Content uploaded by Robert Fahey
Author content
All content in this area was uploaded by Robert Fahey on Oct 15, 2020
Content may be subject to copyright.
From Filter Bubble to Social Divide: Social
Polarisation in Europe and Japan
Robert A. Fahey
Waseda University
Stefano Camatarri
Catholic University of Louvain
September 2020
Draft: Not for citation or circulation without permission of the authors.
Funding Information
This work was financially supported by JSPS Grants-in-Aid for Scientific Re-
search Grant Number 19K23181 (PI: Robert A. Fahey).
1 Introduction
As values re-emerge as a central aspect of political discourse and identity in a
nation, the development and deepening of some form of polarisation appears to
be an inevitable corollary. All social and political value systems define them-
selves, either explicitly or implicitly, in contrast to an opposing value which is
held — or assumed to be held — by some other group within or without the
1
nation. Traditional values, for example, are promoted by those concerned that
what they consider to be the nation’s traditions are being changed or under-
mined by groups with progressive or reformist values. Pluralist values, to give
another example, are held up by those who fear that the nation’s diversity is be-
ing suppressed or discouraged by groups holding values that champion national
homogeneity. The intrinsically polarising nature of value systems is most easily
seen when they bend towards authoritarianism; Altemeyer’s tripartite definition
of right-wing authoritarianism (Altemeyer 1981) includes “authoritarian aggres-
sion” as one of its components, acknowledging that conflict with those who do
not share the same values is an intrinsic aspect of authoritarian values them-
selves. Most definitions of populism, meanwhile, link the phenomenon explicitly
to a conflict of values that divides the nation into two antagonistic groups, “the
good people and the corrupt elite” (Mudde 2004) and frames the clash of their
values as a Manichean struggle between good and evil (Hawkins 2009).
Alongside this values-led approach to understanding polarisation, a more
technical approach to defining the phenomenon has also emerged, rooted in spa-
tial approaches to party competition in which the ideological and value stances
of political actors are represented as points across a uni- or multi-dimensional
space, with the distances between such points being measured as Euclidean dis-
tances with Cartesian coordinates. This is the type of empirical approach upon
which the present study is based. Literature on this topic dating back as far as
Downs (1957) has discussed the existence of “converging” and “diverging” dy-
namics in party competition and public debate. Downs noted that the left-right
preferences of voters may follow either a normal distribution — encouraging
parties to appeal to the middle ground — or a U-shaped distribution, which
creates a centrifugal trend among political actors, leading to increased polar-
isation. More recent research has shown that the more parties diverge along
2
the left-right dimension, the more they tend to strengthen the left-right orienta-
tions of the citizens (Freire 2008), as well as prompting more ideological voting
(Lachat 2008). A similar dynamic also appears to apply to other dimensions of
political conflict. Spanje and Vreese (2011) demonstrated that the more parties
diverged on the question of EU integration, the more individuals’ voice choices
became driven by EU issues. These studies support Downs’ notion of a “cen-
trifugal force” — we might perhaps even think of a “feedback loop” between
voters and parties, in which each party is gradually drawn into an increasingly
extreme and polarised position by the actions or perceived preferences of the
other.
Polarisation emerges and deepens when individuals increasingly come to de-
fine their political identity in line with such a conflict of values, reifying abstract
values into an “us versus them” struggle. As values consequently become a pil-
lar of community identity (rather than a personal or individual belief), social
pressure may be exerted upon community members to conform by homogenis-
ing a wider set of their beliefs with their community, resulting in political and
ideological differences being increasingly aligned along a single dimension be-
tween two polarised values (McCoy et al. 2018; Webster and Abramowitz 2017).
The establishment of values not as a point of legitimate discussion among in-
dividuals but rather as a central pillar of community identity creates a wedge
within society such that people with differing values come to be perceived not as
fellow citizens with legitimately differing views but rather as enemies, members
of a conspiring out-group who seek to undermine the nation and tear down the
values of the good, true in-group — and against whom, consequently, discrimi-
nation and even hatred may be justified (Billig and Tajfel 1973; Gaertner et al.
1993; Iyengar et al. 2012; Iyengar and Westwood 2015).
This worst-case scenario, however, does not always come to pass. While po-
3
litical polarisation — a widening gulf between the political positions of parties
and voters, and a consequent narrowing of the middle ground and reduction
of the scope for compromise — is common in many nations, it does not neces-
sarily follow that their citizens will become deeply divided into entrenched “us
versus them” mindsets. We might characterise this as the difference between
political and social polarisation — the former being a reflection of a widening
and deepening divide in political ideologies and preferences, while the latter
consists in a division that extends outside the political realm and encompasses
basic differences in how people live, with whom they interact, what media they
consume, how they entertain themselves and even what they wear, where they
shop and what they eat. These two forms of polarisation are often connected
to some degree — for example, in nations where social cleavages have provided
the foundations for a political party system divided along cleavage lines. In re-
cent years significant attention has been paid to the hypothesised deepening of
both political and social polarisation in the United States, for example, with a
number of influential works highlighting the extent to which individuals’ social,
consumer and media choices, as well as their beliefs on a broad set of issues,
are increasingly aligned along politically polarised lines (Bishop and Cushing
2008; Campbell 2016; Hetherington and Weiler 2018). Party identification has
been shown to be more polarising even than race for many Americans (Iyengar
et al. 2012) and to underlie a tribalism that fuels “loathing” of rival party sup-
porters (Iyengar and Westwood 2015). The 2016 Brexit referendum in the UK
may have created similar political and social divisions between British “leavers”
and “remainers”, though analysis of this phenomenon is still in its early stages.
However, we can also easily conceive of a nation where strongly polarised po-
litical views exist among a populace that is otherwise fairly homogeneous in its
social behaviour and traits, or conversely of a nation where two or more popu-
4
lations with very different social lives and behaviours are nonetheless joined by
a largely homogeneous set of political ideologies.
The core question posed in this chapter is whether this kind of social polarisa-
tion and its correlation to political polarisation can be detected and quantified,
preferably in a way that permits cross-national comparison — thus allowing,
for example, the conditions of social and political polarisation in European na-
tions to be compared not only with one another, but also with Japan and other
non-European countries. To this end, we turn away from the analysis of party
and political actor positions which has been traditionally employed in studies
of political polarisation (Dalton and Tanaka 2007; Pelizzo and Babones 2007),
and instead employ analysis of large-scale social media data, which allows us to
observe the interests and preferences, political and otherwise, of large numbers
of citizens in different countries.
1.1 Social Media Polarisation
Using social media data to study political and social polarisation is an especially
relevant approach given the extent to which social media itself has been impli-
cated in the rise of polarisation — and the possibly corollary rise in populism
— in many countries. Of particular concern is the existence of “filter bubbles”,
individual- or community-level media environments which prioritise information
and viewpoints that a user will prefer, while suppressing or entirely hiding in-
formation and viewpoints that will challenge their preferences (Pariser 2011).
Much of this process, sometimes called “Cyberbalkanisation”, is a consequence
of algorithms which seek to maximise users’ interactions (sharing or comment-
ing on content) and thus steadily push them towards content which will elicit
such a response from them. Algorithms, however, are not the only force driving
this process. Users’ own choices also create filter bubbles — most notably, their
5
natural preference for following social media accounts with which they tend to
agree or feel positively about results, over the accumulation of dozens or hun-
dreds of following decisions on a network such as Twitter or Facebook, in the
creation of a highly personalised “feed”, or social media environment, that is
tailored to that user’s preferences and positions. This ability for an individual
to implicitly fine-tune their media environment through months or years of fol-
lowing and unfollowing choices, combined with algorithms which further refine
the content to which the user will be exposed, can result in the creation of a
highly rarefied, balkanised media environment and thus in the gradual radical-
isation of individuals whose stances are constantly propped up by exclusively
supportive media and commentary.
The preference of users to follow social media accounts with which they
largely agree (or otherwise feel positively about) has given rise to a new form of
political analysis which uses the social network graph itself to estimate the latent
space positions of political and media actors. Based on earlier work by Hoff et al.
(2002) on calculating latent social positions from network data, Barber´a (2015)
showed that the positions of U.S. political figures along a single-dimensional
scale could be calculated from a network analysis of their overlapping followers
on Twitter. This methodology, while very effective at establishing political divi-
sions on social media — demonstrating both the polarisation between political
actors and the degree of political polarisation among the ordinary citizens who
follow them on Twitter — considers only the political dimension of each user’s
social media environment. It uses the accounts of ordinary citizens exclusively
as a source of network data about the accounts of political actors. In reality,
however, most citizens follow a large number of non-political accounts in ad-
dition to whatever politics-related accounts they choose to follow; aside from
following the accounts of friends and colleagues (their actual “social network”
6
in the most traditional sense), citizens will also follow a variety of social, cul-
tural and commercial accounts related to non-political aspects of their lives —
media, entertainment, arts, sports, food and drink, shopping, consumer brands,
and so on. This raises the possibility that citizens who are deeply polarised
from one another in terms of their political following choices might nonetheless
have very similar social media environments across other dimensions, implying
that citizens’ political differences do not align with a deeper cleavage in other
aspects of their lives — so regardless of which political party a citizen follows,
they may still enjoy the same media and entertainment, watch the same sports
and shop in the same stores as supporters of the opposing party. On the other
hand, if divisions in other dimensions mirror the polarisation along political
lines, this would imply the existence of true social polarisation, wherein citizens
with opposed political views also watch different media, follow different sports,
enjoy different entertainment and perhaps even eat different food, wear differ-
ent clothes and shop in different stores, effectively having very different lived
experiences to their fellow countrymen who live parallel but non-interacting
lives.
This chapter’s methodology focuses on the non-political sphere of citizens’
social media networks, establishing their political leaning by observing the po-
litical accounts they follow and then examining the degree to which followers of
different political parties are similar or different in their choices of non-political
accounts to follow. In essence, our core question is whether dividing Twitter
users up by political leaning reveals divides which carry over from the political
realm into other realms of their social and cultural lives, or whether the divides
disappear once political accounts are removed from the analysis, implying a cit-
izenry that is either largely homogeneous, or at the very least, socially divided
in a manner that does not mirror the nation’s political faultlines.
7
2 Data and Analysis
2.1 Cases
Our research compares four national cases — Japan and three European na-
tions, namely Italy, Belgium and Ireland. These cases provide a wide variety of
different contexts in which to test the relationship between political and social
polarisation, and as such we anticipated that these cases would yield a wide
variation in results from our analyses. Their demographic characteristics vary
significantly in terms of population size — Japan and Italy are both large na-
tions (c.120 million inhabitants and c.60 million inhabitants respectively), while
Belgium and Ireland are relatively small (c.11 million and c.5 million respec-
tively) — but their economic conditions are broadly comparable, with all four
nations being relatively wealthy (2017 GNI per capita ranging from $41,430 for
Italy and $43,490 for Japan up to $50,760 for Belgium and $61,170 for Ireland)
and having relatively low economic inequality scores (the World Bank’s most
recent GINI figures for each nation ranging from 27.4 for Belgium up to 32.8
for Japan, 32.9 for Ireland and 35.9 for Italy (World Bank 2017)).
The cases are divided in terms of political and social homogeneity, especially
in terms of the proliferation of populist political movements which generally
thrive on significant social division. Italy and Belgium both have major pop-
ulist political movements (Movimento 5 Stelle / Five Star Movement / M5S
and Lega Nord / Northern League / LN in Italy, and Vlaams Belang / Flemish
Interest / VB in Belgium), while neither Ireland nor Japan have seen populist
challenger parties make significant headway1. Though each of these countries
1Some scholars have, however, noted populist tendencies in Ireland’s Sinn Fein, which won
significant support in the country’s 2020 general election (Damiani 2020; Suiter 2017), as well
as in aspects of Japan’s political spectrum ranging from niche and regional parties (Hieda
et al. 2019; Klein 2020) to the leadership of the country’s major established parties (Yoshida
2019). Nonetheless, the extent of populist political activity and support in these countries
appears to be on the opposite end of the spectrum from both Belgium and Italy.
8
is homogeneous in terms of religion (Belgium, Italy and Ireland all having large
Catholic majorities, while Japan’s primary religious tradition, a mixture of Bud-
dhism and Shinto, is also held by a large majority), but Belgium, of course, has
a significant national cleavage between its Dutch-speaking and French-speaking
groups, while Italy’s socio-economic divide between the north and south of the
country is long-standing and well-documented (Gonz´alez 2011; Musolino 2018).
By comparison, neither Ireland nor Japan has such a notable socio-political
division between demographic or regional groupings.
The different characteristics of our cases led us to expect that regardless
of their degree of political polarisation, we would find significantly higher non-
political division among social media users in Belgium and Italy (most notably in
Belgium, where the linguistic difference would naturally incline users to engage
with different social media spheres) than in Japan or Ireland.
2.2 Data
Data for this study was collected in a three-stage “snowball” process, with
the data collected in each stage serving as the seed data for the next stage.
Firstly, for each country we identified the Twitter accounts belonging to (a)
every political party and (b) every elected national lawmaker. Each of our
target countries has a bicameral parliament, and Twitter accounts for members
of both the upper and lower house were collected for each. The data collection
model is illustrated in Figure 1, in which these political accounts are the central
nodes labelled p. We next used the Twitter API to download the complete list of
every account which follows one of those “political accounts”, yielding for each
country a second-stage data set containing many millions of accounts (mostly
belonging to ordinary citizens) which follow at least one account belonging to a
national politician or political party. These are the next order of nodes labelled
9
cin Figure 1.
The third-stage data collection involved collecting all of the other Twitter
accounts which each of those “citizen accounts” follows — essentially finding
out what other accounts they follow apart from the political accounts we had
already identified. However, it was impractical to gather the full set of this
data for each country, as we had identified many millions of citizen accounts.
Instead, we filtered the citizen accounts to remove inactive and “bot” accounts2
and removed users who followed less than five political accounts in order to
minimise the number of “orphaned” nodes in the eventual network. We then
randomly selected 10,000 accounts from the filtered user list for each country
and accessed the following list (i.e. the accounts followed by that user) for each
of them, generating the final set of Twitter accounts for our analysis — the
nodes labelled tin Figure 1.
Figure 1: Data Collection Model
p1
c1
c2
p2
c5
c6
p3
c7
c8
c9p4
c10
c11
p5
c14
c15
pk
c16
c17
cnt1
t2
t3
t4
t5
t6
t7
t8
t9
t10 t11
t12
t13
tr
2We based this filtering on simple criteria, removing accounts which had posted zero tweets,
were followed by zero users, or followed an improbably large number of other accounts; these
kinds of accounts are often simply part of a bot network and have the sole function of boosting
an account’s perceived popularity by following and retweeting it.
10
2.3 Basic Political Polarisation
In order to establish the extent of political polarisation in the sample data for
each country, we first ran a set of logistic regression tests to establish how fol-
lowing an account from a given political party would impact a user’s likelihood
of following accounts from other political parties. The results of these tests for
each country are shown in Tables 1, 2, 3 and 4. Each table shows how following
a single account associated with the party in the vertical column influences the
odds of following an account from the party in the horizontal rows. Relatively
high negative values (above -0.1) are highlighted on the tables; these show that
following accounts from this party significantly reduces the odds of a user follow-
ing accounts from another party, implying the presence of political polarisation
between those parties.
The most obvious pattern - if the most predictable - arises from the Bel-
gian data, where there is a significant polarisation between French-speaking
and Dutch-speaking parties, with followers of French parties much less likely to
follow accounts associated with Flemish parties, and vice versa. Outside this
major linguistic divide we found that the following of accounts of the populist
FLemish party VB (Vlaams Belang) is polarised from those of centrist Flemish
party OVld; it also shows a negative correspondence with other Flemish parties
CD&V and sp.a, though in both cases below the -0.10 threshold. Followers of
the Workers party, which spans the country’s linguistic divide, are also polarised
from those of OVld.
High political polarisation was also evident in the Japanese data, suggesting
a deep ideological divide. The primary divide is between the governing parties
— the LDP and Komeito — and the right-wing JIP, and the leftist opposi-
tion parties, the CDPJ, JCP, and assorted Left-Liberal minor parties. Recently
formed right-wing populist party N-Koku (“Protect the People from the NHK”,
11
a reference to the party’s key manifesto promise to abolish the national broad-
caster), too, is strongly polarised from the left- and centre-left parties.
Italy and Ireland, by comparison, showed low political polarisation overall.
Among the Italian parties, the most notable polarisation was that followers of
the right-wing parties Fratelli d’Italia and Lega Nord are much less likely to
follow accounts associated with the Democratic Party, while followers of the
small centrist parties are polarised from followers of small left-wing parties.
Table 1: Political Polarisation: Ireland
Fianna Fail Fine Gael Indep. Left / Green Sinn Fein
Fianna Fail 0.098 0.218 0.053 −0.031
Fine Gael 0.017 0.193 −0.064 -0.223
Indep.10.036 0.009 0.346 0.039
Left / Green2−0.012 −0.010 0.503 −0.062
Sinn Fein −0.013 −0.074 0.171 0.109
Highlighted cells have strong negative values, indicative of political polarisation.
1Indep. includes independent TDs and the Independents4Change group.
2Left / Green group includes Labour, Green, Social Democrat and Solidarity parties.
Table 2: Political Polarisation: Belgium
cdH (F) PS (F) MR (F) D´eFI (F) CD&V (V) sp.a (V) OVld (V) VB (V) N-VA (V) Green Workers
cdH (FR) 0.470 0.600 0.686 0.144 -0.141 −0.055 -0.106 −0.024 0.110 0.192
PS (FR) 1.431 0.324 1.296 −0.019 0.014 -0.160 -0.119 −0.086 0.109 0.201
MR (FR) 1.359 0.249 1.366 0.075 −0.080 0.113 -0.127 0.041 0.043 0.062
D´eFI (FR) 0.350 0.163 0.270 −0.004 -0.117 −0.090 −0.039 −0.097 0.286 0.169
CD&V (VL) −0.009 −0.087 0.031 -0.299 0.249 0.336 −0.072 0.088 0.133 −0.082
sp.a (VL) -0.306 −0.041 -0.203 -0.503 0.282 0.270 −0.072 0.007 0.464 0.243
OVld (VL) 0.001 -0.126 0.110 -0.160 0.477 0.268 -0.166 −0.017 0.098 -0.248
VB (VL) -0.197 -0.150 -0.109 0.073 0.068 −0.024 −0.021 0.141 −0.028 0.119
N-VA (VL) -0.119 -0.196 −0.002 -0.153 0.223 0.003 0.098 0.217 0.014 −0.111
Green 0.187 0.037 −0.073 0.886 0.106 0.741 0.131 −0.068 −0.010 0.297
Workers 0.245 0.151 0.049 0.196 0.021 0.377 −0.066 0.049 −0.017 0.351
Highlighted cells have strong negative values, indicative of political polarisation.
Parties marked (FR) are French-speaking; those marked (VL) are Dutch-speaking.
2.4 Calculating Non-Political Polarisation
The results outlined above and in Tables 1, 2, 3 and 4 pertain exclusively to
“political” polarisation; they look only at the following relationships between
12
Table 3: Political Polarisation: Italy
Minor Centre PD/Viva Five Star Forza Fratelli Minor Left Lega Nord
Minor Centre10.036 0.016 0.253 0.718 -0.386 1.213
PD/Viva20.352 −0.082 −0.049 -0.287 1.390 -0.534
Five Star 0.218 −0.005 −0.087 0.033 0.101 0.092
Forza 0.373 0.018 −0.027 1.525 0.296 0.067
Fratelli 0.421 −0.082 0.014 0.662 0.141 −0.002
Minor Left3-0.199 0.200 −0.008 0.214 0.030 −0.014
Lega Nord 0.690 −0.028 0.013 −0.040 0.161 −0.094
Highlighted cells have strong negative values, indicative of political polarisation.
1Minor Centre parties includes Centrists for Europe (CpE), Union of the Centre (UdC), and the
Sardinian Action Party (PSdAz).
2Italia Viva split from the Democratic Party (PD) only shortly before our data was collected,
making them indistinguishable in this kind of analysis.
3Minor Left parties includes Article 1 (Art.1), Free And Equal (LeU) and Italian Left (SI).
Table 4: Political Polarisation: Japan
LDP Komeito CDPJ DPFP JCP L-Liberal JIP N-Koku
LDP −0.090 -0.124 0.443 -0.248 -0.159 0.570 0.898
Komeito 0.022 −0.041 0.223 −0.057 −0.081 0.062 -0.486
CDPJ -0.102 -0.104 0.613 −0.074 1.693 -0.203 -0.241
DPFP 0.116 −0.065 0.259 −0.060 −0.083 0.129 0.147
JCP -0.101 −0.034 0.172 0.127 0.699 −0.002 -0.127
L-Liberal1−0.001 −0.041 0.311 -0.131 0.043 -0.178 -0.302
JIP 0.141 −0.078 0.006 0.216 −0.056 -0.551 1.402
N-Koku 0.041 -0.199 0.005 0.081 −0.059 -0.103 0.241
Highlighted cells have strong negative values, indicative of political polarisation.
1L-Liberal group includes Social Democratic Party, People’s Life Party and Reiwa Shinsengumi.
13
politicians and citizens (i.e. the inner circle of following relationships between
the cnodes and pnodes in Figure 1), ignoring any other accounts which citi-
zens also follow. This kind of analysis of the overlapping followers of political
accounts, derived from the observation that that people generally choose to fol-
low social media accounts they agree with or feel positively towards, has been
used in the past to establish latent positions of politicians and parties, perhaps
most notably by Barber´a (2015). From the point of view of citizens’ Twitter ac-
counts, however, the politicians they follow represent only a fraction — usually
a small fraction — of the overall set of accounts they follow. Our subsequent
analyses turn the focus away from these political following relationships and in-
stead look at citizens’ following relationships with other, non-political accounts
(i.e. the outer citcle of following relationships between the cand tnodes in
Figure 1).
To carry out this analysis, we first assigned a label to each ”citizen account”
according to their apparent party leaning. We reasoned that a citizen who
followed a majority of accounts from one party could be said to have a leaning
towards that party, in-accordance with the above-mentioned logic that people
primarily follow accounts they agree with (Barber´a 2015; Barber´a et al. 2015).
The precise criteria for this labelling process was that users were labelled as
leaning towards a party when more than 40% of their followed political accounts
were from that party, and no other party amounted to over 20% of followed
accounts. A small number of users in each country (less than 10% of each
national sample) could not be labelled using this process and were excluded
from the analysis.
We next narrowed down our set of target accounts among those followed
by citizens by removing those which were only followed by a small number of
people in our sample. This was done to remove what we might think of as
14
”personal networks” from the sample — the large number of private individuals
who have Twitter accounts with small numbers of followers largely drawn from
their friends, family and other acquaintances. The criteria applied in this step
involved removing any account which was not followed by at least 1% of users
(100 users) in our sample. This made clear a significant difference between the
countries in our sample, which is summarised in Table 5 — the number of ac-
counts followed by 100 users in each country ranged from as low as 3,338 in
Italy up to 12,250 in Ireland. The numbers, which are a rough estimation of
the size of the overall “Twittersphere” for each nation, or the parent population
of accounts from which users selected those they follow, did not scale with the
size of the country, with the smaller countries — Ireland and Belgium — having
significantly larger “Twitterspheres” than the larger nations. We hypothesise
that the key factor here is not the size of the country or the number of Twitter
users in that country, but the number of Twitter users using that country’s
language(s) — hence Ireland, whose English-speaking users can follow popular
accounts from around the enormous English-language Twittersphere, having a
variety of commonly-followed accounts more than twice as broad as Japan, de-
spite Japan’s population being 24 times larger. Belgian users, similarly, can
select accounts to follow from the French-speaking and Dutch-speaking Twit-
terspheres.
Table 5: Accounts followed by at least 100 sampled users
Country Accounts
Ireland 12 250
Belgium 7088
Italy 3338
Japan 5280
Having narrowed down the set of target accounts on which we would focus,
15
we next calculated the Jaccard Index for every pair of citizen accounts in the
sample. This is defined as the intersection over the union of each of their sets
of followed accounts, J(A, B ) = A∩B
A∪B, and is an effective way to quantify the
similarity of the Twitter media environment created by each account’s following
choices. We also used a measure of dissimilarity, Jaccard Distance, for some
analyses - this is derived simply by subtracting the Jaccard Index from 1. Having
calculated this index on a pairwise basis, we could then calculate the mean
Jaccard Index for an entire national level sample, between users with the same
party leaning, or between groups of users with different party leanings.
We first used this index to calculate the “internal consistency” and “external
consistency” for each political party or grouping. The internal consistency of
a party is the mean Jaccard index of those who lean towards that party, with
a higher figure indicating a stronger consistency among party followers in their
choice of non-political accounts to follow. The external consistency is the mean
Jaccard index between followers of that party and all other citizen accounts in
the sample, indicating the degree to which the non-political accounts followed
by that party’s followers are consistent with the general patterns of non-political
account following within the country. Taken together, these two values show us
how distinct from the “average user” the followers of a certain political party
are; a high internal consistency combined with a low external consistency would
imply that the party’s followers exist within a “filter bubble”, with the non-
political accounts they follow being markedly different from those of their fellow
citizens but broadly in line with those of their fellow party followers. The results
of this analysis can be seen in Figures 2, 5, 9 and 13.
Using the same Jaccard index data, we next calculated the mean Jaccard
index between each pair of parties in each country. This provided the basis
for a network analysis, in which each party is a node and the strength of the
16
edge connecting two parties is the mean Jaccard index of that party’s followers
— in other words, the degree to which the non-political following decisions
of the two parties’ followers resemble one another. These networks were then
laid out in two dimensions using spectral embedding (a positioning algorithm
based on the eigenvectors of the graph’s Laplacian matrix), which allows us to
visually distinguish clusters of parties whose followers’ non-political following is
similar. This is especially important since the countries in our study all have
political systems with large numbers of parties, meaning that social polarisation
might be found between one party and all the others, or between two or more
clusters of parties. Other parties, while politically distinct, might show no social
polarisation between them. The network analysis, whose results are shown in
Figures 3, 6, 10 and 14, gave us an a posteriori sense of the party groupings
between which social polarisation is found. The clusters revealed by this network
analysis were then used as a basis for the next stages of our analysis.
2.5 Determinants of Social Polarisation
Political polarisation on a social network, as outlined above, can be determined
by looking at the existence or non-existence of overlapping followers of political
accounts. This approach ignores the remainder of a citizen’s social network
following choices, which the analysis outlined above uses to determine non-
political, or social, polarisation. However, that ”remainder” itself encompasses
a large number of different spheres, some of which might be considered “politics-
adjacent”, such as news media, activists or political commentators, while others
would generally be seen as truly non-political — food and recipe accounts,
weather or traffic information accounts, or sports-related accounts. Musicians,
actors, comedians and authors might all fall somewhere along the spectrum
between those two extremes, depending on the extent to which they make public
17
their political views or participate in social or political activism.
To gain some insights into the kinds of non-political accounts which are most
polarising between members of each of the clusters identified in the network
analysis above, we used the Craig’s Zeta algorithm (Craig and Kinney 2009)
which is more commonly employed in text mining, but is used here to identify
which Twitter accounts appear with unusually high or low frequency in the
following lists of certain parties’ followers compared to those of other parties’
followers. The most polarising accounts revealed for each nation were then
categorised by a native of that nation according to their type - news media,
shopping, musician, author, and so on — giving a snapshot of which types of
non-political accounts proved the most polarising between the clusters identified
in each nation3.
In accordance with the clusters identified in the network analysis stage out-
lined above, this Zeta analysis was carried out for the the divide between Sinn
Fein and all other parties (Figure 4) in Ireland; between the French and main-
stream Flemish parties (Figure 7), and also between the mainstream Flemish
parties and Vlaams Belang (Figure 7) in Belgium; between the Left Niche parties
and the mainstream parties (Figure 11), and Lega Nord and the mainstream
parties (Figure 12) in Italy; and finally, between the progressive parties and
conservative / ruling parties (Figure 15) in Japan.
2.6 Cross-National Comparison
The final analysis is designed to place the four target nations in a comparative
context, showing how the social polarisation among party followers in each na-
tion compares to social polarisation found elsewhere. In order to achieve this,
3This process was actually carried out twice, with the first pass used to identify polarising
accounts which were overtly political — for example, accounts belonging to political candidates
who had not been elected as national representatives and were therefore not included in our
original sample. These accounts were removed from the data set and the analysis re-run, so
the results presented here all exclude the influence of such accounts.
18
we needed to control for the differences in the Twittersphere for each country; as
noted previously, factors such as the language used in each country have a signif-
icant impact on the formation of users’ following networks on Twitter. We took
the mean Jaccard distance among all users to be a reasonable (though rough)
proxy value for the size of the Twittersphere and its impact on the expected
level of polarisation among users in each country, and used it as a baseline by
subtracting this value from the mean Jaccard distances between party clusters
(as used in the network analysis above). By controlling for the Twitter envi-
ronment of each nation in this way, we arrive at a value for the non-political
polarisation between followers of different parties which can be compared across
national borders. The outcome of this final analysis is shown in Figure 16.
3 Results
In this section, we will first discuss the results for each country individually, be-
fore moving on to a discussion of the cross-national comparison and the findings
more broadly.
3.1 Ireland
Although it is the smallest country in our sample by population, Ireland is a
native English speaking country which places it within the single largest lan-
guage group on Twitter. As a consequence, there was a much higher diversity
of popular accounts followed by Irish Twitter users than we found in any other
country — 12,250 accounts were followed by at least 100 of our sampled Irish
users, almost double the number found for the next-highest ranked nation, Bel-
gium. A natural consequence of this diversity was that Irish Twitter users had
the lowest mean Jaccard Similarity figure of the four countries studied, indicat-
ing that Irish Twitter users on average have comparatively few account follows
19
in common with their fellow Irish users.
Turning to analysis of the non-political accounts followed by users in our
sample (Figure 2), Sinn Fein followers stand out from followers of other Irish
political parties for having both the lowest external consistency (meaning that
Sinn Fein followers tend to follow different accounts to other Irish Twitter users)
and the highest internal consistency (meaning that Sinn Fein followers tend
to follow similar accounts to other Sinn Fein followers). Although we found
some degree of polarisation in all of the parties — meaning that every form
of political leaning influences non-political account following choices to some
degree, a factor which turned out to be common across all countries studied - it
was generally quite low. Fine Gael, the ruling party at the time we collected this
data (September 2019), also had relatively high internal / external consistency
figures although they were markedly lower than those seen for Sinn Fein.
Network analysis of the mean Jaccard Similarity between groups of party
followers (Figure 3) confirms that Sinn Fein is an outlier in terms of social
polarisation in Ireland, with all of the other parties being relatively closely
connected in a single “mainstream parties” cluster. There is some divergence
even within this cluster — centre-right party Fianna Fail is somewhat distanced
from the Leftist and Green party group (which includes Labour, the Green
Party and other small leftist groups), for example — but the distance between
Sinn Fein followers and the followers of other parties is by far the most notable
feature of the network graph.
While the extent of the social polarisation between Sinn Fein and the other
Irish parties may to some degree reflect the party’s status as an “outsider” to
the traditional political system of the country (its historical roots as the political
wing of the IRA terrorist group in Northern Ireland having made it into some-
thing of a pariah with other political groups in the Republic of Ireland), the Zeta
20
Figure 2: Ireland: Internal / External Consistency of User Networks
Figure 3: Ireland: Inter-Party Similarity of Non-Political Following
21
analysis of polarising accounts (Figure 4) suggests a more prosaic explanation.
The large majority of very polarising accounts were categorised as “News Me-
dia” or “Media Figure” (meaning journalists, commentators and broadcasters),
with the account most strongly associated with Sinn Fein followers (i.e. most
unbalanced in terms of positive likelihood of being followed by Sinn Fein-leaning
users, versus negative likelihood of being followed by users leaning towards other
parties) being @An Phoblacht, Sinn Fein’s party newspaper, while Irish national
daily newspaper @IrishTimes was the most strongly associated account for non-
Sinn Fein leaning accounts. A large majority of the top 20 strongly polarising
accounts on both sides are either newspapers, TV and radio shows, or journalists
and commentators associated with those shows.
Figure 4: Ireland: Polarising Accounts between Sinn Fein and other parties
While this might suggest strong media polarisation along political lines —
and indeed, An Phoblacht presents a very different political world-view to the
Irish Times — it is also striking that many of the media organisations most
strongly associated with Sinn Fein-leaning accounts are Northern Irish media
22
(including the Irish News, BBC News Northern Ireland and UTV, among oth-
ers), while those associated with the other parties are media from the Repub-
lic of Ireland (including the Irish Times, Irish Independent, and various RTE
television and radio shows). Moreover, among the polarising accounts on the
non-Sinn Fein side are @GardaTraffic, a state-run traffic information service for
the Republic of Ireland, and state economic development bodies @Entirl and
@IDAIRELAND. An obvious conclusion from this divide is that while our data
collection process did not target users in Northern Ireland (as we exclusively
used politicians in the Republic of Ireland as the “seed accounts” for our col-
lection), a large number of the followers of Sinn Fein politicians in the Republic
of Ireland are located in Northern Ireland or, at least, following Northern Irish
media outlets to a far greater extent than followers of other parties in the Re-
public. While it is difficult to assess exactly how much impact this had on the
degree of polarisation measured by our analysis, it is reasonable to conclude
that the most significant line of social polarisation among Irish Twitter users is
one defined not so much by political alignment as by the very different media
spheres which exist on either side of the border between Northern Ireland and
the Republic of Ireland.
3.2 Belgium
As explained above, we anticipated that Belgium would show the most clear-cut
social polarisation of any of the countries in our sample due to the natural social
cleavage which exists between the country’s Dutch-speaking Flemish population
and its French-speaking Walloon population. This social polarisation is indeed
clearly reflected in the network analysis of Belgian party clusters (Figure 6),
which shows a distinct separation between major clusters of French-speaking
and Dutch-speaking parties. This separation is bridged to some extent by the
23
Workers Party and Green Party, both of which sit in the Belgian parliament as
a single caucus of Flemish and Walloon lawmakers. Perhaps more surprisingly,
the populist Vlaams Belang party is also polarised from both groups of parties,
effectively forming an isolated cluster of its own in the network graph.
The large number of political parties in Belgium and the existence of a so-
cial cleavage within the country renders the results for internal and external
consistency of non-political social networks (shown in Figure 5) less useful than
in other countries. While the French-speaking parties appear to be significantly
more polarised than the Dutch-speaking parties, this outcome is readily ex-
plained by Belgium having a larger population of Dutch speakers than of French
speakers, so the “average” Twitter user leans more heavily towards Dutch ac-
counts than French accounts — so those users who follow more French accounts
consequently appear more divergent from their countrymen than those who fol-
low more Dutch accounts. Without identifying the French- or Dutch-speaking
nature of each of the thousands of non-political accounts in our sample (a task
beyond the scope of the current study), it is impossible to control for this factor,
thus limiting the usefulness of the per-party consistency statistics.
We conducted two Zeta analyses of accounts whose following was most
strongly polarised between the different clusters identified on Belgium’s network
graph — one to examine the polarised accounts between the major clusters of
Flemish and Walloon parties (Figure 7) and another to discover the polarised
accounts between the cluster of mainstream Flemish parties and the populist
Vlaams Belang (Figure 8). Unsurprisingly, language was the major feature of
the non-political polarisation found between the Flemish and Walloon parties,
with many of the most polarised accounts belonging to News Media or to Me-
dia Figures such as journalists and broadcasters, reflecting the different media
environments which exist for French- and Dutch-speaking Belgians. A handful
24
Figure 5: Belgium: Internal / External Consistency of User Networks
Figure 6: Belgium: Inter-Party Similarity of Non-Political Following
25
of politically-engaged private individuals and academics were also featured on
these lists. Notably, however, the most polarising accounts were almost exclu-
sively Belgian, with @lemondefr, the account of French newspaper Le Monde,
being the only non-Belgian account to appear in the lists.
A very different picture emerged from the analysis of polarising accounts
between the mainstream Flemish parties and Vlaams Belang, with many of
the accounts most strongly and uniquely associated with Vlaams Belang mem-
bers being those of foreign politicians (who, for the purposes of our analysis,
are coded as “non-political” in the sense that they are not politically active
within the target country, and thus following them arguably represents a cultural
choice rather than a directly political one). These included Dutch right-wing
politicians Geert Wilders (@geertwilderspvv), Thierry Baudet (@thierrybaudet),
Martin Bosma (@Martinbosma pvv), Machiel de Graaf (@GraafdeMachiel) and
Fleur Agema (@FleurAgemaPVV ), as well as Dutch right-wing political party
Forum For Democracy (@fvdemocratie); U.S. president Donald Trump (@real-
DonaldTrump); French right-wing leader Marine le Pen (@MLP officiel ) and
a leader of the UK’s Brexit campaign, Nigel Farage (@Nigel Farage). In ad-
dition, foreign media outlets Voice of Europe (@V of Europe) and The Post
Online (@TPOnl), both of which focus heavily on negative stories regarding
immigration and minorities, also featured among the polarised accounts closely
associated with Vlaams Belang followers. By contrast, the only non-Belgian
account to be strongly polarised towards the mainstream Flemish parties is
that of former U.S. president Barack Obama (@BarackObama). The strong
implication is that at least some part of the polarisation between followers of
mainstream Flemish parties and followers of Vlaams Belang arises from the
heavy engagement of the latter with right-wing and populist politics beyond
Belgium’s borders, especially in the Netherlands but also in other European
26
countries and beyond.
Figure 7: Belgium: Polarising Accounts between French and Flemish parties
Figure 8: Belgium: Polarising Accounts between Flemish parties and Vlaams
Belang
27
3.3 Italy
While Italy does not have the same linguistic and cultural cleavage found in
Belgium, the existence of a North-South divide within the country is well-
documented in past literature. This divide is reflected to some degree in the
political sphere by the Lega Nord (Northern League), but given the lack of a
linguistic divide (as found in Belgium) or a political border creating distinct
media spheres (as we found in Ireland), the extent to which this North-South
divide might be reflected in citizens’ social media environments was a major
point of uncertainty at the outset of our study. Our initial examination of the
internal and external consistency of the different political parties or groupings
(Figure 10) suggested that by far the most socially distinct, at least in terms
of distance from the “average Twitter user”, were the followers of Lega Nord,
closely followed by followers of the Left Niche Party grouping (made up of Ar-
ticle 1 (Art.1), Free And Equal (LeU) and Italian Left (SI)). By contrast, and
uniquely among all of the political parties in all of our target countries, the
followers of the Five Star Movement showed almost no difference between the
internal and external consistency of its followers’ non-political networks, with
the consistency among party supporters’ following choices actually being lower
than the average consistency among Italian Twitter users in general. This im-
plies that the Five Star Movement’s supporters are extremely diverse in their
non-political following choices compared to other groups of voters — meaning
that the party draws its support from a wide range of different types of people,
whose social following choices are more or less independent of their political
following choices.4
4A similar result was found for the Centrist Niche Parties — made up of Centrists for
Europe (CpE), Union of the Centre (UdC), and the Sardinian Action Party (PSdAz) — but
it is significantly less meaningful when applied to a group of parties, as even when they occupy
many similar policy positions, their identities and those of their supporters can be expected
to differ significantly.
28
These results were clearly reflected in the network graph analysis (Figure
10), in which Lega Nord formed an isolated cluster of its own, the Left Niche
Parties formed another isolated cluster, and all of the other parties were broadly
grouped in a single “mainstream parties” cluster. This mainstream cluster is
not entirely coherent — the Five Star Movement and the Democrats (including
recent splinter party Viva Italia) are grouped tightly, while Forza Italia, Fratelli
d’Italia and the Centrist Niche Parties are also grouped tightly, but the simi-
larities between these two sub-groupings are high enough to justify considering
them as a singular cluster for further analysis. One somewhat unexpected result
here was that despite being considered a right-wing populist grouping, Fratelli
d’Italia’s supporters have more in common with mainstream party supporters
than with Lega Nord supporters in terms of non-political account following,
implying that the polarisation being observed here is not simply a proxy for
political preference.
As with the Belgian case, we analysed the non-political Twitter accounts
that were polarised between party clusters using two separate Zeta analyses —
one for the accounts which distinguish followers of Left Niche Parties from those
of the mainstream / core parties (Figure 11), and one for the accounts which
distinguish Lega Nord followers from followers of mainstream parties (Figure
12). In both cases, Italy showed a marked difference from the other European
countries in the study in terms of the number of ostensibly “truly” non-political
accounts which transpired to be polarised between party supporters. Whereas
in both Ireland and Belgium the most polarised accounts tended to be related to
the news media or journalists — and thus arguably possessed of a political align-
ment — in Italy we found a large number of cultural figures such as musicians,
actors, directors and comedians to be among the most polarised accounts. Even
popular TV shows featured in the lists of polarised accounts, with talent show
29
Figure 9: Italy: Internal / External Consistency of User Networks
Figure 10: Italy: Inter-Party Similarity of Non-Political Following
30
The X Factor (@XFactor Italia) being strongly associated with the core party
grouping in the Zeta test against the followers of Left Niche Parties. Many of
the accounts most strongly polarised towards the Left Niche belonged to NGOs
or activists, as well as some comedians, authors and academics, while a large
number of those strongly polarised towards the core parties (in this match-up)
belong to musicians or people involved with TV and film — implying perhaps
that supporters of the Left Niche Parties tend to overtly reject what they per-
ceive as mainstream Italian media and popular culture.
A similar dynamic was also found in the comparison of the core parties’ sup-
porters with those of Lega Nord. A large number of cultural accounts (related to
music, TV, film or even sports) were strongly polarised towards the core party
followers, suggesting a genuine social divide between these two groups of Italian
citizens which transcends their political differences. In terms of the accounts
most strongly polarised towards Lega Nord followers, the regional nature of the
party’s identity was apparent — the state government of the northern region
of Lombardy (@RegLombardia) was among the accounts disproportionately fol-
lowed by Lega Nord supporters, for example. The party’s populist nature may
also be reflected in the only foreign accounts found in this sample, with Lega
Nord followers tending to follow Donald Trump while mainstream party fol-
lowers tended to follow Barack Obama — a similar dynamic to the one found
between mainstream Flemish parties and Vlaams Belang in Belgium.
3.4 Japan
Although Japan has by far the largest population of any of our target nations
— both in terms of the overall population and the number of active Twitter
users — the mean Jaccard similarity amongst its users was broadly in line with
Italy, and higher than Ireland, likely as a consequence of the insular nature
31
Figure 11: Italy: Polarising Accounts between Left Niche and Mainstream par-
ties
Figure 12: Italy: Polarising Accounts between Lega Nord and Mainstream par-
ties
32
of the country’s Twittersphere, with few users following many accounts from
outside Japan. Testing the internal and external consistency of party supporter
groupings (Figure 13), we found that by far the most polarised parties were the
Japan Communist Party (JCP) and the Liberal-Left grouping, encompassing
the Social Democratic Party, People’s Life Party and Reiwa Shinsengumi. Both
of these groupings had extremely high internal consistency, suggesting a strong
tendency for members to follow broadly the same non-political accounts as one
another, as well as having the lowest external consistency in the country, mean-
ing that members tend not to align with other Japanese users in their following
preferences.
This observation was borne out by the network analysis (Figure 14), but
this made clear that the JCP and Liberal-Left supporters’ non-political fol-
lowing patterns are actually quite similar to the major progressive opposition
party — the Constitutional Democratic Party of Japan (CDPJ) — and mod-
erately similar to the centrist Democratic Party for the People (DPFP). Their
polarisation, then, is a contrast with the other cluster which emerges on the
graph, consisting of the two members of the coalition government (the Lib-
eral Democratic Party (LDP) and their junior coalition partner Komeito) and
a centre-right opposition party, the Japan Innovation Party (JIP). The posi-
tioning of parties and formation of clusters in this network analysis largely fits
with a conventional understanding of the Japanese political party system —
which, given that the graph is formed solely from data related to non-political
Twitter accounts, implies a degree of non-political polarisation that may be
considered slightly surprising in a nation so often (somewhat glibly) described
as “homogeneous”.
To analyse the non-political accounts which are most strongly polarised in
Japan, we conducted a single Zeta analysis (Figure 15) contrasting the conser-
33
Figure 13: Japan: Internal / External Consistency of User Networks
Figure 14: Japan: Inter-Party Similarity of Non-Political Following
34
vative grouping (the LDP, Komeito and JIP) with the progressive opposition
parties (CDPJ, DPFP, JCP and Liberal-Left parties). The results of this anal-
ysis showed a wide mixture of different types of accounts which were heavily
polarised between the two sides. Media sphere polarisation plays some role in
the divide — notably the left-leaning newspaper Tokyo Shimbun (@tokyoseijibu)
is polarised towards the progressive parties, while the right-leaning newspaper
Sankei Shimbun (@Sankei news) is largely followed by supporters of conser-
vative parties, and several broadcasters and journalists appear on either side.
However, the lists also includes several authors — notably prolific academic
and author Uchida Tatsuru (@levinassien) and historical non-fiction author Ya-
mazaki Masahiro (mas yamazaki) on the progressive side, while non-fiction
writer Takeda Tsuneyasu (@takenoma) and historical fiction author Hyakuta
Naoki (@hyakutanaoki) are on the conservative side. Notably, several of the
featured authors have written texts related to the Second World War — with
those on the conservative side generally falling into the camp of romanticising
or outright revising the accepted narratives of Japan’s role in the conflict, a
defining feature which is also found in several of the other polarised accounts
on the conservatives side including Takasu Katsuya (@katsuyatakasu), a plastic
surgeon who is well-known in Japan for his right-wing leanings and historical
revisionism, and Kent Gilbert (@KentGilbert01 ), an American actor and his-
torical revisionist who has written several books denying accounts of Japanese
war crimes. The presence of these accounts underlines the extent to which war
remembrance and revisionism is a major dividing line in Japan’s socio-political
landscape.
Another notable feature of the polarising accounts identified in Japan is that
conservative party followers are much more likely to follow certain state-operated
accounts — especially those run by Japan’s self-defence forces (@JGSDF pr,
35
@JMSDF pr, and Ministry of Defence account @ModJapan jp), but also more
general state-run accounts such as the one for the Prime Minister’s Office, @kan-
tei. This reflects the long association between the LDP and the Japanese Gov-
ernment, the party having been in power for all but four years since the mid-
1950s. Additionally, a large number of academics are found on both sides of
the polarisation, significantly more so than in the other countries studied —
which may imply that academics are more likely to play a significant role in
political and social debates in Japan than in Europe, or simply that academia
is itself more of a site of polarisation in Japan than in the European nations we
analysed.
Figure 15: Japan: Polarising Accounts between Conservative and Progressive
Parties
3.5 Cross-National Comparison
Finally, we gathered the different clusters identified in our various national anal-
yses and, controlling for the baseline level of similarity among citizens’ non-
36
political social networks in each nation, compiled a comparison that shows how
each country’s level of non-political polarisation on social media compares to the
others (Figure 16). Anti-clockwise from the right hand side, this figure shows
how the axes of polarisation among the clusters found in Ireland, Italy, Belgium
and Japan compare to one another, with a larger bar indicating a significantly
stronger degree of polarisation between the clusters.
Figure 16: Cross-National Comparison of Non-Political Polarisation
This comparison reveals a significant difference between the degree of social
polarisation found in each target country once the baseline of their citizens’ so-
cial media usage patterns is normalised. While the network analysis carried out
in each country showed notable clusters — indicating that at least to some ex-
tent citizens’ non-political social media following choices are correlated to their
political following choices — placing these clusters in a cross-national context
37
shows that the depth of the social division among citizens located in those clus-
ters varied widely. In line with our expectations, the highest degree of social
polarisation was found in Belgium, which yielded both of the deepest divisions
measured in the analysis — the deepest being the division between the populist
Flemish party Vlaams Belang and the cluster of French-speaking parties, fol-
lowed closely by the division between the French-speaking party group and the
more mainstream Dutch-speaking party group. Given the linguistic divide be-
tween Flemish and Walloon groups in Belgium, this depth of division is largely
to be expected. The divide between Vlaams Belang and the Dutch-speaking
party group which was noted in the network analysis (Figure 6), however, is
shown here to be largely insignificant.
Somewhat surprisingly, the division between the clusters found in network
analysis of the Italian party followers (Figure 10) is almost as deep as the division
noted in Belgium, despite the linguistic homogeneity of Italy. Specifically, the
divisions between Lega Nord and the other two clusters — the minor left-wing
parties and the cluster of core parties — while not quite as deep as those found
between the Flemish and Walloon groups in Belgium are in broadly the same
range, suggesting a strong divide in the socio-cultural identities of Lega Nord
supporters and supporters of other Italian parties. The other division found in
Italy, between the minor left-wing parties and the core party cluster, almost
entirely disappears in this comparative analysis, showing that there are only
minor differences between the non-political following patterns of these groups.
At the other end of the spectrum, the country with the lowest degree of
social polarisation is Ireland, where the divide between followers of Sinn Fein
and of other political parties is seen here to be very minor once we control for the
country’s overall social media conditions. The division found in Japan between
conservative and progressive parties, while deeper than the one seen in Ireland,
38
is also very minor compared to the larger divisions found in Italy and Belgium.
4 Discussion
In the cases of Ireland and Japan, the results of our analysis imply that de-
spite there being enough difference between party followers’ Twitter following
patterns to show up in our network analysis, most citizens on either side of
these divides do not seem to have constructed their non-political Twitter en-
vironment along lines matching their political polarisation. In contrast, that
is exactly what we find in Belgium and Italy, where there is a line of political
polarisation among citizens (between Flemish and Walloon groups of parties
in Belgium, and between Lega Nord and all other parties in Italy) which are
matched by dividing lines in the other non-political (or peripheral to politics)
accounts citizens choose to follow.
An especially notable aspect of this finding is that in many cases, the extent
of the social polarisation we measured did not align with the political polarisa-
tion found in these countries (seen in Tables 1, 2, 3 and 4). The most notable
outliers were Italy, whose social polarisation was much higher than the polit-
ical polarisation had suggested, and Japan, where the opposite was the case.
We also found that the types of popular media accounts that were especially
polarising different significantly among countries; while media-related accounts
were heavily polarised across the board, academics and authors were especially
polarised in Japan, and entertainers (actors, musicians, and comedians) were
strongly polarised in Italy. This implies that social polarisation differs between
countries not only in its extent but also in its nature, with the realms of so-
cial activity which people come to view as part of their political identity being
different from country to country.
These findings reveal how an exclusive focus on polarisation over political
39
matters does not allow us to make reliable inferences about how citizens and/or
elites come together (or diverge) on many other aspects of their life that, as
demonstrated in this chapter, can still be relevant in clarifying the nature of
political competition and social divisions in a country. Future efforts to explore
this phenomenon would benefit from more extensive social media data collec-
tion, involving heterogeneous countries in terms of religious backgrounds, age of
democracy, types of electoral system and so on. Based on a sufficiently extended
series of macro-political variables, it would be possible not only to replicate the
type of exploration performed here on a larger scale, but also to enquire more
in depth into the sources of the (mis)match between political and non-political
polarisation that we have observed. Our study has revealed a number of im-
portant differences just on the basis of a few case studies. In the future, the
integration of spatial empirical strategies with inferential statistics will most
likely be key to combining strictly descriptive contents with more explanatory
information on what causes the (mis)match between political and non-political
polarisation.
40
References
Altemeyer, Bob (1981). Right-wing authoritarianism. Winnipeg: University of
Manitoba Press. 352 pp.
Barber´a, Pablo (Jan. 2015). “Birds of the Same Feather Tweet Together: Bayesian
Ideal Point Estimation Using Twitter Data”. In: Political Analysis 23.1,
pp. 76–91.
Barber´a, Pablo, John T. Jost, Jonathan Nagler, Joshua A. Tucker, and Richard
Bonneau (Oct. 1, 2015). “Tweeting From Left to Right: Is Online Political
Communication More Than an Echo Chamber?” In: Psychological Science
26.10, pp. 1531–1542.
Billig, Michael and Henri Tajfel (1973). “Social categorization and similarity in
intergroup behaviour”. In: European journal of social psychology 3.1, pp. 27–
52.
Bishop, Bill and Robert G. Cushing (2008). The big sort: why the clustering
of like-minded America is tearing us apart. OCLC: ocn180851959. Boston:
Houghton Mifflin. 370 pp.
Campbell, James E. (2016). Polarized: making sense of a divided America.
Princeton: Princeton University Press. 313 pp.
Craig, Hugh and Arthur F Kinney (2009). Shakespeare, computers, and the
mystery of authorship. Cambridge University Press.
Dalton, Russell J. and Aiji Tanaka (Aug. 2007). “The Patterns of Party Polar-
ization in East Asia”. In: Journal of East Asian Studies 7.2, pp. 203–223.
Damiani, Marco (Jan. 2020). Populist Radical Left Parties in Western Europe:
Equality and Sovereignty. en. Google-Books-ID: 1zPNDwAAQBAJ. Rout-
ledge.
Downs, Anthony (1957). An economic theory of democracy. eng. Nachdr. OCLC:
255188691. Boston: Addison Wesley.
41
Freire, Andr´e (Mar. 2008). “Party Polarization and Citizens’ Left—Right Ori-
entations”. en. In: Party Politics 14.2. Publisher: SAGE Publications Ltd,
pp. 189–209.
Gaertner, Samuel L, John F Dovidio, Phyllis A Anastasio, Betty A Bachman,
and Mary C Rust (1993). “The common ingroup identity model: Recatego-
rization and the reduction of intergroup bias”. In: European review of social
psychology 4.1, pp. 1–26.
Gonz´alez, Sara (Jan. 2011). “The North/South divide in Italy and England:
Discursive construction of regional inequality”. en. In: European Urban and
Regional Studies 18.1. Publisher: SAGE Publications Ltd, pp. 62–76.
Hawkins, Kirk A. (Aug. 2009). “Is Ch´avez Populist?: Measuring Populist Dis-
course in Comparative Perspective”. en. In: Comparative Political Studies
42.8, pp. 1040–1067.
Hetherington, Marc J. and Jonathan Weiler (2018). Prius or pickup? how the
answers to four simple questions explain America’s great divide. Boston:
Houghton Mifflin Harcourt. 1 p.
Hieda, Takeshi, Masahiro Zenkyo, and Masaru Nishikawa (May 2019). “Do pop-
ulists support populism? An examination through an online survey follow-
ing the 2017 Tokyo Metropolitan Assembly election”. en. In: Party Politics,
pp. 1–12.
Hoff, Peter D., Adrian E. Raftery, and Mark S. Handcock (Dec. 1, 2002).
“Latent Space Approaches to Social Network Analysis”. In: Journal of the
American Statistical Association 97.460. Publisher: Taylor & Francis eprint:
https://doi.org/10.1198/016214502388618906, pp. 1090–1098.
Iyengar, Shanto, Gaurav Sood, and Yphtach Lelkes (Sept. 1, 2012). “Affect, Not
Ideology”. In: Public Opinion Quarterly 76.3, pp. 405–431.
42
Iyengar, Shanto and Sean J. Westwood (2015). “Fear and Loathing across Party
Lines: New Evidence on Group Polarization”. In: American Journal of Po-
litical Science 59.3, pp. 690–707.
Klein, Axel (May 2020). “Is There Left Populism in Japan? The Case of Reiwa
Shinsengumi”. In: The Asia-Pacific Journal: Japan Focus 18.10, p. 1.
Lachat, Romain (Dec. 2008). “The impact of party polarization on ideological
voting”. en. In: Electoral Studies 27.4, pp. 687–698.
McCoy, Jennifer, Tahmina Rahman, and Murat Somer (Jan. 1, 2018). “Polariza-
tion and the Global Crisis of Democracy: Common Patterns, Dynamics, and
Pernicious Consequences for Democratic Polities”. In: American Behavioral
Scientist 62.1, pp. 16–42.
Mudde, Cas (2004). “The Populist Zeitgeist”. en. In: Government and Opposi-
tion 39.4, pp. 541–563.
Musolino, Dario (June 2018). “The North-South Divide in Italy: Reality or
Perception?” In: European Spatial Research and Policy 25.1, pp. 29–53.
Pariser, Eli (May 2011). The Filter Bubble: What The Internet Is Hiding From
You. en. Penguin UK.
Pelizzo, Riccardo and Salvatore Babones (Jan. 1, 2007). “The Political Economy
of Polarized Pluralism”. In: Party Politics 13.1, pp. 53–67.
Spanje, Joost van and Claes de Vreese (Sept. 2011). “So what’s wrong with the
EU? Motivations underlying the Eurosceptic vote in the 2009 European elec-
tions”. en. In: European Union Politics 12.3. Publisher: SAGE Publications,
pp. 405–429.
Suiter, Jane (2017). “The rise of populism on the left and among independents”.
In: Populist Political Communication in Europe. Ed. by Toril Aalberg, Frank
Esser, Carsten Reinemann, Jesper Str¨omb¨ack, and Claes H. de Vreese. Rout-
43
ledge Research in Communication Studies. Publisher: Routledge. Routledge,
pp. 127–136.
Webster, Steven W. and Alan I. Abramowitz (July 1, 2017). “The Ideological
Foundations of Affective Polarization in the U.S. Electorate”. In: American
Politics Research 45.4, pp. 621–647.
World Bank (2017). GINI index (World Bank estimate) - Ireland, Italy, Bel-
gium, Japan — Data.
Yoshida, Toru (Apr. 2019). “Populism “made in Japan”: A new species?” en.
In: Asian Journal of Comparative Politics, pp. 1–12.
44
