Think about the five people in this class who
are your closest friends. Write up to 5 names on
the lines below, starting with your best friend
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.  
65
66
4.B Roster
66
67
Method 4 (A B) Census or Saturation Sampling

• The most typical kind of network methodology
• Usually what people think of when we say
  networks.
• These data can be graphed and analyzed using
  matrix methods

67
68
2007 Network Class

• dl n25 format nodelist
• row labels
• 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
• Data
• 1 22 3 10
• 2 13 16 22 25
• 3 1 10 22 18
• 4 22 19
• 5 22 2 7 9
• 6
• . . .
• 19 22 4 14
• 22 4 9 13 5 6 7 19
• 23 9 13 25 15 16
• 24
• 25 13 16 23 15 9 2 22

68
69
Graph of 07 class
69
70
2006 Class w/ Depts.
70
71
VNA Netdraw File

• dl n19 format VNA
• node data
• Id DEPT ROLE
• 1 SPPD STUDENT
• 2 PM STUDENT
• 3 PM POST-DOC
• 4 SPPD STUDENT
• 5 PM STUDENT
• .
• 15 OTHER STUDENT
• 16 SPPD STUDENT
• 17 SPPD FACULTY
• 18 PM FACULTY
• 19 PM STUDENT
• tie data
• From to know
• 1 4 1
• 1 6 1

71
72
Pajek File for 96 Class
Here are the nodes and how many

• Vertices 27
• 1 "ALI"
• 2 "AGREE"
• 3 "AMIN"
• 4 "BOULAY"
• 5 "DAVIS"
• 26 "BRAHMBHA"
• 27 "DRENNAN"
• arcslist
• 1 23 10 25 26 6
• 2 12 23 5
• 3 20 8 14 26 4
• 4 23 5 14 8 3
• 5 6 11 13 20 21
• 6 13 3 4 5 12
• 20 4 3 14 5 23
• 21 4 6 23 5 11

Here come the link descriptions
72
73
Nominations vs. Roster

• Nominations
• Less demanding for Respondents
• Rank order of nominations has meaning, 1st is
  strongest, 2nd next strong etc.
• If network is too dense can restrict to fewer
  alters (e.g., only take 1st 3)
• Easier data entry and management (only need
  columns for 5 variables rather than N variables
• Unaided measure

73
74
Roster

• Roster
• Requires respondent to read all names
• May be biased by way list is formed, alphabetical
  for example
• If someone is left off list creates problems
• Cannot rank order choices, though you can put in
  a rating scale
• Primes respondent (sensitized)

74
75
5. Two-Mode Data

• Briegers paper on The Duality of Persons and
  Groups 1974.
• People who attend the same events or belong to
  the same groups are linked through their joint
  participation.
• Matrix manipulation of the affiliation matrix
  provides two matrices indicating dual nature of
  persons and groups.

75
76
Matrix Manipulation for 2-mode
A A B (Person by Person contact matrix) A
A C (Dept. by Dept. count matrix)
76
77
Transposing a Matrix
Matrix A
Matrix A (transpose)
AxA Person by Person AxA Event by Event
77
78
4 Types of Data Collection Strategies
78
79
2 Types of Network Data
Ego-Centric
Sociometric
9
10
1
12
3
1
2
4
11
5
7
2
8
6
79
80
Data Management
Data Collected
Roles or Egocentric
Analyze in SAS/STATA only
Seq./Census/2-mode
Use Netdraw or creates Pajek file
Make Sociograms
1
Create UCINET Import file and Import
Number of Networks
Read Interpret output
3
Save output logs and edit to import into STATA
Program Measures in GAUSS/SAS-IML MATA/R/S-Plus
Merge Variables with SAS/STATA data file
80
81
Data Management (2)

• SAS modules being created so that network
  variables can be created in SAS and easily merged
  with attribute data (Jim Moody _at_Duke)
• I use GAUSS and create network variables in GAUSS
  then merge variables with my STATA file (STATA
  now has MATA).
• Critical when one has multiple networks (e.g., 50
  classes and 3 networks in each class).

81
82
Scott Chapter 4 (pgs. 66-84)

• Graph Theory
• Points connected by lines
• Nodes connected by edges
• Nodes connected by ties
• Individuals connected by relations
• Organizations connected by shared attributes (or
  exchange)
• There is no one "correct" way to draw a graph
  in fact there are an infinite number.
• Types of lines
• non-directed vs. directed
• binary vs. valued
• uni-plex vs. multiplex ( of relations)

82
83
Network Definitions

• adjacent - two points connected by a line
• neighborhood - set of adjacent points
• degree - numerical measure of neighborhood
• walk - sequence of lines
• path - lines in a unique walk
• length - of lines in a path
• distance - length of shortest path (geodesic)
• in-degree - of lines directed toward point
• out-degree - of lines directed away from a point

83
84
Personal Network Density

• maximum of ties n(n-1)/2
• (e.g.., n5--gtmax.10 n6--gtmax.15)
• density total of lines divided by the maximum
  possible
• total should exclude ego's lines
• density lines/(n(n-1)/2
• Example (pg. 75) 2/4(3)/2 .33
• Not clear what the denominator should be for
  valued ego-centric density calculations.
• Density is dependent on the size of a graph and
  this prohibits comparison of density measures
  between graphs. Larger graphs, all things equal,
  will have lower densities than small graphs.
• Density and Community

84
85
Density

• 1. symmetric data
• maximum of ties n(n-1)/2
• (e.g.., n5--gtmax.10 n6--gtmax.15)
• density total of lines divided by the maximum
  possible
• density lines/(n(n-1)/2
• 2. directed data
• maximum of ties n(n-1)
• (e.g.., n5--gtmax.20 n6--gtmax.30)
• density total of lines divided by the maximum
  possible
• density lines/(n(n-1)

85
86
Week 2 (Part 3)Matrix Manipulation (Linear
Algebra)

• Addition add elements (cells) together
• Subtraction subtract elements (cells) from each
  other
• Element by element multiplication multiply
  elements (cells) with each other
• Multiplication Transpose the row of first
  matrix, then multiply corresponding elements and
  then add products.
• Division possible but requires knowing the
  inverse of the matrix which requires a computer

86
87
Symmetrizing Matrices

• Many algorithms require symmetric matrices
• Can symmetrize on different criteria.
• Maximum the higher of the two valueseither
  direction becomes both
• Minimum the lower of the two valuesonly two
  ties are kept.

87
88
Valente Review of NA for PH(adapted from Burt
1980)
88
89
Network Ties Have 3 Components

• 1. Role - How related friend, spouse, co-worker,
  etc.
• 2. Meaning and Content - What do ties talk
  about? That is, love, politics, health, etc.?
• 3. Strength - How well are ties connected? That
  is, not so well or very close strong ties? Do
  they talk often or rarely?

89
90
Week 3 Egocentric Data

• Egocentric network data are collected from
  independently sampled units (people).
• The ties are not connected to one another (at
  least not that we know)
• Rs are asked a name generator and then Rs
  provide data on these alters
• Ego centric data provide some estimate of a
  persons immediate close ties

91
Burt, 1984 GSS Egocentric Survey

• Proposed and tested battery of egocentric network
  survey items
• Recommended name generator Name up to 5 people
  you discuss important matters with
• Data were collected in 1985 and analyzed to
  describe Americans core discussion networks

92
Egocentric Survey (sample)
Please provide the names the first names or
initials of up to 5 people you talk to about
important matters.
93
Graphs of Ego-centric
1
2
3
4
5
94
Egocentric Survey
Does _____ know _____?
95
Graphs of Ego-centric
1
2
3
4
5
96
Ego-centric Networks

• Scores derived from ego-centric Networks with
  named alters
• Typically the following characteristics are
  measured
• 1. Strength of relation (close, acquaintance,
  stranger how long known)
• 2. Frequency of interaction (how often talk to)
• 3. Type of relation (family, friend, coworker)
• 4. Socio-economic characteristics (education,
  religion, wealth)
• 5. Demographic characteristics (age, location)
• 6. Substantive characteristics (practice FP,
  approve of abortion)
• 7. Content of communication (discuss politics,
  health, child rearing)
• From these items two types of variables are
  created
• 1. Compositional
• 2. Heterogeneity (variance)

96
97
Marsden Core Discussion Networks ASR

• Paper provides summary statistics from 1985 GSS
• Survey network data share all the concerns of
  survey research
• GSS considered standard egocentric measure
• GSS provides a national sample to derive national
  estimates for comparison
• GSS measured discusses important matters in
  past 6 months
• Only R knows what was discussed
• Item measuring topic was dropped (Bolivia data
  has)

98
Marsden (cont.)

• Size- number of alters
• Density connectedness of alters
• Heterogeneity diversity of alters
• Age
• Education
• Race/ethnicity
• Sex
• Composition- kin v non-kin

99
Compositional Variance Measures
100
Univariate Distribution of Network Form
Composition
Size 3.01 (SD1.77) Kin 1.53 (55) non-kin
1.40 Density 0.61 (SD0.28) Age Heterogeneity
10.54 years (SD6.39) Education Heterogeneity
1.78 (SD1.37) Race/ethnicity Heterogeneity
0.05 (SD0.18) Sex Heterogeneity 0.68 (SD0.38)
101
Marsden (cont.)

• Data show remarkably high tendency of
  homogeneity- and may be underestimated because of
  presence of kin (table 2)
• These are core networks, they are small,
  centered on kin, comparatively dense, and
  homogenous by comparison to the R population.

102
Subgroup Differences

• Age size drops with age, kin/nonkin varies by
  age
• Education Increases size ( nonkin) and
  negatively associated with density
• Race/ethnicity Whites have larger nets,
  heterogeneity varies by race/ethnicity
  (IQVWhites .03, Blacks .13, Hispanics .22)
• Sex Women have more kin
• Size of Place Urbanites cite more non-kin and
  lower density

103
Conclusions

• GSS data describe relatively small, kin, dense
  homogenous social environments
• Variability is substantial, patterned by
  socio-demo characteristics of Rs
• Networking (diverse ties) best for young
  middle-aged, well educated and urban

104
Ego Centric Data Used to Study 5 Behavioral
Hypotheses

• Degree/Size
• Personal Network Exposure
• Strong Ties
• Concurrency
• Density/Constraint

105
1. Ego-centric degree/size
Lower Risk Or Isolated
Higher Risk Or More Privileged Position
106
Size

• Basic and critical variable in analysis
• Often restricted to name up to 5 or 7 alters so
  the variable is censored.
• Some researchers allow R to name up to 20 or
  more, and then only ask detailed information
  about the first 5 or
• About a random selection of those named

107
2. Personal Network Exposure
Non-FP User
FP User
PN Exposure33
PN Exposure66
PN Exposure100
107
108
Commands for Egocentric Exposure

• egen frndsmk_perc_1rmean(y1_smoke_nom1
  y1_smoke_nom2 y1_smoke_nom3 y1_smoke_nom4
  y1_smoke_nom5)
• egen frndsmk_perc_2rmean(y2_smoke_nom1
  y2_smoke_nom2 y2_smoke_nom3 y2_smoke_nom4
  y2_smoke_nom5)
• egen frnddrk_perc_1rmean(y1_drink_nom1
  y1_drink_nom2 y1_drink_nom3 y1_drink_nom4
  y1_drink_nom5)
• egen frnddrk_perc_2rmean(y2_drink_nom1
  y2_drink_nom2 y2_drink_nom3 y2_drink_nom4
  y2_drink_nom5)
• egen frndsmk_perc_ct_1rsum(y1_smoke_nom1
  y1_smoke_nom2 y1_smoke_nom3 y1_smoke_nom4
  y1_smoke_nom5)
• egen frndsmk_perc_ct_2rsum(y2_smoke_nom1
  y2_smoke_nom2 y2_smoke_nom3 y2_smoke_nom4
  y2_smoke_nom5)
• egen frnddrk_perc_ct_1rsum(y1_drink_nom1
  y1_drink_nom2 y1_drink_nom3 y1_drink_nom4
  y1_drink_nom5)
• egen frnddrk_perc_ct_2rsum(y2_drink_nom1
  y2_drink_nom2 y2_drink_nom3 y2_drink_nom4
  y2_drink_nom5)

109
Contraceptive Behavior Associated with Peer
Behavior
110
Mass Media and Interpersonal Influence in a RH
Communication Campaign in Bolivia

• Valente Saba, Communication Research, 1998
• Evaluation consisted of X-sectional and panel
  surveys of urban men women
• Study attempted to show the joint influence of
  mass and IP communication

110
111
Interpersonal Networks

• Egocentric questionnaire
• Named up to 5 people you discuss important
  matters with
• Collected information on age, gender, education,
  religion, language, SES and topics discussed
• Also asked whether R thought alter used FP

112
Personal Network Thresholds
Non-FP User
FP User
PN Threshold33
PN Threshold66
PN Exposure33
Low Threshold Adopter
High Threshold Adopter
112
113
Table 2 Odds Ratios for the Likelihood of Low
and High-threshold Adoption by Demographic
Characteristics and Campaign Recall
113
114
3. Ego-centric Rank Ordered
Risk Unknown,possible false sense of security
115
Baltimore NEP

• Time Period August 12, 1994 - February 12, 1997
• Repeated interviews with 1,184 respondents at
  baseline, 2-week, 6-month, 1-year, 18-month
• Included ego-centric questions on survey
• Provide the initials or nicknames of up to 5
  your closest friends

116
Schematic of Needle Network Study
Get Needles Return Needles
Return to van periodically and continue about
their business
Registration N3,500
Every Seventh
Participated in Evaluation Study N448
Interviewed at Baseline 2-weeks (364), 6 mos.
(308), 1 year (175), and 18 mos. (88).
116
117
Needle sharing reports decrease
118
Graph of reported syringe sharing by friendship
rank and survey wave
119
4. Ego-centric -Concurrent
Serial
Concurrent
120
5. Dense vs. Radial Personal Networks
Ego
Ego
Integrated/Dense Personal Network Constrained
Radial/Open Personal Network
121
Constraint

• Sum of each alters density (less ego)
• Brokerage
• Structural Holes

122
Dyadic Data

• Ego-centric data are converted to dyadic in which
  case is the person-alter relationships.
• If 100 respondents name an overage of 3.1 alters,
  the dyadic data has 310 cases.
• Makes testing some hypotheses easier (e.g.,
  association between behavior and gender
  homophily).
• Must run random effects models.

123
Dyadic Format
124
Make Dyadic Dataset

• delim
• use c\misc\red\red, clear
• keep y1id y1_nom
• drop if y1id.
• reshape long y1_nom, i(y1id) j(alter)
• drop if y1_nom.
• sort y1_nom
• save c\misc\red\y1_dyad, replace

125
Calculate In- Out-degree

• / Calculate In Degree /
• gen one 1
• collapse (sum) one, by(y1_nom)
• ren one nr_y1
• recode nr_y1 .0
• ren y1_nom y1id
• sort y1id
• save c\misc\red\nr_y1
• / Calculate Out Degree /
• use c\misc\red\y1_dyad
• gen one 1
• collapse (sum) one, by(y1id)
• ren one ns_y1
• sort y1id
• save c\misc\red\ns_y1, replace

126
Snowball Sampling

• Often used for recruitment into clinical trials
  or
• Prevention studies
• Index cases nominate or recruit alters

127
(No Transcript)
128
(No Transcript)
129
Network linkages among indexes and alters in
which at least one alter was enrolled (794 links,
59.2).
D
A
C
B
130
A
B
D
C
131
VPS Study Results

• Tremendous variation in egocentric networks
• Variability across relationships in who they are
  willing to invite

132
Personal v. Sociometric Data

• Several studies have tested the correspondence
  between ego-centric and sociometric data
• Generally find poor agreement on what ego thinks
  alters do, and what alters self report
• Stronger associations with behavior for perceived
  alter behavior than alter self-report.