Social Networks 38 (2014) 115

Contents lists available at ScienceDirect

Social Networks

jo ur nal homepage: www.elsev ier .com

Assessing structural correlates to social capital innetwor

Brandon pea Department ob Oxford Internc School of Infod College of Info

a r t i c l

Keywords:Social capitalFacebookEgo networksTransitivityCommunity detection

muntively correlated with social capital. This research has drawn primarily on Williams (2006) bridging andbonding scales as well as behavioral attributes such as civic engagement. Yet, as social capital is inher-ently a structural construct, it is surprising that so little work has been done relating social capital tosocial structure as captured by social network site (SNS) Friendship networks. Facebook is particularlywell-suited to support the examination of structure at the ego level since the networks articulated on

1. Introdu

With mowidely use2013). Useexisting frienect with o(Ellison et aon the site potential bethe notion tures both

CorresponCommunicatioTel.: +1 517 35

E-mail add

0378-8733/$ http://dx.doi.oFacebook tend to be large, dense, and indicative of many ofine foci (e.g., coworkers, friends from highschool). Assuming that each one of these foci only partially overlap, we initially present two hypothe-ses related to Facebook social networks and social capital: more foci are associated with perceptions ofgreater bridging social capital and more closure is associated with greater bonding social capital. Usinga study of 235 employees at a Midwestern American university, we test these hypotheses alongsideself-reported measures of activity on the site. Our results only partially conrm these hypotheses. Inparticular, using a widely used measure of closure (transitivity) we observe a strong and persistent neg-ative relationship to bonding social capital. Although this nding is initially counter-intuitive it is easilyexplained by considering the topology of Facebook personal networks: networks with primarily closedtriads tend to be networks with tightly bound foci (such as everyone from high school knowing eachother) and few connections between foci. Networks with primarily open triads signify many crosscuttingfriendships across foci. Therefore, bonding social capital appears to be less tied to local clustering thanto global cohesion.

2014 Elsevier B.V. All rights reserved.

ction

re than one billion active users, Facebook is the mostd social network site (SNS) in the world (Facebook,rs employ Facebook to maintain relationships withnds (Ellison et al., 2007; Hampton et al., 2011), recon-ld friends (Smith, 2011), organize social engagementsl., 2013), and seek information from their connections(Lampe et al., 2012; Morris et al., 2010). To assess thenets of Facebook use, researchers have regularly usedof social capitala sociological framework which cap-the potential and actual resources available from an

ding author at: Michigan State University, 404 Wilson Rd. Room 409,n, Arts, and Sciences, East Lansing, MI 48824, United States.5 8372; fax: +1 517 355 1292.ress: brook205@msu.edu (B. Brooks).

actors network (Bourdieu, 1986; Lin, 2001; Putnam, 2000). In par-ticular, there is an expanding body of research that employs thedistinction between bridging and bonding social capital (Gittelland Vidal, 1998; Putnam, 2000) to characterize the potential ben-ets of Facebook engagement. This distinction was popularized byRobert Putnam, who argues that community organizations work asengines of bonding social capital by bringing together individualsfor shared events and group solidarity (2000). Bridging social capi-tal can be traced to Granovetters (1973) articulation of how weakties enable access to novel information (and consequently greaterjob search success). Since Facebook houses both dense clusters ofstrong ties (Gilbert and Karahalios, 2009) and large swaths of weakties, it is plausible that Facebook can be a site for the activation ofboth bonding and bridging social capital.

Although social capital has its roots in structural analysis, thebulk of social capital scholarship in computer-mediated commu-nication concerning Facebook has focused on survey scales thatrelate perceptions of social capital to individual-level metrics such

see front matter 2014 Elsevier B.V. All rights reserved.rg/10.1016/j.socnet.2014.01.002ks

Brooksa,, Bernie Hoganb, Nicole Ellisonc, Cliff Lamf Media and Information, Michigan State University, United Stateset Institute, University of Oxford, United Kingdomrmation, University of Michigan, United Statesrmation Studies, University of Maryland, United States

e i n f o a b s t r a c t

Research in computer-mediated com/ locate /socnet

Facebook ego

c, Jessica Vitakd

ication has consistently asserted that Facebook use is posi-

2 B. Brooks et al. / Social Networks 38 (2014) 115

as self-esteem, messages sent, and attitudes toward Facebook. Inparticular, many researchers have used Williams (2006) InternetSocial Capital Scales (ISCS) to claim that specic characteristics ofusers networks (e.g., Ellison et al., 2011; Vitak, 2012) and usersbehaviors opositively ptake netwosize. On thexplicitly exFacebook. Hsocial capitbroadly (Fri

In this sFacebook nand measurpast researas time on structural-ltions of soFriggeri et a signicantherefore, tConsistent Ellison et alInternet Soccapital.

One of tis that informatically. Testablishedtors, whichties (McPhetend to focuconnectionare only apertheless inconnectionthat the reofine relatbook is abland Karaha

As our Facebook nsequences obe obvious ego networand thus su1984). Thusships as artthrough egoby considerthe cohesioFriggeri et awe suggestmeasure oftriads (relatego networtightly knittions acrossnetwork.

This papsocial capitasocial netwship examinas a personaresearch qu

approach, variable conceptualizations, and descriptive data aboutour participants. We then present the results of a series of bivariateand multivariate analyses and conclude by discussing how networkstructure can partially inuence the perception of social capital in

tworo Faial ca, can

thised andinriads

frag

ratu

ncep

ial cces we or as beeldics, r

tendocialconcnalyoverlr, 200sidererenok isns onvely ve. Tinal ses t

quesetwog soonneform).ert n betion al (1

zatiogageese e outhe poight ciateerneial cits cors ex

timeced srsonaliamethoh whn the site (e.g., Burke et al., 2011; Ellison et al., in press)redict perceptions of social capital. When these studiesrk composition into account they tend to use networke other hand, there is a small body of research thatamines ego-centric measures of network structure onowever, these studies tend to take network structure asal (Brooks et al., 2011) or examine social cohesion moreggeri et al., 2011).tudy we jointly consider the structural properties ofetworks, scales of bridging and bonding social capital,es of site engagement. In doing so, we wish to extendch that has examined individual level variables, suchthe site, while explicitly considering the potential forevel metrics to have an independent effect on percep-cial capital. Consistent with Brooks et al. (2011) andal. (2011), we assume that dense clusters of ties havet bearing on the overall cohesion of the network, andhe likelihood of resource provision from the network.with other work in this vein (e.g., Burke et al., 2011;., in press), we use a modied version of Williams (2006)ial Capital Scale (ISCS) to measure perceptions of social

he attractions of researching Facebook ego networksmation about virtually all alters is available program-his allows us to operate at a scale in between two

strategies for capturing ego networks: name genera- tend to focus mainly on the small number of core socialrson et al., 2006), and enumeration methods, whichs on estimating total network size but forgo alteralters (McCarty et al., 2000). Although Facebook networksproximations of ofine personal networks, they nev-clude large swaths of weak ties and the alteralter

s between these weak ties. Further, past work has shownlationships on Facebook tend to be characteristic ofionships (Ellison et al., 2007), and that activity on Face-e to discriminate ofine strong and weak ties (Gilbertlios, 2009; Jones et al., 2013).ndings suggest, one of the further advantages of usingetworks is that we can assess with high delity the con-f linkages across social groups that may not necessarilyto ego, but still felt as a form of social cohesion. In mostk analysis studies alteralter ties are reported by ego,bject to a host of inaccuracies and biases (Bernard et al.,, the network that is analyzed is not a list of friend-iculated by the friends, but a list of friendships as seens eyes. In this regard, we extend Friggeri et al. (2011),ing the cohesion of the network as a whole, rather thann of distinct clusters within the ego network. Whereasl. use closed triads to signify distinct social groupings;

that the presence of open triads may in fact be a better global cohesion, and that the presence of many closedive to open triads) is in fact a strong indicator that thek is highly fragmented. Each individual cluster might be, but the lack of open triads indicates a lack of connec-

groups, and potentially a lack of social cohesion in the

er is organized as follows: First, we review the use ofl in studies of computer-mediated communication andork analysis. Second, we summarize current scholar-ing Facebook, both as a resource for social capital andl network measurement tool. We then dene our basicestions and hypotheses followed by our methodological

ego netudes tfor socclosureesting,assumless boopen tthat is

2. Lite

2.1. Co

Socresourof mor51)hdemic in politcapitalfrom sof the work abeing Fischeto conto diffFaceboopinioeffectithe aboattitudproceson thetheir nbondinwork cnew incapital

Robtinctiodistincand Vidorganiand enthat thpositivmore, that man assothe Intfor socabout scholatakingenhanthe pe

Wilas a mthrougks on Facebook. In general, we assert that individual atti-cebook usage remain the strongest explanatory factorspital, but that structural measures, particularly triadic

have a strong independent effect. Perhaps most inter- effect of triadic closure is opposite to what would be

higher clustering coefcient is actually associated withg social capital. We argue that this is the result of less

across groups and is experienced by ego as a networkmented rather than globally cohesive.

re

tualization of social capital

apitalthe aggregate of the actual or potentialhich are linked to possession of a durable network

less institutionalized relationships (Bourdieu, 1986, p.en adapted and integrated into a large number of aca-s. Scholars have explored the presence of social capitaleligion, education, family, and culture. In all cases, socials to be a general stand-in for positive social outcomes

interaction. The prominence (and perhaps the dilution)ept of social capital has led members of the social net-sis community to criticize the notion of the concept asy general, instrumental and articial (Kadushin, 2004;5; Fine, 2010). However, there remains a plausible need

how structural features and individual behaviors leadces in perceptions and outcomes of social resources.

not solely a site for sharing music tastes, comparing current affairs, or organizing social events. Rather, itfunctions as a computer-mediated platform for all ofhus, our operationalization of social capital emphasizessentiments, and any descriptions of these resources ashat can be invested or traded are ancillary. We focustion of whether individuals believe they can draw uponrk for emotional and material resources (as a measure ofcial capital) and whether individuals believe their net-cts them to the wider world and provides them withation and experiences (as a measure of bridging social

Putnam is widely regarded as popularizing the dis-tween bridging and bonding social capital (even if theis often attributed to the previously published Gittell998)). In Bowling Alone, Putnam argued that communityns enabled individuals to converge in shared locations

in activities that increase group solidarity. He assertedorganizations were associated with a large number ofcomes, such as greater health and lower crime. Further-stulated that television was among a number of factorsbe responsible for the decline in voluntary activity andd decline in social capital. At the time of its publication,t was only beginning to emerge as an object of studyapital in everyday life and Putnam remained agnosticnsequences for public life. Subsequently, a number ofplored whether the Internet impeded social capital, by

away from ofine activities (cf., Nie et al., 2002) orocial capital, by providing increased connectivity withinl network (Quan-Haase and Wellman, 2004).s (2006) addressed the growing popularity of CMCd of communicationand thereby a separate outletich social capital could be created and exchangedby

B. Brooks et al. / Social Networks 38 (2014) 115 3

constructing scales of social capital and examining online andofine variants of it. Drawing on Putnams articulation of the dis-tinction between bridging and bonding capital (2000) and Resnicksextension toward sociotechnical capital (2001), Williams (2006)developed sfocused on as a sense oitems assessomeone waccess to aing social creported thdifferent kinot attempthey assess in a speciconline an(e.g., Ellisoncontexts ha

Subsequlighted sigbonding soSNS-relatedesteem (BuMendelson

2.2. Facebo

In the lasuse of SNSssocial capittional and sEllison et alBrooks et alerature reveoperational

Much ofcapital and perception-son and coestablishedtics of Facebonding rescompositio2011)whiof a users how diversitively predengagemenmaintenancwish a Frien

Moving by Burke acapitalaga(2006) scaltial study (Bpositively cof directedsages a userelated to bwith a specof access tbridging rethe authorsbonding socauthors sug

through a single good friend and is thus not reliant on changesin Facebook-based communication. However, the authors did ndthat a more sensitive measure of directed communication that onlycaptured inbound messages (i.e., posts, messages, and Likes sent by

s) poatasetiont alse byend whe opnot brk str, Face egorate. Brombeial cary nmber

is thgo. Hunitilly a

et aic r

her a usedlled ok F. TheWouicipathe gted tred tohesmu

ohes 2). Tluateriggemuntogettweenetwus, thident

cebo

eboosuchonteeoplltistrtrand

(19ts pecontee. Fisemb

rks htivitontecales to capture bonding and bridging social capital thattheoretical components associated with social capitalf access to social resources. For bonding social capital,sed the extent to which participants reported havingho could provide emotional support and advice and

scarce resource, such as a nancial loan. For bridg-apital, items assessed the extent to which participantsey had interactions that are consistent with access tonds of people or diverse worldviews. These scales dot to quantify the volume of resources available; rather,respondents perceptions of the availability of resources

context; in Williams (2006) work, this was divided intod ofine contexts, while in subsequent research, local

et al., 2007) and site-specic (e.g., Ellison et al., in press)ve also been used when framing the items.ent research employing Williams scales has high-nicant relationships between both bridging andcial capital and a number of behavioral and attitudinal

factors such as various Facebook activities and self-rke et al., 2010; Ellison et al., 2007; Papacharissi and, 2011; Steineld et al., 2008; Valenzuela et al., 2009).

ok and social capital

t decade, researchers have explored the extent to whichand specically Facebookis associated with variousal constructs, including perceived access to informa-upport-based resources (e.g., Burke et al., 2010, 2011;., 2007, 2011, in press) and network characteristics (e.g.,., 2011; Friggeri et al., 2011). As detailed below, the lit-als a complex relationship between use and the variousizations of social capital.

the research looking at the relationship between socialSNS use has employed an adaption of Williams (2006)based social capital measure. Notably, work by Elli-lleagues (e.g., Ellison et al., 2007, 2011, in press) has

a positive relationship between various characteris-book use and perceptions of access to bridging andources. For example, characteristics of a users networkn, such as the number of actual friends (Ellison et al.,ch was intended to capture a more meaningful measurenetwork than total number of Facebook Friendsande they perceive that network to be (Vitak, 2012) pos-ict users perceptions of social capital, as does userst in communication behaviors that support relationshipe, such as when they respond to a request for advice ord happy birthday (Ellison et al., in press).one step beyond strictly perceptual data, researchnd colleagues has examined perceptions of socialin measured using an adapted version of Williamse and server-level data of Facebook use. In their ini-urke et al., 2010), they found that while Friend countorrelated with both forms of social capital, ones level

communicationmeasured as the number of mes-r exchanged with another Facebook Friendwas onlyonding social capital, such that increases in interactionic Friend were associated with increased perceptionso bonding resources from their network, but not tosources. In a follow-up longitudinal study, however,

(Burke et al., 2011) found no relationship betweenial capital and directed communication over time; thegested that bonding social capital may be generated

Frienddinal dinteracient, bupurposthe Fri

At tsured netwocapitalanalyzto genecapitalthe nuing socfor evethe nudegreewith eopportnormaBrookseconombut rat(2011)tion caFaceborithm)them, If partname then racompathat cthe commore cof 1 orto evaever, Fof comTaken link besocial ual. Thabove

2.3. Fa

Factexts, each csame p(or mumultisFischercontexmany are rarwere mnetwoconnecto all csitively predicted bridging social capital in the longitu-t. Ellison et al. (in press) argue that these more visible

s serve to signal ones relationship, not just to the recip-o to the entire network, and can serve a social grooming

highlighting the relationship and potentially providingith a needed resource.posite end of the spectrum, social capital may be mea-y users perceptions of resources, but through theiructure. While still rarely used in online studies of socialebook provides an ideal environment to measure and

networks. Brooks et al. (2011) utilized Facebooks API personal networks for the purposes of measuring socialoks et al. conceptualized bridging social capital based onr of clusters within an individuals network and bond-pital as the average degree of the network. The degreeode in an undirected ego network can be considered

of friends that alter shares with ego. Thus, the averagee mean number of mutual friendships all alters haveaving many mutual friends, on average, implies manyes for reciprocity, closure and other structural featuresssociated with bonding social capital. Findings froml. suggest that socioeconomic status or more diverseesources was not associated with number of cliques,

larger and more dense network. Likewise, Friggeri et al. an online experiment employing a Facebook applica-Fellows to present users with a visualization of theirriends network (generated using a simple greedy algo-

application showed users a group of Friends and askedld you say that this list of friends forms a group for you?nts answered yes, they were given the opportunity toroup and save it as a Friends list on Facebook. Usershe quality of these suggested groups. When researchershese ratings to the cohesion of the group, they foundion is a strong indicator of users subjective perception ofnity-ness of a set of people as indicated by the fact thative groups received higher ratings (e.g., 4 stars insteadhis work suggests that cohesion is a representative way

a community, at least at a correlational level. How-ri et al. (2011) did not nd support for typical measuresity quality such as density, clustering, or conductance.her, these two papers suggest that there is some missingn personal network measurement and the theoreticalork structure of social capital attached to the individ-is current study examines the association between theied two methods for conceptualizing social capital.

ok personal networks and structural social capital

k networks represent ties from a variety of social con- as school, work, church or the neighborhood. Whilext may be distinguishable from another, some of thee may exist in multiple contexts, leading to multiplexanded) ties. Past work on personal networks has foundedness to be a common occurrence. For example, in82) Northern California study, the average number ofr alter was 1.6. At the same time, most ties do not occupyxts, since ties bridging more than two social contextscher noted that on average only 2.6 alters per networkers of three or more contexts. This suggests personal

ave a structure that is characterized by high degrees ofy across contexts, but not necessarily a core commonxts. It is not the case that a core set of ties stand as a

4 B. Brooks et al. / Social Networks 38 (2014) 115

lynchpin linking all contexts, but different ties link the multitudeof contexts into a cohesive personal network. This assertion is alsoaligned with past analysis of personal networks, such as Wellmanand Wortley (1990) and McCarty (2002).

Structurpotential tHoughton ain that socipresented tencourage Facebook egroups, theand, in manexample, inFriend Listcontent thecontexts ofone hand, suand interacThis consolthus social mation acroto congregasuch contexfor examplesible to chuare made vthat more increased pHoughton aviolations Fvate informmight consnote are a ntal resourceit easier to on a new ptive or taboor family pr

We belidepending interacts wthat are hienables a spotentially Other indivsure betwenetworks aor open (exist side-btion of the sinformation

The struus toward great deal obook netwoQuercia et (Gilbert ansis examiniFriggeri et existing undwork visuabelieve thetypifying ththat capturidentify coh

the Louvain method (Blondel et al., 2008) and assess the globalcohesion of the network across various clusters through averagedegree and transitivity. We discuss our hypotheses and researchquestions based on the above literature.

earc

ndin

earchs of, a ns amers.

s (2en nnetws reasonaigherial soen corahal highl. Sincork,

hesiscial c

andinetw

papetwo

meas higre wrk wtioninkager pren coore

riphe

hesiscial c

h tra the

fewpractsent tivit

he hiivityosed

idgin

ouge mure n

a net byok nal properties of personal networks on SNSs have theo engender context collapse (Binder et al., 2009;nd Joinson, 2010; Marwick and Boyd, 2010; Vitak, 2012)al connections from different life contexts tend to beogether in the same stream and because these sitesbroadcasting content to ones entire network. Whilenables users to segment their connections into sub-

default settings typically reect ones full audiencey cases, few users take advantage of these features; for

one study, just 14% of users reported using Facebooks feature to recreate some of the ofine groups and ltery shared (Vitak, 2012). The net effect of the collapsedten present in SNSs on social capital is unclear. On thech context collapse provides an opportunity to monitort with a diverse set of ties in one convenient location.idation may enhance access to diverse resources andcapital because individuals can learn about new infor-ss social contexts in a single sitting, rather than havingte with each individual cluster of ties. On the other hand,t collapse may engender new privacy concerns, where,, pictures from the bowling teams bar night are acces-rch friends, or political messages meant for ones friendsisible to work colleagues. Binder et al. (2009) founddiverse networks on Facebook were associated witherceptions of tension in their social spheres whereasnd Joinson (2010) documented a number of privacyacebook users experienced, many resulting from pri-ation being shared across contexts. These situationstrain disclosures on the site, which some researchersecessary requirement to accessing specic social capi-s (Ellison et al., 2011). For example, individuals may ndask for information about mundane topics like advicehone or a vacation spot and harder to ask about sensi-o topics like coming out, health issues, nancial woes,oblems (Newman et al., 2011).eve context collapse may be experienced differentlyon the arrangement of the various contexts that egoith on Facebook. Some individuals may have networksghly fragmented, meaning that the default newsfeedingle cross-cutting view of many diverse areas andincongruous social circles of ones personal network.iduals may have networks with a high amount of clo-en social contexts. Whether individuals perceive theirs closed (with many linkages across social contexts)having content from highly differentiated social rolesy-side) ought to make a difference on their percep-ite as a channel for garnering social support and social.ctural properties of Facebook ego networks help guidereasonable social capital metrics for analysis. While af work has focused on the number of ties in a Face-rk, relating it to personality (Golbeck and Robles, 2011;al., 2012), brain size (Kanai et al., 2011) and closenessd Karahalios, 2009), there has been much less analy-ng the topology of personal Facebook networks (c.f.,al., 2011; Brooks et al., 2011). As such, we draw uponerstandings of personal networks, intuitions from net-

lizations, and past work where available. Insofar as wese networks consist of multiple locally dense clusterse social contexts of personal life, we employ measurese these features. Specically, we employ a measure toesive subgroups embedded in a personal network using

3. Res

3.1. Bo

Resfeelingwritesrelatioing othreturnbetwelarger For thiin perwith hpotentalso beand Kaoveraloverala netw

Hypoting so

Expof the In thissonal nwe aretivity istructunetwoconnecmore lwhethbetwethese cthe pe

Hypoting so

Botwithinmakesyet in be preconnecfrom ttransitis uncl

3.2. Br

Althincludthere atexts inattempFaceboh question and hypotheses

g social capital

suggests that network density is strongly related to social support and social capital. For example, Lindenser network with more intimate and reciprocalong members may increase the likelihood of mobiliz-

. .to defend and protect existing resources/expressive001, p. 20). Unfortunately, density scores vary widelyetworks of varying sizes, placing undue demands onorks to have disproportionately more ties per alter.son, we opt to include average degree. Average degreel Facebook networks has previously been correlated

socioeconomic status, suggesting a baseline for highercial resources (Brooks et al., 2011). Alters degree hasrrelated with traditional measures of closeness (Gilbertlios, 2009), although the latter nding did not imply thater average degree meant a presence of more close tiese average degree implies more interconnections withinwe follow Lin in proposing:

1: Average degree is positively related to perceived bond-apital.

ng on average degree we use a measure of the cohesionork rather than simply the density or average degree.er, we are interested in the overall cohesion of a per-rk. By including a measure of triangles over two-paths,suring the presence of sites of local density. If transi-h, then either the network exhibits a coreperipheryith many dense connections in the core, or a multi-coreith few linkages between said cores, but many denses within each core. In either case, transitivity meanses within a set of nodes. Thus, the dense connections,esent in the core or multiple cores with a few tiesres, are a measure of bonding social capital assuminggroups represent stronger personal ties than those onry.

2: High transitivity is positively related to perceived bond-apital.

nsitivity and average degree signify greater connectivitynetwork, but do so in different ways. Average degreeer assumptions about how the network is connected,ice, we assert that a higher average degree is likely toin networks characterized by pockets of high internaly such as high school friends. In this case, more linkagesgh school context to other contexts may actually lower

as each new link between these separate contexts that would lower transitivity.

g social capital

h it is now widely accepted that Facebook networksltiple social contexts (Binder et al., 2009; Vitak, 2012),o established ways to count the number of social con-twork programmatically. Friggeri et al. (2011) make one

using overlapping clusters drawn from a respondentsetwork and asked the respondent to rate the quality

B. Brooks et al. / Social Networks 38 (2014) 115 5

of these clusters. They use a novel cohesion measure based ontriangles inside and outside a cluster. In that work, users tendedto rate highly dense clusters as accurately signifying a group, andless dense clusters as less accurately describing a group. That said,the demandsmall groupof social consignify diffetal. Consequalgorithm (to capture groups of sodistribution

We belisocial capitor bridgingsocial capitsense of conmizes moduthere are feeach clusterleast inform

We thennetwork reters represethrough thethis measur

Hypothesispositively r

3.3. Beyond

Social cathat individhis scales, Wis to measuthe items inuser behaviThis strategwork topolo(2005). Thisdeploy strusures separour foremostructural cthe relationof social capaware of whwork. Put bknow who Facebook rbetter at pasider indiviscales deveship MainteInformation

ISB and with their Fscales attemtains, as wesuch, we hyon bridgingformally we

Hypothesis 4a: Greater FRMB will be associated with both bridgingand bonding capital.Hypothesis 4b: Greater ISB will be associated with both bridging andbonding social capital.

thermese nshipt thake an

que.

hesisural nhesisural n

nsid

ed oned aging ips wlievefrienre frilso m

too bte reothest.

hat ook a

thod

datan onl

in iporte

samn universembeted ts foeted . Sprenterrvey ceboes an

incl Namtionrt ofacebe aprk tom theted t; ofed ths of that algorithm lead to a large number of relativelys in a given ego network rather than a coarse numbertexts such as work, family and high school that wouldrent arenas for the distribution of bridging social capi-ently, we employ a widely-used community detectionthe Louvain method; Blondel et al., 2008) in orderthese larger group structures that represent differingcial cohesion and thus different sites for the potential

of bridging social capital.eve that our approach captures the spirit of bridgingal, as it implies the spread of information across groups

between pockets of dense ties. We consider bridgingal as a metric of access to diverse social resources and anectedness to the wider world. A technique that maxi-larity presents a count of the number of groups wherew connections between the groups, thus ensuring that

would uniquely contribute different information, or atation from a substantively different set of alters.

assume that each unique cluster within a Facebookpresents a specic context. The number of those clus-nts access to possible diverse social resources captured

reported bridging social capital measure. Thus, usinge we propose:

3: More clusters (found using community detection) iselated to perceived bridging social capital.

structure: mobilizing social capital

pital encompasses both actual and potential resourcesuals have access to through their network. In designing

illiams (2006) noted that the intent of these scalesre sentiment as expected outcome. As such, none of

the scales measure either actual network topology orors that would lead to particular structural outcomes.y of explicitly distinguishing social resources from net-gy was also articulated by Van Der Gaag and Snijders

emphasis on perceived outcomes makes it possible toctural, behavioral, demographic and attitudinal mea-ately to assess their relationship to the scales. Whilest emphasis in this paper is on the relationship betweenovariates and social capital, we also wish to addressship between the latent structure as a potential siteital and the behaviors actors employ to become moreat potential social resources are embedded in the net-luntly, what good is having a network if you do nothas what resource? Based on previously mentionedesearch, we pose the question, are some individualsying attention to their network than others? We con-duals engagement with their network using two recentloped specically for this purpose: Facebook Relation-nance Behaviors (FRMB); Ellison et al. (in press)) and-Seeking Behaviors (ISB; Lampe et al., 2012).FRMB are ways of examining individuals engagementacebook network, while Williams (2006) social capitalpt to elicit perceptions of resources that network con-

ll as individuals perceived access to those resources. Aspothesize that these measures have both a direct effect

and bonding social capital and a mediating effect. More propose:

Furthat threlatiosuggeswho taing andcapital

HypotstructHypotstruct

3.4. Co

Basdescribof bridtionshWe beactual tal. Mocould aresentintimaof hypinteres

RQ: WFaceb

4. Me

Thewith auniqueself-re

Thewesterthe unin Novcomplticipancomplhereinto be The suics, Faattitudsurveysion ofinstructhis patheir Fvey. Thnetwo

Frocomplaccounaccessore, with regards to the mediating effect, we suggestscales will attenuate but not eliminate any positive

between structural measures and social capital. Wet given a specic network topology, those individuals

active interest in their network through social groom-stion-asking are likely to report higher levels of social

5a: FRMB will partially mediate the association betweenetwork properties and bridging and bonding capital.

5b: ISB will partially mediate the association betweenetwork properties and bridging and bonding capital.

ering Facebook activity

previous work (Ellison et al., 2011), having more self-ctual friends on Facebook should lead to higher levelscapital, because these represent more meaningful rela-ith ego (as compared to very weak, non actual Friends).

that there are compelling arguments for why moreds could lead to either higher or lower bonding capi-ends could mean a denser set of core ties for bonding. Itean that individuals feel their Facebook networks rep-

road a set of ties to successfully capture the affective andsources associated with bonding capital. Thus, insteadizing we consider its effect to be a research question of

is the relationship between number of actual friends onnd bridging and bonding social capital?

s

collection for this research combined an online surveyine Facebook API network generator. This method wasts technique for comparing ego network metrics andd behavioral data.ple for this study was recruited through a large Mid-iversity within the United States. An email was sent byity to a random sample of 3149 non-faculty staff (1000er 2010 and 2149 in February 2011). Fall participantsa short screener survey which was used to select par-r a more detailed lab session. Lab session participantsthe full survey in addition to other tasks not discusseding survey participants were invited to provide theired into a drawing for one of ten Amazon gift cards.asked a series of questions including user demograph-ok use, Williams (2006) social capital scales, privacyd behaviors, and network characteristics. The online

uded a link to a Facebook application (a modied ver-eGenWeb [Hogan, 2010]), that included study-specic

s and a consent statement. Participants who agreed to the study had to (temporarily) add the application toook account while they completed the rest of the sur-plication saved a copy of the participants Facebook ego

a server maintained by the second author.e fall and spring data collection efforts, 666 participantsa survey, with 534 reporting having an active Facebook

that group, 238 added the Facebook application whicheir Facebook network data; this latter group will be used

6 B. Brooks et al. / Social Networks 38 (2014) 115

in all analyses in this paper. Among the network-only subsample,the average participant was female (66.8%), 45 years old (SD = 10.8)and a college graduate (43.7% had a bachelors degree, 35.3% hadpostgraduate training). These numbers correspond closely to thedemographic composition of the entire sample.

4.1. Measures

The two types of data collected for this studyself-reportedperceptual measures and ego network characteristicsare detailedbelow. Scale items were measured on ve-point Likert-type scalesranging from Strongly Disagree to Strongly Agree unless otherwisenoted. All non-network variables used less than 5% missing val-ues, which were imputed using mean replacement. Items, means,and standard deviations for each scale are presented in AppendixA. Table 1 also provides full descriptive characteristics for variablesused in the study.

4.2. Survey items

Gender was a self-report item with the option of male orfemale with three missing values, which were excluded fromanalyses and descriptive reporting. Age was self-reported in years.Education was asked using an ordinal question with responsesless than high school, high school degree, technical, tradeor vocationdegree, cschool after

Faceboousing an ad(2006). For language wtinguish beFacebook Fnetwork, wwho are nosider the Faasked to thFriends whaccess vario

The bonincludes itein my Facebwith in myme. The b

includes questions such as, Interacting with people in my Face-book network makes me want to try new things and Interactingwith people in my Facebook network reminds me that everyone inthe world is connected.

Facebook Relationship Maintenance Behaviors (FRMB) mea-sures the extent to which Facebook users engage in social groomingand attempt to respond to requests from their Facebook network,which may in turn signal that ego is paying attention to alter (Ellisonet al., in press). The scale (Cronbachs = 0.90, M = 3.72, SD = 0.80)includes ve items, four of which reference users likelihood torespond to requests from other members of their network and afth that captures the common practice of signaling attention toa specic Friend by writing Happy Birthday on their Wall. Pastresearch employing this measure found signicant differences inperceptions of bridging and bonding social capital across differ-ent levels of engagement in FRMB, although measures of networkstructure were not considered (Lampe et al., 2012).

Information-Seeking Behaviors (ISB) examines the extent towhich individuals use Facebooks communication features to seeka range of informational resources from their network (Lampe et al.,2012). This scale (Cronbachs = 0.83, M = 2.34, SD = 0.83) measuresparticipants use of Facebook for getting information or adviceregarding purchases, health, business referrals, and other specicquestions.

Self-esteem was measured using Rosenbergs (1989) seven-itemed sinclue whual frson eany ? Atil exndiv

ofiux awit

ch asple mrks o

the appr frieon the nu

Table 1Sample descri

Nodes Average degTransitivity Clusters Modularity Giant compo

Gender (FemAge Education (oSelf-esteem

Actual friendVisits per daFacebook enInfo-seekingFacebook boFacebook br

N = 235.al school after high school, some college, no 4-yearollege graduate, post-graduate training/professional

college and I dont want to disclose.k bridging and bonding social capital were measuredapted version of Williams bridging and bonding scalesthis study, we replaced Williams (2006) online/ofineith on Facebook and in my social network to dis-tween social capital perceptions associated with theirriends and perceptions associated with their full socialhich includes their Facebook Friends as well as thoset on the site, respectively. In this paper, we only con-cebook-specic responses, in which participants wereink only about their interactions with their Facebooken reporting the extent to which they felt they couldus kinds of resources.ding scale (Cronbachs = 0.88, M = 3.40, SD = 0.73)ms such as, When I feel lonely, there are several peopleook network I can talk to and The people I interact

Facebook network would share their last dollar withridging scale (Cronbachs = 0.90, M = 3.47, SD = 0.66)

validatitems On th

Acton Ellihow mfriendslist unlater, itact. Inis in 1997)gies suas peonetwopeoplerecentactualspent with th

ptive statistics.

Mean Median

217.74 153.00 ree 14.02 12.00

0.57 0.55 6.54 6.00 0.43 0.46

nent percentage 0.82 0.89

ale) 0.70 1.00 44.21 46.00

rdinal) 5.05 5.00 4.32 4.29

s on Facebook 80.90 45.00 y to Facebook 2.21 2.00 gagement (FRMB) 3.73 4.00

on Facebook 2.35 2.25 nding capital 3.40 3.50 idging capital 3.47 3.60 cale (Cronbachs = 0.86, M = 4.32, SD = 0.50). Samplede, I feel that I have a number of good qualities andole, I am satised with myself.iends was measured using a self-report question, basedt al. (2011), in which users were asked, Approximatelyof your TOTAL Facebook Friends do you consider actualt present, Facebook Friends remain in ones Friendsplicitly removed by one of the dyad. Thus many yearsiduals may still be friends on the site despite no con-ne personal networks, however, network membershipnd ties tend to fade over time (Suitor and Keeton,

hout the ease of connection offered by social technolo- Facebook, individuals may lose the ability to re-connectove away, change jobs, etc. Thus, individuals may haven Facebook that do not reect their active ties, or eveny would consider friends. To address this, we followoaches that ask participants to report on the number ofnds within their Facebook network as well as the timee site. As noted in Table 2, actual friends is correlatedmber of Facebook Friends (r = 0.49, p < 0.001). To note,

SD Min Max

227.61 14.00 1950.0010.67 1.00 70.000.12 0.33 0.953.45 1.00 22.000.17 0.00 0.770.17 0.24 0.99

0.46 0.00 1.0010.73 23.00 65.000.95 2.00 6.000.50 2.57 5.00

103.17 0.00 700.001.12 1.00 5.000.79 1.00 5.000.83 1.00 4.750.74 1.00 5.000.66 1.00 5.00

B. Brooks et al. / Social Networks 38 (2014) 115 7

Table 2Bivariate correlations between variables of interest.

Nodes Average degree Transitivity Clusters Modularity Actual friendson Facebook

Self-esteem

Average degTransitivity Number of cModularity 0.40 **

Actual friend 0.37 ** 0.32 **

Self-esteem 0.0 *

Age 0.3Sex 0.1Education 0.0Visits per da 0.4Facebook en 0.3Info-seeking 0.3Facebook br 0.3Facebook bo 0.3

Visitday Face

Sex Education Visits per daFacebook en 0.42Info-seeking 0.41Facebook br 0.35Facebook bo 0.27

N = 235. p < 0.1.* p < 0.05.

** p < 0.01.***p < 0.001.

non-paramtionship (Spaccount forreasonably Individualstheir netwoon the site.

Visits peusing a self-et al., 2010)measured bSuch a meakeep Facebpercent of oand 5% said

4.3. Persona

The pera custom-blink in the on code frapplicationwork formato downloapplicationloaded fromcalculated.

The netwnetwork onstatistics, th(Facebook, The mean nily skewed

rks (367to th

withThat rageltersree 0.82 **

0.40 ** 0.21 **lusters 0.45 ** 0.31 ** 0.52 **

0.27 ** 0.04 0.46 **s on Facebook 0.49 ** 0.33 ** 0.38 **

0.04 0.01 0.04 0.37 ** 0.37 ** 0.29 **0.08 0.00 0.10 0.05 0.03 0.04

y to Facebook 0.47 ** 0.37 ** 0.38 **gagement (FRMB) 0.24 ** 0.21 ** 0.23 **

on Facebook 0.34 ** 0.32 ** 0.28 **idging capital 0.26 ** 0.18 ** 0.21 **nding capital 0.20 ** 0.12 0.37 **

Age Sex Education

0.080.11 0.09

y to Facebook 0.36 ** 0.12 0.00gagement (FRMB) 0.12 0.25 ** 0.01

on Facebook 0.19 ** 0.25 ** 0.15 *idging capital 0.05 0.24 ** 0.01 nding capital 0.26 ** 0.10 0.00

etric correlations indicate a substantially stronger rela-earmans rho = 0.67, p < 0.001) suggesting that when we

the few individuals with very large networks, there is aclose, if not perfect, relationship between these values.

report having a mean of 217.7 nodes (median 153) inrks, but report a mean of 81 (median 45) actual friends

netwo1050, 1closer dense,0.57.1

on aveother ar day to Facebook measured time spent on Facebookreported measure of visits per day. Past research (Burke

has indicated that visits per day was more accuratelyy participants than time spent on the site in minutes.sure also accounts for instances where individuals willook available for chatting on multiple devices. Thirtyur sample said they visit Facebook once a day or less

they visit Facebook ve or more times per day.

l Facebook networks

sonal networks of participants were collected usinguilt Facebook application accessed through a hyper-survey. The algorithm for this application was basedom Hogans (2010) NameGenWeb, a public-facing

for downloading ones friendship relations in a net-t, such as GraphML (Brandes et al., 2002). In order

ad the network, respondents had to approve the from within their Facebook accounts. Data down-

this project was cached while user statistics were

orks used in this study were larger than the average Facebook, as reported by the site. According to the sitese average network contains approximately 130 nodes2011), although this number varies widely by country.umber of nodes in our survey was 217, but was heav-by the presence of a few individuals with very large

Facebooconnected work compnodes in thremaining nof 99% (meaponent, an igroups coexarate socialconnectivittiple socialclusters. Thtechniquesber of compwhen visua

Transitiways, but ttimes the cof all two psecond is thtivity arounthese resul

1 All networand Python 2.7

2 This gurewith clusters d2 0.03 0.132 ** 0.30 ** 0.23 ** 0.061 0.10 0.16 * 0.048 0.05 0.01 0.062 ** 0.28 ** 0.44 ** 0.067 ** 0.20 ** 0.32 ** 0.104 ** 0.18 ** 0.40 ** 0.041 ** 0.19 ** 0.32 ** 0.11

0 ** 0.27 ** 0.34 ** 0.17 *

s pertobook

FRMB IBS Facebookbridgingcapital

**

** 0.49 **** 0.59 ** 0.54 **** 0.44 ** 0.42 ** 0.51 **

the size of the ve largest networks were 843, 1026, and 1950). The median network size was 153, which ise global average of 130. These networks were also very

an average degree of 14 and an average transitivity ofis to say, over half of all potential triads are closed and

each alter in the network is connected to at least 14.

k personal networks in our sample tended to be welloverall, in that the giant component for any given net-rised most of the network. The median percentage ofe giant component was 89 and the mean was 82. Theodes tended to be isolates or isolated dyads. A mediann of 93%) of all nodes was either part of the giant com-solate, or a dyad. Thus, while it is likely that many socialist in ones Facebook network, these groups are not sep-

islands. Rather, they are zones of either greater or lessery where some alters will be highly connected to mul-

groups while others will be primarily tied to specicis also reinforces the use of modularity maximization

for partitioning rather than simply counting the num-onents. Fig. 1 is characteristic of the networks we sawlizing Facebook personal networks.2

vity in a network can be measured in a multitude ofwo conventional measures stand out. The rst is threeount of triangles in the graph divided by the countaths (Davis, 1967; Holland and Leinhardt, 1971). Thee global clustering coefcient, which measures transi-d each node in the network and takes the average ofts (Watts and Strogatz, 1998). Also, these calculations

k metrics were calculated using iGraph 0.6 (Csrdi and Nepusz, 2006).

does not come from the study, but was created by one of the authors,etected using the same algorithms as done for participants.

8 B. Brooks et al. / Social Networks 38 (2014) 115

Fig. 1. Faceboowner has labeswatches deno

exclude egosent 4-cliquinterest herobject, andson for exclisolates, whscore.

Clustersdetection resent distiLouvain mnetwork inmodularitywork that iacross conting to bothalternate mmodularityok network with clusters found through multilevel community detection (the Louvain meled the clusters and veried their accuracy as distinct subgroups. Network captured usingte separate subgroup membership.

, so in the complete graph, triangles actually repre-es, as ego is connected to all three alters. However, oure is in the network that ego perceives as an external

so ego is excluded from calculations. The other rea-uding ego is that it simplies calculations dealing witho would otherwise extremely skew any transitivity

were measured using an automated communityalgorithm. We loosely consider these clusters to rep-nct social groups or contexts. Specically, we used theethod, a highly efcient technique for decomposing ato mutually exclusive clusters that seek to maximize

(Blondel et al., 2008). We acknowledge that in a net-s not completely disconnected some individuals linkexts, and ego may consider this individual as belong-

contexts. We opt for the Louvain method rather thanethods for its efciency and its capacity to maximize

relative to other methods, such as Girvan-Newman

(Girvan andet al., 2004)are identielocal clusteit reaches thcluster hasmost withinassigned a gponent. Wethe degree (2010) suggbe unrealisany node tbetween anFacebook pin commonmodularitydistinctiventhod). Node size corresponds to a log scale of betweenness. Network NameGenWeb and rendered using GUESS 1.04 (Adar, 2006). Different

Newman, 2002), Greedy community detection (Clauset and spectral partitioning (Newman, 2006). The groupsd using a greedy optimization technique that seeks outrs of high density. This iterative process continues untile highest modularity values, thereby ensuring that each

the fewest number of links between each group, the each group, relative to a null model and every node isroup. We only calculate clusters within the giant com-

use the standard conguration null model that xesdistribution while randomizing edges. While Fortunatoests that in some cases the conguration model maytic as a benchmark since it assumes the potential foro connect to any other node. We believe connectionsy two nodes is actually a legitimate assumption withinersonal networks since all alters have at least one friendego. Along with the number of groups, we include the

(or quality) score as a supplementary measure of theess of the groups.

B. Brooks et al. / Social Networks 38 (2014) 115 9

5. Models and analyses

The correlation matrix of our variables of interest indicates amultitude of strong relationships. Results in Table 2 suggest thatmany of the measures discussed above can at least partially explainvariations in bridging and bonding social capital, given the manycorrelations over 0.2. We rst review some of the most notablecorrelations and then proceed to discuss nested OLS regression andmultiple m

5.1. Correla

Bivariatethe topologated with hia higher mo(where p < 0ables. Mostwith more informationThis suggestion sourceproviding aexplore this

Averagecapital as hboth (weakTransitivityboth forms that more cgreater persocial capitbelow and t

Only onticollinearitIn our subseVIF and Tolshowed VIFfrom conceinterestingincreases. Italters are lreturn to thtransitivity

5.2. Multiva

Althoughkey structudard OLS recapital we models whrst as wellmodels [Monumber of and transitivious workexpanded mengagemenFriends, FRM

When coaverage de

3 Despite thof either does

non-signicant. Using only node or average degree (not shown)instead of both does not change this outcome. Thus, we have evi-dence to reject Hypothesis 1. In line with the earlier correlations,transitivity is negatively related to both perceived bridging andbonding casocial capitpersists wit

odeignignic

will ationmenodeica

Hypong a rk mndinters eivityre mome

sultinode

dels y thacapitce. Tral vcanny, we

ediat

xplond s

Preaoder simhiched b

modof theg to dthe rme oyed m

aggr. For tf en

ed soed oncallynding so

e visentiaial toltiple

intependariabdepef theport ediation models.

tions

correlations reinforce our earlier characterization ofy of Facebook ego networks: larger networks are associ-gher average degree, less transitivity, more clusters anddularity score. The network metrics also signicantly.05) relate to the demographic and engagement vari-

notably, networks with more nodes strongly correlateactual friends as well as more visits to Facebook. Also,-seeking is strongly related to the number of clusters.ts individuals who consider Facebook as an informa-

tend to have more social contexts from which to draw, preliminary validation of hypothesis 4b above. We will

relationship further in a multivariate model below. degree is (weakly) positively related to bonding socialypothesized. More clusters and higher modularity arely) related to bridging social capital as hypothesized.

actually shows a signicant negative relationship toof social capital. This goes against the basic hypothesisonnectivity in the network would be associated withception of inclusiveness as suggested by a traditionalal approach. This nding will be explored in the modelshe subsequent discussion.e correlation is high enough to suggest potential mul-y issues: average degree and nodes (r = 0.82, p < 0.001).quent regressions we check for multicollinearity usingerance parameters within SPSS version 19. No models

scores above 10 or Tolerance scores below 0.20. Apartrns with multicollinearity, this correlation in itself isas nodes come into the network, the average degree

would suggest that when people add new alters, theseikely to share friends in common with ego. We willis nding in the discussion as it helps to explain howand bonding social capital can be negatively related.

riate analysis

we wish to test for the independent effects of ourral metrics, we rst consider these metrics in a stan-gression framework. For bonding and bridging socialinclude two models each. These are standard nestedere the second model includes all variables from the

as additional social engagement variables. The smallerdels 1 and 3 in Table 3] include ve structural metrics:nodes, average degree, modularity, number of clustersvity. These models also include controls based on pre-: age, education (ordinal), gender and self-esteem. Theodels [Models 2 and 4 in Table 3] append four Facebookt metrics: visits per day to Facebook, actual FacebookB, and ISB.3

ntrolling for other structural variables, the effects ofgree, modularity and number of nodes is rendered

e high correlation between nodes and average degree, the exclusionlittle to change the model t or the signicance of the variables.

In Monly snon-si4). Wethe relengage

In mno signlation.in havinetwoand boof clustransitsures awith wThe re0.17. M

Monamelsocial varianstructueffect quentl

5.3. M

To ement a(MML;ation mmannemine wexplainoverallcance seekinior on and soemploverbalpowersures oreport

Basspeciand bobridginincludas potpotent

Muable ofthe dedent vto the effect oalso repital. However, it is only signicant where bondingal is used as the dependent variable. This signicanceh the inclusion of engagement variables (model 2).l 3 (bridging social capital), number of clusters is thecant variable, although this relationship is renderedant by the inclusion of engagement variables (Modelexplore this further in a mediation model to help clarifyship between number of clusters (i.e., social contexts),t and bridging social capital.l 1, Hypothesis 1 is not supported; average degree hasnt effect, which was expected based on the weak corre-thesis 2 was also not supported as transitivity persistednegative effect on bonding social capital. Thus, while alleasures were signicantly correlated with the bridgingg scores in the bivariate correlations, only the numbermerges as a signicant predictor for bridging and only

was a signicant predictor for bonding when the mea-odeled jointly. Gender is also a signicant predictor,n reporting greater bridging social capital than men.g Model 1 has a moderate t, with an adjusted R2 of

l 2 has an adjusted R2 of 0.32.2 and 4 indicate support for Hypotheses 4a and 4b,t FRMB and ISB positively predict bridging and bondingal while leading to substantial increases in explainedhese variables also alter the relationship between theariables and social capital. The precise nature of thisot be determined in the current OLS models. Conse-

turn to mediation models.

ion modeling of structural effects

re the relationship between network structure, engage-ocial capital, we employ multiple mediation modelscher and Hayes, 2008), as an extension to classical medi-ls (Baron and Kenny, 1986). These models operate in ailar to structural equation models in that they deter-

part of a variables effect on a dependent variable isy an intermediary variable. That said, they provide anel R2 and employ bootstrapping to assess the signi-

effect of the mediators. Such models are useful whenisentangle the effect of individual perception or behav-elationship between some objectively measured valueutcome variable. For example, Vanbrabant et al. (2012)ediation models in personal networks to assess how

ession was mediated by status and subjective sense ofhis study, we are interested in the extent to which mea-gagement mediate the effects of network structure oncial capital.

the regression models described in Table 3, we focus on disentangling the relationship between transitivityg social capital (Model 5) and number of clusters andcial capital (Model 6). In addition to FRMB and ISB, weits per day to Facebook and number of actual friendsl mediators since these latter variables indicate egos

activate social capital. mediation models articulate three paths from the vari-rest, rather than one single path from all variables toent variable (Fig. 2): the a path, showing the indepen-le to the mediators; the b path, showing the mediatorsndent variable; and the c path, showing the residual

independent variable accounting for the mediators. Wethe c path, which is the total effect of the independent

10 B. Brooks et al. / Social Networks 38 (2014) 115

Table 3OLS regression predicting to the Williams scale of bonding and bridging capital.

DV: Bonding social capital DV: Bridging social capital

Model 1: Networkvariables

Model 2:Network + engagementvariables

Model 3: Network Model 4:

DemographicGender (WAge **

EducationSelf-estee *

Network varNodes Average dModularitNumber oTransitivit **

Engagement Actual frieVisits per ISB ***

FRMB ***

Constant (un ***

Adjusted R2

Standardized cN = 235.

* p < 0.05.** p < 0.01.

*** p < 0.001.

Fig. 2. Exampand mediating

variable onthere is no mc path is sig

Model 5capital andthat transittransitivityFRMB and (p < 0.001), not signicsignicancebelieve tha5b for bondsignicant stherefore nrole in the

4 We emplosomen) 0.06 0.05

0.17 * 0.17 0.02 0.02 m 0.16 ** 0.12iables

0.02 0.03egree 0.01 0.10y 0.07 0.03 f clusters 0.10 0.03 y 0.24 ** 0.22variablesnds on Facebook 0.11 day on Facebook 0.08

0.220.29

standardized) 3.50 *** 2.55F(225) = 6.46*** F(221) = 9.50***

0.17 0.32

oefcients are reported for all numbers unless otherwise noted.le mediation model schema denoting specic paths for independent variables. Control variables are not shown.

the dependent variable. If this c path is non-signicant,eaningful relationship to test in the rst place.4 If this

nicant, we can consider the mediating paths. examines the relationship between bonding social

transitivity (Table 4). The signicant c path suggestsivity has an effect. The signicant c path suggests that

has a direct effect independent of the mediators. TheISB b paths to bonding social capital are signicantbut the a paths from transitivity to FRMB and ISB areant at the p < 0.05 level. However, given that their

values are close to the critical value (p 0.06), wet we can neither reject nor accept Hypotheses 5a anding social capital. Most importantly, the fact that c isuggests that transitivity has an independent effect andetwork structure can play a signicant independent

perceived experience of bonding social capital. The

y the INDIRECT algorithm (version 4.1) written by Hayes in SPSS 18.

coefcient further in t

Model 6ital and numsignicant cHowever, this indirect. ters to FRMto bridgingmediated, pconsideredated. Consesocial capit

In both of actual frthere is no nicantly p

6. Discussi

Results social capitengagemenpredictor inrecent nd2011, in prthat Faceboof more socmaintaininalso reinforlater scale dbook as a sFacebook aa site for msive bondinsocial capitsible to posvariables Network + engagementvariables

0.19 ** 0.050.08 0.100.00 0.060.08 0.04

0.11 0.120.05 0.090.07 0.030.20 ** 0.020.02 0.01

0.010.040.33 ***

0.38 ***

2.21 *** 0.79F(225) = 5.07*** F(221) = 14.040.14 0.42

for transitivity remains negative. We consider thishe discussion section.

examines the relationship between bridging social cap-ber of clusters found using community detection. The

path suggests that the number of clusters has an effect.e non-signicant c path indicates that this relationship

Since there are signicant a paths from number of clus-B and ISB and signicant b paths from FRMB and ISB

social capital we consider this to be an indirect, fullyath. If the a paths were not signicant we would have

this relationship to be spurious rather than fully medi-quently, we accept Hypotheses 5a and 5b for bridgingal.models 5 and 6 there are signicant a paths to number

iends and daily number of visits to Facebook, althoughevidence to support the notion that these variables sig-redict perceived social capital in these models.

on

from the analyses demonstrate that the perception ofal is related to both social structure and patterns oft in complex ways. The fact that FRMB was the strongest

most models, both bridging and bonding, reinforcesings in this area (Burke et al., 2010, 2011; Ellison et al.,ess; Vitak, 2012). In particular, these results indicateok use in itself is not a guaranteed path to perceptionsial capital, but that specic attitudes and strategies forg relationships (FRMB) play a large part as well. This isced by the consistently signicant results for ISB. Thisescribes the extent to which individuals perceive Face-ite for information seeking needs. Those who considers a site for information seeking also tend to report it asore general social capital, both the emotional and inclu-g capital and the more instrumental and broad bridgingal. It is important to note that in this study it is not pos-it a causal direction. As such, we cannot tell if attitudes

B. Brooks et al. / Social Networks 38 (2014) 115 11

Table 4Mediation models for bonding and bridging social capital.

Model 5 Model 6

Bonding (IV: Transitivity) Bridging (IV: Clusters)

p

IV to mediatActual frie 0.0Visits per 0.0Info-seeki 0.0Facebook 0.0

Direct effectActual frie 0.0Visits per 0.2Info-seeki 0.0Facebook 0.0

Total effect oTransitivit 0.0Direct effeTransitivit 0.0

Partial effectGender (w 0.4Age 0.0Education 0.7Self-esteem 0.0Nodes 0.0

Bonding: N = 2 , F(224

lead to behathe accrual

While wated effect of clusters. on Facebooet al. (2011nication thain the case individual tthem to a bonding socmost likely ing close refrom netwobonding socing with fewtheir mass c

The relathas face valtion seekingwho considfound in thsow. Howevmay potenttured in siga clear pattemere handfa network wversity, a fustill be a smsuch as onedistinct clusare closed, closed trianoverlappingone went tothen got a from ones from high sand co-wor

ouldpeopknow

ironal cloon. In

in ea opeivitymorelosurwo petwen tralosenshipnalyn deon ark coCoefcient SE

ors (a paths)nds on Facebook 177.22 52.91 day on Facebook 1.74 0.57 ng on Facebook 0.86 0.45engagement (FRMB) 0.83 0.45 s of mediators on DV (b paths)nds on Facebook 0.00 0.00 day on Facebook 0.05 0.05 ng on Facebook 0.19 0.06 engagement (FRMB) 0.27 0.06f IV on DV (c path)y/number of clusters 1.87 0.41 ct of IV on DV (c path)y/number of clusters 1.42 0.38

of control variables on DVomen) 0.07 0.09

0.01 0.00 0.02 0.04

0.18 0.08 0.00 0.00

35, Adjusted R2 = (0.325), F(224) = 12.29***; bridging: N = 235, Adjusted R2 = (0.424)

viors that reinforce the perceptions of social capital orof social capital changes attitudes and behaviors.e fully expected structural measures to have a medi-on FRMB and ISB, we found no mediation for numberThis may be due to the ways in which people interactk, but it is likely the result of a nding within Burke), who nd that it is direct person-to-person commu-t leads to an increase in bridging social capital. Thus,of bridging social capital, it is more important for ano know of an available job or someone who can linkresource. Further, we agree with Burke et al. in thatial capital generation and maintenance on Facebook isdue to individuals utilizing other methods for maintain-lationships. Our ndings suggest that those individualsrks with low transitivity will be able to experience moreial capital on Facebook because they are communicat-er contexts and do not have to limit the exposure of

ommunication to the least close relationship.ionship between attitudes, behaviors and social capitalidity. People who consider Facebook a site for informa-

there wa few might

Theing loccohesipeopleare fewtransitmean local cmore tping bbetweemore crelatiowork abetweebased netwo and a site for small symbolic practices tend to be peopleer Facebook a place for the positive social resourcese social capital scales. That is, people reap what theyer, a focus on individual attitudes and behaviors aloneially be reductionist. Facebook ego networks are struc-nicantly different ways. Some of these networks showrn of dense pockets of ties from separate contexts with aul that link the network together. For example, considerith separate friendship groups from high school, uni-

ll time job and a neighborhood association. There mayall number of people who link these groups together,

s signicant other or best friend, but the groups remainters. Transitivity would be high because most trianglesbut there are few links between the dense clusters ofgles. Other networks are very diffuse with many ties

between multiple clusters. This might be the case if a local college alongside many people from high schooljob in the same town while living a few blocks downparents. In such a case, it is plausible that many peoplechool who know ones college friends, family memberskers. While the entire network would be very cohesive,

trast, Facebinclusion of

Networkwithin the groups knocially largeThis suggesother. Whebook user cown demanwhere sepasides of theas many nothe graph social capiting ties witbonding soeven if locaet al. (2009because thefore fewer dCoefcient SE p

0 5.06 1.91 0.010 0.07 0.02 0.006 0.05 0.02 0.007 0.07 0.02 0.00

7 0.00 0.00 0.674 0.02 0.04 0.530 0.25 0.05 0.000 0.32 0.05 0.00

0 0.04 0.01 0.00

0 0.01 0.01 0.58

6 0.07 0.08 0.321 0.01 0.00 0.082 0.05 0.04 0.203 0.05 0.07 0.478 0.00 0.00 0.53

) = 18.23***.

be fewer closed triangles. The co-workers might knowle from high school, but not all. The family members

a few friends from college, but not all.y of these two kinds of networks is that by measur-sure, transitivity could be evidence of a lack of global

the rst case, transitivity would be high because mostch of the separate groups knows each other while theren paths between the separate groups. In the second case,

might be lower because the overlapping social circles open two paths between different groups. So whilee is lower, this is due to the fact that there are simplyaths to account for since there are so many ties overlap-en the groups. We initially hypothesized a relationshipnsitivity and bonding social capital since we consideredd paths to be indicative of dense pockets of reciprocals. It is an assumption that is long held in social net-

sis (Feld, 1981; Louch, 2000). However, this relationshipnse reciprocal relationships and social capital may ben untenable assumption in Facebook ego networksamprised primarily of a single cohesive group. By con-

ook networks are almost always characterized by the

multiple social groups that only ever partially overlap.s with lower transitivity mean more open two pathsnetwork. This is evidence that people from separatew each other. Networks with higher transitivity (espe-

networks) tend to have groups that are very distinct.ts that people from separate groups do not know eachn people do not know each other, it suggests that a Face-an be torn between multiple social worlds, each with itsds and expectations. It may be like an awkward partyrate groups unknown to each other stay on opposite

room. In such a case, it is plausible that despite havingdes and edges as a graph with more open two pathsfeels different concerning the perception of bondingal. Thus, somewhat surprisingly we suggest that bridg-hin a Facebook ego network are associated with greatercial capital by making the overall graph more cohesivel network structures are less dense. Similarly to Binder), we suspect individuals will feel less context collapseir network is still relatively well connected and there-istinct subgroups, whereas those individuals with high

12 B. Brooks et al. / Social Networks 38 (2014) 115

transitivity will experience lower amounts of trust and engagementbecause of the lack of privacy (Houghton and Joinson, 2010).

Interpreting these results requires us to make assumptionsabout the sort of relationships that comprise a group (or a clusterfound usingclusters in based on liftion of grouof clusters greater seninformationemerge necgroups repowho have anot considetional mainbridging so

There istionships rawell as a recertain netwclude from relationshipfocus on bastrate that gin bridging However, ndyadic leveoverlap beteffect indep

We haveital can be gnetworks casuch as namabout transsmall netwresearch asindividualsthe wider nbe discoverwhich sociaeach other rGranted, thnovel. For emean path ograms oft-cand Time dracial homodata set haslevel reinfoby major, c2011). Howcan have thsible to stunodes (indiconnectionthat access ethically chanalysis of practical alt

7. Limitati

Facebooindividuals

of the personal network, nor a network that necessarily matchesegos biases, since some ties may exist between alters without egosknowledge. We assert that bonding is related to a sense of the net-work as being globally cohesive or fragmented. It is possible that

y noly. Nrks (suchin hpubliies th

Face to pldetaed incomms.cernunity

theion m

the ds exuvainre cof limunt ere ansidolutny tthe eusiby pring m

but

ond g ad

There this . Nevd sos ea

igheith a ergrr picFurth.S. s

ally, olvitheretwo. Ourntannshipeld e

wled

ackle. Thation community detection). In general, we maintain thatFacebook ego networks represent broad assemblagese course stages and shared activities. This characteriza-ps also helps to explain the full mediation of number

on bridging social capital. More groups should mean ase of connectivity to the wider world and the diverse

resources. Yet this sense of connectedness does notessarily. Those individuals who actively attend to thesert that they draw novel information from them. Those

Facebook network but remain indifferent to it (i.e. dor it a site of information seeking or site for small rela-tenance practices) do not report as high a perception ofcial capital.

a present turn toward considering the quality of rela-ther than mere structure (Aral and Alstyne, 2011) asnewed focus on individual traits that can give rise toork structures (Burt, 2010, 2012). What we can con-

this analysis is that this focus on individual and dyadics makes sense in some arenas. Like Aral and Alsytnesndwidth over diversity, our mediation models demon-reater engagement (i.e., bandwidth) plays a larger role

social capital than multiple social groups (i.e., diversity).ot all structural factors can be reduced to individual orl covariates. Lower transitivity, as evidence of greaterween groups and thus greater global cohesion, has anendent of how individuals approach their networks.

also demonstrated that new insights about social cap-leaned from Facebook ego networks compared to egoptured using traditional respondent driven techniquese generators or enumeration methods. Our ndings

itivity would not have been discovered in the veryorks traditionally employed in core discussion network

it is not an insight about the most closely connected to ego, but about how these individuals are situated inetworks of personal afliation. Similarly, it could noted by merely articulating which individuals belong tol groups since it is about how these groups connect toather than how many individuals exist in which group.e use of Facebook as a social network is not in itselfxample, very large scale analysis has indicated how thef Facebook is approximately 3.74, much lower than Mil-ited six degrees (Backstrm et al., 2012). The Taste, Tiesataset has indicated how students reinforce patterns ofphily (Wimmer and Lewis, 2010). The Facebook 100

revealed how community structure at the university-rces existing networking patterns such as homophilyohort or dorm depending on the school (Traud et al.,ever, this work highlights how Facebook ego networksemselves substantial explanatory power. Thus, it is pos-dy Facebook as a network of networks, consideringviduals), their connections on the site, and how theses and the groups they represent are connected. Givento the total Facebook graph is both extremely limited,allenging and technically formidable, we believe thatsampled ego-centered networks can be a germane andernative in some circumstances.

ons

k personal networks offer a remarkable view into an personal network, but it is neither a complete view

ego maunlikenetwograms ascertathese about tby ourabilityalters includis not nding

Concommrst isdetectond ismethothe Lobe moment oin a cokers wmay cothe resing maGiven it is plaods maprovidbetter,ing.

Beyworkinwest. withinlationssites anconrmwith htion wby unda fullelarge. ing a U2010).

Finhow evor wheboth ncapitalto diserelatioStein

Ackno

WeMelvilFoundt perceive this cohesion. We consider this plausible butevertheless, future research should compare Facebooklike those drawn from NameGenWeb, or similar pro-

as NetVizz) to personal network name generators toow sociocognitive networks on Facebook differ fromcly articulated networks. Conversely, ego may knowat exist in the Facebook network, but were not captured

book app. This is because Facebook offers individuals theace friends on limited prole, meaning ego cannot seeils, but is still alters friend. In this case, alter will not be

the Facebook network as downloaded. We believe thison practice, and thus does not signicantly bias our

ing our methods, we acknowledge several criticisms of detection methods. Two in particular stand out: The

resolution limits of modularity-oriented communityethods (Fortunato and Barthlemy, 2007). The sec-

recent demonstration that most community detectionperience degeneracy near optimal solutions (including

method, cf., Good et al., 2010). These concerns wouldrrosive if this analysis hinged on the correct assign-inal nodes to clusters. However, we were interested

of clusters, without concern for whether certain bro-ssigned to one group or the other (when in reality egoer such brokers as members of both groups). Moreover,ion limit may even work in our favor by not present-iny (arguably trivial) clusters in our larger networks.xplanatory power of the Louvain method in this paperle that other slower but potentially more precise meth-ovide a better t as well as more explanatory power byore accurate results. No method we tested performednew methods for partitioning are continually emerg-

this, we acknowledge the limitations of our sample asults and university employees from the American Mid-

may be cultural norms within this population andgeographical area that do not generalize to larger popu-ertheless, much of the current work on social networkcial capital employs undergraduate samples. Our studyrlier ndings that greater Facebook use is associated

r levels of social capital and expands this to a popula-wider range of ages and life histories than representedaduate students. We believe this enables us to presentture of how Facebook operates in the population ater, the non-student sample is valuable, but only hav-ample poses limitations on the work (Henrich et al.,

we encourage longitudinal work that can disentangleng networks would lead to differences in social capital,

persistent demographic and personality factors driverk structure and the sentiments associated with social

work points toward future research that would attemptgle this research, but does not implicitly inform the

between the works of Burke et al. (2010, 2011) andt al. (2008).

gements

nowledge the advice of Rebecca Gray and Joshuais work was supported in part by the National Science

(HCC 0916019).

B. Brooks et al. / Social Networks 38 (2014) 115 13

Appendix A. Scales

The scales presented in Tables A1 and A2 below are a modi-cation of Williams nal scales, as pruned from a larger questionbank. As in the original paper and subsequent applications, thescales have a tolerable Cronbachs alpha above 0.8. Readers famil-iar with the development of social capital within sociological andsocial network literatures may nd the inclusion of job informationin the bonding social capital scale as somewhat surprising, since jobinformation is presumed to be accessed through weak ties. It inclu-sion in bonding social capital is based on a principal componentsanalysis in Williams (2006). The means refer to a Likert scale fromstrongly disagree (1) to strongly agree (5). Items were standardizedbefore inclusion in the scale, but not weighted by factor loadings.

Tables A3 and A4 are the engagement scales discussed.The nal Table A5 below represents the self-esteem scale dis-

cussed.

Table A1Facebook specic bridging social capital scale.

Items Mean SD

Interacting with people in my Facebook networkmakes me interested in things that happenoutside of my town.

3.75 0.82

Interacting with people in my Facebook networkmakes me want to try new things.

3.59 0.83

Interacting with people in my Facebook networkmakes me interested in what people unlike meare thinking.

3.58 0.79

Talking with people in my Facebook networkmakes me curious about other places in theworld.

3.71 0.86

Interacting with people in my Facebook networkmakes me feel like part of a larger community.

3.68 0.98

Interacting with people in my Facebook networkmakes me feel connected to the bigger picture.

3.54 0.89

Interacting with people in my Facebook networkreminds me that everyone in the world isconnected.

3.71 0.91

I am willing to spend time to support generalFacebook community activities.

3.02 0.92

Interacting with people in my Facebook networkgives me new people to talk to.

3.17 1.01

Through my Facebook network, I come in contactwith new people all the time.

2.96 1.03

Adapted from ).Full scale: M = 3.47, SD = 0.66 ( = 0.90).

Table A2Facebook specic bonding social capital scale.

Items Mean SD

There are several people in my Facebook network Itrust to help solve my problems.

3.30 1.15

There is somturn to fordecisions.

When I feel Facebook

If I needed asomeone i

The people Iwould put

The people Iwould be

The people Iwould sha

The people Iwould hel

Adapted from Full scale: M =

Table A3Facebook Relationship Maintenance Behavior scale (Ellison et al., in press).

Items

When I see anews on F

When I see anews on F

When I see sFacebook,

When a Facepost some

When I see sthat I know

Full scale: M =

Table A4Information-s

Items

I use Facebowant to bu

I use FaceboI use FaceboI use Facebo

Full scale: M =

Table A5Self-esteem sc

Items Mean SD

I feel that Im a person of worth, at least on anequal plane with others.

4.50 0.60

I feel that I have a number of good qualities. 4.50 0.63All in all, I am inclined to feel that I am a failure

(reversed).4.47 0.67

I am able to do things as well as most other people. 4.22 0.65I feel I do not have much to be proud of (reversed). 4.44 0.72I take a positive attitude toward myself. 4.08 0.75On the whole, I am satised with myself. 4.04 0.76

Adapted from Rosenberg (1989).Full scale: M = 4.32, SD = 0.50 ( = 0.86).

References

Adar, E., 2006. GUESS: a language and interface for graph exploration. In: Proceed-ings of the ACM Conference on Human Factors in Computing Systems (CHI 06),http://dx.doi.org/10.1145/1124772.1124889.

Aral, S., Alstyne, M.V., 2011. The diversity-bandwidth trade-off. Am. J. Soc. 117 (1),90171, http://dx.doi.org/10.2139/ssrn.958158.

Backstrm, L., Boldi, P., Rosa, M., Ugander, J., Vigna, S., 2012. Four degrees of sep-aration. In: Proceedings of the 3rd Annual ACM Web Science Conference on WebSci12, pp. 3342, http://dx.doi.org/10.1145/2380718.2380723.

Baron, R.M., Kenny, D.A., 1986. The moderatormediator variable dis-tinction in social psychological research: Conceptual, strategic, andstatistical considerations. J. Pers. Social Psychol. 51, 11731182,

://dx.doi.org/10.1037/0022-3514.51.6.1173., H.R., Killworth, P., Kronenfeld, D., Sailer, L., 1984. The problem of informantracy: the validity of retrospective data. Annu. Rev. Anthropol. 13, 495517,://dx.doi.org/10.1146/annurev.an.13.100184.002431.H., Howes, A., Sutcliffe, A.,2009. The problem of conicting socialres: effects of network structure on experienced tension in socialork sites. In: Proceedings of ACM CHI 2009. ACM, New York, NY,

://dx.doi.org/10.1145/1518701.1518849. V.D., Guillaume, J.L., Lambiotte, R., Lefebvre, E., 2008. Fast unfold-of communities in large networks. J. Stat. Mech. Theor. Exp. P10008,://dx.doi.org/10.1088/1742-5468/2008/10/P10008.u, P., 1986. The forms of capital. In: Richardson, J.G. (Ed.), Handbook of The-and Research for the Sociology of Education. Greenwood, New York, pp.258, http://dx.doi.org/10.1002/9780470755679.ch15.

, U., Eiglsperger, M., Herman, I., Himsolt, M., Marshall, M., 2002.hML progress report structural layer proposal. Graph Drawing,112.B., Welser, H.T., Hogan, B., Titsworth, S., 2011. Socioeconomics updates: family SES and emergent social capital in collegeeone in my Facebook network I can advice about making very important

3.49 1.07

lonely, there are several people in mynetwork I can talk to.

3.52 0.95

n emergency loan of $500, I known my Facebook network I can turn to

3.35 1.16

interact with in my Facebook network their reputation on the line for me.

3.38 0.94

interact with in my Facebook networkgood job references for me.

3.41 0.97

interact with in my Facebook networkre their last dollar with me.

3.08 0.94

interact with in my Facebook networkp me ght an injustice.

3.70 0.84

). 3.40, SD = 0.73 ( = 0.88).

httpBernard

accuhttp

Binder, sphenetwhttp

Blondel,ing http

Bourdieory 241

BrandesGrap109

Brooks, statuMean SD

friend or acquaintance sharing goodacebook, I try to respond.

3.87 0.90

friend or acquaintance sharing badacebook, I try to respond.

3.65 0.95

omeone asking for advice on I try to respond.

3.42 0.95

book friend has a birthday, I try tothing on their wall.

3.87 1.06

omeone asking a question on Facebook the answer to, I try to respond.

3.82 0.85

3.72, SD = 0.80 ( = 0.90).

eeking behavior scale (Lampe et al., 2012).

Mean SD

ok to get advice about something Iy.

2.42 1.02

ok to get business referrals. 2.16 0.98ok to get answers to specic questions. 2.73 1.14ok to ask questions about health issues. 2.04 0.93

2.34, SD = 0.83 ( = 0.83).

ale.

14 B. Brooks et al. / Social Networks 38 (2014) 115

student Facebook networks. Info., Commun., & Soc. 14 (4), 529549,http://dx.doi.org/10.1080/1369118X.2011.562221.

Burke, M., Marlow, C., Lento, T.,2010. Social network activity and social well-being. In: Proceedings of the 28th International Conference on HumanFactors in Computing Systems CHI10. ACM Press, New York, USA,http://dx.d

Burke, M., Krating usesHuman fahttp://dx.d

Burt, R.S., 201Oxford Un

Burt, R., 201tirole evihttp://dx.d

Clauset, A., nity struhttp://dx.d

Csrdi, G., Neresearch. I

Davis, J.A., 196181187, h

Ellison, N., Graexploring InternatioAdvancem

Ellison, N., Stecapital andCommun.

Ellison, N.B., Simplicatiohttp://dx.d

Ellison, N., Vitanetwork ssocial capi

Facebook, 20press/info

Facebook, 201accessed:

Feld, S.L., 1981015103

Fine, B., 2010.London, U

Fischer, C., 198Fischer, C., 20

http://dx.dFortunato, S., 2

http://dx.dFortunato, S., B

Natl. AcadFriggeri, A., Ch

IEEE Inter911, 2011

Gilbert, E., KIn: CHI09on Humahttp://dx.d

Girvan, M., biological http://dx.d

Gittell, R., Vidaopment St

Golbeck, J., CHI11: PHuman Fhttp://dx.d

Good, B.H., dlarity mahttp://dx.d

Granovetter, M(6), l.

Hampton, K.Nand our http://doc

Henrich, J., the worl10.1017/s0

Hogan, B., 20D., Smith,with Nohttp://dx.d

Holland, P.W.groups. S10464964

Houghton, D.J., Joinson, A.N., 2010. Privacy, social network sites,and social relations. J. Technol. Hum. Serv. 28 (12), 7494,http://dx.doi.org/10.1080/15228831003770775.

Jones, J.J., Settle, J.E., Bond, R.M., Fariss, C.J., Marlow, C., Fowler, J.H., 2013. Infer-ring tie strength from online directed behaviour. PLoS ONE 8 (1), e52168,

://dx.dn, C., 2://dx.d., Bahcted i://dx.dC., s Val S://dx.d001. (Eds.)29.H.,

ophil://dx.d, A.Es, con://dx.d, C., 20, C., parin9.on, Jion n://dx.dM.R., g sear912Hillygabilityrnet inn, Mg th://dx.dn, M.E

I donopporACM 23, 20rissi, icatio), Me122

r, K.J.,and cohods 4, R., ://dx.daase, Ah-tec

D., perso

Con://dx.d

P., 2 J. (E476rg, Mersity. (201

Washrts/2d, C., ne soc445, ., Keewome://dx.d.L., Ke

to ch543,

ela, Sork

t, an://dx.dr Gaatal qu://dx.doi.org/10.1145/1753326.1753613.aut, R., Marlow, C., 2011. Social capital on Facebook: differenti-

and users. In: Proceedings of the 2011 annual conference onctors in computing systems, Vancouver, BC,