sociol essay

G

Y a

b

c

a

K S D P N L A T

1

s f i t w b l m p b 1 t a w t M e

C W i W t S m

(

0 d

Social Networks 34 (2012) 73– 81

Contents lists available at ScienceDirect

Social Networks

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / s o c n e t

eography of Twitter networks�

uri Takhteyev a,∗, Anatoliy Gruzd b, Barry Wellman c

University of Toronto, Faculty of Information, 140 St. George Street, Toronto, ON M5S 3G6, Canada Dalhousie University, Canada University of Toronto, Canada

r t i c l e i n f o

eywords: ocial networks

a b s t r a c t

The paper examines the influence of geographic distance, national boundaries, language, and frequency

istance roximity ation-states anguage ir travel witter

of air travel on the formation of social ties on Twitter, a popular micro-blogging website. Based on a large sample of publicly available Twitter data, our study shows that a substantial share of ties lies within the same metropolitan region, and that between regional clusters, distance, national borders and language differences all predict Twitter ties. We find that the frequency of airline flights between the two parties is the best predictor of Twitter ties. This highlights the importance of looking at pre-existing ties between places and people.

. Introduction

Social contact benefits from physical proximity. This fact of ocial life is so basic, that for a long time proximity was often taken or granted: social interaction was understood to mean face-to-face nteraction, for which distance acted as a powerful barrier. The fact hat being near each other facilitated the formation of social ties as for the most part not so much a finding of social research as its

asic premise. Social network analysts were among the first to chal- enge this assumption, showing that the social network approach

ade it possible to follow social ties as they crossed space, map- ing the more distributed communities that were replacing those ased on neighborhoods (Webber, 1963; Wellman, 1979; Fischer, 982). A few decades later, the Internet brought new opportuni- ies for maintaining social ties over distance, as well as greater wareness of such possibilities. Pundits proclaimed that distance as dead (Cairncross, 1997). However, the evidence challenged

his assertion, showing that proximity still made a difference. ost such studies have looked at email, which was shown to help

xtend and maintain existing strong ties (e.g., Mok et al., 2010).

� The authors thank Lilia Smale, MinKyu Kim, Andrew Hilts, Annie Shi, Courtney ardozo, and the anonymous reviewers for their help in preparation of this paper. e also offer special thanks to Joshua Mendelsohn of Duke University for provid-

ng us with the air traffic data and to Erik Zachte of Wikimedia Foundation for the ikipedia usage statistics. We received research support from the Digital Infrastruc-

ures and the Privacy in Networked Environments project of the GRAND NCE, the ocial Science and Humanities Research Council of Canada, and two undergraduate entorship programs of the University of Toronto. ∗ Corresponding author. Tel.: +1416 946 3809.

E-mail addresses: yuri.takhteyev@utoronto.ca (Y. Takhteyev), gruzd@dal.ca A. Gruzd), wellman@chass.utoronto.ca (B. Wellman).

378-8733/$ – see front matter © 2011 Elsevier B.V. All rights reserved. oi:10.1016/j.socnet.2011.05.006

© 2011 Elsevier B.V. All rights reserved.

Recent years have brought new ways of interacting over the Inter- net, some of which seem less tied to strong ties or face-to-face contact. Does proximity still affect these new forms of electronic interaction?

We focus on one such Internet-based system, Twitter, a popu- lar social networking and micro-blogging service that allows users to post and read short messages, limited to 140 characters. Such messages – called “tweets” – are usually public, visible to anyone on the Internet. (Users can make their tweets private, but most do not. Our own sample suggests that only 10 percent of the users protect their tweets.) While tweets can be read anonymously, the preferred method is to create an account and select a set of users that you want to “follow,” so that you would see recent tweets from those accounts whenever you log on to Twitter. A user’s choice of whom to follow is public. Additionally, Twitter users usually spec- ify their geographical location in their profiles. Twitter thus offers us a publicly available, spatially embedded network dataset, a rare luxury in network analysis (Butts and Acton, 2010).

Our analysis shows that distance matters on Twitter, both at short and longer ranges: 39 percent of the ties are shorter than 100 km and ties up to about 1000 km are substantially more com- mon than we would expect if they were formed at random. This result is interesting, considering the ease with which long-distance Twitter connections can be formed. We also look at several other variables that can either impede or facilitate ties while being closely intertwined with distance. We find that national boundaries and a shared language both affect ties but do not explain away the effect of physical proximity. Frequency of airline connections, on

the other hand, predicts non-local Twitter ties better than proxim- ity, with the latter adding relatively little to a model that already includes flight frequency. Thus, the strength of prior ties between places matters more than the simple distance between them.

7 al Net

2

a w 2 m m u i U o

t n ( i c n a b s A o o o I a d t i

c t b w 2 p a h p T

3

o T

s u a t t

d l s (

i d s r t p o

4 Y. Takhteyev et al. / Soci

. Twitter: global reach and weak ties

Several aspects of Twitter make it a particularly valuable case for nalysis. First is Twitter’s popularity and international reach. When e collected our data in the summer of 2009, Twitter (founded in

006) was already attracting tens of millions of unique visitors per onth (Schonfeld, 2009) who were posting and reading millions of essages every day. Our data suggests that over half of the service’s

sers were located outside the United States at the time, which ncluded many users in Australia, Brazil, Indonesia, Japan, and the K.1 This wide distribution of users allows us to explore the effects f distance at different scales: from fairly short to nearly antipodal.

The second relevant aspect is the relative weakness of Twitter ies. Granovetter (1973) defines the strength of a tie as “a combi- ation of the amount of time, the emotional intensity, the intimacy mutual confiding), and the reciprocal services which character- ze the tie” (p. 1361). Even compared to other forms of electronic ommunication, Twitter-based interaction fails Granovetter’s defi- ition on all counts. 140-Character messages take little time to read nd even less time to ignore. The fact that messages are publicly roadcast reduces the level of intimacy and emotional intensity of uch communication. Finally, Twitter ties are often asymmetric: if

follows B, B does not have to follow A. In our sample, 60 percent f the ties are unidirectional.2 This aspect of Twitter contrasts with ther “social networking” sites, such as LinkedIn or Facebook that ften aim to capture pre-existing ties and enforce bi-directionality. nstead, Twitter may be better compared to blogs, where there is

similarly low cost (technically and socially) to establishing a uni- irectional tie by becoming a reader. However, unlike the weak ies between bloggers and their readers, which most often stay nvisible, Twitter ties can be easily observed and analyzed.3

The combination of weak, low-cost ties and global popularity reates an opportunity for people to make links that transcend dis- ances and national borders. Twitter’s ability to support such ties ecame the subject of many news articles in 2009 when the service as actively used by residents of Tehran, Iran, and most recently in

011 by people in Egypt and Tunisia, not only to coordinate local rotests against the national regime but also to inform the world bout these protests (Rainie and Wellman, 2012). We must ask, owever, whether such cases are typical or exceptional. This paper rovides a quantitative investigation of the effect of distance on witter ties.

. Ties, distance and related variables

As discussed below, distance has been shown to have an effect n social ties, including those based on electronic communication. he “weak” nature of Twitter ties may reduce the effect of distance,

1 Some of the other reports, e.g., one by Sysomos (2010) produced around the ame time, suggest that American users accounted for slightly more than half of the sers, while earlier reports by Java et al. (2007) and Krishnamurthy et al. (2008) show

lower share for North American users. Unfortunately, none of such reports describe he data collection and geocoding process in sufficient detail for us to investigate he possible sources of discrepancies.

2 In other words, the majority of Twitter ties are so weak that the followed users o not bother to reciprocate the followers’ interest in their tweets, despite the rather

ow cost of doing so. Compare this with 18–20 percent rate of unreciprocating for a ample of LiveJournal users in Gaudel and Peroni (2010). See also Huberman et al. 2009) on the weakness of Twitter ties.

3 Similar to Twitter ties, LiveJournal “friendship” ties can also be traced, and have n fact been studied, for example, by Liben-Nowell et al. (2005) who found that istance has an effect. LiveJournal “friendship” ties, however, are stronger ties, as uggested by the much lower average number of connections. Liben-Nowell et al. eport that an average user in their sample has eight “friends”. For Twitter, we found hat users on average had around 400 outgoing ties overall, with around 100 ties er user after we excluded those with over 500 ties. (See also the previous footnote n the higher rate of reciprocation in LiveJournal.)

works 34 (2012) 73– 81

but would be unlikely to eliminate it altogether. It is important to ask, however, not only whether distance matters, but also the mechanisms through which distance and ties relate.

It is clear that distance does not usually influence social ties directly. Even in its purest form, distance usually impedes the for- mation of social ties by raising the cost of travel that is required for face-to-face interaction. Consequently, the effect of distance would likely be mediated by the existing transportation infrastructure. Distance is also intertwined with other factors. We focus on two: national boundaries and language differences. Both limit interac- tion while being correlated with distance, and their effects can reduce the average length of ties. We thus focus on four variables in our investigation: the physical distance between the users, the ease of travel measured by the frequency of flights between their cities, whether the users are in the same country, and the match in language.

3.1. Physical distance

In the late 1970s, Wellman’s second study of East York, a local area in Toronto, found expanding social ties: only 22 percent of East Yorkers’ close friends and relatives were in East York and none had most of their active ties living within a mile’s walking distance (Wellman, 1979; Wellman et al., 1988). Yet, East Yorkers’ ties still depended principally on face-to-face interaction, which could now happen at the scale of a metropolitan area thanks to the increasingly widespread use of cars. The telephone was important, but its use was complementary rather than substituting for face-to-face con- tact (Wellman and Tindall, 1993; Mok and Wellman, 2007). A more recent study by Mok et al. (2010) found that today’s East Yorkers maintain more distant connections, some reaching as far as Europe or Pakistan. Technologies such as email and Skype help maintain such ties (see also Boase, 2008). Nonetheless, the number of social ties droped sharply even as the distance increases between 1 and 20 miles. Most of East Yorkers’ email use is also local. Other stud- ies have found similar results (e.g., Wellman et al., 2006). Studies of “friendship” ties on LiveJournal (e.g., Liben-Nowell et al., 2005) also found an effect of distance.

The low cost of Twitter connections can make it less sensitive to distance, giving people an opportunity to “follow” others around the world, without being constrained by the spatial extent of their face-to-face networks. Nonetheless, we can expect that like other forms of electronic communication, Twitter ties may be comple- mentary to face-to-face interaction. We can also expect people to follow friends-of-friends and people they have heard of, who are more likely to be nearer than randomly chosen alters. Users may also select Twitter accounts that distribute information about top- ics they find relevant – again likely displaying a bias towards more proximate sources. We thus expect that Twitter ties will be influ- enced by geographic distance. We do not have, however, an a priori hypothesis as to how strong that effect is likely to be and whether distance is only important at a short range (local vs. non-local) or at different geographic scales.

3.2. Air travel

Perhaps the most important reason why distance limits the formation of social ties is because it reduces the opportunities for face-to-face interactions. The strength of this effect, however, depends on the ease of travel between the places. For longer dis- tances, one of the important components of travel is the availability of airline flights (e.g., Zook and Brunn, 2006). Additionally, fre-

quency of airline connections can be interpreted as a proxy for more general connectedness. Research on global cities has shown, for example, that the cities that are most central in the network of air- line connections are also important in the network of relationships

al Netw

a I n h fl p

3

i N i w N p r i l n o s n n e t d

s a p ( f w T o t s m i b

3

m i q H a l A m t a w

t d a w

d fl

Most of the messages in the original sample (75 percent) had some location value associated with them. They were sent by users who either specified a location in their Twitter profile or, in the

Table 1 Sample size at the different steps in the data preparation process.

Step Sample size

Initial collection 481,248 tweets (7 weeks × 20 messages every 25 s)

Subsampling 3360 egos (20 per 1 h period) Geocoding at the level of country or

smaller 2852 egos (including 2167 egos with precision of <25,000 km2 )

Y. Takhteyev et al. / Soci

mong transnational accounting firms (Beaverstock et al., 1999). n other words, frequent flights between New York and London ot only facilitate travel but also indicate that New York residents ave many reasons to travel to London. The frequency of airline ights may thus be a better predictor of non-local ties than physical roximity.4

.3. National boundaries

Today’s world is organized as a system of nation-states: that s, units that tie together territory, political power, and identity. ational boundaries inhibit social ties in multiple ways. Most triv-

ally, they affect mobility since people usually can move freely ithin their states, but often need visas to move between them. ational boundaries often also define communities of interest: eople usually care about domestic events more than compa- able events abroad. This is in part because some of the most mportant decisions affecting their lives are made at the national evel, but also because access to mass media is often shaped by ational boundaries. At the same time we must avoid “method- logical nationalism” (Wimmer and Schiller, 2002) or “implicit tate-centrism” (Derudder and Witlox, 2005) and avoid taking ations for granted as a unit of analysis. Instead, the extent to which ations matter should be treated as a question to be addressed mpirically. In our case, we expect national boundaries to reduce he likelihood of Twitter ties, separately from the effect of physical istance.

When looking at the effect of national boundaries we must con- ider that not all nations are created equal. National populations ffect the likelihood of a tie between two nations, and core- eriphery structures affect differential attention between countries Smith and Timberlake, 1995). People who live in large and power- ul nations may have more opportunities for domestic connections hile also having less interest in foreign events. We can expect

witter users in such countries to have a disproportionate number f domestic ties. Residents of smaller and less powerful nations, on he other hand, may have a greater interest in what happens abroad ince their lives are quite often affected by foreign events and they ay have fewer opportunities for connecting with like-minded

ndividuals locally. We expect people living in such countries to e more active in following those who tweet from abroad.

.4. Language

Social interaction depends on the two parties’ ability to com- unicate, which nearly always requires that they are competent

n the same language or rely on a bilingual mediator. Conse- uently, language differences can structure social interactions (e.g., utchinson, 2005; Barnett and Choi, 1995). Like national bound- ries, linguistic differences are intertwined with distance. People iving in the same place typically share a language (or several). dditionally, because of the patterns of ancient settlement and ore recent colonization, people in nearby places are more likely

o speak the same or similar languages than people who are far part. We may expect, however, that shared language competency ould have an effect that is separate from the effect of distance.

Yet language stands in a more complicated relationship to dis- ance than national boundaries. While people usually speak the

ominant language of the city or country where they live, they can lso speak other languages. In particular, many people around the orld today learn English in addition to their local and national lan-

4 Note that the frequency of flights is strongly correlated with distance (in our ata we find a correlation of −0.82 between logs of distance and the number of ights), and may serve as a proxy for it, especially in the shorter range.

orks 34 (2012) 73– 81 75

guages. (See for example, Herring et al., 2007, on the role of English in LiveJournal networks.) We can thus hypothesize two separate effects: ties may be more likely between users in places with a strong linguistic connection and between pairs of users who tweet in the same language.

4. Building a sample of Twitter ties

The primary data source used in this article is a sample of dyads of geocoded Twitter accounts connected by a “follow” relation. To assemble this set of dyads, we first collected a sample of ego accounts, then sampled one alter from among the accounts fol- lowed by each ego, resulting in a set of ego-alter pairs in which the ego subscribes to (or “follows”) the Twitter messages authored by the alter. (Additional details are provided in the Appendix.)

4.1. Collecting the sample of egos

To build our sample of egos, we first collected a large number of Twitter messages by querying Twitter’s “public timeline,” a web interface provided by Twitter that returns 20 of the most recent public messages (Twitter, 2011). We used a Python script to query the public timeline every 25 s for a period of seven days in August of 2009, collecting a total of 481,248 messages (Table 1). The tweets included in the public timeline represent a small subset of the mes- sages posted that week. It may not necessarily constitute a random sample, since we do not know exactly what method Twitter uses for selecting tweets that go into the public timeline. Despite this, the public timeline has been commonly used to sample Twitter mes- sages (e.g., Java et al., 2007; Naaman et al., 2010; Golder and Yardi, 2010).

Our sample includes an equal number of messages for each 25- s period of the week, without accounting for diurnal and weekly cycles in Twitter use. Some sources suggest that the rate at which Twitter messages are produced varies throughout the day with twice as many messages produced at the peak time (1:00 pm in New York) than at the quietest time (5:00 am in New York). Our dataset have have consequently undersampled the users who tweet on a New York schedule (which would likely include many of those in North and South America) and oversampled those who tweet when New York sleeps. This distortion, however, would only affect the original sample of egos and not the length of the ties, since the ego’s connections were sampled in a separate step.

4.2. Geocoding and subsampling

Picking alters 2423 dyads Geocoding the alters at the level of

country or smaller 1953 dyads

Picking pairs where both accounts have a location with precision of <25,000 km2

1259 dyads

Spatial clustering 386 clusters Selecting the top 25 clusters for

regression analysis 25 clusters

76 Y. Takhteyev et al. / Social Networks 34 (2012) 73– 81

Table 2 Location precision in the sample of egos. Percentages are based on the 3360 egos, sub-sampled from the original sample.

Location specification Examples Share of the sample Used for

Latitude and longitude ÜT: 34.246769, -118.394672 7.5% Cluster-level and country-level analysis iPhone: 35.498447, -97.477180

A named location with an area of <25,000 km2 Los Angeles, CA 57.0% Cluster-level and country-level analysis Floss Angeles (LA) São Jorge do Ivaí-PR indonesia-bali-depok south wales in the uk P���́��

[“Some place like Roppongi or Chiba”] [“Moscow”]

Not specific, but enough to identify a country U.S.A. 20.4% Country-level analysis indonesia body = midwest, mind = elsewhere

Very broad, non-spatial, humorous, or where trouble is forks 15.10% Not used

m t v n o t p o u a a g H t “

e e n e T d o c i o l d i ( a w ( t r d b m h t a

l u e b

between sampled egos, and instead sampled an additional user – an alter – for each ego whose location was identified at the level of country or better and who followed between 1 and 500 Twitter

5 This value roughly represents the size of the largest of the metropolitan regions frequently named in our sample. (For example, the San Francisco Bay Area has

undecipherable 127.0.0.1 Hogwarts

inority of cases, used a Twitter client that automatically updated he location field in their profile. The format of those descriptions aried. Some of them provided specific addresses or even coordi- ates. Many identified cities. Some named a country, a continent, r the planet. Some referred to fictional locations or did not appear o refer to any locations at all. We found, however, that about 85 ercent of such descriptions referred to a real place, at the level f a country or smaller, while 65 percent referred to a geographic nit the size of a major metropolitan area or smaller (Table 2). We ssume that descriptions referring to real places represent users’ ctual locations, that is, either places where they tend to be in eneral or where they were at the time the message was posted. owever, we were careful to classify correctly location descriptions

hat suggested that they were wishful or outdated, for example, America i wish, England:(,” or “From Dallas but live in ATL.”

A small minority of location descriptions (6 percent) provided xact geographic coordinates, which were nearly always prefixed ither with “iPhone:” (about 39 percent of locations with coordi- ates) or “ÜT:” (57 percent), suggesting that they were submitted ither by a Twitter application running on the iPhone or by Über- witter, a popular Twitter application for BlackBerry. The rest escribed their location using a variety of ways. While US locations ften followed the city-comma-state pattern, many alternative onventions were employed. For example, locations in Brazil often ncluded dashes instead of commas (“Maceió-AL”) or put the name f the state before the name of the city (“RJ - Petrópolis”). Many ocations were identified with just the name of a city. While most escriptions used Roman letters, some used other scripts, includ-

ng those, like Japanese, that do not separate words with spaces Table 2). The language in which the location was named did not lways match the language of the location. We also encountered ide use of nicknames, such as “L.A.,” “Floss Angeles,” “Floss Town

LA), CA,” as well as idiosyncratic spelling (“LosAngeles”). We found hat automatic geocoding of such data was quite error prone; it esulted in either false positives or a failure to correctly locate place escriptions. Furthermore, this appeared to introduce a geographic ias, as locations outside the United States were more likely to be isidentified. For this reason, we decided to code the locations by

and, using a variety of reference materials, including Google Maps, o resolve place names that we were not familiar with, but avoiding pplying any of them blindly.

The need for manual processing made it impossible to geocode

ocations for all collected tweets. Instead, we took a sub-sample of sers who provided a location description, drawing 20 users from ach one-hour segment of the seven-day period. In the small num- er of cases where we drew a user who had already been included in

a sample based on an earlier hour, we drew a replacement, result- ing in a sample of 3360 unique users. We refer to these users as egos.

Table 2 summarizes the precision with which we were able to identify the location of the 3360 egos. We divide location descrip- tions into three classes. We considered a description to be specific if it identified a place with an area of up to 25,000 km2.5 Egos who provided such specific locations or actual coordinates comprised 65 percent of the sample of 3360 egos and were used to investigate the extent of spatial clustering. An additional 20 percent of the egos provided locations sufficiently narrow to determine the country. We use those cases for our analysis of Twitter use by country.

Our success rate in identifying users’ locations translates into an overall response rate of 48 percent at the level of metropolitan area and 63 percent at the level of a country. Both numbers account for the cases where no location was specified at all. It is possible that a user’s location influences the likelihood of them reporting the location (or the precision with which they report it), potentially creating a non-response bias. Such a bias, however, is unlikely to affect the main conclusions of the paper, which focus on the effect of distance and related variables on Twitter ties rather than on the geographic distribution of Twitter users per se. When we analyze the effect of distance, we do so in relation to the observed distri- bution. The only kind of non-response bias that would undermine our conclusions would be if users were substantially less likely to specify their location when they have long-range ties. We cannot rule out this possibility, but we do not see reasons to expect such an effect on a substantial scale.

4.3. Sampling the alters

Since our sample of egos included a relatively small number of users picked from among hundreds of millions of user accounts, the sampled egos were predominantly connected to users outside our sample of egos. For this reason, we did not attempt to analyze ties

22,000 km2 .) It also roughly corresponds to the upper limit on commuting dis- tance. In most cases, such descriptions included names of cities or metropolitan agglomerations. For consistency, however, we applied the same criteria to all named places (states, provinces, and countries) that fell under the 25,000 km2 threshold, for example “Wales” and “Jamaica”.

al Networks 34 (2012) 73– 81 77

a l a t t t w

4

u s a s s e a l c t F “

5

T e t d d n A l v M (

n p w a r b i c c p 9 t c o a t o

5

a a

T

Table 3 Top clusters.

Rank Clustera Share of egos (%)b

Share of egos (%) for egos in dyadsc

Share of alters (%)d

Localitye

1 “New York” 8.5 8.3 10.2 54.3 2 “Los Angeles, CA” 5.1 5.6 10.4 53.3 3 “ ” (Tokyo) 4.1 4.8 5.0 62.9 4 “London” 3.6 3.3 4.9 48.8 5 “São Paulo” 3.5 3.0 3.6 78.4 6 “San Francisco” 2.8 2.7 4.1 41.2 7 “New Jersey”f 2.5 2.8 2.1 20.0 8 “Chicago” 2.2 2.0 1.7 32.0 9 “Washington, DC” 2.1 2.8 2.6 34.3 10 “Manchester, UK” 1.9 2.0 1.1 30.8 11 “Atlanta” 1.7 2.1 2.1 46.2 12 “San Diego” 1.5 1.5 1.1 26.3 13 “Toronto, Canada” 1.3 1.1 1.5 42.9 14 “Seattle” 1.3 1.4 1.2 58.8 15 “Houston” 1.2 1.2 1.0 40.0 16 “Dallas, Texas” 1.2 1.0 1.4 61.5 17 “Rio de Janeiro” 1.2 1.0 1.1 30.8 18 “Boston, MA” 1.2 1.2 1.1 20.0 19 “Amsterdam” 1.1 1.1 0.9 50.0 20 “Jakarta, Indonesia” 1.1 0.6 0.3 42.9 21 “Austin, TX” 1.0 1.0 1.3 50.0 22 “Sydney” 0.9 1.0 0.8 38.5 23 “Orlando, Forida” 0.9 1.0 0.6 16.7 24 “Phoenix, AZ” 0.8 0.7 0.6 11.1 25 “ ” (Hyōgo)g 0.8 1.0 1.0 25.0

a Each cluster is labeled with the name most frequently used for locations assigned to the cluster.

b Out of the 2167 egos located with precision of <25,000 km2 . c Out of the 1259 egos included in dyads with both parties located with precision

of <25,000 km2 . d Out of the 1259 alters included in dyads with both parties located with precision

of <25,000 km2 . e Defined as the share of local of ties among all ties for egos in a cluster. f Centered between Philadelphia and Trenton, NJ and includes all locations iden-

tified as just “New Jersey”.

tribution of egos in our sample, only two percent of the ties would be local if the ties were formed randomly. (An average user’s cluster accounts for two percent of the total number of egos.)

Y. Takhteyev et al. / Soci

ccounts, by randomly drawing an account from among those “fol- owed” by each of those egos. We then coded the locations of the lters using the same procedure as we did for the egos, removing hose pairs where the alter could not be assigned to a country. In he end, we obtained a sample of 1953 ego-alter pairs with both he ego and the alter assigned to a country, including 1259 pairs ith “specific” locations for both parties (Table 1).

.4. Aggregating nearby locations

Since specific locations vary substantially in precision and since sers can often choose between a range of specific names for the ame place (e.g., “Palo Alto” vs. “Silicon Valley” vs. “SF Bay”), we ggregated nearby locations within each country, by assigning a et of coordinates (obtained from Google Maps) to each location maller than 25,000 km2 and then merging nearby locations within ach country by replacing their coordinates with a weighted aver- ge of the coordinates of the merged locations. This reduced our ocation descriptions to a set of 386 regional clusters, which are omparable in size to metropolitan areas. We labeled each clus- er with the most common name associated with it in our sample. or example, the cluster centered on Manhattan is referred to as New York.”

. Analysis

In this section we analyze the factors affecting the formation of witter ties. We first look at the effect of each variable identified arlier based on theoretical considerations: the actual physical dis- ance, the frequency of air travel, national boundaries, and language ifferences. In addition to presenting the descriptive statistics emonstrating the effects of each variable and investigating the ature of such effects, we correlated the effects using the Quadratic ssignment Procedure (QAP, Krackhardt, 1987; Butts, 2007). In the

ast subsection we also examined the relationship between the ariables using QAP regression (Double Dekker Semi-partialling RQAP). All statistical calculations were done using UCINet 6.277

Borgatti et al., 2002). For correlation and regression analysis we used networks with

odes representing the 25 largest regional clusters of users (see revious section). The edges of each network were then assigned eights based on an operationalization of the corresponding vari-

ble. For the dependent variable network the weight of the edges epresented the natural logarithm of the number of Twitter ties etween users in the two clusters. The weights for the edges in the

ndependent variable networks are described below, when we dis- uss each variable. We have found that the network of 386 Twitter lusters was extremely sparse, since the number of ties in the sam- le was small relative to the number of nodes. As a result, more than 9 percent of cluster pairs had zero Twitter connections between hem, leading to low correlation (between 0.05 and 0.1) with the omparison networks, with the only exception being the network f airline connections.6 For this reason, we limited our correlation nd regression analysis to the ties between just the 25 largest clus- ers, which allowed for a much denser Twitter network (an average f 0.76 ties per pair).

.1. Physical distance

The use of Twitter is concentrated in the United States, which ccounts for 49 percent of our sample of egos, 54 percent of the lters, and 6 of the 10 largest clusters (Table 3). At the same time,

6 Note that the airline network was very sparse, much like the Twitter networks. he other networks, by comparison, had non-zero values for all pairs.

g Centered near the boundary between Hyōgo and Osaka prefectures, in the Kansai region of Japan.

over half of the egos are in other countries, as are 4 of the 10 largest clusters: Tokyo, São Paulo, and two clusters in the United Kingdom. In this sense, Twitter users are distributed quite widely around the globe. In addition to the relative concentration of users in certain countries, however, we also observe a very substantial concentration of users in a relatively small number of specific local clusters. 25 clusters account for 54 and 61 percent of the egos and alters respectively. This level of concentration exceeds the general concentration of the population in major urban agglomerations.7

Being in the same cluster also has a strong effect on the forma- tion of ties: 39 percent of the ties between egos and alters fall within the same regional cluster. The large share of in-cluster ties can be partly explained by the substantial degree of clustering: when users are concentrated in a handful of places, a large share of ties would be local even if ties were formed randomly, disregardling location. The share of local (in-cluster) ties, however, is substantially higher than what we would expect just due to clustering. Considering the dis-

7 For example, the New York cluster in our sample accounts for 17 percent of US-based egos, while the New York Metropolitan Area (which exceeds the size of our “New York” cluster) accounts for only 6 percent of the United States population. For the two main clusters located outside North America and Europe, the degree of concentration is even more substantial: the São Paulo cluster accounts for 37 percent of egos located in Brazil, while Tokyo accounts for 64 percent of those located in Japan.

78 Y. Takhteyev et al. / Social Networks 34 (2012) 73– 81

F the n 5 the tw

t i d t d f f w 2 t a u t t a c

( w d ( t i a t l d s N b t i

l a d i a t p d a

w

they may continue to follow people back home.) Another inter- pretation suggests that flight connections themselves reflect the structure of the world city system, and that Twitter ties are influ- enced by this structure. Our data does not allow us to disambiguate

Table 4 QAP correlations, top 25 clusters. Distance, the number of Twitter ties, and the number of flights are logged.

Twitter Flights Language Domestic

Distance −0.448 −0.817 −0.617 −0.720

ig. 1. Histogram of physical distances between egos and alters. The graph shows 590 km, are counted towards the 5400 km bin). The total number of ties in each of

Fig. 1 shows the distribution of distances between egos and heir alters, comparing it to two simulated baselines and show- ng that distance also has an effect on non-local ties. The observed istribution is shown as the thick solid line. When analyzing the dis- ribution of tie lengths, it is again important to consider the uneven istribution of the users’ locations around the globe. If ties were ormed by picking random points on the surface of the planet (with ull disregard for uneven distribution of land mass and population), e would expect a symmetric distribution on the range from 0 to

0,000 km, with a peak at 10,000 km, represented by the smooth hin line in Fig. 1 (labeled “simulation 1”). Twitter users, however re not distributed evenly around the globe. (Nor is human pop- lation in general.) This uneven distribution substantially skews he expected distributions of distances between egos and alters owards shorter ties. Further, since the users are concentrated in

few clusters, we can expect the distribution to peak at values orresponding to distances between major clusters.

This distribution is demonstrated by the second simulation Fig. 1, medium line, labeled “simulation 2”), in which egos, located here they are in our sample, form ties among each other at ran- om. The graph shows a substantial number in the very first bin 0–200 km), followed by a decline in bins representing longer dis- ances. The count goes up, however, as we approach bins that nclude distances that span the two coasts of the United States, with

particularly sharp peak for the 3800–4000 km bin, which catches he distance between New York and Los Angeles. This peak is fol- owed by another valley, corresponding to not-quite-transatlantic istances, and then a rise as we reach Europe. The simulation hows another large peak corresponding to the distance between ew York and São Paulo, followed by one matching the distance etween New York and Tokyo. We see relatively few ties longer han 12,000 km, since the antipodal points of all major clusters fall n the ocean.

Compared with this baseline, the observed distribution of tie engths shows a clear surplus of ties for distances up to 1000 km,

somewhat mixed record from there to 5000 km and a consistent eficit of ties at greater distances. We note, though, that the peak

n the number of ties at the New York–Los Angeles distance is actu- lly higher than we would expect if ties were formed randomly. On he other hand, several other expected peaks remain unrealized. In articular, we observe no peaks at the values corresponding to the

istances between New York and São Paulo, New York and Tokyo, nd Tokyo and São Paulo.

For network comparison we created a “distance” network in hich the weight of edges was set to a natural logarithm of the

umber of ties by distance, in 200 km bins (for example, New York–London ties, at o simulations is the same as in the observed data. Based on 1259 dyads.

great-circle distance between the two clusters, calculated using the standard haversine formula. The comparison of this network to the network of Twitter ties for the top 25 clusters shows a correlation of −0.45 for the top 25 clusters, with p < 0.001 (Table 4). We note that our dependent network (“Twitter”) is based only on ties that connect users in different clusters, omitting the 39 percent of the ties that fall within clusters. Therefore, the correlation with the dis- tance network cannot be explained simply by the large number of local ties, but rather, shows a further constraining effect of distance on non-local ties.

5.2. Air travel

To investigate the effect of the ease of travel on Twitter ties we obtained a dataset showing a number of direct flights between pairs of 3023 airports on five different days in 2008 and 2009 (Mendelsohn, unpublished data). We assigned those flights to pairs of clusters by matching each cluster to the airports located within 100 km from its center. We then constructed a network by giving each pair of clusters a weight based on the natural logarithm of the observed number of flights between the airports assigned to each of them.

Comparing the air travel network with the network of Twit- ter ties shows a correlation of 0.51 for the top 25 clusters, with p < 0.001 (Table 4). The network of flights is thus a better predictor of non-local Twitter ties than physical distance. One interpreta- tion of the predictive power of flight frequency is that frequent flights facilitate travel, which allows for formation of face-to-face ties and increases the likelihood of Twitter connections. (This may, for example, include the fact that when people travel or move

Domestic 0.440 0.723 0.709 Language 0.418 0.637 Flights 0.510

All p-values are ≤0.005.

Y. Takhteyev et al. / Social Networks 34 (2012) 73– 81 79

Table 5 Top countries.

Share of egos (%)a

Share of egos (%) for egos in dyadsb

Share of alters (%)c

Percentage of domestic tiesd

Percentage of domestic ties among non-local tiesd

Following foreign alters/being followed from abroad

Country named explicitly (% of egos)

USA 48.5 45.7 54.5 91.6 89.3 0.3 8.1 Brazil 10.6 12.1 10.5 83.5 72.5 4.9 55.4 UK 7.6 8.3 7.6 50.6 33.3 1.2 45.3 Japan 5.5 6.5 6.3 92.1 86.0 1.4 25.0 Canada 3.7 3.8 2.9 33.3 23.1 1.6 58.5 Australia 2.7 2.7 1.9 50.0 32.0 2.2 69.7 Indonesia 2.6 1.8 1.2 60.0 25.0 7.0 83.3 Germany 2.1 1.8 1.3 62.9 58.8 3.2 58.6 Netherlands 1.4 1.4 1.2 66.7 22.2 1.5 54.3 Mexico 1.2 1.3 0.7 44.0 8.3 7.0 56.7

a Out of the 2852 egos located at the level of country or better. country or better.

all ties for egos in that country.

b c ( b i r 1

5

n p a t d

p o s w d o n f e c t s c E t b a t p

a l t t U I m s

o c

Table 6 The most common languages. Based on 2852 egos.

Language % of egos

English 72.5 Portuguese 10.1 Japanese 5.4 Spanish 3.1 Indonesian 1.8 German 1.7 Dutch 1.0 Chinese 0.9 Korean 0.4

b Out of the egos included in 1953 dyads with both parties located at the level of c Out of the 1953 alters located at the level of country or better. d The number of ties with the ego and the alter in the given country as a share of

etween those two interpretations. We also note that top Twitter lusters intersect only to an extent with Alderson and Beckfield’s 2004) ranking of world cities based on multinational corporations’ ranch headquarters. (Of Alderson and Beckfield’s top 25 cities by

n-degree or “prestige,” 13 appear in the top 25 Twitter clusters anked by in-degree centrality, with another 6 appearing in top 00.)

.3. National borders

Of the ties that were matched to countries, 75 percent con- ect users in the same country. This prevalence of domestic ties is artly explained by the high frequency of local connections, since ll local ties are domestic. Looking at just the non-local ties (i.e., ies between users in different clusters), we find that the share of omestic ties is lower but still substantial: 63 percent.

As with distance, the high frequency of domestic ties can be artly explained by the concentration of users in a small number f countries, with nearly half of them in the United States. The hare of domestic ties, however, substantially exceeds what we ould expect if users formed connections randomly while being istributed as they are now, which would result in only 26 percent f the ties being domestic. Further, the surplus of domestic con- ections holds for all major countries, including those that account

or just a small fraction of the egos, as shown in Table 5. (The ffect is somewhat reduced for countries that have only one major luster, since in those cases removing local ties means removing he majority of the domestic ties.) The table also shows that the hare of domestic ties is generally higher for non-English-speaking ountries (as long as they have several clusters), yet even the nglish-speaking countries show a higher share of domestic ties han would be expected from their share of egos. A comparison etween the network of Twitter ties between the top 25 clusters nd the “domestic” network (where edges were set to 1 for domes- ic ties and 0 for international) shows a correlation of 0.44, with

< 0.001 (Table 4). The substantial share of the United States in the sample warrants

comparison with other countries. The share of domestic ties is ower for egos located outside the United States: 62 percent of all ies and 42 percent of non-local ties. However, the share of domestic ies is higher for pairs where both parties are located outside the nited States: 80 percent of all ties and 65 percent of non-local ones.

n other words, Twitter users outside of the US have a somewhat ore international orientation than American users, but only in the

ense that they tend to follow users in the US. It is also important to note the differences in the pattern of

utgoing ties (following) and in-coming ties (being followed). As olumn 7 in Table 5 shows, the majority of the US’s international

Swedish 0.4 Russian 0.4

ties are incoming: Twitter users in the United States are often fol- lowed from abroad, with over three incoming ties for each outgoing tie. For some of the other countries, on the other hand, international ties are overwhelmingly outgoing. For example, the ratio of incom- ing ties to outgoing is nearly one to five for users Brazil, who actively follow foreign accounts, but receive little attention in return.

The more domestic orientation of the American users also reflects itself in how they describe their locations. When coding the locations we noted whether the country was stated explicitly or implied (e.g., “São Paulo, Brasil” vs. just “São Paulo”). As shown in column 8 of Table 5, only eight percent of US location descriptions explicitly name the country, compared to, for example, 55 percent of locations in Brazil. This may suggest that American users of Twit- ter either see their audience as exclusively domestic (even though it is not), expect foreign users to know the names of American cities, or simply do not think about Twitter users abroad. The United States is closely followed by Japan, where only 25 percent of location descriptions identify the country explicitly. However, in the case of Japan, this may be explained by the fact that in the overwhelm- ing majority of cases, locations in Japan were identified in Japanese (using kanji or kana), which makes them intelligible only to people who know Japanese and would be familiar with Japanese cities. Additionally, Japanese users are followed almost exclusively by others in Japan: ties from foreign egos account for a relatively small fraction (10 percent) of the ties received by Japanese alters. Note that the Brazilian users have proportionally even fewer incoming foreign ties. This does not, however, stop them from identifying their country explicitly.

5.4. Language

A large majority of egos (62 percent) and alters (68 percent) are located in countries where English is the dominant language.

80 Y. Takhteyev et al. / Social Networks 34 (2012) 73– 81

Table 7 Language combinations. Based on 1768 dyads in which both ego’s and alter’s language was identified.

Language combinations As a percentage of:

All ties In-cluster ties X-cluster ties Int’l ties

Same language (total) 88.4 91.6 88.4 75.5 English–English 67.5 63.3 74.4 70.1 Same language, non-English 20.9 28.3 14.0 5.4

Cross-language Other–English 7.4 2.6 8.5 20.5 English–Other 3.1 4.5 2.1 3.1 Different languages where neither is Englisha 1.1 1.2 1.0 0.9

a The most common combinations were Japanese–Chinese, Spanish–Italian, and Portuguese–Spanish.

Table 8 QAP regressions of the number of Twitter ties on distance, being in the same country (“domestic”), language similarity and the number of flights. Distance, the number of twitter ties and the number of flights are logged. Standardized coefficients are shown in parentheses.

Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7 Model 8 Model 9

Intercept 2.346 0.139 1.450 −0.048 1.503 0.012 −0.015 0.451 0.070 Domestic 0.522*** (0.440) 0.290* (0.244) 0.342** (0.288) 0.093 (0.079) Language 0.600*** (0.418) 0.329* (0.229) 0.307 (0.214) 0.171 (0.119) Flights 0.118*** (0.510) 0.101** (0.435) 0.082* (0.356) R2 0.201 0.193 0.229 0.175 0.233 0.216 0.261 0.263 0.278 Adj. R2 0.201 0.193 0.228 0.175 0.232 0.215 0.261 0.262 0.274 Number of observations 600 combinations of 25 nodes

A c c o l 9 E w p r c

m s i p p t f S w l w t m f p w t h

e s E n r t

* p ≤ 0.05. ** p ≤ 0.01.

*** p ≤ 0.001.

lmost all egos (96 percent) located in the English-speaking ountries follow alters who are also located in English-speaking ountries. This number, of course, reflects in part the large share f domestic ties within the United States. However, even for egos ocated in English-speaking countries other than the United States, 1 percent of ties are to English-speaking countries. For non- nglish speaking countries, the share of ties to users in countries ith the same dominant language is lower but still significant: 69 ercent. (For the most part, however, this latter number simply epresents the share of domestic ties for the non-English-speaking ountries.)

For the purposes of correlation analysis we built a language atch network using a dataset of access to language-specific ver-

ions of Wikipedia from each country. For example, the dataset ndicated that requests for the English Wikipedia accounted for 94 ercent of all requests coming from the United States and for 15 ercent of requests coming from Brazil, while requests for the Por- uguese Wikipedia accounted for 83 percent of requests coming rom Brazil and 0.16 percent of requests coming from the United tates. To get a measure of proximity between a pair of clusters e summed the products of the two countries’ preferences for

anguages. For example, the New York–São Paulo pair received a eight of 0.14, reflecting the match in English (0.94 × 0.15 = 0.14),

ogether with negligible terms for other languages (0.001 for the atch in the preference for Portuguese and about the same amount

or the match in preference for Spanish). The New York–Tokyo air received 0.03, while the New York–Amsterdam pair received a

eight of 0.39, reflecting primarily the much lower preference for

he English Wikipedia among the requests from Japan and the much igher preference among requests coming from the Netherlands.8

8 We also constructed an alternative network based on the languages spoken in ach clusters and the proximity between the languages in the hierarchical clas- ification of languages (for example, assigning a higher degree of similarity to the nglish–Dutch pair than to English–Japanese). We have found that the two language etworks had a correlation of 0.95 and produced nearly identical results. For this eason we avoid the discussion of the alternative language metric, focusing just on he network produced from the Wikipedia dataset.

The resulting language network shows a correlation of 0.42 with the 25-cluster Twitter network, which is slightly smaller than our results for the country match network (Table 4).

It is important to consider, however, that users do not necessar- ily tweet in the language that is dominant in their location. For this reason, we coded the language of messages from egos and alters who could be located at least at the level of a country to see whether egos and alters use the same language. Table 6 shows the most common languages used in the tweets. English is by far the dominant language, accounting for 73 percent of egos’ messages – higher than the percentage of egos located in English-speaking countries. Portuguese is the only other language accounting for more than 10 percent. Japanese, Spanish, Indonesian and German each account for 1–10 percent, with all other languages being under 1 percent. Table 7 shows the most common combinations of lan- guages between egos and alters. In 88 percent of the cases, the ego and the alter tweet in the same language – slightly higher than the 86 percent of ties that connect users located in the same country or countries with the same dominant language. Over three-quarters of those (68 percent of all ties) are cases where both are using English, with slightly over one-quarter being cases where both use a differ- ent language, most often Portuguese or Japanese. Cross-language ties are relatively rare.

The share of same-language ties in languages other than English is substantially higher for local ties (28 percent) and substantially lower for ties between clusters (14 percent). It falls even further if only international ties are considered (5 percent). The total share of same-language ties drops somewhat as well: from 92 percent for local ties, to 88 for ties between clusters, to 76 percent for inter- national ties. This loss is made up almost exclusively by the share of ties in which a non-English-tweeting ego follows an English- tweeting alter.

Looking at the languages used by egos in each cluster or coun- try, we found a somewhat imperfect match between the language used by individual users and the dominant language of the clus-

ter. For example, while Portuguese is unambiguously the dominant language of Brazil, 13 percent of the tweets from users located in Brazil are in other languages, including 8 percent in English. An

al Netw

i t f

5

e t s t s t n d 8 o

6

a a w t p d h t t t w t o t u

b t s h l m m T o

n f i a p fi s

A

t

R

A

B

Y. Takhteyev et al. / Soci

nformal analysis of the profiles suggests that many of the English- wittering users located in Brazil are Brazilians rather than travelers rom English-speaking countries.

.5. Multivariate analysis

Having found that all four variables that we considered have an ffect on Twitter ties, we used regression analysis to see whether heir effects are independent. The results of the regressions are pre- ented in Table 8. Comparing model 3 with models 1 and 2, we see hat distance and being in the same country have independent and ignificant effects. A comparison of models 1, 4 and 5 shows that he same is true for distance and language. The effect of language is ot significant when we control for country. Similarly, the effect of istance is no longer significant when we control for flights (model ). In a model combining all four variables (model 9), only the effect f flights remains significant.

. Conclusions

Looking at the network of ties in Twitter we find that distance nd related variables (language, country, and the number of flights) ll have an effect on Twitter ties despite the seeming ease with hich long range ties can be formed. As a lightweight system that

akes little effort to join and can be used from either personal com- uters or mobile devices, Twitter offers a promise of transcending istance, connecting everyone with anyone. Our analysis shows, owever, that distance considerably constrains ties. Two fifths of ies (39 percent) connect users within the same regional cluster, ypically the size of a metropolitan area. All such ties are domes- ic and connect users in the same linguistic area. Most of them fall ithin easy driving distance. Even for the remaining longer range

ies between different clusters, distance matters. Ties at distances f up to 1000 km are more frequent than what we would expect if he ties were formed randomly, while ties longer than 5000 km are nderrepresented.

For the longer ties, distance, language differences, country oundaries, and ease of travel can vary independently, even as hey remain strongly correlated. This warrants a comparison of uch variables. We find that country and the frequency of flights ave independent effects in pair-wise comparisons. The effect of

anguage is no longer significant when country is included in the odel. A closer look at language suggests that the language effect ight be weakened by the wide use of English as a lingua franca.

he effect of distance is no longer significant when the frequency f airline travel is included.

The number of airline flights emerges as the best predictor of on-local Twitter ties. This likely reflects the role of air travel in

acilitating long-distance face-to-face interaction, which in turn nfluences the formation of electronic ties. Air travel can also stand s a proxy for other kinds of pre-existing connections between laces, which in turn influence formation of electronic ties. These ndings highlight the importance of considering structural con- traints on ties rather than simple physical distance.

ppendix A. Supplementary data

Supplementary data associated with this article can be found, in he online version, at doi:10.1016/j.socnet.2011.05.006.

eferences

lderson, A., Beckfield, J., 2004. Power and position in the world city system. Amer- ican Journal of Sociology 109, 811–851.

arnett, G.A., Choi, Y., 1995. Physical distance and language as determinants of the international telecommunication network. International Political Science Review 16, 249–265.

orks 34 (2012) 73– 81 81

Beaverstock, J.V., Smith, R.G., Taylor, P.J., 1999. A roster of world cities. Cities 16, 445–458.

Boase, J., 2008. Personal networks and the personal communication system: using multiple media to connect. Information, Communication and Society 11 (4), 490–508.

Borgatti, S.P., Everett, M.G., Freeman, L.C., 2002. UCINet for Windows: Software for Social Network Analysis. Analytic Technologies, Cambridge, MA.

Butts, T.C, 2007. Permutation models for relational data. Sociological Methodology 37, 257–281.

Butts, T.C., Acton, R., 2010. Spatial modeling of social networks. In: Nyerges, T., Couclelis, H., McMaster, R. (Eds.), Sage Handbook of GIS and Society Research. Sage, Thousand Oaks, CA.

Cairncross, F., 1997. The Death of Distance: How the Communications Revolution is Changing Our Lives. Harvard Business School Press, Boston.

Derudder, B., Witlox, F., 2005. An appraisal of the use of airline data in assessing the world city network: a research note on data. Urban Studies 42, 2371–2388.

Fischer, C., 1982. To Dwell Among Friends. University of California Press, Berkeley, CA.

Gaudel, A., Peroni, C., 2010. Reciprocal attention and norm of reciprocity in blogging networks. Jena Economic Papers Report 2010-020, Friedrich Schiller University Jena, Jena, Germany.

Golder, S.A., Yardi, S., 2010. Structural predictors of tie formation in Twitter: transi- tivity and mutuality. In: Proceedings of the Second IEEE International Conference on Social Computing, August 20–22. IEEE Press, Minneapolis.

Granovetter, M., 1973. The strength of weak ties. American Journal of Sociology 78 (6), 1360–1380.

Herring, C.S., Paolillo, C.J., Ramos-Vielba, I., Kouper, I., Wright, E., Stoerger, S., Scheidt, A.L., Clark, B.,2007. Language networks on LiveJournal. In: Proceedings of the Hawaii International Conference on System Sciences. IEEE Press, Los Alamitos, CA.

Huberman, B., Romero, D., Wu, F., 2009. Social networks that matter: Twit- ter under the microscope. First Monday 14 (1), http://firstmonday.org/ htbin/cgiwrap/bin/ojs/index.php/fm/article/view/2317/2063.

Hutchinson, K.W., 2005. Linguistic distance as a determinant of bilateral trade. Southern Economic Journal 72, 1–15.

Java, A., Song, X., Finin, T., Tseng, B., 2007. Why we twitter: understanding microblog- ging usage and communities. In: Proceedings of the 9th WebKDD and 1st SNA-KDD 2007 Workshop on Web Mining and Social Network Analysis, San Jose, CA.

Krackhardt, D., 1987. QAP partialling as a test of spuriousness. Social Networks 9, 171–186.

Krishnamurthy, B., Gill, P., Arlitt, M., 2008. A few chirps about Twitter. In: Proceed- ings of the First Workshop on Online Social Networks, Seattle.

Liben-Nowell, D., Novak, J., Kumar, R., Raghavan, P., Tomkins, A., 2005. Geographic routing in social networks. Proceedings of the National Academy of Sciences of the United States of America 102 (33), 11623–11628.

Mendelsohn, J. Local cities, global influence: an inquiry into the relationship between the global city and its region, unpublished data.

Mok, D., Wellman, B., 2007. How much did distance matter before the Internet? Social Networks 29, 430–461.

Mok, D., Wellman, B., Carrasco, J.A., 2010. Does distance still matter in the age of the Internet? Urban Studies 46 (13), 2743–2783.

Naaman, M., Boase, J., Lai, C.,2010. Is it really about me? Message content in social awareness streams. In: Proceedings of the 2010 ACM Conference on Computer Supported Cooperative Work (CSCW ‘10). ACM, New York.

Schonfeld, E., 2009. Twitter Reaches 44.5 Million People Worldwide in June (com- Score), TechCrunch, Available from: http://www.techcrunch.com/2009/08/03/ twitter-reaches-445-million-people-worldwide-in-june-comscore/.

Rainie, L., Wellman, B., 2012. Networked: The New Social Operating System. MIT Press, Cambridge, MA.

Smith, D., Timberlake, M., 1995. Conceptualizing and mapping the structure of the world systems city system. Urban Studies 32, 287–302.

Sysomos, 2010. Exploring the Use of Twitter Around the World, Available from: http://www.sysomos.com/insidetwitter/geography.

Twitter, 2011. Twitter API Documentation, Available from: http://apiwiki.twitter.com/w/page/22554679/Twitter-API-Documentation.

Webber, M., 1963. Order in diversity: community without propinquity. In: Wingo, L. (Ed.), Cities and Space: The Future Use of Urban Land. Johns Hopkins Press, Baltimore, MD.

Wellman, B., 1979. The community question: the intimate networks of East Yorkers. American Journal of Sociology 84 (5), 1201–1231.

Wellman, B., Carrington, P., Hall, A., 1988. Networks as personal communities. In: Wellman, B., Berkowitz, S.D. (Eds.), Social Structures: A Network Approach. Cam- bridge University Press, Cambridge, MA, pp. 130–184.

Wellman, B., Hogan, B., with Berg, K., Boase, J., Carrasco, J.-A., Côté, R., Kayahara, J., Kennedy, T.L.M., Tran, P., 2006. Connected lives: the project. In: Purcell, P. (Ed.), Networked Neighbourhoods: The Online Community in Context. Springer, Guildford, UK, pp. 157–211.

Wellman, B., Tindall, D., 1993. How telephone networks connect social networks. Progress in Communication Science 12, 63–94.

Wimmer, A., Schiller, N.G., 2002. Methodological nationalism and beyond: nation-

state building, migration and the social sciences. Global Networks 2 (4), 301–334.

Zook, M., Brunn, S.D., 2006. From podes to antipodes: positionalities and global airline geographies. Annals of the Association of America Geographers 96 (3), 471–490.