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])

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

8/16/2013

2

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

Trade relations among nationsRoad system

Organization chart

Plumbing system

Networks are all around us

Food web

Proportion of all articles indexed in Google Scholar with “social network” in the title, by year

4

0

0.005

0.01

0.015

0.02

0.025

1963 1968 1973 1978 1983 1988 1993 1998 2003 2008 2013

Shar

e

Borgatti, Brass & Halgin 2013

8/16/2013

3

Basic Concepts: Nodes

5

Persons

Organizations

Countries

Families

Animals

Web pages

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.

Installers

© David Krackhardt

8/16/2013

6

Abe

Bob

Carl

Dale

Ev

Fred

Gary

Hal

Ivo

JackKen

Len

Nan

Ovid

Pat

Quincy

Robin

Steve

Tom

UptonVic

Walt

Rick

Zoe

Alex

Ben

Chris

Dan

Earl

Fran

Gerry

Hugh

Irv

Jim

York Mel

Advice network

© David Krackhardt

Friendship network

Mostcentral

Abe

Bob

Carl

Dale

Ev

FredGary

Hal

IvoJack

Ken

Len

Nan

OvidPat

Quincy

Robin

Steve

Tom

Upton

VicWalt

York

Zoe

Alex

Ben

Chris

Dan

Earl

Fran

Gerry

Hugh

Jim

© David Krackhardt

8/16/2013

7

Abe

Bob

Carl

Dale

Ev

Fred

Gary

Hal

Ivo

Jack

Ken

Len

Mel

Nan

Ovid

Pat

Quincy

RobinSteve

Tom

Upton

Vic

Walt Rick

York

ZoeAlex

Ben

ChrisDan

Earl

Fran

Gerry

Hugh

Irv

Jim

Chris’s perception of the friendship network

© David Krackhardt

Ev’s perception of the friendship network

Dale Ev

Ivo

Ken

Pat

Robin

Steve

Tom

Rick

Zoe

Alex

Chris

Gerry Hugh

Irv

Jim

© David Krackhardt

8/16/2013

8

Abe

Bob

Carl

Dale

Ev

FredGary

Hal

IvoJack

Ken

Len

Nan

Ovid Pat

Quincy

Robin

SteveTom

Upton

VicWalt

York

Zoe

Alex

Ben

Chris

Dan

Earl

Fran

Gerry

Hugh

Jim

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?





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

PAM 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 0 0PAT 1 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0JEN 0 0 0 1 1 0 0 1 0 0 0 0 0 0 0 0 0 0

PAU 0 0 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0ANN 0 0 0 1 0 1 1 0 0 0 0 0 0 0 0 0 0 0

MIC 1 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0BIL 0 0 0 0 0 0 0 0 1 0 0 1 0 1 0 0 0 0LEE 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0

DON 1 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0JOH 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 1HAR 1 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0GER 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 1STE 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1 1BER 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 1RUS 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0

45 Halgin & Brass Links 2013

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

1 2 3 4 5- - - - - -

1 0 1 0 3 52 1 0 1 4 3 2 3 0 44 1 0 3 0 35 1 1 1 4 0

Symmetrizing 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?





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

Sample univariate statistics outputUNIVARIATE STATISTICS

--------------------------------------------------------------------------------

Dimension: MATRIX

Diagonal valid? NO

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


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



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



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


Node Level

Node level network property



Actor attribute

Actor attribute Node level network property

8/16/2013

42

83

Examples of node level studies


Node Level



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



Node Level


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



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


Group Level

Group level network property



Other group attribute

Other group attribute Group level network property

90

Examples of group level studies


Group Level



• 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



Group Level


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



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

http://www.analytictech.com/ucinet6/UCINET_list.htm Ethical considerations

http://www.analytictech.com/mgt780/topics/ethics.htm

8/16/2013

51

Popular software packages

101

Analysis UCINET (http://www.analytictech.com/ucinet.htm)

ENET (http://analytictech.com/e-net/e-net.htm)

Pajek (http://pajek.imfm.si/doku.php?id=pajek)

RSIENA R NodeXL

Visualization NetDraw (2D, embedded in UCINET, see above)

Mage (3D, embedded in UCINET, see above)

visit www.netvis.org/resources.php for more

Data collection of personal networks EGONET (http://sourceforge.net/projects/egonet)

Thank You

102

Dan Halgin ([email protected])

Joe Labianca ([email protected])

Tejas Channagiri & Theresa Floyd

OMT Division