Ego-centered Network Analysis Meredith Rolfe, Oxford University Using Sample Surveys to Study Social Networks Connect to wireless Download file .

Ego-centered Network Analysis

Meredith Rolfe, Oxford University

Using Sample Surveys to Study Social Networks

Connect to wirelessDownload filehttp://dl.dropbox.com/u/7685173/polnet2011.R

Whole Networks….S

ourc

e: C

hris

taki

s an

d F

owle

r, 2

008

….vs. EgoNets

Pros and Cons of Egonets

Whole Networks Often limited samples

(college/high school students, work groups)

Political attitudes are often tack-ons, if collected at all

Subjects enjoy Lots of well-developed

analysis methods Nice pictures Recall/Missing Data

Ego Networks General Sample Included in widely available

attitude and election studies (NES, GSS, South Bend, ILS, NCCS, BES, CCAP)

Respondent burden Recall/Reporting errors Sample attrition

EGO NETWORK MEASURES: 5 POLITICAL DISCUSSANTS

The Traditional Approach

Surveys with Political Networks data

Columbia School (Elmira) Detroit Area Studies South Bend 1985 (Huckfeldt and Sprague 1985) Indianapolis-St. Louis (1996 ILS) (Huckfeldt and Sprague

2000) ANES (years) Spencer Foundation 2000 (Mutz) General Social Survey (1985/1987) CNES 1992

Great Britain, Germany, Japan, Spain, and the United States Snowball sample of Spouses (D2) and Other Discussants (D3)

CCAP (US, UK, Germany)

Political Discussion Name Generator

Structure of the Data: Original (Wide)

Main Resp

Name 1 Name 2 … Name 5 D1.Male D2.Male …

1 John Dean … NA 1 1 …

2 Sue Pete … Aya 0 1 …

3 Kate Sara … Ewan 0 0 …

Example: CNES & CCAP

Read in Datacnes.d1<-read.dta("http://dl.dropbox.com/u/7685173/cnes.d1.dta")

cnes.d2<-read.dta("http://dl.dropbox.com/u/7685173/cnes.d2.dta")

cnes.d3<-read.dta("http://dl.dropbox.com/u/7685173/cnes.d3.dta")

ccap<-read.dta(http://dl.dropbox.com/u/7685173/ccap.dta)

Summary of Datasummary(cnes.d1)colnames(cnes.d1)summary(cnes.d2)summary(cnes.d3)summary(ccap.nets)

Structure of the Data: Transformed to Dyadic (Long)

Main Resp

R.Vote D.Num Name

D.Male D.Educ D.Vote …

1 Dem 1 John 1 Coll Rep …

1 Dem 2 Dean 1 HS Dem …

2 Rep 1 Sue 0 HS Rep …

2 Rep 2 Pete 1 Coll DK

2 Rep … …

2 Rep 5 Aya 0 Coll Rep

3 DK 1 Kate 0 HS Dem

3 DK 2 Sara 0 HS DK

3 DK …

3 DK 5 Ewan 1 HS Dem

Transform the Data to Long Format

Step 1 - “line up” the variables correctlyshape<-c(seq(2,10,by=2),seq(3,11,by=2))for(i in 12:23) shape<-c(shape,seq(i,i+48, by=12))temp<-NULLfor(i in seq(1,66,5)) temp<-c(temp,list(names(cnes.sm)[shape[i:(i+4)]]))

Step 2 – reshapecnes.long<-reshape(cnes.d1, varying=temp, idvar="caseid", timevar="discnum", v.names=c("d.given", "d.name", "d.relate", "d.cowork", "d.church", "d.nghbr", "d.friend", "d.close", "d.educ", "d.discpol", "d.disagree", "d.male", "d.expert", "d.vote"), direction="long", time=1:5)

Step 3 – Do any variable recoding (see polnet20011.R)

Personal Network Size

Discussant Name Given? (d1.given – d5.given)cnes.d1$netsize<-rowSums(cnes.d1[,grep(“d[1-5]_given”, colnames(cnes.d1))]==1, na.rm=TRUE)egen netsize = neqany(d1.given d2given d3given d4given d5given), values(1)

Analysisattach(cnes.d1)table(netsize)hist(netsize, breaks=6)##also see lattice versionsumm netsizetab netsize

Non-response

No one to talk totapply(married, netsize==0,mean.na)35% are married!

Nothing to talk abouttable(sp_talkpol, married)

Low political interestprop.table(table(polint, netsize==0),2)chisq.test(polint, netsize)

Forgetting Non-compliance

Online survey issues:Invalid responses

Telephone and Face to Face surveys, no invalid answers that could be verified

Online – there could be many invalid answers!attach(ccap)table(b2.pn1.cat)table(b2.pn1[b2.pn1.cat=="missing (sure)"])

Invalid answers can increase/decrease depending on non-response format and forced/semi-forced choice optionsprop.table(table(b2.pnum==b2.pnum.orig))

Pol

itica

l Dis

cuss

ants

Nam

ed

Network size:Political Discussion sub-network

Need to identify political discussants from important matters discussantstable(d.discpol)table(tapply(d.discpol%in%c("often", "sometimes", "rarely"), caseid, sum.na))

Assign back to “wide” format filecnes.d1$pnum[order(cnes.d1$caseid)]<-tapply(d.discpol%in%c("often", "sometimes”), caseid, sum.na)

Can set different thresholds for discussiontable(tapply(d.discpol%in%c("often", "sometimes”), caseid, sum.na)

Graphic: Network Size of Impt. Matters and Political Discussion Networks

par(mfrow=c(2,2))

hist(cnes.d1$netsize, breaks=6, main="Important Matters", xlab="Network Size")

hist(cnes.d1$pnum, breaks=6, main="Talk Politics Subnetwork", xlab="Network Size")

hist(tapply(d.discpol%in%c("often", "sometimes"), caseid, sum.na), breaks=6, main="Talk Politics Sometimes Subnetwork",xlab="Network Size")

hist(tapply(d.discpol%in%c("often"), caseid, sum.na), breaks=6, main="Talk Politics Often Subnetwork", xlab="Network Size")

###also see lattice alternative

EGO NETWORK DESCRIPTION:

NETWORK COMPOSITION

The Traditional Approach

Personal Network Composition

Who does R discuss politics with? Family, friends, coworkers?

Is political discussion primarily a male activity?

How politically interested are R’s discussants?

Does R disagree about politics with discussants?

Traditional Name Interpreters

How is [name 1] connected to you? spouse or partner other relative [specify] unrelated

Is [name 1] a coworker?Is [name 1] a member of same church?Is [name 1] a neighbor?

CCAP Online SurveyPolitical Discussion Name Interpreter

Structure of the Data: Original (Wide)

Main Resp

Name 1 Name 2 … Name 5 D1.Male D2.Male …

1 John Dean … NA 1 1 …

2 Sue Pete … Aya 0 1 …

3 Kate Sara … Ewan 0 0 …

Structure of the Data: Transformed to Dyadic (Long)

Main Resp

R.Vote D.Num Name

D.Male D.Educ D.Vote …

1 Dem 1 John 1 Coll Rep …

1 Dem 2 Dean 1 HS Dem …

2 Rep 1 Sue 0 HS Rep …

2 Rep 2 Pete 1 Coll DK

2 Rep … …

2 Rep 5 Aya 0 Coll Rep

3 DK 1 Kate 0 HS Dem

3 DK 2 Sara 0 HS DK

3 DK …

3 DK 5 Ewan 1 HS Dem

Non-response: a potential issue

Respondents who name NO discussants

Omit Rs from all composition measures

Divide by netsize=0 or NA will omit R

Respondents who don’t provide one or more composition variables

Omit discussants with invalid or missing information from R’s netsize (must adjust manually for each characteristic)

table(d.given[is.na(d.relate)])

Personal Network Composition:Who does R discuss politics with? Family, friends, etc?

Number of discussants that are family memberstable(tapply(d.relate%in%c(”spouse", ”family”), caseid, sum.na))

Proportion of discussants that are family membersx<-tapply(d.relate%in%c("spouse", "family"), caseid, sum.na)/tapply(d.given==1 & (d.relate%in%c("dk","rf"))==FALSE, caseid,sum.na)

summary(x)

mean.na(x==0)

mean.na(x==1)

Personal Network Composition:Who does R discuss politics with? (continued)

Proportion of discussants that are coworkersx<-tapply(d.cowork=="yes", caseid, sum.na)/tapply(d.given==1 & (d.relate%in%c("dk","rf"))==FALSE & (d.cowork%in%c("dk","rf"))==FALSE, caseid,sum.na)

summary(x)

mean.na(x==0)

mean.na(x>.5)

Personal Network Composition:Is political discussion primarily a male activity?

Proportion of discussants that are maleprop.table(table(d.male)[1:2])

Proportion of POLITICAL discussants that are maleprop.table(table(d.male[d.discpol%in%c("often", "sometimes", "rarely")])[1:2])

prop.table(table(d.male[d.discpol%in%c("often", "sometimes”)])[1:2])

x<-tapply(d.male=="male" & d.discpol%in%c("often", "sometimes"), caseid, sum.na)/tapply(d.given==1 & (d.male%in%c("dk","rf"))==FALSE & d.discpol%in%c("often", "sometimes"), caseid, sum.na)

Personal Network Composition:Is political discussion primarily a male activity? (cont.)

summary(x)

prop.table(table(x==0))

prop.table(table(x>.5))

prop.table(table(x==1))

Frequency of political discussion with male discussantsprint(prop.table(table(d.discpol, d.male)[,1:2]), digits=2)chisq.test(table(d.discpol, d.male)[,1:2])t.test((4-unclass(d.discpol))~d.male, data=cnes.long[d.male%in%levels(d.male)[1:2],])

Personal Network Composition:How politically informed are R’s discussants? (cont.)

Distribution of expertise among all or political discussantsprint(prop.table(table(d.expert)[1:3]), digits=2)print(prop.table(table(d.discpol, d.expert)[,1:3],1), digits=2)

Do R’s prefer informed discussants?chisq.test(table(d.discpol, d.expert)[,1:3])

How many R’s have highly informed discussants?x<-tapply(d.expert=="Great deal", caseid, sum.na)/tapply(d.given==1 & (d.expert%in%c("dk", "rf"))==FALSE,caseid, sum.na)table(x==0)table(x>.5)

Pol

itica

l Net

wor

k C

ompo

sitio

n

Network composition:Do R and discussants disagree about politics?

Self-reportedDisagreement

When you discuss politics with [name], do you disagree

table(d.disagree)

print(prop.table(table(d.disagree, d.vote!=vote)[1:4,],1), digits=2)

Self-reported Vote Choice Differences

Which candidate do you think [name] supported in the presidential election this year?

table(d.vote, vote)prop.table(table(d.vote== vote))[2]

print(prop.table(table(d.disagree, d.vote!=vote)[1:4,],2), digits=2)

Misperception of Political Preferences of Discussants: Snowball Sample

Huckfeldt & Sprague (1987)Non voter

Reagan (Disc)

Mondale (Disc)

Nonvoter

.222 (9)

.790 (19)

.818 (11)

Reagan .400 (20)

.912 (170)

.662 (65)

Mondale .333 (15)

.547 (53)

.992 (90)

Merge data files tp replicatedetach(cnes.long)

temp<- merge(cnes.d3,cnes.d2, by=c(colnames(cnes.d3)), all=TRUE)

cnes.match<-merge(cnes.long, temp,by=c("caseid","discnum"), all=FALSE)

cnes.match$correct<-cnes.match$d.vote== cnes.match$act.both.vote

attach(cnes.match)

Misperception: Analysis

Does misperception depend on discussant vote?prop.table(table(d.vote==act.vote,d.vote),2)

Does misperception depend on respondent vote?prop.table(table(d.vote==act.vote,vote),2)

Does misperception depend on agreement?print(tapply(correct, list(vote, act.vote), mean.na), digits=2)

EGO NETWORK ANALYSIS: PARTICIPATION AND INFLUENCE

The Traditional Approach

Personal Network Analysis:Traditional Approaches One time wave (with snowball)

Instrument for discussant vote choice (H&S, 1991) Two or more survey waves (change)

Kenny (1994) Nested/hierarchical models

van Duijn, Busschbach and Snijders 1999 Lubbers et al 2010 de Miguel Luken and Tranmer 2010

Respondent driven sampling (snowball) Gile and Handcock (2010), Goel and Salganik (2010)

Use in UCInet/whole network software if enough ties

Changes in Personal Networks(from Feld, Suitor and

Existence of Ties Nature of ties

Dyadic Tie Which ties that come and go

Examples: Selection

How characteristics of ties change

Example: Persuasion

Personal Network

Expansion and contraction of network

Example: Stop discussing politics

Change in overall characteristics of network

Example: increase in average political disagreement

Political Participation

McClurg 2003 South bend, dyadic transformed DV: index of political participation Specification: negative binomial regression model3

Controls for socioeconomic status, politically-relevant attitudes, generalized civic engagement, and political mobilization

See Also: McClurg 2006 (“wide” data) See Also: La Due Lake and Huckfeldt, 1998

Dyadic P

articiapation:S

pouses, Fam

ily, and F

riends

“Wide” Participation model

Influence

Kenny (1994) “Microenvironment of Attitude Change” Journal of Politics

OLS on transformed dyadic data DV is directional change in pid from pre-election to post-election (-6

to 6) Includes national and local level context variables Discussion network in 3rd wave

Influence model

EGO NETWORK MEASURES: NETWORK MECHANISMS

The Traditional Approach

Network Mechanisms & Measures

Weak ties Neighbors or

Coworkers Strong ties

Contact frequency transitive ties reciprocated ties

(snowball) Centrality

Self-report

Social capital Organisational

memberships Network size

Network diversity Do you know a… % who disagree % with diff

demographics

Network Styles: Activities

Network Style: Time Use

DRAWBACKS AND SIGNIFICANT ISSUES

The Traditional Approach

Threats to analysis and inference

Mistaken perceptions of discussant political views (minimal)

Sample issues with snowball Selection vs. influence (difficult to disentangle) Loosely defined traditional mechanisms Very narrow conception of how networks might impact

political behavior No network structure Little thought to social cleavages & social groups (old

school)

Solutions?

Longitudinal studies with shorter measures Formal modeling (mathematical and simulation)

identifies new network mechanisms New questions to tap into the mechanisms & cleavages Other test implications if we think interaction matters

(usually with formal models) Network structure measures