8/16/2013 1 Introduction to Social Network Analysis 1 Dan Halgin ([email protected]) Joe Labianca ([email protected]) University of Kentucky Introduction 2 What is network analysis? Network theory and research How we analyze network data? Software demonstration Network analysis resources
51
Embed
Introduction to Social Network Analysis - Dan Halgindanhalgin.com/.../assets/docs/AOM_Intro_to_SNA_PDW_2013_Final.23… · Introduction to Social Network Analysis 1 Dan Halgin ([email protected])
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Nodes are 1000 scientists at RAND. Ties are collaboration on research projects
Types of Ties
State
Similarities
Co-locationPhysicaldistance
Co-membershipSame boards
Shared AttributesSame race
Social Relations
KinshipCousin of
Other roleBoss of; Friend of
Cognitive / Affective
Knows; Dislikes
Event
Interactions
Email to, lunch with, sex with
Flows
Information Money, Disease
See Borgatti & Halgin 20116
8/16/2013
4
Ties and Nodes: Other termsTies Sometimes called edges Sometimes called links Sometimes called arcs Sometimes called lines
Nodes Sometimes called vertices Sometimes called actors Sometimes called egos
Halgin & Brass Links 20137
How do we use the term “network”?
8
Unfortunately, too many different ways
It can refer to a whole network of nodes and ties
It can refer to the same set of nodes, but each type of tie is a different “network” Examples: the communication network, the friendship network,
the advice network)
It can also refer to each individual node’s personal network
Important to clarify what you mean by network
8/16/2013
5
What is Unique about SNA?
9
Phenomenon studied distinctive type of data, it’s about relations
How we study it Distinctive tools, “traditional” statistical methods don’t always
apply
How we understand it Mechanisms and processes that interact with network structures
to yield certain outcomes for individuals and groups Antecedents of network position
Relations Matter: Where’s the power?
Mel
Earl
York
Ben
Pat
Robin
Dale
Rick
Alex
Hal
Chris
Tom
Gary
Ovid
Walt
Fran
Ken
Bob
Gerry
Nan
Ev
Ivo
Dan
Jack
Vic
Hugh
Quincy Upton
Zoe
Carl
Len
Abe
Fred
Steve
Irv
Jim The boss, Steve
The managers Ev, has most reports
Formal power isn’t everything Informal network
10
Case drawn from:
Krackhardt, D. 1992 “The Strength of Strong Ties: The Importance of Philos in Organizations.” In N. Nohria & R. Eccles (eds.), Networks and Organizations: Structure, Form, and Action: 216–239. Boston, MA: Harvard Business School Press.
What happened? Hal was a bad choice. Had the drive, but not the social capital
Chris had the influence and was pro-union
The union didn’t know the (social) territory
STRUCTURE MATTERS
16
“I would never have conceived my theory, let alone have made a great effort toverify it, if I had been more familiar with major developments in physics thatwere taking place. Moreover, my initial ignorance of the powerful, falseobjections that were raised against my ideas protected those ideas from beingnipped in the bud.”
– Michael Polanyi (1963), on a contribution to physics
A CB
8/16/2013
9
Learning relationships between U.S. business schools
17
How do we use the term “network”?
18
Unfortunately, too many different ways
It can refer to a whole network of nodes and ties, but where there are multiple components Brings up the issue of whether nodes that are completely
disconnected are really part of the “network”
It can refer to the same set of nodes, but each type of tie is a different “network” (e.g., the communication network, the friendship network, the advice network)
It can also refer to each individual node’s personal network
8/16/2013
10
Steps to a SNA study
19
1. Identify the population Bounding, sampling, gaining access
2. Determine the data sources Archival, interviews, observations, surveys
3. Collect the data Survey design
4. Preparing data for network analysis Missing data? Symmetrize? Dichotomize?
5. Gathering descriptive statistics Using univariate statistics, visualizing the data
6. Testing hypotheses
Sources of Network Data
20
Experiments Rumor planting; Milgram small world; experimental exchange
networks
Observation Western-Electric Hawthorne plant studies Ethnographic studies Gary Alan Fine story telling; Whyte street corner etc
Surveys Telephone, web, paper, etc.
8/16/2013
11
Additional sources: records
21
Memberships in groups Facebook “networks”; google groups, listservs Boards of directors
Participation in events Listserv message threads; DGG deep south data Voting records, e.g. supreme court data White house diaries
Text analyses Copdab, KEDS Crawdad, automap
Other Email records, purchase/sale records, marriage records, alliances, etc
From Records
22
8/16/2013
12
Archival Data
Corporate Board Interlocks
Seagan (2012) : http://blog.kiwitobes.com/23
Archival data: Bernie
24
8/16/2013
13
Archival data: Steroids Supply Network in Baseball
25
Archival data: Steroids Supply Network in Baseball NY Yankees Roster
26
8/16/2013
14
Archival: “Diss Songs” in Rap Music (1970-2008)
27
Halgin, Borgatti, & Huang, 2012Please direct all correspondence to [email protected]
Each nodes represent a rap artist. A directed tie from A to B indicates that rapper A recorded a diss song about rapper B. Nodes are colored and grouped by geography.
Levels of Analysis
28
8/16/2013
15
Are you focused on ego networks or whole networks? Ego networks: centered around a particular actor. Includes
the “ego” and direct tie “alters,” and ties among those alters. One actor’s network.
Whole networks: attempt to get data from all members of a bounded network.
Ego/Personal vs. Whole Network
30
1
23
4
5
Chris
1
23
4
5
Chris
8/16/2013
16
Ego networks within and outside of whole networks Some research focuses on ego networks within a whole
network Example: focusing on how many structural holes a particular
individual spans
That same research can be conducted on random ego networks Example: studying individuals’ structural holes and personality
by collecting data through MTurk or e-Rewards
Important to keep track of this because it determines how you’ll analyze the data Big consideration is whether you need to account for network
autocorrelation
Example whole network surveyHow frequently do you communicate with this person?
How do you generally feel about this person?
Would you go to this person for personal advice? (including family situations, relationships with non-church members, or church members in non-professional setting)
Would you go to this person for professional advice? (including relationships with other clergy, church members in course of duties, or other job-related duties)
Check if this minister has provided material support to you or your church (e.g., financial support, staff support, member referrals, guest preaching)
ID N
um
ber From
Roster
On
ce/ year
Several tim
es/year
On
ce/ mon
th
On
ce/ week
Daily or m
ore
Dislike a lot
Dislike a little
Neu
tral
Like a little
Like a lot
Stron
gly D
isagree
Disag
ree
Neu
tral
Ag
ree
Stron
gly A
gree
Stron
gly D
isagree
Disag
ree
Neu
tral
Ag
ree
Stron
gly A
gree
Ch
eck if app
licable
# RM11 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
# 99991 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
8/16/2013
17
Example ego network survey
A B C D E F G HAbigail x x xBetty x xChad x x x xDavid xEphraim x xFlora xGeorge xHoward
Ego Network Analysis
Combine the perspective of network analysis with the data of mainstream social science
WholeNetworkAnalysis
MainstreamSocial Science
EgoNetworks
perspectivedata
8/16/2013
18
Recall vs. rosterIf collecting via survey, ask people to: List names open (e.g., name generator) relies on recall often used for ego network studies
Circle names on a roster Bounded, researcher defines network relies on recognition often used for whole network studies
Reliability Concerns About Social Network DataActors are not very good about remembering specific
interactions.
Bernard et al. 1984
But they are good about remembering recurrent, repeated interactions or on-going relationships.Freeman et al. 1987
8/16/2013
19
Valued vs. Binary DataWe can collect valued data as well as binary data.
Binary – yes or no, 1 or 0
Valued – example: on a scale from 1-5
Notice both valued and binary dataHow frequently do you communicate with this person?
How do you generally feel about this person?
Would you go to this person for personal advice? (including family situations, relationships with non-church members, or church members in non-professional setting)
Would you go to this person for professional advice? (including relationships with other clergy, church members in course of duties, or other job-related duties)
Check if this minister has provided material support to you or your church (e.g., financial support, staff support, member referrals, guest preaching)
ID N
um
ber From
Roster
On
ce/ year
Several tim
es/year
On
ce/ mon
th
On
ce/ week
Daily or m
ore
Dislike a lot
Dislike a little
Neu
tral
Like a little
Like a lot
Stron
gly D
isagree
Disag
ree
Neu
tral
Ag
ree
Stron
gly A
gree
Stron
gly D
isagree
Disag
ree
Neu
tral
Ag
ree
Stron
gly A
gree
Ch
eck if app
licable
# RM11 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
# 99991 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
8/16/2013
20
Collecting Node Attribute DataWe can also collect attribute data.Examples: Gender Race Personality
We will deal with attribute data in one of two ways (more on this later): To examine whether the attribute affects some outcome (e.g., women have more
negative ties) To examine whether dyadic similarity on the attribute affects some outcome (e.g.,
men or women are more likely to have negative ties with someone of their own gender)
Mainstream Logical Data Structure 2-mode rectangular matrix
in which rows (cases) are entities or objects and columns (variables) are attributes of the cases
Analysis consists of correlating columns Emphasis on explaining one
variable
ID Age Education Salary
1
2
3
4
40 Halgin & Brass Links 2013
8/16/2013
21
Adjacency Matrix
We have created an actor by actor square “adjacency” matrix (one mode matrix)
Each cell in the matrix indicates if the actors are related (1,0) or the extent of the relationship (1-5).
Data are, by convention, “directed” from rows (i) to columns (j), so that every cell (x) reports the relationship from the perspective of i to j.
Cells are also referred to by row and column (cell 3,4 is row 3, column 4)
Network Logical Data Structures
Ed Sue Jim Bob
Ed - 1 0 0
Sue 0 - 1 1
Jim 0 0 - 0
Bob 1 0 0 -
Ed Sue Jim Bob
Ed - 4 0 2
Sue 0 - 5 1
Jim 0 0 - 0
Bob 3 0 4 -
Friendship
Email Communication• Individual characteristics only half the
story...RELATIONS MATTER!
• Values are assigned to pairs of actors
• Hypotheses can be phrased in terms of correlations between relations
42
8/16/2013
22
Relational Data & Attribute DataEd Sue Jim Bob
Ed - 1 0 0
Sue 0 - 1 1
Jim 0 0 - 0
Bob 1 0 0 -
Gender Education Salary
Ed 0 14 50000
Sue 1 15 99000
Jim 0 12 65000
Bob 0 8 15000
Relational Data Attribute Data
SNA provides the ability to combine relational data with attribute data (e.g., homophily, heterogeneity, etc)
43Halgin & Brass Links 2013
Whole network
Data: Campnet44 Halgin & Brass Links 2013
8/16/2013
23
HO BR CA PAM PAT JEN PAU ANN MIC BIL LEE DO JOH HAR GER STE BER RUSHO 0 0 0 1 1 0 0 0 0 0 0 1 0 0 0 0 0 0BR 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 1 0CA 0 0 0 1 1 0 1 0 0 0 0 0 0 0 0 0 0 0
Edgelist Data EntryIn a file, enter the following:
Ace Buddy
Buddy Chuck 5
Chuck Ace 4
Dave Ace 3
Esther Buddy 4
Note: If you omit a number after the edge, it is automatically assigned a 1Note: Edge is another term for tie
8/16/2013
24
Nodelist Data EntryIn a file, enter the following:
Ace Buddy Dave Esther
Buddy Chuck
Chuck Ace Dave Esther
Dave Ace Buddy Chuck
Esther Buddy Dave
Note: These data are ONLY binary
Preparing Network Data for Analysis
48
8/16/2013
25
Three major decisions during network data preparation phase1. How do I handle missing data?
2. Do I symmetrize or leave the data non-symmetrical?
3. Do I dichotomize or leave the data valued?
Missing data: The nemesis of network analysis
If I hand out a psychometric survey and 60% of the population responds, I’ve got an incredible opportunity to have a wonderful representative sample.
If I hand out a sociometric survey (sociometric is the term for a network survey) and get the same response rate for a whole network study, I’ve only managed to collect 36% of the network (0.6 squared).
General rule of thumb for management publications – you need a 70% response rate (which gets you about half the network)
You’ll get more leeway from reviewers for a field study, less for a study of students (e.g., MBA students in a cohort) – degree of difficulty score
8/16/2013
26
Missing data choices Leave the missing data as they are
Enter the letter “n” in UCINET in cells where there is no data This is the preferred method
Network imputation – still in its infancy, not considered “trustworthy”
Another possibility is to symmetrize away the missing data
Directed data provides measures such as: in-degree: number of links coming in to the actor (this is reported by
all of the other respondents) out-degree: number of links going out from the actor (this is reported
by the respondent)
Directed data can be symmetrized into non-directed data (top and bottom half of matrix are identical).
Thus, if persons B, C, and D all say they are friends of A, even if you don’t get A to respond to your survey, you might decide to symmetrize away the missing data.
But BEWARE! Reviewers are evaluating your choices carefully.
8/16/2013
27
Symmetrizing data: Understanding the nature of the relationship
Symmetrizing data only makes sense for certain type of network relationships that can be reasonably considered non-directed, but not for others that are inherently directed. Example:
Advice network is inherently directed – difficult to justify symmetrizing.
Communication network is inherently non-directional –symmetrizing is easier to justify.
Other networks – check reciprocation rate. Follow up to resolve discrepancies. Make a case for symmetrizing.
Ways to Symmetrize
If you decide to symmetrize, you need to decide how to symmetrize. Below are three typical ways to symmetrize:
Maximum Xij = 5, Xji = 1 Symmetrizes to Xij, ji = 5 Can also be done with binary data or missing data
Minimum Xij = 5, Xji = 1 Symmetrizes to Xij, ji = 1 Can also be done with binary data or missing data
Average Xij = 5, Xji = 1 Symmetrizes to Xij, ji = 2.5
8/16/2013
28
Dichotomizing: Notice the choices for the communication variable
How frequently do you communicate with this person?
How do you generally feel about this person?
Would you go to this person for personal advice? (including family situations, relationships with non-church members, or church members in non-professional setting)
Would you go to this person for professional advice? (including relationships with other clergy, church members in course of duties, or other job-related duties)
Check if this minister has provided material support to you or your church (e.g., financial support, staff support, member referrals, guest preaching)
ID N
um
ber From
Roster
On
ce/ year
Several tim
es/year
On
ce/ mon
th
On
ce/ week
Daily or m
ore
Dislike a lot
Dislike a little
Neu
tral
Like a little
Like a lot
Stron
gly D
isagree
Disag
ree
Neu
tral
Ag
ree
Stron
gly A
gree
Stron
gly D
isagree
Disag
ree
Neu
tral
Ag
ree
Stron
gly A
gree
Ch
eck if app
licable
# RM11 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
# 99991 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
Four types of scales
Nominal scales Categories, but no order (e.g., gender, department)
Ordinal scales Categories, but order (e.g., rank, communication in the previous
example)
Interval scales Continuous data having equal intervals but no zero point (e.g., time) Generally not seen in social network analysis
Ratio scales Continuous data having equal intervals and a zero point (e.g.,
number of email msgs exchanged)
8/16/2013
29
Dichotomizing nominal and ordinal scales
Nominal scales Categories, but no order (e.g., gender, department) Expected that these will be dichotomized or trichotomized –
inherent in the scales
Ordinal scales Categories, but order (e.g., rank, communication in the previous
example) Sometimes necessary to dichotomize (e.g., weekly communication
vs. more infrequent communication) When constructing a communication question, you need to know
what the appropriate behavioral anchors will be
Dichotomizing ratio scales
Ratio scales Continuous data having equal intervals and a zero point
(e.g., number of email msgs exchanged) Could dichotomize (e.g., more than 10 email messages in a
week vs. less) Be extremely cautious Why are you reducing rich data?
Reviewers will not generally be positively predisposed to this
You need to justify
8/16/2013
30
Gathering descriptive statistics in network analysis
Understanding Your Data
Before doing anything else, get to know your data
Examine means and standard deviations Use univariate statistics command in UCINET to better understand your
data
Create a network graph to see visually what is happening with your data UCINET has an embedded visualization called NetDraw that is very easy to
use Many other network visualization programs are available
Create a correlation matrix
8/16/2013
31
Univariate statistics Open UCINET
Type “Control-U”
Or hunt for it through the menu TOOLS > UNIVARIATE STATS
Input dataset: C:\Documents and Settings\jlabi2\My Documents\LaGrange Network Data\MinistersCom
Descriptive Statistics
1
--------
1 Mean 2.605
2 Std Dev 0.963
3 Sum 409.000
4 Variance 0.927
5 SSQ 1211.000
6 MCSSQ 145.516
7 Euc Norm 34.799
8 Minimum 1.000
9 Maximum 5.000
10 N of Obs 157.000
Statistics saved as dataset C:\Documents and Settings\jlabi2\My Documents\LaGrange Network Data\MinistersCom-Uni
----------------------------------------
Running time: 00:00:01
Output generated: 28 May 08 14:50:22
Copyright (c) 1999-2005 Analytic Technologies
8/16/2013
32
Visualize your data Open NetDraw and open your matrix, as well as any
attribute files in which you are interested
Examine the network visually to understand the overall social geography prior to beginning more formal analysis
Learning relationships between U.S. business schools
8/16/2013
33
Smokers in Framingham, MA: 1972-2002 Content.nejm.org/content/vol358/issue21/images/data/2
249/DC1/NEJM_Christakis_2249a1.shtml?sid=ST200
Creating a correlation matrix: QAP correlations
Network data are autocorrelated, so we can’t use standard OLS regression. One technique developed to handle network autocorrelation is called Quadratic Assignment Procedure Correlation, or QAP correlation (e.g., Krackhardt, 1988).
Involves correlating two matrixes cell-wise. Pearson correlation coefficient is saved. Then the procedure randomly permutes the rows and the columns thousands of times, each time computing new coefficients.
The procedure computes the proportion of coefficients generated from the random permutations that are as extreme as the coefficient between your two matrices of interest to determine whether the relationship is significant.
8/16/2013
34
Sample QAP correlation output
Assembling network measures
68
8/16/2013
35
Multiple levels of network analysis
69
Dyad (relationship) level Who is friends with whom in an office Distance in meters between people’s desks Marriage ties among families in Renaissance Florence Business ties among the same families
Node (actor) level Aggregate these dyads to the node level (e.g., # of friends) Or measure aspects of a node’s position in the network
Group (network) level Aggregation to the group or whole network level (e.g., # of
ties within group/# of potential ties within group) Or measure aspects of network shape (e.g., centralization)
Examples of measures at each level of analysis
70
Subgroupidentification
Clique
Geodesicdistance
Proximity
Adjacency Simmeliantie
Equivalence
Roleidentification
Structuralequivalence
Regularequivalence
Block
Cohesion
DensityAvg. distance
Shape
Degree distribution
Core-peripheriness
Centrality and Constraint
DegreeCloseness
Constraint
Betweenness
Group level(properties of network structures)
Node level(properties of node positions)
Dyad level(properties of relationships)
Small-worldedness
8/16/2013
36
Levels of Analysis
71Cross, Parker, & Borgatti, 2002. Making Invisible Work Visible. California Management Review. 44(2): 25-46
9 Months Later
Cross, Parker, & Borgatti, 2002. Making Invisible Work Visible. California Management Review. 44(2): 25-46 72
8/16/2013
37
Examples of measures at each level of analysis
73
Cohesion
Centrality and Constraint
Proximity Equivalence
Subgroupidentification
Roleidentification
Clique
Adjacency Simmeliantie
Geodesicdistance
Structuralequivalence
Regularequivalence
Block
DensityAvg. distance
DegreeCloseness
Shape
Degree distribution
Core-peripheriness
Constraint
Betweenness
Group level(properties of network structures)
Node level(properties of node positions)
Dyad level(properties of relationships)
Network style hypotheses
74
8/16/2013
38
Types of network hypotheses
75
Using these three levels of analysis (dyad, node, group) and the network ties and nodal attributes we can investigate numerous types of hypotheses.
76
Examples of dyad level hypotheses
Independent Variable Dependent Variable
Dyad Level
Network tie Network tie
Network tie Attribute similarity
Attribute similarity Network tie
8/16/2013
39
77
Examples of dyad level studies
Independent Variable Dependent Variable
Dyad Level
Network tie Network tie
Brass (1981) formal required interaction creates the opportunity for friendship
Note: This is why collecting the formal network when doing intra-organizational networks studies is crucial
Potential analytic methods: MRQAP, ERGM
Multiple regression: MRQAP
Multiple Regression QAP is similar to QAP Correlation Specify the dependent matrix Enter the independent matrixes
Involves regressing the dependent matrix on the independent matrixes cell-wise. Regression coefficients are saved. Then the procedure randomly permutes the rows and the columns thousands of times, each time computing new coefficients.
The procedure computes the proportion of coefficients generated from the random permutations that are as extreme as the coefficients between your matrices of interest to determine whether the variables are significant.
Use the Dekker version of MRQAP
8/16/2013
40
ERGM
Exponential random graph model
Take your observed network Now imagine a random network and begin to introduce parameters that
acts as “decision rules” Examples:
A is friends with B, B’s chances of being friends with A are higher (reciprocity) A and B are both friends with C, thus they are likely to become friends with each other
(transitivity) If A and B share the same attribute (homophily), they are more likely to become friends If A has a certain attribute (e.g., competence), do they attract more friends (receiver effect) and do
they name more friends (sender effect)
Allows you to see if the simulated networks you generate via the parameters you added look like the observed network, and which of the parameters is driving this
80
Examples of dyad level studies
Independent Variable Dependent Variable
Dyad Level
Network tie Attribute similarity
Erickson (1988) – Relational basis of attitudes theory
Network ties lead to attitude similarity: Task-related attitudes (Burkhardt, 1994) Justice perceptions (Umphress, et al., 2003) Leadership perceptions (Pastor, et al., 2002) Companies with which to interview (Kilduff, 1990) New technology (Rice & Aydin, 1991)Potential analytic methods: MRQAP, ERGM
8/16/2013
41
81
Examples of dyad level studies
Independent Variable Dependent Variable
Dyad Level
Attribute similarity Network tie
Homophily eases the formation of network ties
Brass, 1985; McPherson & Smith-Lovin, 2001; Ibarra, 1992, 1993 similarity in attributes (age, sex, education, prestige, social class, tenure, function, religion, professional affiliation, & occupation) leads to the formation of network ties
Potential analytic methods: MRQAP, ERGM
82
Examples of node level hypotheses
Independent Variable Dependent Variable
Node Level
Node level network property
Node level network property
Node level network property
Actor attribute
Actor attribute Node level network property
8/16/2013
42
83
Examples of node level studies
Independent Variable Dependent Variable
Node Level
Node level network property
Node level network property
Being central in one type of network (e.g., advice ties related to technology) might lead to being central in a different type of network (e.g., friendship or power networks) (cf., Burkhardt & Brass, 1990).
Potential analytic methods: Node level regression
84
The procedure is principally used to model a vector using values of other vectors.
The algorithm proceeds in two steps. In the first step, it performs a standard multiple regression across corresponding cells of the dependent and independent vectors.In the second step, it randomly permutes rows the elements of the dependent vector and recomputes the regression, storing resultant values of r-square and all coefficients. This step is repeated hundreds of times in order to estimate standard errors for the statistics of interest. For each coefficient, the program counts the proportion of random permutations that yielded a coefficient as extreme as the one computed in step 1.
Node level regression
8/16/2013
43
85
Examples of node level studies
Independent Variable Dependent Variable
Node Level
Node level network property
Actor attribute
Promotion Brass (1984, 1985) – centrality increases promotion likelihood Burt (1992) – white males with structural holes in networks
promoted quickly
Networks Actor attribute
86
Individual in-role performance Betweenness centrality related to performance (Mehra, et al.,
2001)
Extra-role performance People perform ICBs for friends, powerful others, and friends
of powerful others (Bowler & Brass, 2006)
Creativity Ideas from managers with lots of structural holes judged to be
more creative (Burt, 2004)Potential analytic methods: • Collected in one whole network? Node level regression• Collected in independent ego networks? OLS regression
8/16/2013
44
87
Examples of node level studies
Independent Variable Dependent Variable
Node Level
Actor attribute Node level network property
• High self-monitoring personality type related to betweenness centrality and to occupying structural holes (Mehra, et al., 2001; Sasovova, et al., 2010)
• Big Five related to networks (Klein, et al., 2004)
Potential analytic methods: • Collected in one whole network? Node level regression• Collected in independent ego networks? OLS regression
88
Longitudinal studies
• Any of the prior questions can also be asked longitudinally
• Example: Does having a negative tie with someone at Time 1 lead to negative gossip with a third party about that person at Time 2?
• Potential analytic methods:• RSiena• Tie churn statistics in UCINET
8/16/2013
45
89
Examples of group level hypotheses
Independent Variable Dependent Variable
Group Level
Group level network property
Group level network property
Group level network property
Other group attribute
Other group attribute Group level network property
90
Examples of group level studies
Independent Variable Dependent Variable
Group Level
Group level network property
Group level network property
• These types of studies don’t tend to be done much in organizational research
• Tend to be seen more when examining extra-organizational networks (e.g., comparisons of multiple communities of practice, and how their founding conditions affect their subsequent development)
8/16/2013
46
91
Examples of group level studies
Independent Variable Dependent Variable
Group Level
Group level network property
Other group attribute
Group performance increased by:• Internal density and external range in knowledge
sharing network (Reagans, et al., 2004)• Number of bridging relationships in informal
socializing network; inverted-U relationship with internal density (Oh, et al., 2004; see also Oh, et al., 2006) Potential analytic methods:
• Collected in one whole network? Node level regression• Collected in independent ego networks? OLS regression
92
Examples of group level studies
Independent Variable Dependent Variable
Group Level
Other group attribute Group level network property
Not an area that is often studied in organizational network research; might be an interesting way to marry the research on TMTs and boards with network research
Potential analytic methods: • Collected in one whole network? Node level regression• Collected in independent ego networks? OLS regression
8/16/2013
47
93
Structural Determinism vs. Agency
Subjective vs. Objective
Method vs. Theory
Dynamic Networks
Relational Content of Ties Multiplexity Valence
Current Issues in Social Network Analysis
Defense Alliance & Conflict networksCorrelates of War data-- conflicts 1946-2001-- alliances 2001
8/16/2013
48
Defense Alliance network
Defense Alliance & Conflict
8/16/2013
49
Software Demonstration
97
Network Analysis Resources
98
8/16/2013
50
Networks Training Workshops
99
LINKS Center at University of Kentucky 1 week, June, 2014, UCINET and SIENA/ERGM http://linkscenter.org/workshops/
University of Essex, UK 3 short courses offered between July-August, 2014, UCINET and SIENA/ERGM
CARMA 2.5 days, May , 2014, UCINET
ICPSR (various locations, including Indiana U., UNC, UMass Amherst) 2.5 days to 1 week
Sunbelt social networks conference 2 day workshops, Feb 2014 in St. Pete Beach, Florida
More resources
100
www.LinksCenter.org The SAGE Handbook of Social Network Analysis (Eds. Scott and
Carrington, 2011) Analyzing Social Networks by Borgatti, Everett, & Johnson
Companion website: sites.google.com/site/analyzingsocialnetworks/ Exponential Random Graph Models for Social Networks: Theory, Methods,
and Applications by Lusher, Koskinen, & Robins Professional association
Int'l Network for Social Network Analysis - www.insna.org Conferences
Sunbelt (http://www.insna.org/sunbelt/) Listservs & Groups
SOCNET listserv REDES listserv UCINET user’s group