Journal of Public Economics 206 (2022) 104578
Contents lists available at ScienceDirect
Journal of Public Economics
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 / j p u b e
Broadband Internet and social capitalA
Andrea Geracia
, Mattia Nardottob
*, Tommaso Reggianic
, Fabio Sabatinid
'European Commission, Joint Research Centre (JRC), Italy
b
KULeuven, CEPR, CESifo, Belgium
c
Cardiff University - Cardiff Business School, Masaryk University, IZA, United Kingdom
d
Sapienza University of Rome, IZA Italy
A R T I C L E I N F O
Article history:
Received 9 December 2019
Revised 22 October 2021
Accepted 8 December 2021
JEL Classification:
D91
L82
Z13
Keywords:
ICT
Broadband infrastructure
Networks
Internet
Social capital
Civic capital
A B S T R A C T
We study the impact of broadband penetration on social capital in the UK. Our empirical strategy exploits
a technological feature of the telecommunication infrastructure that generated substantial variation in
the quality of Internet access across households. The speed of a domestic connection rapidly decays with
the distance of a user's line from the network's node serving the area. Merging information on the topology
of the network with geocoded longitudinal data about individual social capital from 1997 to 2017, we
show that access to fast Internet caused a significant decline in civic and political engagement. Overall,
our results suggest that broadband penetration crowded out several dimensions of social capital.
© 2021 Elsevier B.V. All rights reserved.
1. Introduction
Social capital plays a role in many desirable outcomes such as
access to credit and loan repayment (Karlan, 2005), financial development
(Guiso et al., 2004), innovation (Knack and Keefer, 1997),
productivity (Bloom et al., 2012), mitigation of agency problems
(Costa and Kahn, 2003), and economic growth (Algan and Cahuc,
2010). However, there are signals that some social capital dimen*
We are grateful to the Editor, Maria Petrova, and three anonymous referees for
their comments and suggestions that helped us improve the paper substantially.
The manuscript also benefited from comments by Luigino Bruni, Andrea Fazio,
Milos Fisar, Karsten Gareis, Martin Guzi, Dylan Henderson, Melanie Jones, Stepan
Mikula, Rainer Michael Rilke, Matteo Rizzolli, Valentina Rotondi, Margaret
Samahita, and seminar participants at Cardiff University, CERGE-EI Prague, Masaryk
University, Queen's University Belfast, Sapienza University of Rome, Seinan Gakuin
University, University of Oslo, University of Stavanger, Technical University of
Ostrava, and University of Verona. Support from Masaryk University is gratefully
acknowledged (project MUNI/G/0985/2017).
* Corresponding author.
E-mail addresses: andrea.geraci@ec.europa.eu (A. Geraci), mattia.nardotto@kuleuven.be
(M. Nardotto), reggianit@cardiff.ac.uk (T. Reggiani), fabio.sabatini@uniromal.it
(F. Sabatini).
sions are eroding. In his bestseller Bowling Alone, Putnam (2000)
documented that a decline in civic engagement and trust began
in the United States in the 1960s, with a sharp acceleration in
the 1990s. The author suggested that domestic entertainment such
as television, video players, and computer games displaced relational
activities in individuals' leisure time. If television, a unidirectional
mass medium, can crowd out face-to-face interaction, it
stands to reason that broadband Internet, which provides ondemand
content and allows for interactive communication, might
induce an even more powerful substitution effect.
Given the pervasiveness of the Internet and the economic outcomes
of social capital, the effect of broadband penetration should
be put under scrutiny by economic research. This paper uses panel
data to compare the behavior of households before and after the
diffusion of fast Internet. To address endogeneity concerns, we
exploit the exogenous variation in the quality of the households'
Internet connections due to the technical features of the access
technology. In the UK, broadband penetration long relied on the
digital subscriber line (DSL) infrastructure. This technology allows
the high-speed transmission of data over the old copper telephone
network. However, connection to the voice grid is a necessary but
https://doi.org/10.1016/j.jpubeco.2021.104578
0047-2727/O 2021 Elsevier B.V. All rights reserved.
A Ceraci, M. Nardotto, T. Reggiani et al.
not sufficient condition for accessing fast Internet. The actual speed
of a domestic connection rapidly decays with the distance of a final
user's telephone line from the network's node serving the area, also
called local exchange (hereafter LE). When the network was
designed in the 1930s, the length of the copper wire connecting
houses to the LEs did not affect the quality of voice communications.
However, the introduction of the DSL technology unpredictably
turned distance from the LE into a critical determinant
of Internet access and quality in the 1990s. Proximity to the network's
node serving the area thus resulted in actual broadband
access, while higher distance prevented households from using fast
Internet.1
We exploit this feature of the telecommunication infrastructure
by combining unique information about the topology of
the voice network provided by Ofcom - including the geolocation
of the LEs and the city blocks served by each of them - with geocoded
individual survey data from the British Household Panel Survey
(BHPS) over 1997-2008 and the UK Household Longitudinal Study
(UKHLS) from 2009 to 2017.
First, we use our detailed map of the topology of the network to
calculate the distance of the LE serving each Lower Layer Super
Output Area (LSOA) from the centroid of the area. LSOAs are the
second-narrowest geographical areas in the UK census, comprising
on average 650 households and 1,500 inhabitants. In densely populated
metropolitan areas, they correspond to portions of city
blocks. We then match this information with the geographic coordinates
of the households surveyed in the BHPS and the UKHLS.
Since the physical distance between the LE and the premises long
affected the connection's speed, we then employ an intention-totreat
approach to assess the impact of fast Internet on different
social capital dimensions, such as participation in voluntary organizations,
political engagement, social trust, and various forms of
social interaction.
We find that, after broadband's take-up, civic and political
engagement started to decline with proximity to the node of the
network, suggesting that the use of fast Internet displaced social
capital. The effect is sizable, as a one standard deviation reduction
in the distance from the LE (equal to 1.8 kilometers), resulting in a
higher connection speed, caused a statistically significant decrease
in the likelihood of participation in civic organizations by 4.7 percent
over the period 2005-2017.
Back of the envelope calculations based on the population's distribution
across the distance from local exchanges suggest that
proximity to the LE shifted the social participation of approximately
450,000 residents. The types of organizations that Putnam
(1995) defined as supportive of collective action, such as volunteering
associations, suffered the most from broadband penetration.
We find that living one standard deviation closer to the LE
caused a statistically significant reduction in the likelihood of participation
in these organizations by 7.8 percent.
For organizations supporting redistributive goals such as political
parties (Knack and Keefer, 1997), broadband availability
caused a statistically significant reduction in the likelihood of participation
by 5.1 percent.
By contrast, we find no significant impact of broadband penetration
on cultural consumption and respondents' relationships
with friends. Overall, our results suggest that broadband Internet
displaced the time-consuming activities oriented to pursuing the
common good that Putnam et al. (1993) referred to as "bridging
social capital". However, the crowding-out effect spared the relationships
between people who know each other well, generally
referred to as "bonding social capital" (Putnam et al., 1993; Gittel
and Vidal, 1998). The panel structure of the dataset allows us to
1
Ofcom (2011).
Journal of Public Economics 206 (2022) 104578
confirm that distance from the LE is not associated with changes
in individual social capital before fast Internet's take-up.
The treatment effects do not change significantly across the two
main stages of broadband penetration we address in the empirical
analysis. This evidence suggests firstly that distance from the networks'
nodes - and thus the quality of Internet connection remained
a relevant driver of fast Internet access and use in recent
years, and secondly that, in different ways, broadband access continued
to have a strong effect on social capital when new types of
content appeared on the Internet landscape.
Our paper connects to three strands of literature. The first
broadly includes empirical research on the sources of social capital.
Several studies in this field investigated the role of historical experiences,
such as slave trade (Nunn and Wantchekon, 2011), political
independence (Guiso et al., 2016), and inherited culture (Algan
and Cahuc, 2010). We add to this research by investigating how ICT
progress can induce a more rapid, though persistent, change in the
stock of social capital. We particularly relate to contributions
assessing the response of civic engagement and trust to contingent
stimuli such as conflict (Guriev and Melnikov, 2016), corruption
(Barenjee, 2016), teaching practices (Algan et al., 2013), and regulation
(Aghion et al., 2010).
The second strand of literature includes studies empirically
testing the relationship between ICTs and political participation.
Several authors assess the impact of Internet use by exploiting discontinuities
in broadband access. Falck et al. (2014) and Gavazza
et al. (2019) provide evidence that fast Internet crowded out voter
turnout in Germany and the UK. Campante et al. (2018) consistently
find that broadband availability displaced political engagement
in Italy at the initial stages of Internet penetration.
However, the advent of social media later triggered renewed interest
in public affairs, encouraging disintermediated forms of political
participation. Enikolopov et al. (2020) show that the
penetration of VK, the most used Russian social networking site
(SNS), supported the 2011 political protests in Russia. Engagement
in SNSs is also associated with lower corruption in Russia
(Enikolopov et al., 2018) and an increased ability of U.S. congress
candidates to raise funds across their supporters (Petrova et al.,
2017). On the negative side, social media have been found to support
the spreading of political misinformation (Allcott and
Gentzkow, 2017), increase polarization (Levy, 2021), erode social
trust (Antoci et al., 2019), encourage hate crime against refugees
in Germany (Miiller and Schwartz, 2020a), and Muslims in the U.
S. (Miiller and Schwartz, 2020b), and undermine users' wellbeing
(Allcott et al., 2020).2
Despite the growing knowledge on the political impact of fast
Internet, there still is limited evidence on how broadband penetration
affects social capital. The closest paper to ours is
Bauernschuster et al. (2014), who exploit the incompatibility
between the DSL technology and the glass fiber infrastructure
installed in some areas of East Germany 3
to study the impact of
Internet use on social capital over 1999-2009. The authors find that
Internet use did not significantly affect social capital. Our finding
that broadband penetration instead displaced some social capital
dimensions in the UK allows us to catch another fragment of a larger
and complex picture. As we document in Section 5, Germany displayed
significantly higher social and civic engagement than the
UK in the early 1990s. This difference in the level of participation
reflects the different patterns of Internet use in the two countries.
We show that, in the years of broadband's take-up, UK surfers used
2
For a review of the literature on the social and political impact of social media, see
Zhuravskaya et al. (2020) and Campante et al. (2021).
3
Glass fiber was deemed to improve the quality of voice communication, but later
turned out to be incompatible with the DSL technology, thereby preventing cabled
areas from accessing fast Internet.
2
A Ceraci, M. Nardotto, T. Reggiani et al.
the Internet for commercial purposes, such as private entertainment
and the purchase of goods and services, much more than German
users. By contrast, Germans were more eager to use the Internet to
engage in community affairs and keeping in touch with others. This
descriptive evidence suggests that the baseline levels of social capital
and the most popular online activities could help explain the differential
impact of broadband penetration across countries. In
Germany, citizens eager to be involved in community affairs took
advantage of the increase in the connection speed to exploit new
tools for participation, resulting in strengthened civic engagement.
In the UK, where bridging social capital was lower than in Germany
and users mainly exploited the web for commercial purposes, fast
Internet may have diverted users from time-consuming relational
activities oriented to the common good.
More in general, our work connects to studies assessing the
impact of broadband penetration on macroeconomic outcomes,
such as employment (Hjort and Poulsen, 2019) and growth
(Czernich et al., 2011), and human behaviors such as sex crime
(Bhuller et al., 2013), sleep (Billari et al., 2018), self-image concerns
(McDool et al., 2020), enrollment in colleges (Dettling et al., 2018),
educational attainment (Faber et al., 2015), and demand for healthcare
services (Amaral-Garcia et al., 2021).
The remainder of the paper proceeds as follows. Section 2 provides
background about the broadband infrastructure in the UK.
Section 3 presents the data and our empirical strategy. In Section 4,
we present our empirical analysis of broadband Internet's effect on
social capital. Section 5 discusses our results in light of the previous
literature. Section 6 concludes.
2. The broadband infrastructure in the UK
Fast Internet access in the UK took its first steps at the end of
the 1990s, relying on an infrastructure built several decades earlier:
the telephone network. This network was designed and rolled
out in the 1930s by the former state monopolist British Telecom
(BT) to enable voice communication. The network consists of
nodes, the Local Exchanges (LEs hereafter) connected to each other
to ensure global connectivity, each serving all premises located in
the respective catchment area. Internet service providers operating
over this network have been by far the dominant supplier of broadband
services until the first decade of the 2000s (the only competitor
being the cable operator). However, providers exploiting the
old voice network remained the primary source of broadband
access also in the second decade of the 2000s when mobile operators
started to offer fast Internet services.
The voice network, which was designed to transmit the analog
signal of voice communication, can also transmit digital signals,
thus enabling all forms of digital communication and Internet
access. The main limitation of the pre-existing network lies in its
copper wires. A digital signal transmitted on a copper wire suffers
from substantial decay with the distance traveled, making the
length of the copper section of the network a crucial determinant
of local Internet access conditions.
In the 1990s, the copper wires of the network provided a lowspeed
connection to the Internet via dial-up (i.e., a modem connecting
to a service provider by dialing a telephone number).
Around 1995 the introduction of the Digital Subscriber Line technologies
(DSL) and later of the Asymmetric Digital Subscriber Line
(ADSL), which use a wider range of frequencies over the copper
line, made it possible to provide Internet access at a low cost
through the voice network. Although the first versions of ADSL
could achieve very limited speed (and did not qualify as broadband),
technological improvements allowed reaching a speed of
2Mbit/s (which qualifies as broadband) at the beginning of the
2000s'. The maximum download speed that could be reliably
Journal of Public Economics 206 (2022) 104578
provided to users increased rapidly during the first decade of the
2000s, reaching 24Mbit/s in 2008, and continued to increase during
the second decade of the 2000s with the roll out of Very
high-speed Digiral Subscriber Line technologies (VDSL) , which
enabled a theoretical maximum speed of 56Mbit/s. However, not
every Internet user could immediately enjoy the maximum connection
speed. DSL technologies required substantial network
upgrades and faced a fundamental local limitation: the so-called
last mile, i.e., the section of the network connecting premises with
the LE serving the area. As the digital signal suffers from a substantial
decay when transmitted over a copper wire, with its strength
declining more than proportionally with the distance traveled, BT
upgraded all the connections between the LEs, accounting for most
of the distance between the final user and the content provider/ISP,
with fiber optic wires. Still, the state monopolist did not replace
the connections between the LEs and the premises, which continued
to rely on the old copper infrastructure.4
This choice allowed to
reach a decent Internet speed at a relatively low investment cost but
generated local differences in the quality of access between households
depending on the length of their last mile. Having been
installed in the 1930s for different purposes than the transmission
of the digital signal, the LEs have not been located in order to minimize
the average decay. The LEs' catchment areas are irregularly
shaped, and LEs are often not located in their center. Thus, local
access conditions can vary substantially in relatively small areas.
In a 2011 report, Ofcom, the UK's communications regulator, says:
"A characteristic of ADSL broadband is that performance degrades
due to signal loss over the length of the telephone line. This characteristic
implies that the speeds available to different customers vary
significantly, with those with shorter line lengths (i.e., who live closer
to the exchange) typically able to achieve higher speeds than
those with longer line lengths. [...] We found that the average
download speed received for up to 20 Mbit/s or 24 Mbit/s ADSL packages
was 6.6 Mbit/s, and 37% of customers had average speeds of 4
Mbit/s or less" Ofcom, 2011, p. 7).5
Broadband Internet penetrated slowly in the UK between the
end of the 1990s and the early 2000s (Deshpande, 2014) for two
primary reasons. First, the average navigation speed was relatively
low due to the incomplete development of the DSL technology.
Second, the institutional setting was not favoring private investments
in the sector. The breakthrough in broadband penetration
occurred in 2005 after the implementation of the so-called "local
loop unbundling" (LLU) required by the European Union's policy
on competition in the telecommunications sector and introduced
in the UK between 2003 and 2004. LLU is the process whereby
the incumbent makes its local voice network available to other
companies, and it has been the cornerstone of the open access legislation
in European countries. Entrants are allowed to put their
equipment inside BT's exchanges to supply customers with an
upgrade of their voice lines to DSL services (Nardotto et al.,
2015). The number of Internet service providers increased rapidly,
together with the market share of LLU operators, which went from
only 2.2% at the end of 2005 to almost 40% at the end of 2009. The
4
Connecting the LEs with premises through fiber cables (so-calledfiberto the home,
FTTH, connection) allows faster navigation speed as the signal never travels on
copper, but requires to upgrade also the line between the LEs and the premise,
resulting in a much higher cost (especially in the case of a large-scale roll-out). Until
2010, FTTH connections were deployed exclusively for large companies or institutions
which had dedicated high-speed lines. After 2010, investments in FTTH remained
very limited and confined to the center of the main cities.
5
In line with the measurements by Ofcom, the data on the speed test used in
Ahlfeldt et al. (2017), (see page 604), show the strong, nonlinear, decay in speed with
distance. In particular, the first 2 km reduce speed by approximately 50%. Therefore,
adopting 2Mibt/s as the speed threshold for a broadband connection implies that
beyond 2 km to the LE, virtually all ADSL installed until 2008 would not qualify as
broadband.
3
A Ceraci, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
process substantially boosted broadband penetration in the UK, as
shown in Fig. 1, which reports the evolution of broadband penetration
between 2000 and 2018. In the five years from 1999, which
can be considered the starting year of its take-up, to 2004 broadband
penetration went from 0 to 18%, while in the following six
years it almost reached the threshold of 80%. After that, with the
market becoming almost saturated, broadband Internet penetration
continued to grow slowly but constantly, passing the threshold
of 90% in the second half of the 2010s.
DSL was not the only option to access fast Internet in our period
of analysis, although it was the most popular choice. Until 2010,
approximately 80 percent of broadband accesses in the UK relied
on DSL, making the old telephone network the primary determinant
of broadband penetration. The remaining 20 percent of broadband
accesses used the cable network, while less than 0.1 percent
relied on fiber and mobile operators.6
After 2010, mobile operators gained importance thanks to the
development of 3G and then 4G technologies that could finally
provide Internet access at a sufficient speed to be considered as
broadband.7
However, the fixed-line market remained dominated
by DSL operators.8
In light of the evolution of the Internet market, it is natural to
divide our period of analysis into three parts: i) A pre-Internet
phase until 2004, characterized by limited take-up and low
access speed; ii) A first post-Internet phase from 2005 until
2010 that saw the massive entry of ISPs and considerable
improvements in Internet speed, resulting in significant take-up
by final users; iii) A second post-Internet phase, after 2010,
when the market consolidated and broadband penetration
peaked, but the length of the last mile remained an important
determinant of the actual connection speed for domestic connections
(Ofcom, 2017).
A crucial element in our identification strategy is the length of
the last mile. As explained, despite the breakthrough in broadband
penetration, not every household connected to the voice network
could access fast Internet with the DSL technology. Because of this
technological limitation, the advent of DSL unpredictably turned
distance from the LE into a critical determinant of broadband penetration,
thereby creating an exogenous source of variation in
access to fast Internet.
3. Data and empirical strategy
In this section, we first present the data, and we then detail our
identification strategy. In a nutshell, the empirical analysis exploits
6
The cable network, initially deployed to offer cable-TV, could also be upgraded to
supply Internet access. The cable company. Virgin Media, made this conversion in
parallel to the DSL market and saw its market share declining from 29% in 2005 to
22% in 2010, mainly due to the increased quality of DSL connections over time (see
Nardotto et al., 2015 for more details).
7
Customers of mobile operators purchasing voice services could include data
services in their mobile plans since the mid-2000s. However, mobile technologies
lagged behind in terms of speed compared with fixed-line. A typical 2G connection
(by far the most diffused type of mobile connection until the end of the 2000s) could
offer a download speed of O.lMbit/s. Early versions of 3G, which started to gain
market share at the end of the 2000s, despite being a significant improvement to the
previous standard, barely passed the threshold of IMbit/s. Consistently with these
technological limitations, the yearly reports by Ofcom (see, for instance, the 2010
Ofcom Report on the Communication Market available at: https://www.ofcom.org.uk/
data/assets/pdf_file/0010/20305/uk-telecoms.pdf) listed the poor connection quality,
the data volume limitations, the network congestion in some moments of the day,
and thus the overall value for money of the service, as the main reason for consumers'
dissatisfaction with mobile Internet connections. It is only with the diffusion of later
versions of 3G (in particular the HSPA+ that started to be deployed in 2011) that the
mobile market finally took off, thanks to faster and more reliable connections,
although limitations to data volumes persisted.
8
The cable operator saw a slight decline in its market share in the fixed-line
market but remained between 15 and 20 percent throughout the decade.
s
1999 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019
Year
Fig. 1. Broadband penetration between 2000 and 2018 (source: Eurostat).
individual differences in the actual quality of Internet access to
identify the effect of broadband penetration on social capital. For
this purpose, we combine two sources of data: [I] a dataset with
detailed geographical information on the topology of the telephone
network provided by Ofcom (Office of Communications), the regulatory
and competition authority for the telecommunications
industries in the UK, which we match with [II] geocoded individual
data from the British Household Panel Survey (BHPS) over the period
1997-2009 and the UK Household Longitudinal Study (UKHLS)
over 2010-2017. The resulting dataset is a 20 years panel including
information on each survey respondent's distance from the relative
node of the telephone network.
3.0.1. Understanding Society and Harmonised BHPS
The British Household Panel Survey (BHPS) and the UK Household
Longitudinal Study (UKHLS or Understanding Society) are two
panel surveys containing detailed information on several aspects of
UK households' life. The BHPS started in 1991 based on a representative
sample of the British population (Taylor et al., 2010) including
more than 5,000 households, approximately totaling 10,000
individual interviews. It is household-based, and every adult in
the household is interviewed. The BHPS ended after 18 waves of
annual interviewing, with the 18th and last wave completed in
2008. The UKHLS panel is a direct development of the BHPS. It
started in 2009 and interviewed around 40,000 households, including
approximately 8,000 of the initial BHPS households. Although
the two surveys share many characteristics, BHPS being the predecessor
of UKHLS, they also present several differences. In 2016, the
Institute for Social and Economic Research started a project harmonizing
the content of the two surveys to support their combined
use.9
In this paper, we make use of the last available update of the
harmonised data file.1 0
To match BHPS-UKHLS data with the map
of the telephone network provided by Ofcom, we employ the Lower
Layer Super Output Areas (LSOA) Special License version of the
BHPS-UKHLS data that contains spatial LSOA references. Since former
BHPS respondents were interviewed for the first time during
the second wave of UKHLS (2010), our analysis covers the years
1991-2008 (waves 1-18) and 2010-2017 (waves 20-27). Impor-
9
For more information on the harmonization between the two surveys, see: http://
repository.essex.ac.uk/21094/1/bhps-harmonised-user-guide.pdf
1 0
University of Essex, Institute for Social and Economic Research. (2020). Understanding
Society: Waves 1-10, 2009-2019 and Harmonized BHPS: Waves 1-18,
1991-2009: Special Licence Access, Census 2001 Lower Layer Super Output Areas.
12th Edition. UK Data Service. SN: 6670, http://doi.org/10.5255/UKDA-SN-6670-12
4
A Ceraci, M. Nardotto, T. Reggiani et al.
tantly, we exclude from the analysis new UKHLS respondents who
were not part of the BHPS sample.
3.0.2. Social capital
Social capital is generally referred to as all the features of
social life, such as "networks, norms, civic engagement and
trust" that enable individuals to act together more effectively
to pursue shared objectives (Putnam, 1995). The literature has
provided so many definitions of social capital that clarifying
the dimensions of the concept has long been a research priority.
Uphoff (1999) proposed a distinction between structural and
cognitive dimensions: structural social capital refers to individuals'
behaviors and consists of social participation and civic
engagement (e.g. meetings with friends and membership in
organizations). Cognitive social capital derives from individuals'
perceptions resulting in trust, values, and beliefs that promote
pro-social behavior.
In measuring social capital, we follow Uphoff (1999) and consider
both its structural and cognitive dimensions. We use three
sets of measures of structural social capital. The first set captures
the frequency of specific forms of cultural consumption that are
usually enjoyed in company during leisure time, such as watching
movies at the cinema and attending concerts and theatre
shows, on a scale ranging from never, once a year or less, several
times a year, at least once a month to at least once a week. We
transform responses into a dichotomous variable taking value 0
if the attendance is less than once a month and 1 if at least once
a month.
The second set includes aspects of social connectedness, such as
the frequency with which respondents meet friends and talk to
neighbors, on a scale ranging from never, less than once a month,
once or twice a month, once or twice a week, to most days. Again,
we recode responses into a binary variable equal to 0 if the frequency
is less than once or twice a week and 1 otherwise.
The third set captures political and civic engagement through
dummies revealing whether respondents are members of political
parties, trade unions, professional associations, environmental
groups, and other organizations. The BHPS reports information
on membership and active participation in the form of yes/no
answers. In our empirical analysis, we consider a respondent as
participating in an organization if she is a member of it or declares
to participate in its activities. 1 1
We consider six types of organization, which we partition
into two groups based on our elaboration of the literature about
the "Olson-Putnam controversy": Olson-type organizations and
Putnam-type organizations1 2
. Following Knack and Keefer
(1997), we define participation in Olson-type organizations as
either membership or active participation in political parties, trade
unions, or professional associations. Though being different in
scope and structure, these organizations serve similar purposes,
as they represent "groups with redistributive goals" (Knack and
Keefer, 1997, p. 1273). Thus, we include all of them into our indicator
of Olson-type organizations. We also estimate the impact of
fast Internet on political parties, trade unions, and professional
associations separately in Section 5. Similarly, we define participation
in Putnam-type organizations as either membership or active
participation in environmental associations, voluntary service
groups, or scout/guides groups. These organizations were credited
1 1
Notice that not every member actively engages in the organization's activities,
and not all those reporting to take part in the activities are members. Thus, we include
both behaviors to get a comprehensive measure of civic engagement.
1 2
In addition to the seminal works of Olson (1971) and Putnam et al. (1993), see for
example Knack and Keefer (1997) and, for a review of the literature, Degli Antoni and
Grimalda (2016).
Journal of Public Economics 206 (2022) 104578
by Putnam et al. (1993) with the ability to instill habits of cooperation,
solidarity, and public-spiritedness in their members
(Putnam et al., 1993, pp. 89-90). Knack and Keefer (1997) defined
Putnam groups as those "least likely to act as distributional coalitions
but which involve social interactions that can build trust and
cooperative habits" (Knack and Keefer, 1997, p. 1273).
Our indicator of cognitive social capital captures trust towards
unknown others, also called "social trust". We measure social trust
through responses to the question: "Generally speaking, would you
say that most people can be trusted, or that you can't be too careful
in dealing with people?", developed by Rosenberg (1956), possible
answers being: depends, cannot be too careful, or most people can be
trusted. We transform responses into a binary variable that takes
value 1 if the respondent believes that most people can be trusted
and 0 otherwise.
We also draw from the BHPS and UKHLS information on respondents'
socio-demographic characteristics, including age, income,
employment status, type of occupation, the household's and the
dwelling's features, and the housing contract.
3.0.3. Broadband infrastructure
Information on the broadband infrastructure consists of a
detailed map of the topology of the telephone network provided
by Ofcom and previously used in Nardotto et al. (2015), Faber
et al. (2015), and Ahlfeldt et al. (2017). The data report the geographic
coordinates of each LE and all the 7-digits postcodes served
by the node of the network.1 3
Using this information, we can reconstruct
the exact catchment area of each LE and, thus, compute the
(linear) distance between each household's LSOA and the respective
3.1. Empirical strategy
To identify the causal effect of broadband Internet on social
capital, we exploit individual differences in the actual speed of
the connection determined by the variation in the distance
between respondents' LSOA and the LE serving their area of residence,
as explained in Section 2. We estimate an intention-totreat
effect assuming that broadband penetration resulted in
the household's access to fast Internet depending on the distance
from the LE. This approach is driven by the lack of reliable
individual-level survey data about the actual use of fast Internet.
Information on Internet access is available only for a limited
number of BHPS and UKHLS waves and does not distinguish
between fast and slow connections (see Section 3.3 for further
details and for an empirical analysis of how the topology of
the network affected the Internet take-up).
Using information on the broadband infrastructure allows us to
exploit the panel dimension of the data, as the time span is long
enough to observe the sampled individuals before and after the
introduction of broadband Internet. The longitudinal structure of
the data proves useful in two respects. First, it allows us to control
for unobserved characteristics that might be correlated with both
Internet access and our measures of social capital. Second, since
the distance from the LE should not be associated with changes
in social capital before the broadband penetration, we can perform
1 3
There are approximately 1.7 million active postcodes in the UK. On average, a
postcode covers an area with a radius of 50 meters, but it is often smaller (i.e., a
building) in urban areas.
1 4
Ideally, one would like to use the exact address of the household. However, the
BHPS and Understanding Society report, even under restricted access, only the LSOA.
Thus, we compute the distance between the (exact) coordinates of the LE and the
geographical centroid of the LSOA
5
A Ceraci, M. Nardotto, T. Reggiani et al.
a Difference-in-Differences exercise that further strengthens a causal
interpretation of results.
In our main specification, we classify our panel waves into two
distinct periods according to the stage of broadband Internet diffusion.1
5
The Pre-Broadband period corresponds to the BHPS waves
between 1991 and 2004 (waves 1-14), referring to years in which
broadband access was minimal among British households (overall
broadband penetration in 2004 was lower than 20% and mainly confined
to the more affluent parts of large cities). The Post-Broadband
period (labeled Post in equations and tables) covers the years from
2005 onward, witnessing broadband's take-up in the UK. In our setting,
this period spans from 2005 to 2017, encompassing all the harmonized
waves of the BHPS and UKHLS surveys in which our
outcome variables were available (waves 15-27).
yk = yDistancek x Postt + /3Xk + MSOAk x r + Wavet + j/(
+ £« (1)
In our second specification, we unpack the Post-Broadband period
into two different sub-periods: Post-Broadband(l) - labeled
Post(\) in equations and tables -, which ranges from 2005 to
2008 (covered by the BHPS waves 15-18), and captures the stage
of rapid diffusion of broadband Internet in the UK; PostBroadband^)
- labeled Post(2) in equations and tables - which
ranges from 2010 to 2017 (covered by the UKHLS harmonized
waves 20-27). In these years, broadband coverage approximately
reached 92 percent in England and between 85 and 89 percent in
the rest of the UK, with distance remaining a significant determinant
of fast Internet access (see Ofcom, 2017).1 6
In Section 4.2,
we exploit UKHLS data to document that the time spent in online
activities such as gaming and surfing social media significantly
decreases with distance in the Post(2) period (see Table A.4 in
the Appendix).
Given that our outcomes are binary variables, we estimate linear
probability models for the effect of the distance from the LE,
determining the quality of Internet access (lower distance —> faster
Internet access —> higher exposure to the treatment), on our
measures of social capital. Our regression model is reported in (1):
yk = y^Distancek x Post(Y)c + y2Distancek x Post(2)c + flXk+
+MSOAk xt + Wavec + Y]{ + ek
(2)
In the above equation, yit is the outcome of interest for person i in
year t. Distance,t x Postc is the (reversed) treatment intensity for
individual i in year t. This variable is the product of Distance^, which
is the average distance from the LE (expressed in kilometers) serving
the LSOA where individual i lives in year t, and Postt, a binary
variable set to 1 if year t is in the Post-Broadband period, and 0
otherwise. X,t is a set of observed time-varying characteristics:
age, log of real annual household income, indicators for (i) occupation
type, (ii) occupational status, (iii) type of housing tenure, (iv)
type of household, (v) education level. Since respondents might
change address between waves, the vector Xit also includes an indicator
variable for change of address1 7
, and the variable Distancek.
We also include a set of monthly indicators for the month in which
the interview took place. ^, is the individual's fixed effect, which
allows us to control for time-invariant unobserved characteristics
that might be systematically associated with the propensity for
1 5
Fig. A.4 in the Appendix reports the timeline of the data used in the empirical
analysis and the division in periods.
1 6
As mentioned in the data description, 2009 was a transition year for former BHPS
respondents, who entered the new UKHLS sample only in the second wave of 2010.
1 7
The survey reports whether the respondent has a different address than the one
of the previous wave. We include this variable because moving to a new neighborhood
can have a direct effect on our measures of social capital.
journal of Public Economics 206 (2022) 104578
Internet use and social capital.1 8
As in standard Diff-in-Diffs, causal
interpretation of parameter y in Eq. (1) requires parallel trends in
pre-treatment outcomes. In our setup, this requires differences in
distance from the LE not being associated with differential trends
in the considered outcome in the Pre-Broadband period, conditional
on the considered individual characteristics, time, and individual
fixed effects. This assumption might be valid if the average distance
from the LE proxies other time-varying characteristics of the local
area. To address this concern, we include MSOA-specific linear
trends, MSOAk x t, in all our main regressions.19
In the Appendix, we also provide evidence suggesting that deviations
from the parallel trends assumption are not likely in our
context. We conduct an event study analysis in the spirit of
Autor (2003). The term yDistanceit x Postt is replaced by
Y^tytDistancek x Wavet, a set of interactions between our treatment
variable and wave dummies for all years except the first in
which information about yit is available.2 0
Fig. A.l in the Appendix
contains plots of the estimated coefficients yt for each of the dependent
variables considered. The graphs suggest that our continuous
treatment variable is not associated with statistically significant patterns
in the dependent variables in the Pre-Broadband period. This
result is confirmed by a battery of tests of joint significance of the
coefficients in the Pre-Broadband period, which do not reject the null
hypothesis of 0 (p-values are reported in the Appendix together with
the graphs).
Finally, we explore the possibility to follow the identification
strategy of Gavazza et al. (2019), who exploit exogenous weather
conditions to instrument for broadband take-up, as rainfall affects
investments, Internet quality and take-up. However, the sample
size of the BHPS-UKHLS, and the noise in the answers to the questions
related to broadband, make this exercise not conclusive,
although results are in line with our main findings (see Table A.6
in the Appendix).
3.2. Summary statistics
Tables 1 and A l (the latter in the Appendix) report the descriptive
statistics of the dependent and control variables, respectively.
To compute the summary statistics, we consider each individual/
1 8
In the Appendix, we report the result of a related exercise where we split the
households into 2 groups - Low - Distance and High - Distance - based on the
distance between the LSOA of residence and the LE. We set a threshold of 3.1 km, and
we assign households at a distance of 3.1 km or less to the Low - Distance group. We
then estimate the model in (3):
yk = '^ylLow - Distancek x Wavec + pXk + MSOAk x t
t
+ Wavet + r\i + ek (3)
where we control for the same set of controls as in (2), thus including
time FE, individual FE, and MSOA linear trends. Due to the substantial
attrition at the end of the sample and the split of the
households into two groups, we exclude the last wave of the data
(i.e. wave 27 corresponding to year 2017) from the sample. We plot
the estimated coefficients, together with confidence intervals (standard
errors are clustered at the level of the LSOA), in Fig. A.5 in the
Appendix. Results are discussed in Section 4.2 together with the
results of our main identification strategy.
1 9
The MSOA is the next narrowest geographical areas in the UK census after the
LSOA While LSOAs have on average 1500 households and correspond to blocks in
urban areas, MSOAs have on average 6000 households, corresponding to neighborhoods
in urban areas. On average, an MSOA contains four LSOAs. An example is in
Fig. A.3, which reports a map of the LSOAs and of the MSOAs in Colchester.
2 0
The model specification is identical to the one in Eq. (1) where the interaction
term Distance^ x Postt is replaced by a full set of interaction terms between wave
dummies and the variable Distance, i.e., ^2tytDistancek
x
Wavet. The set of controls is
identical to the one used in the main regressions and includes time FE, individual FE,
and MSOA linear trends. Standard errors are clustered at the level of the LSOA
6
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Table 1
Descriptive statistics.
Variable Mean St.Dev. Min Max Obs
Any organization 0.28 0.45 0.00 1.00 138054
Olson organizations 0.22 0.41 0.00 1.00 138054
- Political parties 0.03 0.16 0.00 1.00 138054
- Trade unions 0.15 0.36 0.00 1.00 138054
- Professional organizations 0.08 0.27 0.00 1.00 119344
Putnam organizations 0.09 0.29 0.00 1.00 138054
- Environmental organizations 0.03 0.18 0.00 1.00 138054
- Voluntary organizations 0.05 0.22 0.00 1.00 138054
- Scout organizations 0.02 0.13 0.00 1.00 119344
Cinema attendance 0.15 0.36 0.00 1.00 78592
Theater and concerts attendance 0.05 0.22 0.00 1.00 78606
Talks to neighbors 0.39 0.49 0.00 1.00 142268
Meets people 0.46 0.50 0.00 1.00 142264
Most people can be trusted 0.38 0.48 0.00 1.00 68300
Changed neighborhood 0.19 0.39 0.00 1.00 260403
Broadband internet 0.75 0.43 0.00 1.00 82358
Daily use of internet 0.57 0.50 0.00 1.00 73081
Distance house-LE (km) 2.43 1.84 0.01 28.98 260403
Note: the number of observations depends on the BHPS and the UKHLS waves considered.
wave observation that we use in our empirical analysis. Since not
all the questions related to our outcomes of interest are available
for all waves, the sample size differs across variables.2 1
Table 1 reports the summary statistics of the indicators
employed as dependent variables. In our sample, which is representative
of the UK's population, 15% of people regularly go to
the cinema, and 5% attend concerts and theatre shows. A large
majority declares meeting friends and regularly talking with
neighbors, while approximately 40% believe that most people can
be trusted. Less than 30% of the population is a member or participates
in the organizations we consider, with Olson-type organizations
attracting more participation (22%) than Putnam-type ones
(9%). Finally, the average household lives 2.4 km from the LE providing
Internet access, with a significant variation, the standard
deviation being equal to 1.8 km.
Table A . l in the Appendix reports summary statistics of the
socio-demographic information collected in the BHPS and the
UKHLS that we employ in our empirical analysis.2 2
3.3. Effect of distance on Internet take-up
Our identification strategy relies on the fact that, in the years
we consider, the distance between the geographical location of
the households and the respective LE crucially affected the quality
of broadband access. The use of technological factors as a source of
exogenous variation in Internet access conditions is grounded in a
large body of evidence provided by public institutions and the previous
literature. Technical reports produced by the industry and
the UK's communications regulator emphasize the role of distance
as a crucial determinant of connection speed (Ofcom, 2010; Ofcom,
2011; Ofcom, 2012; Ofcom, 2013). Recent studies also exploit factors
affecting the quality of Internet access (such as the distance of
the dwellings from the LE) to identify the causal effect of broadband
penetration on a range of hypothetical outcomes (see, for
instance, Falck et al., 2014; Miner, 2015; Campante et al., 2018;
Billari et al., 2018; Amaral-Garcia et al., 2021; McDool et al., 2020).
Most of those studies took the effect of the distance between
the premises and the LEs on fast Internet take-up for granted,
2 1
Fig. A.2 in the Appendix describes the availability of information about each of the
considered outcomes in the BHPS waves
2 2
Notice that these variables are a subset of the socio-demographic information
collected in the BHPS and the UKLS. We focus on key demographics as estimates
account for individual fixed effects.
although, due to the lack of suitable data, they were not able to test
it empirically (e.g. Ahlfeldt et al., 2017; Falck et al., 2014). In this
section, we use self-reported information on individual Internet
access collected in the BHPS and UKHLS to provide evidence that
distance from the LE indeed affected broadband access in the UK.
To this purpose, we use a question on broadband Internet
access included in some waves of the BHPS. Respondents were
asked whether they had a fixed-line broadband connection at
home at the time of the interview and, in the affirmative case,
how much time they spent online on average per day. On the
one hand, answers to these questions are worth examining
because they can provide evidence on the effect of distance on
take-up. On the other hand, they have certain limitations, which
discourage exploiting this information in the main empirical analysis.
The first limitation is that this variable is likely to be subject to
measurement error. Most users hardly knew whether a connection
could be defined as broadband and what the broadband speed/
technology actually was, with some of them giving a positive
answer even though they had a slow DSL connection (such as an
ISDN) or an only in theory broadband connection that suffered
from substantial decay due to the distance from the LE.2 3
In addition,
respondents might give a positive answer even when the connection
was not at home (which is crucial in the allocation of leisure
time) but, for example, at the workplace.
The second limitation relates to the waves in which this information
was collected and the sample of households interviewed,
which is not entirely consistent with the other samples employed
in the rest of the empirical analysis.2 4
2 3
In our period of analysis, broadband packages were typically advertised referring
to their theoretical maximum speed, while providers omitted information about the
actual or average speed. As a result, consumers largely ignored the fact that the actual
speed depended on the distance from LEs. According to Ofcom reports focusing on
consumer protection issues, this has been a source of significant confusion for UK
Internet users (Ofcom, 2006; Ofcom, 2009). The remarkable 60% gap between
advertised and actual speed registered in the UK was far above the EU average of 40%
(EuropeanCommission, 2012). Still in 2016, an Ofcom research found that even
"business consumers -particularly small or medium-sized enterprises- are confused
about how the actual speed of their broadband service compares to the headline
maximum speed used in advertising" (Ofcom, 2016; HouseCommons, 2017). For these
reasons, consumer self-reported survey data on Internet connections are considered
not wholly reliable (most likely upward biased) and therefore only weakly informative
(OECD, 2008).
2 4
As reported in Fig. A.2 in the Appendix, questions on broadband Internet access
started to be included only in wave 16 of the BHPS.
7
A Ceraci, M. Nardotto, T. Reggiani et al.
Although these limitations prevented us from using information
on individual broadband access in a TSLS strategy, it is worth
investigating the link between distance and Internet access, keeping
in mind that the following results are not fully conclusive. The
empirical model we employ, reported in Eq. (4), is similar to (1)
that we employ in the main empirical analysis.
Internet^ = yDistancei + pXit + Wavec + r\i + e,t (4)
Here, the outcome Internet^ is in one case the presence of a broadband
connection at home, and in the other case the time spent
online. Distancej is the distance between the household and the
respective LE, and, as in previous models, X,t is a set of timevarying
controls including income, household type (categorized),
employment status, type of occupation, age and tenure; ^, is the
individual's random effect; and e,t is the error term.
The main difference with respect to model (1) is that we do not
employ a fixed-effect estimator but a random-effect estimator
(with controls) because our main variable of interest, the distance
between the respondent's telephone line and the LE, is constant
over time (it only changes for those who move) and the Internet
questions were asked only starting from the last waves of the
BHPS, i.e., only in the Post-Internet period.
A clear negative relationship between the distance and broadband
access emerges from the data, and it holds both for the presence
of a broadband connection at home (the extensive margin)
and for the time spent online (the intensive margin). We illustrate
this negative relationship in Fig. 2, which results from the estimation
of our model, and a straightforward application of the FrischWaugh-Lovell
theorem (see, for instance Davidson et al., 2004),
which enables us to partial-out the role of socio-demographics
when we examine the relation between take-up and distance.
We split regressors into two groups: the geographical distance
between the households and the respective LEs, and the sociodemographic
control variables. We report on the y-axis the residuals
of model 4 where we employ as independent variables only
the socio-demographic control variables (i.e., we exclude the Distance),
and on the x-axis the residuals of a regression of Distance
on the socio-demographic control variables. Residuals are grouped
in 100 bins for which we take the average value.
As explained, Fig. 2 is strictly related to the main model in Eq.
(4), whose estimates are reported in Table 2.
Column (1) of the table shows the estimated effect of distance
between the house and the respective LE on the adoption of a
broadband connection. As expected, the coefficient is negative
and statically significant, thus indicating that distance influences
Internet access (the extensive margin). Furthermore, column (2) of
the table reports the estimated impact of distance on the time
spent online (the intensive margin), showing a strong, negative
effect of distance. Thus, we can conclude that the distance between
dwellings and the LE serving the area was crucial for broadband
take-up and the time users spent online. Distance exerted its effect
by affecting Internet providers' running costs, the quality and reliability
of the connection, and possibly its very availability since the
connection speed could be so low beyond 4.2 km that broadband
connections could not be available until 2008.
4. The effect of fast Internet on social capital
In this section, we present the results of the estimations of
model (1). We start by analyzing how the quality of Internet access
(increasing with proximity to the LE) affects two forms of cultural
consumption that people usually enjoy in company (Section 4.1).
We then present results regarding participation in organizations
(4.2), trust, and some forms of offline interaction (4.3). Finally,
Journal of Public Economics 206 (2022) 104578
we discuss our findings in light of the previous literature in
Section 5.
4.1. Fast Internet and cultural consumption
As discussed in Section 3, the BHPS and UKHLS neither provide a
precise measure of the time spent on leisure activities nor report
detailed information on the composition of individuals' leisure
time. We thus focus on two forms of cultural consumption that
usually entail social interaction: watching movies at the cinema
and attending concerts or theater shows. Our dependent variable
is a dichotomous indicator taking value 0 if the attendance is less
than once a month and 1 if at least once a month.
Since these outcome variables are only available in the BHPS,
we cannot perform our main specification based on the interaction
term between distance and the Post dummy, covering years from
2005 to 2017. Therefore, we only rely on our second specification
and interact the distance between respondents' dwelling and the
network's node serving the area with a dummy for the period
2005-2008. Results are illustrated in Table 3, where column (1)
refers to cinema attendance and column (2) to the attendance of
concerts or theater shows.
The sign of the main coefficient of interest, Distance x Post,
which captures the quality of Internet access, is positive, and its
size implies that one standard deviation increase in the distance
from the LE (equal to 1.8 km), resulting in a slower connection, is
associated with an increase in the likelihood of attending cinema
shows of 3.31 percent.2 5
However, the estimated coefficient is not
statistically significant. Thus our evidence does not support the
claim that fast Internet crowded out cultural consumption in the
UK, in line with the findings of Bauernschuster et al. (2014) for Germany
over the period 2007-2009. The lack of a statistically significant
relationship may be due to a double-sided impact of the
Internet on offline cultural activities. On the one hand, the better
quality of the Internet access may encourage users to replace the
attendance of cinema, theater, and music shows with online activities.
However, the availability of richer and faster information on
the offline supply of cultural events may well encourage consumption,
compensating for the potential substitution effect.
4.2. Fast Internet and civic engagement
We now study the effect of broadband access on civic engagement,
measured as participation in voluntary organizations. Since
its introduction in the pioneering work of Putnam et al. (1993),
membership in organizations is commonly considered as one of
the most reliable indicators of social capital because it captures
individuals' interest in public affairs and their propensity for contributing
to the public good (see for example Knack and Keefer,
1997; Guiso et al., 2004; Guiso et al., 2016). We define participation
as formal membership in the organization or the active
engagement in its initiatives. We include both behaviors to get a
more comprehensive measure of civic engagement, as not every
member declares having being involved in the organization's activities,
and not all those declaring to have taken part in the activities
are members. As explained in Section 4, we distinguish between
Olson- and Putnam-type organizations in line with the social capital
literature (e.g. Knack and Keefer, 1997). Our dependent variables
are dichotomous indicators taking value 1 if the respondent
is a member of the organization or declares to participate in its
activities. Tables 4-6 report estimation results. In Table 4, columns
(1) and (2) report the coefficient when the outcome variable is participation
in any organization with no distinction. Columns (3) and
The percent effect is computed as:
8
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Fig. 2. Internet access and distance between the house and the LE. Both panels report on the x-axis the residuals of a regression of the distance between the house and the
respective LE on the socio-demographic control variables. The y-axis of the left panel reports the residual of a regression for the presence of broadband at home on the sociodemographic
control variables. The y-axis of the right panel reports the residual of a regression for the time spent online on the socio-demographic control variables.
Table 2
Effect of distance from the LE on fast Internet take-up.
Dep. Variables: Broadband internet
(1)
Daily use of internet
(2)
Distance -0.50*" -0.58*"
(0.14) (0.13)
Controls w w
Time FEs \S \S
Individual REs \S \S
Mean of Dep.Var. 73.67 55.17
Observations 99034 87715
R2
0.33 0.28
Num. PIDs 17447 15938
Dependent variables in (1) and (2) are indicators for having a broadband internet
connection at home and the daily use of the Internet, respectively. Dependent
variables are re-scaled such that 1 = 100. The variable Distance is the average distance
from the LE in the respondent's LSOA, measured in km. The sample includes
all respondents in the BHPS waves for which the relevant information was collected
and available (waves 16-17-18 in column 1 and 18 in column 2). Controls included
in each regression: wave indicators, average distance from the LE in the LSOA, age,
log of real annual household income, indicators for (i) occupation type, (ii) occupational
status, (iii) type of housing tenure, (iv) type of household, (v) education
level, (vi) change of address from previous year, (vii) a set of monthly indicator
variables for the month of the interview. Standard errors in parentheses are clustered
at the level of the LSOA. Standard errors in parentheses are clustered at the
level of the LSOA. *, **, and *** denote significance at the 10, 5, and 1 percent level,
respectively.
(4) refer to Olson-type organizations, and columns (5) and (6) refer
to Putnam-type organizations.
Our results suggest that access to fast Internet (shorter distance
from the LE) significantly and strongly reduces civic engagement.
When we consider all the organizations without distinction, the
estimated effect reported in column (1) implies that a reduction
of one standard deviation in the distance from the LE (equal to
1.8 km), resulting in a higher connection speed, causes a reduction
in the likelihood of participation of 4.7 percent over the period
2005-2017 (Post). This result is also in line with the evidence
reported in Fig. A.5 in the Appendix, which plots the coefficients
of an alternative model, discussed in footnote 18 of Section 3.1.
Under this approach, we split the households into two categories
- Low Distance and High Distance - based on the distance between
the LSOA of residence and the LE, and we estimate the model in (3).
Our findings confirm that the effect of fast Internet on participation
started to appear in 2005 when households close to the respective
LE went online.
Table 3
Effect of fast Internet on cultural consumption.
Dep. Variables: Go to cinema
(1)
Go to concerts or to theater
(2)
Distance x Post(l) 0.27 0.03
(0.17) (0.13)
Controls w w
Time FEs \S \S
Individual FEs \S \S
MSOA linear trend \S \S
Mean of Dep.Var. 14.10 5.02
Observations 73441 73459
R2
0.61 0.55
Num. PIDs 16071 16073
Dependent variables in (1) and (2) are indicators for cinema attendance at least
once a month, and the attendance of concerts and theatre shows at least once a
month, respectively. Dependent variables are re-scaled such that 1 = 100. The
variable Distance is the average distance from the LE in the respondent's LSOA,
measured in km. Post(l) takes value 1 for waves from 15 (2005) to 18 (2008),
representing the post-internet period captured by the BHPS (the variable has not
been collected in UKHLS). The sample includes all respondents in the BHPS waves
for which the relevant information was collected and available. Controls included in
each regression: wave indicators, average distance from the LE in the LSOA, age, log
of real annual household income, indicators for (i) occupation type, (ii) occupational
status, (iii) type of housing tenure, (iv) type of household, (v) education level, (vi)
change of address from previous year, (vii) a set of monthly indicator variables for
the month of the interview. Standard errors in parentheses are clustered at the level
of the LSOA. Standard errors in parentheses are clustered at the level of the LSOA. *,
**, and *** denote significance at the 10, 5, and 1 percent level, respectively.
To provide a more meaningful estimate of the effect size, we
calculate the population's distribution across 100 meters intervals
of distance from the LE throughout the UK. This information allows
us to estimate that the increase in connection speed reduced the
likelihood of participation of 447,442 residents. This negative
effect also holds after splitting the period of analysis into the two
sub-periods encompassing the first stage of rapid broadband diffusion
over 2005-2008 (Post(l), covered by BHPS data) and the following
completion of fast Internet penetration between 2010 and
2017 (Post(2), covered by UKHLS data). The estimated coefficients
in Column (2) imply that one standard deviation decrease in the
distance from the LE, resulting in a faster connection, causes a
reduction in the likelihood of participation by 4.9 percent over
2005-2008 - Post(l) - and 4.1 over 2010-2017 - Post(2) -.
In columns (3)-(4) and (5)-(6), we report results for
Olson- and Putnam-type organizations separately. In our main
9
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Table 4
Effect of fast Internet on civic engagement.
Dep. Variables: Any organization Olson organizations Putnamorganizations
(1) (2) (3) (4) (5) (6)
Distance x Post 0.71***
(0.19)
0.61***
(0.18)
0.38"
(0.16)
Distance x Post(l) 0.74***
(0.20)
0.58***
(0.19)
0.45"
(0.18)
Distance x Post(2) 0.62**
(0.30)
0.72***
(0.26)
0.11
(0.21)
Controls j-* w j-* w j-* w
Time FEs j-* \S j-* \S j-* \S
Individual FEs j-* \S j-* \S j-* \S
MSOA linear trend j-* \S j-* \S j-* \S
1-i = 72 (P-val) 0.70 0.60 0.14
Mean of Dep.Var. 28.17 28.17 22.02 22.02 9.31 9.31
Observations 131085 131085 131085 131085 131085 131085
R2
0.59 0.59 0.62 0.62 0.51 0.51
Num. PIDs 19953 19953 19953 19953 19953 19953
Dependent variables in (l)-(2), (3)-(4), and (5)-(6) are indicators capturing membership or participation in the activities of, respectively, any organization, Olson organizations,
and Putnam organizations. Dependent variables are re-scaled such that 1 = 100. The variable Distance is the average distance from the LE in the respondent's LSOA,
measured in km. The indicator variable Post takes value 1 for wave 15 and after. Post(l) takes value 1 for waves from 15 (2005) to 18 (2008). The indicator variable Post(2)
takes value 1 for waves from 19 (2010) to 26 (2017). The sample includes all respondents in the BHPS and UKHLS waves for which the relevant information was collected and
available. Controls included in each regression: wave indicators, average distance from the LE in the LSOA, age, log of real annual household income, indicators for (i)
occupation type, (ii) occupational status, (iii) type of housing tenure, (iv) type of household, (v) education level, (vi) change of address from the previous year, (vii) a set of
monthly indicator variables for the month of the interview. Standard errors in parentheses are clustered at the level of the LSOA. Standard errors in parentheses are clustered
at the level of the LSOA. *, **, and *** denote significance at the 10, 5, and 1 percent level, respectively.
Table 5
Effect of fast Internet on Olson organizations.
Dep. Variables: Political parties Trade unions Professional organizations
(1) (2) (3) (4) (5) (6)
Distance x Post 0.31***
(0.09)
0.29*
(0.15)
0.14
(0.11)
Distance x Post(l) 0.29***
(0.09)
0.30*
(0.16)
0.13
(0.11)
Distance x Post(2) 0.38***
(0.12)
0.24
(0.22)
0.20
(0.19)
Controls j-* j-* j-* w j-* V
Time FEs j-* j-* j-* \S j-* \S
Individual FEs j-* j-* j-* \S j-* \S
MSOA linear trend j-* j-* j-* \S j-* \S
1-i = 72 (P-val) 0.31 0.78 0.66
Mean of Dep.Var. 2.76 2.76 15.37 15.37 7.79 7.79
Observations 131085 131085 131085 131085 113038 113038
R2
0.63 0.63 0.64 0.64 0.59 0.59
Num. PIDs 19953 19953 19953 19953 18780 18780
Dependent variables in (1 )-(2), (3)-(4), and (5)-(6) are indicators for membership or participation in the activities of, respectively, political parties, trade unions, professional
organizations. The main sample includes all respondents in the BHPS and Understanding Society waves for which the relevant information was collected and available.
Dependent variables are re-scaled such that 1 = 100. The variable Distance is the average distance from the LE in the respondent's LSOA, measured in km. The indicator
variable Post takes value 1 for wave 15 (year 2005) and after. The indicator variable Post(l) takes value 1 for waves from 15 (2005) to 18 (2008). The indicator variable Post(2)
takes value 1 for waves from 19 (2010) to 26 (2017). Controls included in each regression: wave indicators, average distance from the LE in the LSOA, age, log of real annual
household income, indicators for (i) occupation type, (ii) occupational status, (iii) type of housing tenure, (iv) type of household, (v) education level, (vi) change of address
from previous year, (vii) a set of monthly indicator variables for the month of the interview. Standard errors in parentheses are clustered at the level of the LSOA. *, **, and ***
denote significance at the 10, 5, and 1 percent level, respectively.
specification with the Post-Broadband period covering the years
from 2005 onward, the estimated coefficients are positive and
highly statistically significant for both types of organization, suggesting
that participation decreases with connection speed
(shorter distance from the LE). In the case of Olson-type organizations,
the estimates imply that a one-standard deviation
decrease in the distance from the LE, resulting in a faster connection,
reduces the likelihood of participation by 5.1 percent.
The effect amounts to 4.9 percent in the Post(l) period covering
the initial stage of broadband penetration and is slightly bigger
in magnitude (6.0 percent) in the Post(2) period covering
2010-2017. The estimated effect is larger for Putnam-type organizations
and amounts to 7.8 percent, though slightly less statistically
significant. Estimates in column (4) are entirely consistent
with the main specification. For Putnam-type organizations, estimates
in column (6) suggest that the overall impact of broadband
penetration observed over 2005-2017 is mainly driven by
the years of rapid broadband penetration between 2005 and
2008 (the Post(l) period). In this period, a one standard deviation
reduction in the distance from the LE, resulting in a faster
connection, causes a 9.2 percent reduction in the likelihood of
participation in Putnam organizations. In the following years
(the Post(2) period from 2010 to 2017), the effect of proximity
to the LE is still negative, though losing statistical significance.
In Tables 5 and 6, we report disaggregated results for specific
types of Olson and Putnam organizations. We first focus on three
forms of Olson organizations in Table 5: political parties, trade
unions, and professional associations. Table 6 reports results
about three forms of Putnam organization: environmental, voluntary
service, and scout associations.
10
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Table 6
Effect of fast Internet on Putnam organizations.
Dep. Variables: Environmental organizations Voluntary organizations Scout organizations
(1) (2) (3) (4) (5) (6)
Distance x Post 0.07
(0.10)
0.28"
(0.13)
0.09
(0.07)
Distance x Post(l) 0.11
(0.10)
0.32"
(0.14)
0.08
(0.08)
Distance x Post(2) -0.09
(0.16)
0.12
(0.18)
0.11
(0.10)
Controls j-* j-* w j-* w
Time FEs j-* j-* \S j-* \S
Individual FEs j-* j-* \S j-* \S
MSOA linear trend j-* j-* \S j-* \S
1-i = 72 (P-val) 0.13 0.32 0.83
Mean of Dep.Var. 3.41 3.41 5.18 5.18 1.89 1.89
Observations 131085 131085 131085 131085 113038 113038
R2
0.55 0.55 0.45 0.45 0.51 0.51
Num. PIDs 19953 19953 19953 19953 18780 18780
Dependent variables in (l)-(2), (3)-(4), and (5)-(6) are indicator variables for membership or participation in the activities of, respectively, environmental organizations,
voluntary service organizations, and scout organizations. Dependent variables are re-scaled such that 1 = 100. The variable Distance is the average distance from the LE in the
respondent's LSOA measured in km. The indicator variable Post takes value 1 for wave 15 (year 2005) and after. The indicator variable Post(l) takes value 1 for waves from 15
(2005) to 18 (2008). The indicator variable Post(2) takes value 1 for waves from 19 (2010) to 26 (2017). Controls included in each regression: wave indicators, average
distance from the LE in the LSOA age, log of real annual household income, indicators for (i) occupation type, (ii) occupational status, (iii) type of housing tenure, (iv) type of
household, (v) education level, (vi) change of address from the previous year, (vii) a set of monthly indicator variables for the month of the interview. Standard errors in
parentheses are clustered at the level of the LSOA. Standard errors in parentheses are clustered at the level of the LSOA. *, " , and * " denote significance at the 10, 5, and 1
percent level, respectively.
Column (1) of Table 5 shows that the exposure to broadband
Internet has a sizable and statistically significant impact on participation
in political parties. The estimated coefficient implies that a
reduction of one standard deviation in the distance from the LE,
resulting in a faster connection, reduces the likelihood of participation
by 19 percent. In the years of fast broadband penetration from
2005 to 2008 (the Post(l) period in our analysis), a one standard
deviation reduction in the distance from the LE, resulting in a faster
connection, leads to a 17.8 percent decrease in political participation.
In the Post(2) period, the effect is still significant and slightly
bigger in magnitude (23.3 percent). Column (3) shows that the
negative impact of fast Internet also affects participation in trade
unions, even though to a lesser extent. A reduction of one standard
deviation in the distance from the LE, resulting in a faster connection,
causes a 3.6 percent reduction in the likelihood of participation
in trade unions over the PostBroabdband period (from 2005
onward). The effect is significant at the 10 percent level. Estimates
in column (4) show that the effect is driven by the years of rapid
broadband penetration (2005-2008, corresponding to the Post(l)
period). In the following years, the effect of proximity is still negative,
though losing statistical significance. The different extent
to which participation in the two kinds of organizations declined
with access to fast Internet may be connected with the strength
of the distributional action they exert. While political parties only
indirectly safeguard their supporters' particular interests, trade
unions have a stronger and more explicit commitment to advocate
for distributive measures in favor of their members. Thus, the
incentive to participate in trade unions may have been more
weakly affected by the displacement effect of fast Internet.
On the other hand, access to fast Internet does not affect participation
in professional organizations (columns 5 and 6), which is
self-interestedly aimed at pursuing particular goals and mostly
takes place in the context of individuals' professional life, instead
of during their leisure time.
As it concerns Putnam-type organizations, Table 6 shows that
broadband penetration substantially displaced participation in
service-oriented associations providing social care through volunteering
activities. In this case, broadband availability causes a statistically
significant reduction of participation by 10.3 percent.
Coefficients in Column 4 show that the effect of proximity to the
LE is negative in both periods, though losing statistical significance
in Post(2).
Columns (l)-(2) and (5)-(6) show no statistically significant
effect of broadband penetration on environmental and scout organizations.
However, coefficients are oriented in the expected direction,
thus indicating a general tendency towards a decrease in
participation with faster Internet access.
Overall, the pattern of results does not meaningfully change
between the two periods. Columns 2, 4, and 6 of Tables 4, 5 show
that the coefficients of the treatment interaction with Post(l) and
Post(2) are not significantly different. This finding has two major
implications. First, proximity to the LE likely continued to play a
critical role in access to fast Internet in more recent years, when
broadband coverage extended to approximately 90 percent of the
British population. Ofcom, 2017 explains that broadband speed
was still low in many areas in 2017, as the last mile still primarily
relied on copper wires, and full-fiber investments for the supply of
FTTH connections were only starting to happen.
To delve deeper into the persistent role of distance over the second
stage of broadband penetration (Post(2)), we exploit UKHLS
data to explore the correlation between proximity to the LE and
aspects of Internet use that require a fast connection, such as
online gaming and the active use of social media. We derive a
dummy for social media use from the question "How many hours
do you spend chatting or interacting with friends through social
websites on a typical week day, that is Monday to Friday?", available
in three waves of the UKHLS (2011-12, 2014-15, and 2017-
18), with possible answers being none, less than one hour, between
one and three hours, between four and six hours, and seven hours or
more. The social media use dummy is equal to 1 if respondents
declare using them one hour or more per day. We derive the online
gaming dummy from the question "Do you ever play multi-player
online games?", available in four waves (2010-11, 2012-13, 2014-
15, and 2016-17). Both dummies are re-scaled so that 1 = 100 as in
the main analysis. Table A.4 in the Appendix shows that online
gaming and the active use of social media significantly decrease
with distance from the LE.
The second implication of the consistency of results across the
two periods lies in the persistence of the treatment effects despite
the advent of social media likely introduced significant changes in
i i
A Ceraci, M. Nardotto, T. Reggiani et al.
the way people use the Internet in Post(2) years. Even if the most
popular social networking sites launched in the second half of
the 2000s, social media usage substantially took up around the second
half of the 2010s in the UK, with the share of adults aged 16 or
more that subscribed to one or more social platforms rising from
45 percent in 2011 to 66 percent in 2017 (ONS, 2020). In Section 5,
we discuss this issue more in depth in light of the literature on the
societal impact of social media.
4.3. Fast Internet, social interaction and trust
We conclude the analysis by focusing on social interaction and
the cognitive dimension of social capital. We estimate our empirical
model on three different outcomes: trust in others, the frequency
with which respondents talk to neighbors, and the
frequency of meetings with other people. We collect our results
in Table 7.
Our measure of trust is a binary variable taking value 1 if the
respondent says that most people can be trusted and 0 otherwise.
As for social interaction, our dependent variable is a dichotomous
indicator equal to 0 if the frequency of the talks or meetings is less
than once or twice a week and 1 otherwise.
Since these outcome variables are only available in the BHPS,
we cannot perform our main specification based on the interaction
term between distance and the Post dummy, covering the period
from 2005 to 2017. Therefore, we only rely on our second specification
and interact the distance between respondents' dwelling
and the network's node serving the area with a dummy for the period
2005-2008. Given the lack of information about cultural consumption
in the UKHLS panel, we cannot estimate the coefficient
relative to the Post(2) period (from 2010 to 2017).
We find a positive but negligible and not statistically significant
association between broadband Internet and the frequency of conversations
with neighbors (column 1). The association between
fast Internet and the frequency of meetings with friends is negative
but not statically significant (column 2). On the other hand, chats
with neighbors generally occur occasionally and incidentally, for
example when leaving or coming back home. They thus seem to
Table 7
Effect of fast Internet on social interactions and trust.
Dep. Variables: Talks to neighbors Meets people Trusts people
(1) (2) (3)
Distance x Post(l) 0.04 0.02 -0.21
(0.22) (0.24) (0.29)
Controls w w w
Time FEs \S \S \S
Individual FEs \S \S \S
MSOA linear trend \S \S \S
Mean of Dep.Var. 39.41 46.01 37.97
Observations 138629 138622 63553
R2
0.54 0.47 0.63
Num. PIDs 18577 18578 15281
Dependent variables in (1), (2), and (3) are indicators for, respectively, more than
weekly conversations with neighbors, more than weekly meeting with friends, and
social trust. Dependent variables are re-scaled such that 1 = 100. The variable
Distance is the average distance from the LE in the respondent's LSOA, measured in
km. The indicator variable Post(l) takes value 1 for waves from 15 (2005) to 18
(2008), representing the post-internet period captured by the BHPS (the variable
has not been collected in UKHLS) (the variable has not been collected in UKHLS).
The sample includes all respondents in the BHPS waves for which the relevant
information was collected and available. Controls included in each regression: wave
indicators, average distance from the LE in the LSOA, age, log of real annual
household income, indicators for (i) occupation type, (ii) occupational status, (iii)
type of housing tenure, (iv) type of household, (v) education level, (vi) change of
address from the previous year, (vii) a set of monthly indicator variables for the
month of the interview. Standard errors in parentheses are clustered at the level of
the LSOA. Standard errors in parentheses are clustered at the level of the LSOA. *, **,
and *** denote significance at the 10, 5, and 1 percent level, respectively.
Journal of Public Economics 206 (2022) 104578
be less or not at all vulnerable to the displacement effect possibly
caused by Internet use.
Column (3) shows no evidence that fast Internet has a significant
impact on social trust. Social trust is a cognitive phenomenon
that depends on individuals' values and perceptions, unrelated to
time constraints and less sensitive to the risk of crowding out.
Our result is consistent with the economic literature suggesting
that trust is a persistent trait mainly inherited from ancestors
and only limitedly affected by life experiences (Guiso et a l ,
2006; Guiso et al., 2009; Algan and Cahuc, 2010).
5. Discussion
Our empirical analysis shows that fast Internet substantially
displaced several dimensions of social capital in the UK. Timeconsuming
activities oriented to the pursuit of the common good
suffered the most from broadband penetration. Putnam et al.
(1993) labeled these activities as forms of bridging social capital,
with the term "bridging" referring to the ability to create bridges
that connect groups of people with different backgrounds. The
effect is statistically significant and sizable. Considering all associations
together, one standard deviation reduction in the distance
from the LE, resulting in a faster connection, caused a 4.7 percent
decline in the likelihood of participation in our period of analysis.
Back of the envelope calculations based on the population's distribution
across the distance from local exchanges suggest that the
increase in connection speed reduced the likelihood of participation
for 447,442 residents. The displacement effect spared relationships
between friends, which are generally considered as a form of
bonding social capital (Putnam et al., 1993; Gittel and Vidal, 1998).
While bridging social capital is commonly credited with a positive
role in the diffusion of information and trust, supporting economic
activities and development, bonding social capital is a potential
obstacle to cooperation (Putnam et al., 1993).
Our finding that broadband penetration displaced civic engagement
must be understood in relation to the work of
Bauernschuster et al. (2014), who find no statistically significant
effect of fast Internet on political engagement and volunteering
in East Germany. The authors exploit a wrong technological choice
made by the state-owned provider of Internet services that slowed
down substantially the supply of fast Internet in some areas of East
Germany to study the impact of broadband penetration on social
capital. Reduced form estimates show that people living in areas
excluded from the broadband network are less socially engaged
than the control group over 1999-2009. However, the difference
is not statistically significant for any social capital indicator but
the attendance of theaters and exhibitions. Two stages estimates
using treated areas as a binary instrument for Internet access suggest
that the broadband's impact is never statistically different
from zero. Our finding of a significant and sizable detrimental
effect of broadband penetration in the UK catches another fragment
of a larger and complex picture, suggesting that fast Internet
may have differentially affected social and civic engagement across
countries depending on the baseline levels of social capital and the
most popular online activities. As shown in Table A.2 in the Appendix,
German citizens reported significantly higher social and civic
engagement than the UK in the early 2000s. In the years of fast
Internet take-up, Germany reported higher levels of voluntary
work for cultural, sport, or hobby associations, participation and
unpaid work for political parties, and membership in humanitarian
organizations. Compared to British respondents to the European
Social Survey (Eurostat, 2002), Germans were significantly more
likely to declare that politics and volunteerism are important in life
and that voting in elections and being active in voluntary organizations
and political parties is essential to be a good citizen. At the
12
A Ceraci, M. Nardotto, T. Reggiani et al.
same time, Internet use also differed between the two countries.
Table A.3 in the Appendix shows the most popular online activities
among Internet users in the UK and Germany in the year of broadband's
take-up.2 6
In 2005, British users browsed the Internet for
commercial purposes significantly more than Germans. In the UK,
people were significantly more likely to consider the Internet as a
tool for e-shopping, online gaming, and other forms of private entertainment
than the Germans. By contrast, German users were significantly
more eager than Britons to use the Internet for online
learning, keeping in touch with others, gathering health information,
keeping themselves posted on public affairs, and interacting with
public institutions.
This descriptive evidence helps make sense of the differential
impact of broadband penetration, suggesting that the most popular
online activities and the initial endowments of social capital may
affect the societal impact of broadband penetration across countries.
In Germany, where bridging social capital was significantly
higher than the UK, and transitioning regions were catching up
with the Western levels of participation, social and civic engagement
may have taken advantage of the availability of fast Internet,
which offered new tools for participation to citizens eager to be
involved in community affairs. Conversely, in the UK, where bridging
social capital was lower than in Germany and users were more
inclined to exploit the web for commercial purposes, fast Internet
may have diverted users away from time-consuming relational
activities oriented to the common good.
The treatment effects do not significantly differ across the two
periods we address in the empirical analysis (Post(l) and Post(2)),
strikingly suggesting that the take-up of social media in the
2010s did not appreciably alter the impact of fast Internet. However,
this result does not necessarily imply that social media did
not influence bridging social capital. Instead, online networking
could affect offline connectedness in conflicting ways, leaving the
displacement effect of fast Internet substantially unchanged in
more recent years.
On the positive side, social networking sites can strengthen
civic engagement by facilitating coordination and supporting collective
action. Zhuravskaya et al. (2020) suggest that low entry barriers
make it easier to spread political information and exchange
ideas on public affairs, thereby triggering interest in politics and
possibly political engagement. Dissatisfaction with specific public
policies or the government circulates more rapidly, raising people's
eagerness for political participation. Recent experimental results
provide evidence of these mechanisms. For example, Allcott et al.
(2020) and Mosquera et al. (2020), use randomized experiments
to show that social media support the diffusion of political information
and interest in politics.
In addition, the multitude of interaction tools typical of social
media can help solve coordination issues and make it easier to
organize and participate in protests (Zhuravskaya et al., 2020).
For example, Enikolopov et al. (2020) exploit the exogenous variation
in the penetration of VK created by a peculiar subscription rule
to show that the social network significantly reduced coordination
costs and increased participation in protests against a massive
electoral fraud in Russian parliamentary elections. Still, the network's
penetration overall increased government support, consistently
with previous findings on the role of Sina Weibo in China
(Qin et al., 2017). Social media may also trigger peer effects more
widely than offline interaction, as social image concerns create
compelling incentives to join political protests (Enikolopov et al.,
2017; Larson et al., 2019; Qin et al., 2019).
However, the ability to spread information at unprecedented
speed and extensiveness across horizontal networks also makes
The table is based on Gareis et al., 2005.
Journal of Public Economics 206 (2022) 104578
social media an ideal vehicle for diffusing false information and
extremist views (Allcott and Gentzkow, 2017). Moreover, social
media's capacity to create filter bubbles can also limit the exposure
to counter-attitudinal news and opinions, fostering polarization
(Levy, 2021). The polarizing effect has proven to be more potent
for individuals or groups already having extreme views. For example,
Bursztyn et al. (2019) show that the penetration of VK
increased hate crimes in cities with a pre-existing high level of
nationalist sentiments, suggesting that the spreading of racist ideas
online has tangible implications offline. Similarly, Muller and
Schwartz (2020a,b) exploit exogenous Facebook and Internet
outages to find that the exposure to social media fed anti-refugee
sentiments, which in turn predict crimes against refugees in Germany
and Muslim in the US.
Overall, the literature suggests that social media facilitate the
spreading of false information, increase polarization, and exacerbate
anti-minority sentiments, with real implications for hate
crime. These phenomena may be detrimental to bridging social
capital to the extent that they could erode trust and undermine
the willingness to contribute to the common good. There is evidence
that polarization causes a decline in social capital and public
goods in offline communities (Bazzi et al., 2019) and reduces political
participation (Rogowski, 2014). Still, it is not clear if these conclusions
can be generalized. The existing evidence suggests that
social media's impact may differ depending on the specific characteristics
of users and the institutional and cultural background. In
general, a speculative interpretation of the non-significant impact
of social media as resulting from the balance between opposing
forces strengthening and weakening social capital at the same time
finds some support in previous studies. However, turning this
speculation into empirical evidence requires further research
retrieving exogenous sources of variation in specific online activities
to identify the causal impact of social media on civic
engagement.
The multiple effects of social media on human relations could
be as conflicting as their political outcomes. For example, Allcott
et al. (2020) identify a crowding-out effect of Facebook use on offline
activities such as face-to-face socialization, playing sports and
carrying out physical exercise, suggesting that active participation
in social media may compete with offline social engagement in
users' time choices. This evidence contrasts with Bauernschuster
et al. (2014), who find no significant effect of broadband access
on similar outcomes in Germany. Our finding of the lack of a statistically
significant impact of fast Internet on the frequency of meetings
with friends and chats with neighbors is consistent with
Bauernschuster et al. (2014).
However, our analysis on bonding social capital is limited to the
Post(l) period covered by BHPS data, as the UKHLS does not collect
information on respondents' relational habits. This limitation does
not allow us to speculate further on the potential role of social
media. The statistical insignificance of the relationship between
distance and bonding social capital in the Post(l) period could
result from opposite effects rather than the lack of any effect.
Early Internet studies suggest that surfing the web for leisure
necessarily steals time from social activities such as communicating
with friends, neighbors, and family members, thereby leading
ICTs to crowd out face-to-face interaction (e.g., DiMaggio et al.,
2001). In Bowling Alone, Putnam (2000) builds on the concept of
"community without propinquity" (Webber, 1963) to suggest that
Internet penetration may have triggered a broader societal transformation
that weakens the incentives for face-to-face interaction.
In a famous paper, Wirth (1938) claimed that any increase in the
heterogeneity of the urban environment would provoke the
cooling-off of "intimate personal acquaintanceship" and would
result in the "segmentation of human relations" into those that
were "largely anonymous, superficial, and transitory" (Wirth,
13
A Ceraci, M. Nardotto, T. Reggiani et al.
1938, p. 1). According to Putnam (2000), this line of reasoning can
be applied to the Internet, which has the potential to fragment
local communities into new virtual realities of shared interests that
may negate the necessity of face-to-face encounters.
From this perspective, even e-commerce and online banking
may be detrimental to socialization, as they decrease the need
and chances for face-to-face interaction (see for example Antoci
et al., 2011; and Conrads and Reggiani, 2017).
Though focused on social media, the field experiment by Allcott
et al. (2020) provides causal evidence of online entertainment's
ability to displace offline meetings. On the other hand, fast Internet
helps users preserving existing relationships despite time and distance
constraints, thanks to the supply of a wide range of communication
tools. Case studies in applied psychology suggest that
social media could also support social relationships in specific
communities in the early stages of their penetration. For example,
Ellison et al. (2007) and Steinfield et al. (2008) find significant associations
between Facebook use and offline interactions with
friends and family among groups of college students in 2006 and
2007.2 7
The differential impact of broadband penetration on bridging
and bonding social capital may also relate to their different weight
in users' utility functions. Online activities are unlikely to undermine
the marginal utility of self-interested behaviors like meetings
with family and friends, which are crucial for individual well-being
(Heliwell and Huang, 2013). On the other hand, time-consuming
tasks more oriented to the public good such as civic engagement
are likely more vulnerable to time constraints and the competition
with alternative ways of spending leisure time.
In any case, broadband access outcomes depend on the most
popular ways of using fast Internet. Further research is needed to
understand how the specific online activities that rely on fast Internet
interact with different institutional and cultural contexts. More
detailed knowledge of Internet users' online activities is necessary
to understand better the mechanisms channeling the impact of
broadband and social media penetration on social capital and
derive reliable implications for policy and regulation.
6. Conclusions
In this paper, we studied how the penetration of broadband
Internet affected social capital in the UK. Matching unique information
on the topology of the old voice communication infrastructure
with geocoded survey data on individual behaviors, we could
exploit discontinuities in Internet connection speed to test
whether the availability of fast Internet displaced various forms
of social capital. Our results paint a complex picture. We do not
find evidence that broadband access displaced bonding social capital
in the forms of meetings with friends, conversations with
neighbors, and cultural activities that people usually enjoy in company.
By contrast, our empirical analysis shows that fast Internet
significantly crowded out time-consuming activities oriented to
the pursuit of the common good, commonly referred to as bridging
social capital, in the form of Putnam-type associations that pursue
other-regarding, non-particularistic goals, and Olson-type associations
mainly devoted to redistributive goals. While bonding social
capital seems resilient to technological change, bridging social capital
proves fragile and vulnerable to the pressure of technologyintensive
entertainment on the agents' time allocation choices.
This result is disturbing as it suggests that progress in ICT can
undermine an essential factor of the economic activity and the
2 7
The Post( 1) period in our analysis spans from 2005 to 2008. Facebook launched in
2004, and activated photo features and news feeds in 2005 and 2006, respectively.
Twitter and VK launched in 2006, FriendFeed and Tumblr in 2007.
Journal of Public Economics 206 (2022) 104578
well-functioning of democratic institutions (Putnam et al., 1993).
However, the pattern we have detected in the case of the UK
may not necessarily have general validity. Our results must be
understood in connection with previous, conflicting evidence on
the outcomes of broadband penetration. The behavioral and societal
impact of fast Internet may vary depending on the institutional
background, the initial stock of social capital, and the most popular
online activities. Countries with higher endowments of bridging
social capital, where people are eager to use the Internet to connect
with others and engage in public affairs, may suffer less from the
displacement effect we detect for the UK. These considerations also
urge caution in extrapolating implications from ours and previous
studies on the societal impact of broadband penetration. This body
of research only assesses the outcomes of fast Internet use or availability.
Further research is needed to understand the potential
impact of specific ways of exploiting faster connection speeds,
especially in light of the growing role of the few social media platforms
that monopolize the online discourse, such as Twitter and
Facebook.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared
to influence the work reported in this paper.
Appendix A
Additional Figures
A.1,A.2,A.3,A.4,A.5
Additional Tables
A1,A2,A3,A4
Alternative identification based on rainfall
We also tested the possibility to employ the same IV approach
as in Gavazza et al., 2019. This identification strategy does not
prove to be suitable in the context of this study for the following
reasons: i) Noise in the response to the question on broadband
access (see the discussion in Section 3.3); ii) Small sample size
due to the inclusion of the question on broadband access in only
3 waves (16th to 18th) which overlap with the question on participation
in organizations only in 1 case (17th wave); iii) Small sample
size in terms of households sampled. The combination of these
elements weakens the IV strategy and makes it inconclusive,
although the results are in line with what we find with our current
identification strategy.
The first stage of the IV, i.e., the effect of rain on internet,
implies to regress the self-declared broadband access status (connected
with a broadband connection) on the level of rainfall. Thus,
the equation we seek to estimate is:
Internets = yRainm + /}X,-t + Wavec + Area,t + e,t (5)
where Internet^ is an indicator variable for declaring to have a
broadband connection by household i living in LSOA / at time t.
The regressor we are interested in is RainU[ which can be computed
at the LSOA level in each year and it's the amount of rainfall (in millimeters)
during the previous year. As we report in Table A.5, we
have estimated the model with several functions of the rainfall.
Then Xit,Wavet, and Regionit are controls for observed household
characteristics, and wave/area fixed-effects. We introduced the latter
both at the level of the Local Authority District (smaller area,
14
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Org: Political Parties
2 3 4 5 7 9 11 13 15 17 21 24 27
Wave
Org: Trade Org: Professional
2 3 4 5 7 9 11 13 15 17 21 24 27
Wave
Org: Environmental
2 3 4 5 7 9 11 13 15 17 21 24 27
Wave
Org: Voluntary
2 3 4 5 7 9 11 13 15 17 21 24 27
Wave
Org: Scout
5 7 9 11 13 15 17 21 24 27
Wave
Org: Any
2 3 4 5 7 9 11 13 15 17 21 24 27
Wave
Org: Any Olson Org: Any Putnam
2 3 4 5 7 9 11 13 15 17 21 24 27 2 3 4 5 7 9 11 13 15 17 21 24 27
Wave Wave
Talking to Neighboors Meeting with Friends
- - , ° - - <."
v - 2Cinema
Attendance
12 W
Wave
Theatre and Concerts Trusting People
Fig. A.l. Event Study Analysis. Each panel plots the coefficients and 90% confidence intervals associated with the interactions between wave dummies and the variable
Distance, using the first available wave as baseline. Plots are presented separately for each of the dependent variables considered in the analysis. The model specification is
identical to the one in Eq. (1) where the interaction term Distance^ x Post, is substituted by a full set of interaction between wave dummies and the variable Distance, i.e.,
J2,7,Distanceit x Wave,. The set of controls is identical to the one used in the main regressions and includes time FE, individual FE, and MSOA linear trends. Standard errors are
clustered at the level of the LSOA. We test the null hypothesis of joint significance for the coefficients in the Pre-Broadband period for participation in Any organization, Olson
organizations and Putnam organizations, and in all cases we do not reject the null hypothesis of 0 (p-values are 0.738, 0.568, and 0.1.47 respectively.
corresponding to counties) and at the level of the Region (much larger
areas, such as West Midlands or Scotland).
For this regression, we can use all waves from the 16th to 18th
(the first stage that makes use of only wave 17th is reported in
Table A.7 which shows the TSLS estimates). Results in columns
(1) and (5) indicate a negative effect of rainfall on the probability
that households declare to have internet at home, conditional on
a large set of characteristics and on county (or region) fixed effects.
When we introduce further powers of the rainfall we do not find
support for a non-linear relationship and the estimated coefficients
become all insignificant. Thus, we use the linear term of the rain as
our instrument in the subsequent analysis.
The ITT of the IV is estimated as follows:
ym = yRainm + fiXit + Wavec + Areait + eit (6)
where ym is the outcome of interest (participation into Putnam or
Olson organizations, or the other measures of social capital that
we consider in the paper) for the household i living in LSOA / at time
t. The regressor we are interested in is Rainut. Then, as before,
X,t , Wavec, and Regionk are controls for observed household characteristics,
and wave/area fixed-effects. As the question about participation
into organizations is asked only in the 17th wave, we
restrict the analysis to this wave. Results are reported in Table A.6.
The estimated coefficients are, as expected, all positive, indicating
that more rainfall induces a larger participation into organizations,
possibly through a reduction in Internet use. However, the
coefficient is statistically significant only when we use region FEs
(although in that case it is smaller in size, suggesting a negative
bias), and only at the 10% level.
Finally, Table A.7 reports the 2SLS estimates for the IV regression
of the outcome (here we report the participation into organizations
but we do not find any evidence also for the other
outcomes that we consider in the paper) on Internet access, which
we instrument using (lagged) rainfall. The table is divided in 3 panels
for: i) Any organization; ii) Olson organizations; iii) Putnam
organizations. Within each panel, the first pair of columns are
15
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Pre Post (1) Post (2)
BHPS BHPS U K H L S
Broadband internet
Daily use of internet
•Mem./act. any org.
• •• • • • • • • • • • •Mem./act. any Putnam org.
• •• • • • • • • • • • •Mem./act. scout org.
• • • • • • • • • • •Mem./act. vol. org.
• •• • • • • • • • • • •MemVact. env. org.
• •• • • • • • • • • • •Mem./act. any Olson org.
• •• • • • • • • • • • •MemVact. prof. org.
• • • • • • • • • • •Mem./act. trade union
• •• • • • • • • • • • •Mem./act. political party
• •• • • • • • • • • • •Conc./theat. attendance
• • • •Cinema attendance
• • • •Talk to neighbors
Meet with friends
People can be trusted
i i i I I I I I I I I I i • •
1 1 1
1 1 1 1 1 1 1 1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Wave
Fig. A.2. Event Study Analysis.
Fig. A.3. LSOAs and MSOAs of Colchester. Borders of MSOAs are reported in red while borders of LSOAs are reported in blue.
Post
Pre Post(l) Post(2)
1991 1992 ••• 2004 2005 2006 2007 2008 2009 2010 2011 2012 ••• 2017 Year
I 1 1 1 1 1 1 1 1 1 1 1 1 1 >•
1 2 ••• 14 15 16 17 18 19 20 21 22 ••• 27 B H P S / U K H L S
harmonized wave
14 15 16 17 18
B H P S w a v e
B H J S respondents enter U K H L S sample in wave 2
1 1 1 1
2 3 4
U K H L S w a v e
Fig. A.4. Timeline.
16
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Fig. A.5. Participation in Any organization splitting households by High and Low Distance to the LE and estimating model (3). The figure reports the estimated coefficients of
the interaction between the indicator variables Low - Distance (which is 1 if the household's LSOA is at 3.1 km or less to the respective LE) and the Wave. Standard errors are
clustered at the LSOA level and 90% confidence intervals are reported.
Table A.1
Descriptive statistics - Demographic information BHPS and UKHLS.
Variable Mean StDev. Min Max Obs
Age
Log of real household income
Household type:
- Single: elderly
- Single: non-elderly
- 1 Adult with kids
- 1 Couple: no kids
- 1 Couple: with kids
- 2 + Adults (no couple): no kids
- 2 + Adults (no couple): with kids
- 3 + Adults (at least one couple): no kids
- 3 + Adults (at least one couple): with kids
- Other
Occupation:
- Large employers
- Higher management
- Intermediate management
- Lower management
- Small employers and own account
- Lower supervisory and technical
- Semi-routine
- Routine
- Unemployed
- Retired
- Inactive
Housing tenure:
- Owned Outright
- Owned with Mortgage
- Local Authority rented
- Housing Assoc. rented
- Rented from Employer
- Rented private unfurnished
- Rented private furnished
- Other rented
Education:
- Degree
- Other higher degree
- A-level etc.
- GCSE etc.
- Other qualification
- No qualification
46.13
8.08
0.08
0.06
0.01
0.29
0.21
0.07
0.04
0.14
0.10
0.00
0.02
0.04
0.09
0.16
0.06
0.06
0.09
0.07
0.04
0.22
0.17
0.28
0.46
0.13
0.05
0.01
0.05
0.03
0.00
0.14
0.09
0.21
0.24
0.10
0.21
18.74
0.78
0.27
0.24
0.09
0.45
0.41
0.25
0.20
0.35
0.30
0.02
0.15
0.20
0.28
0.36
0.23
0.23
0.28
0.25
0.19
0.41
0.38
0.45
0.50
0.33
0.21
0.09
0.21
0.18
0.03
0.34
0.29
0.41
0.43
0.30
0.41
15.00
-2.35
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
102.00
11.75
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
260403
Note: the number of observations depends on the BHPS and the UKHLS waves considered. We reports the summary statistics of the socio-demographic variables taking each
individual/wave (only once) that enters any of the regression models estimated in Section 4.
17
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Table A.2
GER vs UK social capital.
Survey question UK GER UK vs GER Variable description
Sports/outdoor activity club, last 12 months: member 0.261 0.296 <* * * 0,1: 0 = no, 1 = yes
(0.439) (0.456)
Sports/outdoor activity club, last 12 months: participated 0.211 0.192 >* 0,1: 0 =no, 1 = yes
(0.408) (0.393)
Sports/outdoor activity club, last 12 months: voluntary work 0.055 0.102 <* * * 0,1: 0 =no, 1 = yes
(0.229) (0.302)
Cultural/hobby activity organisation, last 12 months: member 0.165 0.156 > 0,1: 0 =no, 1 = yes
(0.363) (0.371)
Cultural/hobby activity organisation, last 12 months: participated 0.159 0.134 >** 0,1: 0 =no, 1 = yes
(0.366) (0.340)
Cultural/hobby activity organisation, last 12 months: voluntary work 0.047 0.068 <* * * 0,1: 0 =no, 1 = yes
(0.213) (0.253)
Trade union, last 12 months: member 0.151 0.139 > 0,1: 0 =no, 1 = yes
(0.357) (0.346)
Trade union, last 12 months: participated 0.025 0.034 <* * * 0,1: 0 =no, 1 = yes
(0.157) (0.181)
Trade union, last 12 months: voluntary work 0.005 0.012 <*** 0,1: 0 =no, 1 = yes
(0.069) (0.110)
Business/profession/farmers organisation, last 12 months: member 0.127 0.079 >* * * 0,1: 0 =no, 1 = yes
(0.333) (0.269)
Business/profession/farmers organisation, last 12 months: participated 0.053 0.029 >* * * 0,1: 0 =no, 1 = yes
(0.224) (0.170)
Business/profession/farmer organisation last 12 months: voluntary work 0.014 0.013 > 0,1: 0 =no, 1 = yes
(0.120) (0.113)
Humanitarian organisation etc., last 12 months: member 0.033 0.052 <* * * 0,1: 0 =no, 1 = yes
(0.033) (0.223)
Humanitarian organisation etc., last 12 months: participated 0.028 0.026 > 0,1: 0 =no, 1 = yes
(0.165) (0.160)
Humanitarian organisation etc., last 12 months: voluntary work 0.021 0.017 > 0,1: 0 =no, 1 = yes
(0.144) (0.128)
Environmental/peace/animal organisation, last 12 months: member 0.059 0.056 > 0,1: 0 =no, 1 = yes
(0.237) (0.229)
Environmental/peace/animal organisation, last 12 months: participated 0.031 0.032 < 0,1: 0 =no, 1 = yes
(0.175) (0.176)
Environment/peace/animal organisation, last 12 months: voluntary work 0.016 0.015 > 0,1: 0 =no, 1 = yes
(0.125) (0.123)
Political party, last 12 months: member 0.028 0.031 < 0,1: 0 =no, 1 = yes
(0.172) (0.167)
Political party, last 12 months: participated 0.009 0.032 <* * * 0,1: 0 =no, 1 = yes
(0.092) (0.178)
Political party, last 12 months: voluntary work 0.006 0.020 <* * * 0,1: 0 =no, 1 = yes
(0.079) (0.139)
Other voluntary organisation, last 12 months: member 0.048 0.069 <* * * 0,1: 0 =no, 1 = yes
(0.214) (0.253)
Other voluntary organisation, last 12 months: participated 0.042 0.034 > 0,1: 0 =no, 1 = yes
(0.201) (0.182)
Other voluntary organisation, last 12 months: voluntary work 0.042 0.018 >* * * 0,1: 0 =no, 1 = yes
(0.202) (0.134)
Important in life: politics 3.74 5.091 <* * * 0-10:0 = not important at
(2.499) (2.473) all - 10 = extremely important
Important in life: voluntary organisations 3.594 4.002 <*** 0-10:0 = not important at
(2.953) ( 2.830) all - 10 = extremely important
Discuss politics/current affairs, how often 4.209 3.158 >* * * 1-7 : 1 == every day, 7 = never
(2.114) (1.817)
To be a good citizen: how important to vote in elections 7.189 7.465 <* * * 0-10:0 = not important at
(2.790) (2.683) all - 10 = extremely important
Good citizen: how important to be active in voluntary organisations 5.171 4.775 <* * * 0-10:0 = not important at
(2.530) (2.667) all - 10 = extremely important
To be a good citizen: how important to be active in politics 3.438 4.200 <* * * 0-10:0 = not important at
(2.422) (2.451) all - 10 = extremely important
Total survey sample 2,052 2,919
Source: European Social Survey 2002- EUROSTAT. "*p<0.01, "p<0.05, *p<0.10.
the two stages of the IV where we employ LAD FEs while the second
pair are the two stages of the IV where we use the Region FEs.
The first thing to notice is the lack of identification power coming
from the first-stages (reported in the even columns) when we
use only 1 wave. The coefficient of the rain is negative as expected
but it is never significant, which is reflected by the very low value
of the first-stage F-test. The second stages, reported in the even
columns, are negative in all cases, but also never statistically
significant.
Overall, we consider this exercise useful, but inconclusive. The
estimated coefficients suggest that our main findings could be confirmed
using the IV identification strategy of Gavazza et al. (2019),
but the lack of identification power makes this strategy less reliable
that the approach we employ in the main empirical analysis.
18
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Table A.3
GER vs UK Internet use.
Survey question UK GER UK vs GER Variable description
E-commerce use (last 3 months) 0.611 0.305 >* * * 0,1: 0 = no, 1 = yes
(0.019) (0.017)
E-commerce: annoyed by online advertising? 3.097 4.081 <** 1-5:0 = does not annoy at
(1.314) (1.191) all - 5 = annoys me very much
Internet: important e-commerce motives 3.544 3.095 >* * * 0-5: 0 = not important at
(1.344) (1.298) all - 5 = very important
Internet: important for leisure (online games, music) 2.368 2.174 >** 0-5: 0 = not important at
(1.413) (1.288) all - 5 = very important
E-mail: importantto keep in touch with people 3.760 3.875 <* * * 0-5: 0 = not important at
(1.300) (1.339) all - 5 = very important
Internet: important for information gathering 4.346 4.411 <* 0-5: 0 = not important at
(0.872) (0.864) all - 5 = very important
Internet: used foronline learning (last 12 months) 0.389 0.488 <* * * 0,1: 0 = no, 1 = yes
(0.488) (0.501)
Internet: search for health-related information 0.366 0.429 <** 0,1: 0 = no, 1 = yes
(0.495) (0.495)
Internet: would like to use more e-gov services 0.458 0.474 <** 0,1: 0 = no, 1 = yes
(0.498) (0.499)
Total survey sample 1,001 1,001
Source: eUSER 2005 - EUROSTAT/Empirica. ~p<0.01, **p<0.05, *p<0.10. Standard deviation in parentheses.
Table A.4
Distance and online activities.
Dep. Variables: Use of social media Online gaming
(Adults) (Kids 10-15)
(1) (2) (3) (4)
Distance -0.40*** -0.26" -1.61" -1.29"
(0.14) (0.11) (0.74) (0.65)
Controls j-* w
Time FEs V j-* j-* \S
MSOA FEs j-* \S
Mean of Dep.Var. 21.22 19.91 51.36 51.46
Observations 108356 102794 12406 10940
R2
0.18 0.31 0.33 0.50
Num. PIDs 53857 51296 7771 6975
Dependent variable in (1), (2) is an indicator for the use of social media to chat and interact with friends more than one hour per day. Information on social media use for
adults is available for waves 3 (2011-2012), 6 (2014-2015), and 9 (2017-2018) of Understanding Society. Dependent variable in (3), (4) is an indicator for playing multiplayer
online games. Information on online gaming for kids (10-15 years) is available for waves 2 (2010-2011), 4 (2012-2013), 6 (2014-2015), and 8 (2016-2017) of
Understanding Society. Dependent variables are re-scaled such that 1 = 100. The variable Distance is the average distance from the LE in the respondent's LSOA, measured in
km. Controls included in (1), (2) are wave indicators, sex, age dummies, log of real annual household income, indicators for (i) occupation type, (ii) occupational status, (iii)
type of housing tenure, (iv) type of household, (v) education level, (vi) month of interview. Controls in (3) and (4) include: wave indicators, sex, age dummies, log of real
annual household income, indicators for (i) occupation type of the mother and the father (if present), (ii) education of the mother and the father (if present), (iii) type of
housing tenure, (iv) type of household, (v) month of interview. Standard errors in parentheses are clustered at the level of the LSOA. *, **, and *** denote significance at the 10,
5, and 1 percent level, respectively.
Table A.5
Rainfall and broadband internet access.
Dependent variable: Internet
Local authority FEs Region FEs
(1) (2) (3) (4) (5) (6) (7) (8)
Rain -0.063* 0.025 0.115 -0.076* -0.054" 0.066 0.247 -0.067"
(0.035) (0.112) (0.284) (0.040) (0.021) (0.088) (0.279) (0.029)
Rain2
-0.036
(0.046)
-0.118
(0.268)
-0.051
(0.038)
-0.214
(0.250)
Rain3
0.023
(0.079)
0.045
(0.070)
Rain max 0.072
(0.105)
0.073
(0.108)
LAD FEs Yes Yes Yes Yes No No No No
Region FEs No No No No Yes Yes Yes Yes
Time FEs Yes Yes Yes Yes Yes Yes Yes Yes
Controls Yes Yes Yes Yes Yes Yes Yes Yes
R2
0.341 0.341 0.341 0.341 0.294 0.294 0.294 0.294
Observations 32276 32276 32276 32276 32799 32799 32799 32799
Dependent variable is an indicator for having broadband Internet at home. Rain is the (lagged) yearly amount of rainfall in the LSOA where the household lives. Rain max is the
rainfall of the month of the previous year that experienced the largest rainfall.The sample includes all respondents in the BHPS waves for which the relevant information was
collected and available. *, **, and *** denote significance at the 10, 5, and 1 percent level, respectively.
19
A Ceraci, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Table A.6
Rainfall and broadband Internet access.
Dep variable: Participation in organizations
Any organization
(1)
Olson organization
(2)
Putnam organization
(3)
Any organization
(4)
Olson organization
(5)
Putnam organization
(6)
Rain 0.051 0.047 0.023 0.031* 0.026* 0.004
(0.039) (0.034) (0.031) (0.017) (0.016) (0.012)
LAD FEs Yes Yes Yes No No No
Region FEs No No No Yes Yes Yes
Time FEs Yes Yes Yes Yes Yes Yes
Controls Yes Yes Yes Yes Yes Yes
R2
0.215 0.241 0.095 0.127 0.157 0.023
Observations 10958 10958 10958 10952 10952 10952
Dependent variable is an indicator participation in Olson, Putnam or any organization. Rain is the (lagged) yearly amount of rainfall in the LSOA where the household lives.*,
**, and *** denote significance at the 10, 5, and 1 percent level, respectively.
Table A.7
Rainfall and broadband Internet access.
Dep variable: Internet Any Internet Any Internet Olson Internet Olson Internet Putnam Internet Putnam
org. org. org. org. org. org.
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
Internet -1.159 -0.876 -1.173 -0.742 -0.206 -0.026
(2.251) (0.889) (2.155) (0.770) (0.940) (0.379)
Rain -0.034 -0.033 -0.034 -0.033 -0.034 -0.033
(0.054) (0.024) (0.054) (0.024) (0.054) (0.024)
LAD FEs Yes Yes No No Yes Yes No No Yes Yes No No
Region FEs No No Yes Yes No No Yes Yes No No Yes Yes
Time FEs Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
R2
0.349 0.217 0.286 0.128 0.349 0.243 0.286 0.159 0.349 0.096 0.286 0.023
F-test 0.409 1.778 0.409 1.778 0.409 1.778
Observations 10834 10834 10828 10828 10834 10834 10828 10828 10834 10834 10828 10828
IVFirst-stages. Dependent variable in columns (1), (3), (5), (7), (9) and (11) is an indicator for having broadband Internet at home. Rain is the (lagged) yearly amount of rainfall
in the LSOA where the household lives. IVSecond-stages. Dependent variable in columns (2), (4), (6), (8), (10) and (12) is an indicator participation in Olson, Putnam or any
organization.*, **, and *** denote significance at the 10, 5, and 1 percent level, respectively.
References
Aghion, P., Algan, Y., Cahuc, P., Shleifer, A., 2010. Regulation and distrust. Quart. J.
Econ. 125 (3), 1015-1049.
Ahlfeldt, C, Koutroumpis, P., Valletta, T., 2017. Speed 2.0: Evaluating access to
universal digital highways. Journal of the European Economic Association 15
(3), 586-625.
Algan, Y., Cahuc, P., 2010. Inherited trust and growth. American Economic Review
100 (5), 2060-2092.
Algan, Y., Cahuc, P., Shleifer, A, 2013. Teaching practices and social capital.
American Economic Journal: Applied 5 (3), 189-210.
Allcott, H., Braghieri, L, Eichmeyer, S., Gentzkow, M., 2020. The welfare effects of
social media. American Economimic Review 110 (3), 629-676.
Allcott, H., Gentzkow, M., 2017. Social media and fake news in the 2016 election.
Journal of Economic Perspectives 31 (2), 211-236.
Amaral-Garcia, S., Nardotto, M., Propper, C, Valletta, T, 2021. Mums Go Online: Is
the Internet Changing the Demand for Healthcare? The Review of Economics
and Statistics, https://doi.org/10.1162/rest_a_01033.
Antoci, A., Bonelli, L., Paglieri, F., Reggiani, T, Sabatini, F., 2019. Civility and trust in
social media. Journal of Economic Behavior and Organization 160, 83-99.
Antoci, A., Sabatini, F., Sodini, M., 2011. The solaria syndrome: Social capital in a
growing hyper-technological economy. Journal of Economic Behavior and
Organization 81 (3), 802-814.
Autor, D.H., 2003. Outsourcing at will: The contribution of unjust dismissal doctrine
to the growth of employment outsourcing. Journal of Labor Economics 21 (1),
1-42.
Barenjee, R., 2016. Corruption, norm violation and decay in social capital. Journal of
Public Economics 137, 14-27.
Bauernschuster, S., Falck, 0., Woessmann, L., 2014. Surfing alone? the internet and
social capital: Evidence from an unforeseeable technological mistake. Journal of
Public Economics 117, 73-89.
Bazzi, S., Gaduh, A., Rothenberg, A.D., Wong, M., 2019. Unity in diversity? how
intergroup contact can foster nation building. American Economic Review 109
(11), 3978-4025.
Bhuller, M., Havnes, T, Leuven, E., Mogstad, M., 2013. Broadband internet: an
information superhighway to sex crime? Rev. Econ. Stud. 4 (1), 1237-1266.
Billari, F.C., Giuntella, 0., Stella, L., 2018. Broadband internet, digital temptations,
and sleep. Journal of Economic Behavior & Organization 153, 58-76.
Bloom, N., Sadun, R., Van Reenen, J., 2012. The organization of firms across
countries. Journal of Public Economics 127 (4), 1663-1705.
Bursztyn, L, Egorov, G., Enikolopov, R., and Petrova, M. (2019). Social media and
xenophobia: evidence from Russia. NBER Working Paper N. 26567.
Campante, F., Durante, R., Sobbrio, F., 2018. Politics 2.O.: The multifaceted effect of
broadband internet on political participation. Journal of the European Economic
Association 16 (4), 1094-1136.
Campante, F., Durante, R., and Tesei, A. (2021). Media and social capital. NBER WP
29230.
Conrads, J., Reggiani, T., 2017. The effect of communication channels on promisemaking
and promise-keeping: experimental evidence. J. Econ. Interac. Coord. 12
(3), 595-611.
Costa, D., Kahn, M.E., 2003. Cowards and heroes: group loyalty in the american civil
war. Quart. J. Econ. 118 (2), 519-548.
Czernich, N., Falck, 0., Kretschmer, T, Voessmann, L, 2011. Broadband
infrastructure and economic growth. Econ. J. 121 (552), 505-532.
Davidson, R., MacKinnon, J.G., et al., 2004. Econometric theory and methods, volume
5. Oxford University Press, New York.
Degli Antoni, G., Grimalda, G., 2016. Groups and trust: Experimental evidence on
the olson and putnam hypotheses. Journal of Behavioral and Experimental
Economics 61, 38-54.
Deshpande, A. (2014). An examination of the evolution of broadband technologies
in the UK. PhD thesis, The Open University.
Dettling, L.J., Goodman, S., Smith, J., 2018. Every little bit counts: The impact of
high-speed Internet on the transition to college. Rev. Econ. Stat. 100 (2), 260-
273.
DiMaggio, P., Hargittai, E., Russel Neumann, W., Robinson, J.P., 2001. Social
implications of the internet. Annual Review of Sociology 27, 307-336.
Ellison, N.B., Steinfield, C, Lampe, C, 2007. The benefits of facebook friends: social
capital and college students' use of online social network sites. Journal of
Computer-Mediated Communication 12, 1143-1168.
Enikolopov, R., Makarin, A., Petrova, M., 2020. Social media and protest
participation: Evidence from russia. Econometrica 88 (4), 1479-1514.
Enikolopov, R., Makarin, A., Petrova, M., and Polishchuk, L. (2017). Social image,
networks, and protest participation. Available at SSRN: https://doi.org/10.2139/
ssrn.2940171.
Enikolopov, R., Petrova, M., Sonin, K., 2018. Social media and corruption. American
Economic Journal: Applied 10 (1), 150-174.
EuropeanCommission (2012). Quality of broadband services in the eu. Technical
report, European Commission.
Eurostat (2002). European Social Survey - wl/2002. Technical report, EUROSTAT.
20
A Gerau, M. Nardotto, T. Reggiani et al. Journal of Public Economics 206 (2022) 104578
Faber, B., Sanchis-Guarner, R., and Weinhardt, F. (2015). let and education: evidence
from student home addresses. NBER WP 21306.
Falck, 0., Gold, R., Heblich, S., 2014. E-lections: Voting behavior and the internet.
American Economic Review 104 (7), 2238-2265.
Gareis, K., Cullen, K., and Korte, W. B. (2005). eUSER project: Putting the user at the
centre - what it means for the provision of online public services in the EU.
Technical report, EUROSTAT & Empirica.
Gavazza, A, Nardotto, M., Valletta, T., 2019. Internet and politics: Evidence from UK
local elections and local government policies. Rev. Econ. Stud. 86 (5), 2092-
2135.
Gittel, R.J., Vidal, A., 1998. Community organizing: Building social capital as a
development strategy. Sage Publications, Los Angeles, CA.
Guiso, L, Sapienza, P., Zingales, L, 2004. The role of social capital in financial
development. American Economic Review 94 (3), 526-576.
Guiso, L, Sapienza, P., Zingales, L, 2006. Does culture affect economic outcomes?
Journal of Economic Perspectives 20 (2), 23-48.
Guiso, L, Sapienza, P., Zingales, L, 2009. Cultural biases in economic exchange?
Quart. J. Econ. 124 (3), 1095-1131.
Guiso, L, Sapienza, P., Zingales, L., 2016. Long-term persistence. Journal of the
European Economic Association 14 (6), 1401-1436.
Guriev, S., Melnikov, N., 2016. War, inflation, and social capital. American Economic
Review 106 (5), 230-235.
Heliwell, J.F., Huang, H., 2013. Comparing the happiness effects of real and on-line
friends. PLoS ONE 8 (9)(e72754).
Hjort, J., Poulsen, J., 2019. The arrival of fast Internet and employment in Africa.
American Economic Review 109 (3), 1032-1079.
HouseCommons (2017). Broadband speeds and advertising. Technical report, House
Commons - CDP 2017/0072.
Karlan, D., 2005. Using experimental economics to measure social capital and
predict financial decisions. American Economic Review 95 (5), 1688-1699.
Knack, S., Keefer, P., 1997. Does social capital have an economic payoff? Quart. J.
Econ. 112 (4), 1251-1288.
Larson, J.M., Nagler, J., Ronen, J., Tucker, J.A., 2019. Social networks and protest
participation: Evidence from 130 million Twitter users. American Journal of
Political Science 63 (3), 690-705.
Levy, R., 2021. Social media, news consumption, and polarization: Evidence from a
field experiment. American Economic Review 111 (3), 831-870.
McDool, E., Powell, P., Roberts, J., Taylor, K., 2020. The internet and children's
psychological wellbeing. Journal of Health Economics 69.
Miner, L., 2015. The unintended consequences of internet diffusion: Evidence from
malaysia. Journal of Public Economics 132, 66-78.
Mosquera, R., Odunowo, M., McNamara, T., Guo, X., Petrie, R., 2020. The economic
effects of Facebook. Experimental Economics 23, 575-602.
Müller, K., Schwartz, C, 2020a. Fanning the flames of hate: Social media and hate
crime. Journal of the European Economic Association, https://doi.org/
10.1093/jeea/jvaa045.
Müller, K. and Schwartz, C. (2020b). From hashtag to hate crime: Twitter and antiminority
sentiment, mimeo.
Nardotto, M., Valletta, T., Verboven, F., 2015. Unbundling the incumbent: evidence
from UK broadband. Journal of the European Economic Association 13 (2), 330-
362.
Nunn, N., Wantchekon, L, 2011. The slave trade and the origins of mistrust in africa.
American Economic Review 101 (7), 3221-3252.
OECD (2008). Broadband growth and policies in oecd countries. Technical report,
OECD.
Ofcom (2006). Wireless last mile - final report. Technical report, Ofcom.
Ofcom (2009). UK broadband speeds 2009: Consumers experience of fixed-line
broadband performance. Technical report, Ofcom.
Ofcom (2010). UK broadband speeds 2010: The performance of fixed-line
broadband delivered to UK residential consumers. Technical report, Ofcom,
London, UK.
Ofcom (2011). UK fixed-line broadband performance 2011: The performance of
fixed-line broadband delivered to UK residential consumers. Technical report,
Ofcom, London, UK.
Ofcom (2012). Infrastructure report: 2012 update. Technical report, Ofcom, London,
UK.
Ofcom (2013). UK fixed-line broadband performance 2013: The performance of
fixed-line broadband provided to UK residential consumers. Technical report,
Ofcom, London, UK.
Ofcom (2016). Clarity for businesses baffled by broadband speeds. Technical report,
Ofcom.
Ofcom (2017). Connected nations 2017. Technical report, Ofcom, London, UK.
Olson, M., 1971. The Logic of Collective Action. Harvard University Press,
Cambridge, MA.
ONS (2020). Internet access - households and individuals. Technical report, Office of
National Statistics, London, UK.
Petrova, M., Sen, A., and Yildirim, P. (2017). Social media and political donations:
New technology and incumbency advantage in the United States. Technical
report, CEPR Discussion Paper No. DPI 1808.
Putnam, R.D., 1995. Bowling alone: America's declining social capital. Journal of
Democracy 6 (1), 65-78.
Putnam, R.D., 2000. Bowling Alone: The Collapse and Revival of American
Community. Simon & Schuster, New York.
Putnam, R.D., Leonardi, R., Nanetti, R.Y., 1993. Making Democracy Work: Civic
Traditions in Modern Italy. Princeton University Press.
Clin, B., Strömberg, D., Wu, Y, 2017. Why does China allow freer social media?
protests versus surveillance and propaganda. Journal of Economic Perspectives
31 (1), 117-140.
Qin, B., Strömberg, D., and Wu, Y. (2019). Social media, information networks, and
protests in China, mimeo.
Rogowski, J.C., 2014. Electoral choice, ideological conflict, and political
participation. American Journal of Political Science 58 (2), 479-494.
Rosenberg, M., 1956. Misanthropy and political ideology. Am. Sociol. Rev. 21, 690-
695.
Steinfield, C, Ellison, N.B., Lampe, C, 2008. Social capital, self-esteem, and use of
online social network sites: A longitudinal analysis. Journal of Applied
Developmental Psychology 29, 434-445.
Taylor, M.F., Brice, J., Buck, N., and Prentice-Lane, E. (2010). British household panel
survey: User manual. Technical report.
Uphoff, N., 1999. Understanding social capital: Learning from the analysis and
experience of participation. The World Bank, Washington, DC.
Webber, M.M., 1963. Order in diversity: Community without propinquity. Johns
Hopkins Press, Baltimore.
Wirth, L, 1938. Urbanism as a way of life. Am. J. Sociol. 44, 3-24.
Zhuravskaya, E., Petrova, M., Enikolopov, R., 2020. Political effects of the internet
and social media. Annual Review of Economics 12, 415-438.
21