Journal Pre-proof

THE INTERNET AND CHILDREN’S PSYCHOLOGICAL WELLBEING

Emily McDool, Philip Powell, Jennifer Roberts, Karl Taylor

PII: S0167-6296(18)31140-8
DOI: https://doi.org/10.1016/j.jhealeco.2019.102274
Reference: JHE 102274

To appear in: Journal of Health Economics

Received Date: 12 December 2018

Revised Date: 31 July 2019
Accepted Date: 7 December 2019

Please cite this article as: McDool E, Powell P, Roberts J, Taylor K, THE INTERNET AND
CHILDREN’S PSYCHOLOGICAL WELLBEING, Journal of Health Economics (2019),
doi: https://doi.org/10.1016/j.jhealeco.2019.102274

This is a PDF file of an article that has undergone enhancements after acceptance, such as
the addition of a cover page and metadata, and formatting for readability, but it is not yet the
definitive version of record. This version will undergo additional copyediting, typesetting and
review before it is published in its final form, but we are providing this version to give early
visibility of the article. Please note that, during the production process, errors may be
discovered which could affect the content, and all legal disclaimers that apply to the journal
pertain.

© 2019 Published by Elsevier.

THE INTERNET AND CHILDREN’S
PSYCHOLOGICAL WELLBEING

Emily McDool 1, Philip Powell 1, 2, Jennifer Roberts 1*and Karl Taylor 1, 3

1
Department of Economics, University of Sheffield, UK;
2
School of Health and Related Research, University of Sheffield, UK; and

f
3
IZA Bonn

oo
pr
Abstract
Late childhood and adolescence is a critical time for social and emotional development. Over the past
two decades, this life stage has been hugely affected by the almost universal adoption of the internet
e-
as a source of information, communication, and entertainment. We use a large representative sample
of over 6,300 children in England over the period 2012 to 2017, to estimate the effect of
neighbourhood broadband speed, as a proxy for internet use, on a number of wellbeing outcomes,
Pr
which reflect how these children feel about different aspects of their life. We find that internet use is
negatively associated with wellbeing across a number of domains. The strongest effect is for how
children feel about their appearance, and the effects are worse for girls than boys. We test a number
of potential causal mechanisms, and find support both for the ‘crowding out’ hypothesis, whereby
internet use reduces the time spent on other beneficial activities, and for the adverse effect of social
al

media use. Our evidence adds weight to the already strident calls for interventions that can reduce the
adverse effects of internet use on children’s emotional health.
n

JEL Codes: D60; I31; J13

ur

Key Words: digital society, social media, wellbeing, children, happiness.

Jo

Acknowledgements: This research was funded by the UK EPSRC research grant EP/L003635/1
Creating and Exploring Digital Empathy (CEDE). The main Understanding Society data set is an
initiative by the Economic and Social Research Council, with scientific leadership by the Institute for
Social and Economic Research, University of Essex, and survey delivery by the National Centre for
Social Research and TNS BRMB. The data was accessed via the UK Data Service; some of it under
Special Licence. We would like to thank Cristina Sechel for her help and expertise with ArcGIS. We
are also grateful to Sarah Brown, Andy Dickerson and Gurleen Popli for comments on an earlier draft
of this paper; also seminar participants at Clark University, Monash University and the University of
York; and to attendees at the 2017 Royal Economic Society annual conference, 2017 Understanding
Society scientific conference and 2018 International Society for Quality of Life Research conference
for their input. We are also extremely grateful to the comments received from two anonymous referees
and the editor.
1. INTRODUCTION

Late childhood and adolescence is a critical and potential vulnerable time for social and emotional

development. One aspect of this life-stage that has changed dramatically in the past 20 years is the

almost universal adoption of the internet as a source of information, communication, and

entertainment. United Nations research has estimated that 3.5 billion people (47 per cent of the world

population) use the internet globally; one third of these are under 18. 1 In the UK, today’s teenagers

have grown up with the internet and now spend more time online than they do watching television

f
(Ofcom, 2015). Almost all 12-15 year olds (98%) use the internet; with 96% accessing it at home via

oo
a fixed broadband (BB) connection, and a large proportion also using a mobile network signal

(Ofcom, 2015). “The internet is not just something children access when they want certain bits of

pr
information; it is an essential and intrinsic part of the world they inhabit” (House of Lords, 2017).

Internet use can have both beneficial and detrimental effects on children’s wellbeing because
e-
of the wide range of activities that are undertaken online.2 Jackson (2008) describes a ‘utopian’ view,
Pr
where the internet provides a chance to develop the skills needed for the modern workplace, as well

as new opportunities for self-expression, communication and access to information. However, she

contrasts this with a ‘dystopian’ perspective where time online ‘crowds out’ other beneficial activities
al

such as reading, playing sports and face-to-face interaction with friends and family, as well as
n

exposing children to potentially inaccurate and harmful content, and sexual and commercial
ur

exploitation. Early empirical studies, largely based on cross section surveys of teenagers, lend support

to an overall negative effect of internet use on wellbeing (Brenner, 1997; Kraut et al., 1998).
Jo

However, more recent evidence is inconclusive, with, for example, Bauernschuster et al. (2014)

1
www.un.org/apps/news/story.asp?NewsID=54931#.WLlbn8pXUpE
2
The same argument can be made for adults. Fujiwara et al. (2018), in their analysis of the wellbeing effects of the UK
Superfast BB programme, find a negative association between BB speed and adult life satisfaction using Annual
Population Survey data; although they find a positive association using UKHLS data.
2
finding a positive association between internet use and social capital, but Lohmann (2015) finding a

negative influence of internet use of wellbeing, operating largely via relative income effects.

The potentially detrimental effects of online activity have prompted concern in the UK among

a number of bodies responsible for the mental health and wellbeing of young people. 3 A recent

enquiry on ‘children and the internet’ by the House of Lords (2017) concluded that the current regime

of self-regulation very often put commercial considerations first, with scant regard for wellbeing.4

The enquiry acknowledged the association between increasing numbers of unhappy and anxious

children and the growth in internet use, but also called for more robust research into the possible

causal relationships.5 Our study is an attempt to respond to that call; we explore the association

f
oo
between children’s emotional health and internet use, and also consider a number of potential causal

channels, which we investigate empirically.

pr
Children’s use of the internet is an important topic for study because it is a significant

component of time use, and its effects on wellbeing and mental health are ambiguous (Kalmus et al.,
e-
2014). Most of the existing evidence comes from samples of adults, or students in higher education
Pr
(see for example Cotton, 2008; Bhuller et al., 2013). There is very little evidence for children, and

that which does exist tends to come mainly from cross section data, and relatively small, bespoke

samples (see for example Gross et al., 2002; Jackson, 2008).

al

In this study we use neighbourhood BB speed (as reported by Ofcom) as a proxy for internet
n

access. There is increasing evidence to suggest that children do not regard the internet as a separate,
ur

distinct entity that they access; rather it is woven ubiquitously throughout all aspects of their life. For

this reason, self-reported data on time spent on the internet may be misleading. In addition, children
Jo

3
The National Society for the Prevention of Cruelty to Children recently cited social media as a major cause of the
dramatic increase in the numbers of children admitted to hospital as a result of self-harming (www.nspcc.org.uk/). An
inquiry into cyberbullying by the charity Young Minds and The Children’s Society was carried out in 2017
https://youngminds.org.uk/resources/policy/cyberbullying-inquiry/. Barnardo’s, the children’s charity, produced a ‘youth
and the internet’ guide for policymakers in 2015 (Barnardo’s, 2015).
4
In 2019, the government published a White Paper setting out their recommendations for reducing online harms,
particularly among children (HM Government, 2019).
5
Recently, a cross-parliamentary enquiry on ‘the impact of social media and screen-use on young people’s health’, again
cited the lack of evidence of causal relationships: www.parliament.uk/business/committees/committees-a-z/commons-
select/science-and-technology-committee/inquiries/parliament-2017/impact-of-social-media-young-people-17-19/
3
use the internet for a wide range of diverse activities, and it is likely that the impact on their wellbeing

depends on the quality and type of activities they are engaged in. Our data contains only limited

information on what children are doing online. Instead, we assume that faster BB speeds mean that

children can access more content in a given amount of time. Hence, we might expect to observe an

association between BB speed and child wellbeing because of the enhanced internet access enabled

by faster BB speeds.

We improve on the majority of existing evidence on the effects of internet use, firstly, using

a large nationally representative sample of children, rather than a small survey of a selective group.

Secondly, we utilise measures of emotional health across different domains, which enable us to

f
oo
explore which aspects of children’s lives are most affected. Thirdly, we explore a number of potential

causal mechanisms that can help to explain why internet use affects emotional health. Fourthly, we

pr
use longitudinal data and fixed effects (FE) models, in order to eliminate endogenous selection bias

arising from time-invariant unobserved variables, such as childhood circumstances or neighbourhood

e-
characteristics, which influence both wellbeing and BB speed. We also explore in detail the
Pr
assumption of quasi-random assignment of BB speed, which is necessary for us to identify the effect

of BB speed on psychological wellbeing. Finally, we also consider which groups of children are most

affected by use of the internet.

al

Our results are worrying for anyone concerned with children’s emotional health. We find a
n

negative relationship between internet use and wellbeing domains. In our most stringent specification,
ur

children feel worse about their schoolwork, appearance, friends and the school they attend, as internet

use increases. For example, a 1% increase in BB speed reduces how children feel about their
Jo

appearance by approximately 0.6 per cent. The adverse effects of the internet are worse for girls than

boys, and the strongest effect (for both sexes) is for how they feel about their appearance. Our

evidence also suggests that these results can be explained via the ‘crowding out’ hypothesis, whereby

internet use reduces the time spent on other beneficial activities, and from the adverse effect of social

media use.

4
 Jo
ur
nal

5
Pr
e-
pr
oo
f
 2. BROADBAND SPEED AND CHILDREN’S WELLBEING

The potential influence of the internet on the emotional health and wellbeing of children is attracting

increased attention across a range of disciplines. Castellacci and Tveito (2018) review this disparate

literature and classify the relevant mechanisms through which the internet can affect wellbeing into

four distinct channels. Their review does not consider children explicitly, but all of these channels

have particular implications for children, which we attempt to draw out here; we also test a number

of these causal channels in our empirical work. Children’s wellbeing can be affected directly in the

ways described below, but also indirectly via intergenerational effects; parents’ behaviours and

wellbeing can be shaped by internet use and this will in turn affect their children (see Pfeffer and

f
oo
Schoeni, 2014). Further, there are reasons to expect that children may be particularly vulnerable to

some of the mechanisms described below. As well as being at a critical stage of social and emotional

pr
development, children often adopt technological innovations before parents, schools or policy makers

can consider the lasting implications of this technology. This problem is exacerbated by the fact that
e-
internet content is increasingly consumed via a tablet or phone, and this move to smaller, portable
Pr
devices is making adult supervision much more difficult than when the internet was largely accessed

via a single family computer.

The internet can influence wellbeing firstly because it changes time use patterns. On one hand,
al

this can make existing tasks, like shopping, coursework and job search, more efficient, thus freeing
n

up time for other activities. For example, by aiding job search the internet could improve parents’
ur

labour market outcomes, helping them to find jobs that better match their skills or exit unemployment

spells more quickly, which in turn affects the child’s wellbeing. Gürtzgen et al. (2018) found that, in
Jo

Germany, digital technologies helped the unemployed find a job. On the other hand, there is evidence

that the internet can crowd-out other activities known to be beneficial for wellbeing, such as playing

sport and having face-to-face interactions with friends and family (Moreno et al., 2013; Wallsten,

2013). Further, in evidence given to the House of Lords (2017) enquiry, Barnardo’s, the children’s

charity, warns that children are far from being fully informed, rational consumers when it comes to

6
online commercial transactions. Ofcom research suggests that less than half of 12-15 year olds were

aware of paid endorsements by vloggers or personalised advertising; they also could not spot

advertising in online search results (Ofcom, 2015).

Secondly, the internet facilitates new activities, which can have positive or negative effects

on wellbeing. The internet has enabled online gaming and digital social networks, which are now

almost ubiquitous in their use among children; it also facilitates access to on demand entertainment

via traditional media sites such as online television, as well as specialist online sites such as YouTube

and music streaming apps like Spotify. While these activities can have important positive effects such

as enabling children to develop creativity and social skills, and providing opportunities both for

f
oo
stimulation and relaxation, there is increasing concern about the potentially negative effects such as

addiction, commercial exploitation and increased chances of exposure to inappropriate content (Kuss

pr
and Griffiths, 2012). In addition, there is some evidence that new activities such as gaming and

streaming entertainment displace time spent on schoolwork (Kaiser Family Foundation, 2010).
e-
Thirdly, the internet enables greater access to information; this can contribute to social and
Pr
educational development. However, the proliferation of inaccurate content, ‘fake news’ and

inappropriate sexual and/or violent content on the internet can be damaging to children’s wellbeing.

Further, while there is no conclusive evidence yet, there are suggestions that overwhelming exposure
al

to information in itself may be affecting concentration and decreasing attention spans, in adults as
n

well as children (Carr, 2010). Also, in common with many adults, children do not have an
ur

understanding of how search engines work and have limited knowledge to enable them to judge the

accuracy or providence of online information (Ofcom, 2016). In a review of the effects of computers,
Jo

software and the internet on education, Bulman and Fairlie (2016) argue that, while increased

availability is associated with increased use, few studies find positive effects on educational

outcomes. Faber et al. (2015) explore the effects of internet speed on test scores for English primary

and secondary school students over the period 2002-2008, and find no significant effects on student

time spent studying online or offline, or on their educational outcomes.

7
 Much of the material that children are exposed to online can be classified as inappropriate

(Martellozzo et al., 2016), and pornography has emerged as a particular concern. The British Board

of Film Censors argue that “This has led to the normalisation of largely unfettered access to the

strongest, sometimes unlawful, pornography by children online”.6 Young people themselves are also

expressing their concerns about pornography on the internet; in a survey of 500 young people, 80%

said that it was too easy to access pornography online, and 72% felt that this was leading to unrealistic

views about sex, particularly among boys (Institute for Public Policy Research, 2014).7

Fourthly, the internet provides new communication tools, such as email, instant messaging,

social media, Skype and Facetime. These tools have the potential to increase both the scope and

f
oo
intensity of social interactions, both of which are among the strongest predictors of wellbeing

(Kahneman et al., 1999). Social media use has grown extremely rapidly, and is a core part of young

pr
people’s lives. A survey in 2015 revealed that, in the UK, 92% of 16 to 24 year olds had used social

networks, such as Facebook, Snapchat, WhatsApp and Instagram, in the last three months.8 Younger
e-
children are also increasingly users of social media; while most sites stipulate a minimum user age of
Pr
13, few require any validation, and a survey for the children’s BBC television channel found that

more than three quarters of 10 to 12 year olds had social media accounts.9 Social networks are

children’s primary interface with the internet. These portals are generally used in an ‘always on’ state,
al

often via smartphones and tablets, such that many children are permanently connected to their virtual
n

social network, continually receiving and checking feed, and regularly posting their own updates
ur

(Boyd, 2014). For this reason, we believe that social media use is potentially a very important

mechanism through which internet use can influence children’s emotional health. The results that we
Jo

6
House of Lords (2017) para. 130.
7
Online bullying (discussed below) is also intertwined with the normalisation of online sexual content. Almost half of 18
year-olds questioned in the same survey stated that sending naked pictures to each other (known as ‘sexting’) was part of
everyday life; and that these pictures were often shared more widely in an attempt to bully and shame individuals.
8
www.ons.gov.uk/ons/guide-method/method-quality/specific/business-and-energy/e-commerce-and-ict-activity/social-
networking/index.html
9
www.bbc.co.uk/news/education-35524429
8
present in section 4 below lend support to this view. For this reason, we spend some time here

considering this mechanism in more detail.

Digital social networks serve a multiplicity of functions. They are a tool for developing and

maintaining interpersonal relationships, a real-time portal for accessing information, news, advice

and social support, as well as a canvas for sketching a selective and idealised self-portrait. While it is

generally acknowledged that social media can have a positive impact on social capital, for example

by enhancing friendships and decreasing loneliness (Franzen, 2003; Antoci et al., 2012), there are

concerns that ‘excessive’ time spent on social media is associated with low self-esteem, common

mental health problems, and socioemotional difficulties (e.g., Beardsmore, 2015; Kross et al., 2013).

f
oo
Evidence is also emerging demonstrating a detrimental effect of social media use on sleep (Levenson

et al., 2016).10 There is also evidence that online communication substitutes for face-to-face

pr
interactions (Sabatini and Sarracino, 2017). Helliwell and Huang (2013) compare face-to-face friends

with online social networks in an adult Canadian sample; they find a positive correlation between the
e-
size of real and online social networks, but that only increases in the number of face-to-face friends
Pr
are associated with improved wellbeing.

There are two additional complementary theories that can help to explain why extensive social

media use may have a negative effect on children’s emotional health, on top of the ‘crowding out’
al

hypothesis discussed above in relation to internet access in general. First, ‘social comparison’ theory,
n

posits that increased social media use is linked to more frequent social comparisons, which are more
ur

likely to be ‘upward’ (negative) in direction. The material people choose to present online represents

selectively idealised versions of their true lives (Mendelson and Papacharissi, 2010), and there is
Jo

evidence that young social media users act naively, in that they fail to understand that the material is

not representative (Royal Society for Public Health, 2017).11 Sabatini and Sarracino (2016) find that

social network users in Italy have a higher probability of making social comparisons than non-users,

10
‘Fear of missing out’ is linked to obsessive checking of social media feeds and sleep problems, with almost of half of
pupils questioned in a survey admitting to checking their mobile devices after going to bed. www.hmc.org.uk/wp-
content/uploads/2016/10/Mobile-Device-Media-Brief-FINAL.pdf
11
Enke (2017) presents experimental evidence that this irrationality is also present in adult behaviour.
9
and that this tendency is greatest in younger people. In related work, Clark and Senik (2010) found

that, in Europe, people with internet access attach more importance to income comparisons than those

without, and Lohmann (2015) finds that people who regularly use the internet as a source of

information derive less satisfaction from their income. Chou and Edge (2012) found that students

who spent more time on Facebook were more likely to think that other people were happier and had

better lives than their own. Furthermore, a growing body of research attests to the mediating role of

envy in the relationship between Facebook use and decreased affective wellbeing (e.g., Tandoc Jr. et

al., 2015; Verduyn et al., 2015). Issues of body image and self-esteem have been raised as a

particularly negative aspect of greater social media use, especially among girls (Kleemans et al., 2016;

f
oo
Children’s Society, 2018a). While, these concerns are not a new development (and have been linked

in the past to women’s magazines for example), the internet and social media increase the

pr
accessibility and immediacy of unrealistic body images, thus intensifying their effect.

Secondly, ‘cyberbullying’ theory, relates to the fact that children who spend more time on
e-
social networks have a greater chance of being the victim of direct attacks from others on their sense
Pr
of self, wellbeing, and self-esteem (Cowie, 2013). Childline counselling services reported a 12

percent increase in the number of cases related to cyberbullying in 2016/17 compared to the previous

year (Children’s Society, 2018b). Sampasa-Kanyinga and Hamilton (2015) reported a significant
al

increase in the odds of being victimised for every hour spent using social networking sites. While
n

cyberbullying often overlaps with traditional ‘offline’ bullying, the former may be particularly
ur

pernicious because children’s continual connectedness means they cannot escape (Slonje et al.,

2012).12 A recent study for the US using micro data from the Youth Risk Behaviour Survey, found
Jo

that cyberbullying influences suicidal behaviour (Nikolaou, 2017), and there have been a number of

high profile cases involving teenagers taking their own lives in part because of being harassed over

the internet (Hinduja and Patchin, 2010).

12
A number of economic studies have illustrated the negative and persistent effects of being bullied in childhood. For
example, Eriksen et al. (2014) find detrimental effects of being bullied on educational attainment in a large Danish sample.
Brown and Taylor (2008) find that being bullied at school in the UK has an adverse effect on human capital accumulation
both at and beyond school.
10
 3. DATA AND METHODOLOGY

In this paper, we use data from Understanding Society – The UK Household Longitudinal Study

(UKHLS), a representative sample of over 40,000 households across the UK (University of Essex,

2018). Eight waves are currently available, starting with wave 1 in 2009, which provided data on over

50,000 individuals, and the latest wave 8 (at the time of writing) where over 39,000 individuals were

interviewed between 2016 and 2018. All adult members of each household are interviewed, along

with children aged 10 to 15 years old. In this analysis, waves 3 to 8 are used; these provide data on

just over 6,300 children residing in England, who are the focus of the empirical analysis.13

Children’s data comes from the Youth Self-completion Questionnaire, which is used

f
oo
alongside data from the adult surveys, giving information on household characteristics such as

income, homeownership and parental education. The outcomes of interest are obtained by asking

pr
children how they feel about different aspects of their life, specifically: school work; appearance;

family; friends; school attended; and life as a whole (see Appendix, Table A1). Internet use may affect
e-
these domains of children’s wellbeing differently, and perhaps even in opposing directions, hence it
Pr
is useful to be able to disaggregate wellbeing in this way. We order the responses to range from

“1=not happy at all” through to “7= completely happy”. We interpret these measures as indicators of

psychological wellbeing or emotional health (see Clark et al., 2018). After conditioning on missing
al

values, we create an unbalanced panel of 6,310 children for the period 2012 (wave 3) through to 2017
n

(wave 8). Out of this sample 36% of children are observed once, 27% twice, 17% three times, 11%
ur

four times, 6% five times and 3% in all waves, giving a total number of observations 13,938 across

children and over time.

Jo

Data are available from Ofcom, the UK communications regulator, on the average

synchronisation speed of existing BB connections, measured in megabits per second (mbps). This is

defined at the neighbourhood level, where neighbourhoods are classified via Lower Layer Super

13
We do not use UKHLS waves 1 and 2 because Ofcom data on BB speed are not available for those years at the level
of disaggregation required. Similarly, the Ofcom data is available for LSOAs in England only.
11
Output Areas (LSOA).14 LSOAs are very small areas; there are 32,844 LSOAs in England, with an

average size of 650 households; and the children in our sample reside in 3,765 LSOAs. The UKHLS

provides LSOA identifiers for each household, enabling us to match neighbourhood BB speed to our

sample of children. The BB speed data are available across all years 2012 to 2017 and are matched

to the LSOA-year in which the child was interviewed. The sample of children we draw on live in

households that do not change address during the period, which is important as the analysis is based

upon location FE.

Reduced form approach

For ease of interpretation we standardize each outcome to have a mean of zero and standard deviation

f
oo
of unity,15 and condition on a set of covariates and BB speed. The reduced form models we estimate

are of the following form:

pr
𝑦𝑖𝑗𝑤 = 𝑿′𝑖𝑗𝑤 𝜷 + 𝜙𝑍𝑗𝑤 + 𝛼𝑖 + 𝜓𝑗 + 𝜆𝑤 + 𝜀𝑖𝑗𝑤 (1)

where 𝑖(= 1, . . ,6310), 𝑗(= 1, . . ,3765), 𝑤(= 3, . . ,8) denote the child, the neighbourhood in which
e-
the child lives (LSOA), and wave of interview respectively; the outcome of interest is denoted by
Pr
𝑦𝑖𝑗𝑤 and 𝑍𝑗𝑤 is the average neighbourhood BB speed. The error term is normally distributed

𝜀𝑖𝑗𝑤 ~𝑁(0, Σ) and 𝛼𝑖 , 𝜓𝑗 , 𝜆𝑤 denote child, neighbourhood and time specific FE, respectively.

Equation (1) is estimated using FE analysis, where the inclusion of FE eliminates endogenous
al

selection bias arising from time-invariant unobserved variables, e.g. childhood circumstances or
n

neighbourhood characteristics, which influence both wellbeing and BB speed. Our interest lies in the
ur

sign and statistical significance of the estimate 𝜙̂, which, based upon the quasi-random assignment

of BB (considered in detail below) gives the intention-to-treat (ITT) effect of BB speed on

Jo

psychological wellbeing, capturing the causal effect of being assigned to treatment (see Angrist and

Pischke, 2009).

14
The average synchronisation speed is the average speed at which the modem connects with the internet; it is a key BB
performance metric used by Ofcom. The data are available from www.ofcom.org.uk
15
Ferrer-i-Carbonell and Frijters (2004) show that assuming cardinality for ordinal measures of wellbeing is acceptable
in models where individual effects are included, as is the case in our analysis.
12
 The covariates in vector 𝑿𝑖𝑗𝑤 control for individual child, parent, household and local area

characteristics and comprise: age, specifically whether aged 10, 11, 12, 13 or 14 (with aged 15 as the

omitted category); whether male; the number of children aged 0-2, 3-4, 5-11 or 12-15 in the

household;16 whether either of the child’s parent(s) own their house, are employed, or have a degree

or equivalent qualification; whether the child lives in a single parent household; the natural logarithm

of real equivalized net household income; and local area characteristics (including the unemployment

rate, gross value added (GVA) per capita, the share of females, the share of the population over 65,

and the share of the population of working age) defined at the Local Authority District (LAD) level,

to proxy for local economic conditions.17

f
oo
Full variable definitions are given in the Appendix, Table A1. Summary statistics are shown

in Table 1. Figure 1 shows histograms of the distribution of each dependent variable prior to

pr
standardization. Clearly, across each of the psychological wellbeing outcomes, on average children

report towards the upper of the scale, although for feelings about their school work and appearance
e-
the mean response is lower and the standard deviation higher in comparison to the other domains;
Pr
approximately 20% of respondents state that they are ‘completely happy’ with their school work

and/or appearance.

Approximately 51% of children are aged 13 to 15 and just under half are male. In terms of
al

family background, 77% of children have at least one parent who is either an employee or self-
n

employed, 33% of parents have a degree; 22% of children live in a single parent household; the
ur

average real net equivalized household income is £1,326 per month; and 67% of parents own their

home either outright or with a mortgage. Table 1 shows that the average BB speed over the period
Jo

was just under 26 mbps. Figure 2 provides kernel density plots of the neighbourhood average BB

speeds for each year, where clearly the average speed, and also the variance, has increased over time.

16
For the categories 5-11 and 12-15 the number of children excludes the respondent.
17
In the UKHLS there are 330 LADs in England. The LAD variables are obtained from www.nomisweb.co.uk which is
a service provided by the Office for National Statistics containing official labour market statistics.
13
Figure 3 shows the distribution in growth in BB at the LSOA level for each wave as a complement

to Figure 2, where growth occurred between all waves.

Figure 4 shows a map of England split into LSOAs where white areas are not included in our

sample. The map shows London along with the eight core cities of England with boundaries given in

blue. By considering quartile ranges of average BB speed over time at the LSOA level, the map

highlights where the variation in BB speed comes from. It would appear that BB speeds are highly

dispersed, not concentrated in any one particular region, and not dominated by large cities, e.g.

London (noticeably there is evidence of variation in BB speed across quartiles within London).

Random allocation

f
oo
The empirical approach adopted relies only on an assumption of quasi-random assignment of BB and

does not require an exclusion restriction. BB rollout in England has been a complex mix of

pr
commercial and government initiatives. The bulk of commercial rollout, which was largely demand

driven, was completed before our analysis period starts in 2012 (Department for Digital, Culture
e-
Media and Sport, 2018). Existing service quality was a result of a number of factors including poor
Pr
home wiring, long telephone lines and random network effects (Department for Business Innovation

and Skills, and Department for Culture Media and Sport, 2009). The government subsidised superfast

BB programme was announced in 2010 in response to concerns that commercial deployment would
al

fail to reach many parts of the country where installation would not be profitable for a number of
n

reasons (Department for Business Innovation and Skills, and Department for Culture Media and
ur

Sport, 2010).18 The government provided £780m of public resources for a substantial programme

developed in partnership with Local Authorities, who were expected to match central government
Jo

funds on a one-to-one basis. Local Authorities were also responsible for procurement and

management of contracts with local suppliers, resulting in a large amount of local variation

(Department for Digital, Culture Media and Sport, 2018). Recently published evaluation evidence

suggests that the areas that benefitted from this subsidised rollout had features that were expected to

18
Superfast BB is usually interpreted as meaning download speeds of at least 24 megabits per second.
14
reduce the commercial viability of upgrading local infrastructure; these included, being further from

the exchange, having a higher share of ‘exchange only’ lines and having a low density of premises

(Department for Digital, Culture Media and Sport (2018).19 The evaluation also reveals that schemes

were primarily concentrated in rural areas with comparatively low population densities but any

differences in economic performance were less apparent (Department for Digital, Culture Media and

Sport (2018).

Overall, the subsidised programme, which was rolled out in the period coinciding with our

analysis period, substantially distorted commercial demand driven BB installation. As a result, we do

not expect BB speed in our analysis period to be systematically related to local socio-economic

f
oo
characteristics. We also explore this assumption empirically following Bhuller et al. (2013), who

investigate whether increased internet access triggers sex crimes, and Akerman et al. (2015), who

pr
estimate whether or not the adoption of BB internet by firms enhances labour productivity. Firstly,

we regress BB speed on neighbourhood (LSOA) and time FE, and area (LAD) level time varying
e-
characteristics (unemployment, population shares, education and GVA per head). We find that 83
Pr
percent of the variation in BB speed is attributable to time-invariant neighbourhood characteristics

and common time effects, whilst less than 1 percent is due to time-varying covariates (which might

pick up demand and supply effects associated with improving BB speeds). Secondly, we consider the
al

timing of increasing BB speed, as revealed in Figure 2, and baseline characteristics, which are
n

evaluated at the start of the sample period, by estimating the following at the neighbourhood level:
ur

̅𝑗,2012 ]′𝜓𝑤 + 𝜖𝑗𝑤

∆𝑍𝑗𝑤 = 𝛾𝑗 + [𝛿𝑤 × 𝑿 (2)

where Δ𝑍𝑗𝑤 is the first difference in average BB speed across waves, i.e. ∆𝑍𝑗𝑤 = 𝑍𝑗𝑤 − 𝑍𝑗𝑤−1 , shown
Jo

̅𝑗,2012 contains the 2012 mean of all

in Figure 3, 𝛿𝑤 is a vector of time fixed effects and the vector 𝑿

child-family and area level covariates averaged at the neighbourhood level. Table 2 reports estimates

of 𝜓𝑤 and joint tests for each covariate of 𝜓𝑤 across waves. It would appear that generally there is

19
BB connections are made over a copper cable, which can either be connected to a roadside cabinet near the premises,
or directly to the local telephone exchange. This latter type of ‘exchange only’ connection requires a more expensive
upgrade procedure to enable the premises to access superfast BB.
15
no systematic relationship between improving BB performance and the control variables. More

importantly, there is little correlation with area-level background variables such as the unemployment

rate, GVA per capita,20 and population shares. This suggests that service expansion was not

concentrated in those locations where the economy was growing fastest (or slowest). Moreover, the

variation in the growth of BB speed shown in Figure 3 did not just occur in years where the covariates

predict the change in BB connection, see Table 2, and indeed perhaps not surprisingly the predictive

power of the model is weak where the overall model F-statistic only has a p-value of 0.42.

Furthermore, we require that increasing BB speed is uncorrelated with the time-varying

factors that influence psychological wellbeing. Following the approach of Altonji et al. (2005a,

f
oo
2005b) we also test this by considering selection on observables as an indication of selection on

unobservable effects. This is undertaken for the full sample of children and sub-samples of potential

pr
interest, e.g. males and females, since the effect of gender cannot be recovered in the FE framework

of equation (1), where we explore the extent to which the BB speed is correlated with the part of the
e-
outcome explained by the covariates. Specifically we estimate the following for each outcome:
Pr
𝑐𝑜𝑣(𝑿′𝑖𝑗𝑤 𝜷, 𝑍𝑗𝑤 |𝛼𝑖 )/𝑣𝑎𝑟(𝑍𝑗𝑤 |𝛼𝑖 ) (3)

Table 3 shows the results of this where it is clear that for each domain across the full sample and the

sub-groups, the correlation between 𝑿′𝑖𝑗𝑤 𝜷 and 𝑍𝑗𝑤 is very low and practically indistinguishable from
al

zero. This result is consistent with the assumption that increasing BB speed is uncorrelated with the
n

time-varying factors that influence psychological wellbeing. To summarise we conclude that quasi-
ur

random assignment of BB in our data is a reasonable assumption, whereby the estimate of 𝜙, from

equation (1), provides the intention-to-treat (ITT) effect of BB speed on psychological wellbeing.
Jo

20
The only exception, albeit at the 10 per cent level, is for GVA per head, consistent with the finding of Bauernschuster
et al. (2014) for Germany. Moreover, contrary to expectations, results show that where individually statistically significant
the effect is associated with lower growth in internet speed, e.g. between 2012-13, 𝜓̂3 , a 1% increase in GVA per head is
associated with 0.086 per cent reduction in the growth of BB speed.
16
 4. RESULTS

In this section, we firstly estimate the reduced form model of equation (1) for the full sample of

children, before considering heterogeneity across a number of sub-groups. Finally, we explore

potential mechanisms that may be capable of explaining the results.

Reduced form analysis

Table 4 presents the coefficients on 𝑍𝑗𝑤 , i.e. 𝜙̂ the ITT, showing the effect of BB speed on each of

the wellbeing domains reported in each column. We report four different specifications of equation

(1): (i) Panel A with no control variables, just child FE; (ii) Panel B incorporating covariates, 𝑿𝑖𝑗𝑤 ,

f
which include child and family characteristics as well as local economy controls (unemployment rate;

oo
GVA per head; share of females; share of population over 65; and the share of the population aged

16-65) and child FE; (iii) Panel C also incorporating time FE; and (iv) Panel D, incorporating the full

pr
set of covariates, time, area and child FE.

Focusing initially on the FE models in Panels A and B, across each domain there is a negative
e-
association with BB speed, i.e. 𝜙̂ < 0. The inclusion of covariates generally reduces the parameter
Pr
estimate on 𝑍𝑗𝑤 but the effect remains statistically significant. Looking at Panel B a 1% increase in

BB speed decreases happiness in each of the domains ranging from 0.008 standard deviations (around
al

0.8 per cent) for appearance to 0.005 standard deviations for life overall. In Panel D, which is the

most stringent specification, BB speed remains statistically significant for four out of the six domains,
n

where a 1% increase in BB speed decreases happiness with appearance by 0.006 standard deviations
ur

(i.e. an ITT of around 0.6 per cent).

We compare the effect of BB speed to the other control variables in Table 5, which reports
Jo

the full results from Panel C of Table 4. Interestingly, where statistically significant, younger children

generally feel happier than those aged 15 (the omitted category) across each domain, whilst there are

no effects stemming from household income. In related work, Anand and Roope (2016), who consider

child wellbeing in Germany, also found no income effect, whilst, analysis from the US reveals that

parental earnings are positively associated with a number of childhood health outcomes (Mazumder

17
and Davis, 2013). Those children residing in a single parent household have lower psychological

wellbeing for appearance and life overall. What is evident is that in a FE framework few child level

and area level covariates are statistically significant, yet BB speed has a negative effect on four of the

wellbeing domains. Focusing upon appearance the coefficient on 𝑍𝑗𝑤 is on average approximately a

tenth of the magnitude of the coefficient associated with whether the child lives in a single parent

household, so arguably not a trivial effect. This is consistent with concerns about the negative

influence of the internet on body image.

In equation (1) BB speed is assumed to have a linear effect on each domain. We now relax

this by replacing 𝑍𝑗𝑤 in equation (1) with a set of binary indicators for whether the BB speed is

f
oo
between: 10 and less than 20 megabits per second (mbps); 20 and less than 30 mbps; 30 and less than

40 mbps; 40 and less than 50 mbps; 50 and less than 60 mbps; and 60 mbps and above, respectively.

pr
The results of this analysis are shown in Table 6, which incorporates full controls, area, year and child

FE. Clearly, there is evidence of linearity in that the higher BB speed is associated with worse
e-
wellbeing across all domains where statistically significant and the effect is monotonic, i.e. increasing
Pr
in magnitude. For example, considering appearance having a BB speed between 20 and 30 mbps is

associated with 0.11 lower standard deviations, but this rises in magnitude to a 0.36 standard deviation

reduction in wellbeing for BB speeds at 60 mbps and above.

al

Sub-group analysis
n

By looking at the entire sample of children, we may be missing important differences across the
ur

distribution of children – i.e. the effects of BB speed may have heterogeneous impacts across different

groups of children. Hence, we also consider the following sub-groups: male and females; children
Jo

aged below 13 and those aged 13 and above; and urban and rural areas. Equation (1) is re-estimated

for each of these groups. Table 7 shows the results splitting the sample of 6,310 children by gender

(Panels A and B), age (Panels C and D), and urban-rural location (Panels E and F). All estimates

incorporate controls as well as area, time and child FE. Considering gender differences, where

statistically significant, the largest effects are apparent for girls, hence it would seem that BB speed

18
is more detrimental for wellbeing across each domain for girls – with the largest effect stemming

from how girls feel about their appearance; again consistent with body image concerns. In Panels C

and D, the sample of children is split by age, specifically whether aged below 13 or aged 13 and

above.21 The largest effects are all for those children aged 13 and over. The only significant effects

of BB speed for those children under the age of 13 are on appearance and family, where a 1% increase

in BB speed reduces happiness in these domains by around 0.007 and 0.006 standard deviations

respectively.

Although we have controlled for area (LAD) level covariates and area FE, it is possible that

children in more geographically isolated localities have worse internet access, and also fewer

f
oo
opportunities to engage in activities (such as sports, or interacting with friends) that could affect their

psychological well-being. To investigate this, in Table 7 Panels E and F we split the sample according

pr
to whether the child lives in an urban or rural location.22 No significant effects are found for children

in rural households, but the negative effect of BB speed holds across each domain, for those children
e-
living in urban areas.
Pr
Potential mechanisms

This section explores the potential mechanisms that may explain the negative association between

BB speed and wellbeing domains, discussed in Section 2. Firstly, for children living in a household
al

with unemployed parent(s) a faster internet connection may help parents find jobs that better match
n

their skills or exit unemployment spells more quickly. Secondly, a faster internet connection may
ur

affect the happiness of the parents, which in turn may influence the happiness of their children.

Thirdly, an improved internet connection may be associated with children's academic outcomes.
Jo

Fourthly, the internet may crowd-out other activities that the child would otherwise have undertaken

which in turn reduces wellbeing. Finally, we consider whether BB speed is associated with children’s

use of social media.

21
Given that BB deployment increases over time (see Figure 2) as does age of the children, splitting into age categories
may help to disentangle these two effects.
22
Urban areas are defined as settlements with a population of 10,000 or more according to Department for Environment,
Food and Rural Affairs www.gov.uk/government/collections/rural-urban-classification.
19
 To investigate each of the above channels we take a two-step approach. Firstly, we explore

whether each of the mechanisms has a direct influence on each domain of children’s wellbeing, to do

this we estimate the following:

𝑦𝑖𝑗𝑤 = 𝑿𝑖𝑗𝑤 ′𝜸1 + 𝜃𝑚𝑒𝑐ℎ𝑎𝑛𝑖𝑠𝑚𝑖𝑗𝑤 + 𝛼1𝑖 + 𝜓1𝑗 + 𝜆1𝑤 + 𝜈1𝑖𝑗𝑤 (4a)

where the results are shown in Table 8, which, for brevity, only report 𝜃̂. Then in the next stage we

investigate whether BB speed (𝑍𝑗𝑤 ) is associated with the mechanism in question, by estimating:

𝑚𝑒𝑐ℎ𝑎𝑛𝑖𝑠𝑚𝑖𝑗𝑤 = 𝑿𝑖𝑗𝑤 ′𝜸2 + 𝜋𝑍𝑗𝑤 + 𝛼2𝑖 + 𝜓2𝑗 + 𝜆2𝑤 + 𝜈2𝑖𝑗𝑤 (4b)

The results of this analysis are shown in Table 9, where for brevity we only report 𝜋̂. From an

f
empirical perspective, for a mechanism to be capable of explaining the link between BB speed and

oo
children’s wellbeing one would expect the estimates of both 𝜃 and 𝜋 to be statistically significant.

Initially, focusing upon on transitions in labour market state, we consider transitions of either

pr
parent from unemployment in the previous time-period into employment in the current year, giving
e-
a sample size of 5,668 based on 2,987 children. Table 8 Panel A shows that there is no association

between parental transition into employment and any child wellbeing domain, and in Table 9 Panel
Pr
A clearly BB speed has no significant effect on this potential mechanism (where equation 4b is

estimated as a linear probability model). Next turning to the duration of the parents unemployment
al

spell, based upon a sample of 5,047 covering 2,716 children who currently have an unemployed

parent, there is some evidence of an inverse relationship between the length of time that the parent
n

has been unemployed and their child’s feelings about school work and school (Table 8 Panel B).
ur

However, it would appear that parental labour market experience is not a potential mechanism as BB

speed has no significant effect on the duration of parental unemployment. Moreover, the coefficient
Jo

is positive where a priori we would expect it to be negative, with faster internet connection allowing

parents to exit unemployment quicker, perhaps due to enabling a better job match (maybe due to

improved job search through online resources or through being able to obtain additional formal job

interviews, e.g. see Gürtzgen et al., 2018).

20
 We now consider the wellbeing of the parent as a potential mechanism for the effect on child

wellbeing. We measure parental psychological wellbeing using the General Health Questionnaire

(GHQ-12), which covers various dimensions, including: depression; anxiety; somatic symptoms;

feelings of incompetence; difficulty in coping; and sleep disturbance (Goldberg and Williams,

1988).23 The GHQ-12 score is on the scale 0 (the least distressed) through to 12 (the most distressed).

After removing missing values on parental GHQ-12 the focus is upon a sample of 13,227 observations

comprising 6,069 children. The results shown in Table 8 Panel C reveal an inverse relationship

between the parents GHQ-12 score and children’s wellbeing across the domains school work,

appearance and life overall, i.e. parents in more distressed states have children with lower wellbeing

f
oo
in these domains, 𝜃̂ < 0. This is consistent with an intergenerational correlation in wellbeing.

However, whether this is the mechanism through which BB speed is operating is doubtful given the

pr
estimate of the association of the speed of the internet connection with GHQ-12; 𝜋̂, is statistically

insignificant, as can be seen from Table 9 Panel C.24

e-
A key potential mechanism through which BB speed may be operating is that of the child’s
Pr
educational progress. As part of the National Curriculum assessment in England, statutory Standard

Attainment Tests (SATs) are undertaken throughout a child’s school years. We explore attainment

via two alternative outcome measures relevant for children of different ages. Key Stage 2 tests are
al

taken when children are aged 10-11, and Key Stage 4 (GCSE) tests are usually taken when children
n

are aged 15/16.25 For Key Stage 2 our outcome is the percentage of children obtaining level 4 or more
ur

in reading, writing and mathematics SATs where the test score ranges from 1 to 100.26 For Key Stage

4 our outcome is the percentage of children who achieve five or more A* to C grades, including
Jo

23
The GHQ-12 is a widely used screening instrument for common mental disorders, in addition to being a general measure
of psychiatric wellbeing.
24
In terms of the mechanisms considered so far: labour market transitions are based upon either the mother or father
exiting unemployment into a state of employment; the duration of the unemployment spell is the average across both
parents; and the GHQ-12 score is also averaged across both parents, where applicable (i.e. couple households). The same
results hold if we consider the father only, the mother only, or single parent households. Results are available upon request.
25
Note that information on Key Stage 3, SATs taken between ages 11 and 14, stopped being collected in 2011 with
information from 2009 based solely upon on-going teacher assessment.
26
https://www.gov.uk/government/collections/statistics-key-stage-2. The national expectation for children in English
schools is that children reach a level 2/3 by the end of Year 2 (Key Stage 1 SATs) and a level 4/5 by the end of Year 6
(Key Stage 2 SATs).
21
Mathematics and English.27 Unfortunately, these educational outcome measures are not available at

the child level. Instead, we use LAD area level outcomes (for the LAD of pupil residence), which are

defined consistently over the period 2013-2015 for Key Stage 2 and 2012-17 for Key Stage 4.28 For

this area level analysis we average child wellbeing scores (for the appropriately aged children) for

each domain at the LAD-year level and all other covariates are also averaged accordingly. These data

are then matched at the LAD-year level to the Key Stage test results. For the Key Stage 2 results we

focus on children aged 10 and 11, yielding a sample of 778 observations. For the Key Stage 4 results

we focus on children aged 15 giving a sample of 1,096 observations. For this analysis, the regressions

in equations (4a) and (4b) are based at the LAD level conditional on annual mean characteristics

f
oo
incorporating area FE.

The results are shown in Table 8 Panels D (Key Stage 2) and E (Key Stage 4); these reveal no

pr
relationship between test scores in reading, writing and mathematics (Key Stage 2 performance), or

GCSE attainment (Key Stage 4 performance) and any of the child wellbeing domains. However,
e-
interestingly BB speed is strongly positively associated with Key Stage 2 attainment, and negatively
Pr
associated with Key Stage 4 attainment, as can be seen from Table 9 Panel D (E). This evidence

suggests that the effect of faster internet speeds on child wellbeing domains is not operating through

educational attainment. This is consistent with the findings of Faber et al. (2015) who report that even
al

very large changes in available internet speeds have an estimated zero effect on educational
n

attainment (i.e. when summed across different Key Stages).

ur

Next, we consider whether a faster internet connection potentially crowds out beneficial

activities that the child might otherwise engage in. In waves 4, 6 and 8 of the UKHLS children were
Jo

asked a series questions regarding the activities that they undertake, including: playing sports; face-

to-face interaction with friends and family; going to youth clubs or other organised events;

undertaking voluntary or community work; and attending out of school classes such as art, music etc.

27
https://www.gov.uk/government/collections/statistics-gcses-key-stage-4
28
Ideally, the analysis would be at the child level. However, currently this is not possible by matching in the National
Pupil Database (NPD) to the UKHLS as the matched NPD data is only available for those children interviewed in UKHLS
wave 1, and the NPD data for these children is only available up to 2012/13.
22
Focusing upon a sub-sample of 5,150 children (7,389 observations) present in these waves, we sum

the number of activities that the child undertakes at least once per week. The number of activities

ranges from zero (13.5%) to six or more (2.5%) where on average children undertake two activities

at least once a week. Table 8 Panel F shows that the number of activities that the child undertakes is

positively associated with wellbeing, and significantly so for four out of six domains. In Table 9 Panel

F, we explore whether BB speed is potentially crowding out face-to-face interaction. The estimate on

BB speed is negative and statistically significant, which is consistent with the crowding out

hypothesis whereby the internet reduces the number of activities the child undertakes, which

ultimately results in lower wellbeing (Moreno et al., 2013; Wallsten, 2013).

f
oo
The final mechanism we explore is the role of social media. Given that social media sites are

children’s primary interface with the internet, we believe that they may represent a key channel for

pr
the effects of internet use on wellbeing. In the Youth Self-completion Questionnaire of the UKHLS

children are asked: Do you belong to a social website such as Bebo, Facebook or Myspace? 65% of
e-
the respondents were members of a social network and were subsequently asked: How many hours
Pr
do you spend chatting or interacting with friends through a social website like that on a normal school

day? The responses to this question are coded into the variable range from “1=none”, “2=less than

an hour”, “3=1-3 hours”, “4=4-6 hours”, and “5=7 or more hours”. We replace the 1-5 scale with
al

“midpoints” of the ranges (i.e. 0, .5, 2, 5, and 8) since arguably this is a better approximation of the
n

linear effect and also enables interpretation of the impact an additional hour spent online.29 In Tables
ur

8 and 9 Panels G to I we consider: the number of hours spent online; non-linearity by exploring

whether children spend 5 or more hours online; and then whether the child is a member of a social
Jo

media site, respectively. For hours spent online and intense time spent using social media (i.e. 5 hours

or above), the sample size is 9,948 comprising 5,235 children; whilst for membership of a social

network the corresponding observations are 13,938 and 6,310 respectively (the sample size for time

spent online is slightly lower due to missing values on hours).

29
We are grateful to an anonymous referee for this suggestion.
23
 Focusing on hours spent online and the association with children’s wellbeing, Table 8 Panel

G reveals that time spent online is inversely related to each domain and statistically significant, with

the exception of friends. For example, an extra hour spent online decreases the wellbeing score of

appearance and school work by around 0.03 standard deviations (i.e. approximately 3 per cent).

Similarly, spending a large amount of time online (5 hours or more), has a large inverse association

with all wellbeing domains apart from friends, see Table 8 Panel H; thus it appears that the negative

effects of social media use are intensified with high use.

Another aspect which we investigate is the decision to use social media, where again there is

an inverse relationship across three of the six domains, see Table 8 Panel I, although perhaps not

f
oo
surprisingly, the magnitude of the effect is not as large as that seen in Panel G from extremely high

daily time spent online. It would appear that time spent online using social media sites is a plausible

pr
mechanism, but empirically this rests upon 𝜋 > 0, i.e. faster internet speed should a priori result in

more time spent using online sites, and moreover whether the estimate is statistically significant. This
e-
is investigated by estimating equation (4b), with the results reported in Table 9 Panels G to I, where
Pr
greater internet speed is associated with more time spent online, intense usage (five or more hours)

and membership of social media sites. This is evidence in favour of social media being a key

mechanism that could link BB speed to child wellbeing;30 in contrast, evidence of other mechanisms
al

possibly operating through parental outcomes, e.g. in the labour market or intergenerational
n

transmission of wellbeing, or the child’s educational progress, are not evident in the data.
ur

5. CONCLUSION

In this paper we have explored the effect of internet use on the psychological wellbeing of children
Jo

aged 10 to 15, measured by the way they feel about five different aspects of their life, and their life

overall. Internet use is proxied by the BB speed available in the neighbourhood. We employ a FE

framework where, based upon the quasi-random allocation of BB speed, we recover the ITT effect,

30
These results might also explain why in the sub-group analysis the effects of BB speed were only observed for older
children, given the minimum age stipulation on social media use; and similarly, when we considered gender differences
the wellbeing of girls was more adversely affected by BB speed (consistent with Booker et al., 2018), see Table 7.
24
which shows that even in the most stringent specifications estimated there is evidence of a negative

causal relationship between faster BB speed and domains of children’s wellbeing. The largest effect

from a 1% increase in BB speed is for how children feel about their appearance, decreasing the score

by approximately 0.6 per cent on average. A number of potential channels are investigated as possible

mechanisms capable of explaining this phenomenon. The empirical analysis provides support for both

the ‘crowding out’ hypothesis (whereby beneficial activities are sacrificed for more time spent on the

internet) and also for the adverse effect of increased social media use.

The internet, and social media in particular, are hugely important phenomena of the past

decade. Given the extent of use among children and adolescents, concern with the potential adverse

f
oo
(and long-term) effects on children’s emotional health is increasing. The results of our analysis are

important given the central role of these platforms in children’s lives, and the fact that childhood

pr
wellbeing has been shown in previous research to have persistent effects into adult life (e.g.

Lindeboom et al. (2010), Conti and Heckman (2014) and Bertoni (2015)). Our results suggest that
e-
interventions to appropriately limit internet and social media use during childhood may help to
Pr
improve emotional health.

REFERENCES
al

Akerman, A., Gaarder, I., and Mogstad, M. (2015). The skill complementarity of broadband internet.
Quarterly Journal of Economics, 130, 1781-824.
n

Altonji, J. G., Elder, T. E., and Taber, C. R. (2005a). An evaluation of instrumental variable strategies
for estimating the effects of Catholic schooling. Journal of Human Resources, 40, 791-821.
ur

Altonji, J. G., Elder, T. E., and Taber, C. R. (2005b). Selection on observed and unobserved variables:
Assessing the effectiveness of Catholic schools. Journal of Political Economy, 113, 151-84.
Jo

Anand, P., and Roope, L. (2016). The development and happiness of very young children. Social
Choice and Welfare, 47, 825-51.
Angrist, J., and Pischke, J-S. (2009). Mostly Harmless Econometrics: An Empiricists Companion.
Princeton University Press.
Antoci, A., Sabatini, F., and Sodini, M. (2012). See you on Facebook! A framework for analyzing
the role of computer-mediated interaction in the evolution of social capital. Journal of Socio-
Economics, 41, 541-47.

25
Barnardoʹs (2015). Youth and the Internet: A Guide for Policy Makers. www.barnardos.org.uk/youth_
and_the_internet_report.pdf.
Bauernschuster, S., Falck, O., and Woessmann, L. (2014). Surfing alone? The internet and social
capital: Evidence from an unforeseeable technology mistake. Journal of Public Economics,
117, 73-89.
Bertoni, M. (2015). Hungry today, unhappy tomorrow? Childhood hunger and subjective wellbeing
later in life. Journal of Health Economics, 40, 40-53.
Beardsmore, R. (2015). Measuring National Well-being: Insights into Children’s Mental Health and
Well-being. London: Office for National Statistics.
Bhuller, M., Havnes, T., Leuven, E., and Mogstad, M. (2013). Broadband internet: An information
superhighway to sex crime? Review of Economic Studies, 80, 1237-66.

f
Booker, C., Kelly, Y., and Sacker, A. (2018). Gender differences in the associations between age

oo
trends of social media interaction and well-being among 10-15 years olds in the UK. BMC
Public Health, 18:321
Boyd, D. (2014). It’s Complicated: The Social Lives of Networked Teens. Yale University Press.

pr
Brenner, V. (1997). Psychology of computer use: XLVII. Parameters of internet use, abuse, and
addiction: The first 90 days of the internet usage survey. Psychological Reports, 80, 878-82.
e-
Brown, S., and Taylor, K. (2008). Bullying, education and earnings: Evidence from the National
Child Development Study. Economics of Education Review, 27, 387-401.
Pr
Bulman, G., and Fairlie, R. W. (2016). Technology and education: Computers, software, and the
internet. In E. A. Hanushek, S. Machin and L. Woessmann (Eds.), Handbook of the Economics
of Education, Volume 5 (pp.239-80). Elsevier.
Carr, N. (2010). The Shallows: How the Internet is Changing the Way we Think, Read and Remember.
al

Atlantic Books.
Castellacci, F., and Tveito, V. (2018). Internet use and well-being: A survey and a theoretical
n

framework. Research Policy, 47, 308-25.

ur

Children’s Society (2018a). Safety Net: Cyberbullying’s Impact on Young People’s Mental Health:
Inquiry Report. www.childrenssociety.org.uk/
Jo

Children’s Society (2018b). The Good Childhood Report. www.childrenssociety.org.uk

Chou, H. G., and Edge, N. (2012). “They are happier and having better lives than I am”: The impact
of using Facebook on perceptions of others’ lives. Cyberpsychology, Behavior, and Social
Networking, 15, 117-21.
Clark, A. E., Flèche, S., Layard, R., Powdthavee, N., and Ward, G. (2018). The Origins of Happiness:
The Science of Well-being Over the Life Course. Princeton University Press.
Clark, A. E., and Senik, C. (2010). Who compares to whom? The anatomy of income comparisons in
Europe. Economic Journal, 120, 573-94.
26
Conti, G., and Heckman, J. J. (2014). Economics of child well-being. In A. Ben-Arieh, F. Casas, I.
Frones and J. E. Korbin (Eds.), Handbook of Child Well-being: Theories, Methods and
Policies in Global Perspective (pp.363-401). Springer.
Cotton, S. R. (2008) Students' technology use and the impacts on well-being. New Directions for
Student Services, 124, 55-70.
Cowie, H. (2013). Cyberbullying and its impact on young people’s emotional health and well-being.
The Psychiatrist, 37, 167-70.
Department for Business Innovation and Skills and Department for Culture Media and Sport (2009).
Digital Britain: Final report. Cm7650.
Department for Business Innovation and Skills and Department for Culture Media and Sport (2010).
Britain’s Superfast Broadband Future.

f
Department for Digital, Culture Media and Sport (2018). Evaluation of the Economic Impact and

oo
Public Value of the Superfast Broadband Programme.
Enke, B. (2017). What you see is all there is. Mimeo. Harvard University.
Eriksen, T., Nielsen, H., and Simonsen, M. (2014). Bullying in elementary school. Journal of Human

pr
Resources, 49, 839-71.
Faber, B., Sanchis-Guarner, R., and Weinhardt, F. (2015). ICT and education: Evidence from student
e-
home addresses. CEP Discussion Paper No. 1359.
Ferrer-i-Carbonell, A., and Frijters, P. (2004). How important is methodology for the estimate of the
Pr
determinants of happiness? Economic Journal, 114, 641-59.
Franzen, A. (2003). Social capital and the Internet: evidence from Swiss panel data. Kyklos, 56, 341-
60.
Fujiwara, D., Houston, R., Keohane, K., Gramatki, I., Maxwell, C. (2018). Subjective Wellbeing
al

Analysis of the Superfast Broadband Programme. Simetrica.

Goldberg, D. P., and Williams, P. (1988). A User’s Guide to the GHQ. NFER-Nelson, Windsor.
n

Gross, E. F., Juvonen, J., Gable, S. L. (2002). Internet use and well-being in adolescence. Journal of
ur

Social Issues, 58, 75-90.

Gürtzgen, N., Nolte, A., Pohlan, L., and van den Berg, G. (2018). Do digital information technologies
Jo

help unemployed job seekers find a job? Evidence from the broadband internet expansion in
Germany. IZA Discussion Paper No. 11555.
Helliwell, J., and Huang, H. (2013). Comparing the happiness effects of real and on-line friends. PLoS
One, 8, 1-17.
Hinduja, S., and Patchin, J. W. (2010). Bullying, cyberbullying, and suicide. Archives of Suicide
Research, 14, 206-21.
HM Government (2019). Online Harms White Paper. CP57.
House of Lords (2017). Growing Up with the Internet. HL 130.
27
Institute for Public Policy Research (2014). Young People, Sex and Relationships: The New Norms.
IPPR, London.
Jackson, L. A., Fitzgerald, H. E., Zhao, Y., Kolenic, A., von Eye, A., and Harold, R. (2008).
Information Technology (IT) use and children’s psychological well-being. Cyberpsychology
and Behavior, 11, 755-8.
Kahneman, D., Diener, E., Schwarz, N. (1999). Well-being: The Foundations of Hedonic Psychology.
Russell Sage Foundation.
Kaiser Family Foundation (2010). Generation M2: Media in the Lives of 8- to 18-Year Olds. Kaiser
Family Foundation Study.
Kalmus, V., Siibak, A., and Blinka, L. (2014). Internet and child well-being. In A. Ben-Arieh, F.
Casas, I. Frones and J. E. Korbin (Eds.), Handbook of Child Well-being: Theories, Methods

f
and Policies in Global Perspective (pp.2093-133). Springer.

oo
Kleemans, M., Daalmans, S., Carbaat, I., Anshutz, D. (2016). Picture perfect: The direct effect of
manipulated Instagram photos on body image in adolescent girls. Media Psychology, 21, 93-
110.

pr
Kraut, R., Patterson, M., Lundmark, V., Kiesler, S., Mukopadhyay, T., and Scherlis, W. (1998). A
social technology that reduces social involvement and psychological wellbeing? American
Psychologist, 53, 1017-31.
e-
Kross, E., Verduyn, P., Demiralp, E., Park, J., Lee, D. S., and Lin, N. (2013). Facebook use predicts
Pr
declines in subjective well-being in young adults. PLoS One, 8, 1-6.
Kuss, D. J. and Griffiths, M. D. (2012). Online gaming addiction in children and adolescents: A
review of empirical research. Journal of Behavioural Addictions, 1, 3-22.
Levenson, J. C., Shensa, A., Sidani, J. E., Colditz, J. B., and Primack, B. A. (2016). The association
al

between social media use and sleep disturbance among young adults. Preventive Medicine,
85, 36-41.
n

Lindeboom, M., Portrait, F., and van den Berg, G. J. (2010). Long-run effects on longevity of a
ur

nutritional shock early in life: The Dutch potato famine of 1846-1847. Journal of Health
Economics, 29, 617-29.
Jo

Lohmann, S. (2015). Information technologies and subjective well-being: Does the internet raise
material aspirations? Oxford Economic Papers, 67, 740-59.
Martellozzo, E., Monaghan, A., Adler, J. R., Davidson, J., Leyva, R., and Horvath, M. A. H. (2016).
“I wasn’t sure it was normal to watch it…” A Quantitative and Qualitative Examination of
the Impact of Online Pornography on the Values, Attitudes, Beliefs and Behaviours of
Children and Young People. NSPCC. London: Middlesex University

28
Mazumder, B., and Davis, J. M. V. (2013). Parental earnings and children’s well-being: An analysis
of the Survey of Income and Programme Participation matched to social security
administrative earnings data. Economic Inquiry, 51, 1795-808.
Mendelson, A. L., and Papacharissi, Z. (2010). Look at us: Collective narcissism in college student
Facebook photo galleries. In Z. Papacharissi (Ed.), The Networked Self: Identity, Community
and Culture on Social Network Sites (pp. 251-73). New York: Oxford University Press.
Moreno, M. A., Jelenchick, L. A., Koff, R., Eickhoff, J. C., Goniu, N., Davis, A., Young, H. N., Cox,
E. D., and Christakis, D. A. (2013). Associations between internet use and fitness among
college students: an experience sampling approach. Journal of Interaction Science, 1:4.
Nikolaou, D. (2017). Does cyberbullying impact youth suicidal behaviors? Journal of Health
Economics, 56, 30-46.

f
Ofcom (2015). Children and Parents: Media Use and Attitudes Report. www.ofcom.org.uk/

oo
Ofcom (2016). Children and Parents: Media Use and Attitudes Report. www.ofcom.org.uk/
Pfeffer, F., and Schoeni, R. (2014). Intergenerational transmission of well-being. Focus, 31, 39-43.
Royal Society for Public Health (2017). #StatusofMind: Social Media and Young People’s Mental

pr
Health and Wellbeing. London.
Sabatini, F., and Sarracino, F. (2016). Keeping up with the e-Joneses: Do online social networks raise
e-
social comparisons? Munich Personal RePEc Archive 69201.
Sabatini, F., and Sarracino, F. (2017). Online social networks and subjective wellbeing. Kyklos, 70,
Pr
456-80.
Sampasa-Kanyinga, H., and Hamilton, H. A. (2015). Use of social networking sites and risk of
cyberbullying victimization: A population-level study of adolescents. Cyberpsychology,
Behavior, and Social Networking, 18, 704-10.
al

Slonje, R., Smith, P. K., and Frisén, A. (2012). The nature of cyberbullying, and strategies for
prevention. Computers in Human Behavior, 29, 26-32.
n

Tandoc Jr., E. C., Ferrucci, P., and Duffy, M. (2015). Facebook use, envy, and depression among
ur

college students: Is Facebook depressing? Computers in Human Behavior, 43, 139-46.

University of Essex (2018). Institute for Social and Economic Research, NatCen Social Research,
Jo

Kantar Public. Understanding Society: Waves 1-8, 2009-2017 and Harmonised BHPS: Waves
1-18, 1991-2009. [data collection]. 11th Edition. UK Data Service. SN: 6614.
Verduyn, P., Lee, D. S., Park, J., Shablack, H., Orvell, A., Bayer, J., Ybarra, O., Jonides, J., and
Kross, E. (2015). Passive Facebook usage undermines affective well-being: Experimental and
longitudinal evidence. Journal of Experimental Psychology: General, 144, 480-8.
Wallsten, S. (2013). What are we not doing when we’re online? NBER Working Paper No. 19549.

29
FIGURE 1: Distribution of dependent (wellbeing) variables
f
oo
pr
e-
Pr
n al
ur
Jo
FIGURE 2: Density plots of neighbourhood (LSOA) broadband speed (mbps) by wave, 𝑍𝑗𝑤

f
oo
pr
e-
FIGURE 3: Density plots of neighbourhood (LSOA) growth in broadband speed over time, ∆𝑍𝑗𝑤
Pr
n al
ur
Jo
FIGURE 4: Map showing the average broadband speed by quartile (produced in ArcGIS using Ofcom data)

f
oo
pr
e-
Pr
n al
ur
Jo
TABLE 1: Summary statistics – dependent variables, key explanatory variables and broadband speed

MEAN ST. DEV. MIN MAX

DEPENDENT VARIABLES, 𝑦, Level of Happiness with: @
School work 5.5177 1.233 1 7
Appearance 5.3448 1.431 1 7
Family 6.3804 1.021 1 7
Friends 6.2409 1.026 1 7
School 5.6366 1.445 1 7
Life 5.8497 1.183 1 7

EXPLANATORY VARIABLES, 𝑿
Child Aged 10 0.1566 0.363 0 1
11 0.1653 0.371 0 1
12 0.1706 0.376 0 1
13 0.1732 0.378 0 1

f
14 0.1695 0.375 0 1

oo
Child male 0.4999 0.500 0 1
Parent(s) employed 0.7726 0.419 0 1
Parent(s) has degree 0.3295 0.470 0 1
Single parent household 0.2233 0.417 0 1

pr
Natural logarithm of real equivalized net monthly household income 7.0404 0.512 3.24 12.07
No. of other children in household aged 0-2 0.0761 0.289 0 3
3-4 0.0884 0.298 0 3
e-
5-11 0.9310 0.880 0 5
12-15 1.0753 0.703 0 5
Parent(s) own home 0.6675 0.471 0 1
Pr
Natural logarithm of unemployment rate # 1.8261 0.444 0.26 2.80
Natural logarithm of GVA per capita # 10.0033 0.366 9.40 11.51
Natural logarithm of share of females # 3.9356 0.023 3.83 4.01
Natural logarithm of share of population 16-65 # 4.2903 0.084 4.06 4.51
al

Natural logarithm of share of population 65+ # 3.2378 0.382 1.92 4.24

Average synchronization speed (mbps), 𝑍𝑗𝑤 $ 25.6419 14.783 1.84 87.89

n

Number of children (N) 6,310

ur

Observations (NT) 13,938

@
Denotes variables that are categorical; in the empirical analysis these variables are standardized to have a mean zero and standard
Jo

deviation of unity. # Defined at the local authority district (LAD) level. $ Defined at the neighbourhood (LSOA) level. See Appendix
Table A1 for full definitions.
 TABLE 2: Relationship between changing broadband speed and baseline average neighbourhood (LSOA) characteristics
MEAN OF COVARIATE 2012 𝜓̂3 𝜓̂4 𝜓̂5 𝜓̂6
f 𝜓̂7 𝜓̂8 𝐻0 : 𝜓3 = ⋯ = 𝜓8 = 0
oo
Male -2.05 (1.80) -0.18 (0.85) 0.20 (0.87) -0.01 (0.02) -0.05 (0.12) -0.44 (0.71) 0.93, p=[0.475]
Age10 3.02 (1.35) 0.19 (0.56) -0.13 (0.39) 0.16 (0.42) 0.11 (0.21) -0.64 (0.72) 0.53, p=[0.790]
pr
Age 11 0.42 (0.16) -0.18 (0.56) -0.32 (0.89) -0.16 (0.43) 0.19 (0.35) -0.39 (0.46) 0.24, p=[0.965]
Age 12 0.78 (0.37) 0.17 (0.56) 0.25 (0.79) -0.01 (0.03) 0.55 (0.95) 0.03 (0.05) 0.32, p=[0.929]
Age 13 -1.69 (0.68) -0.29 (1.01) -0.44 (1.35) 0.31 (0.75) 0.18 (0.29) -0.44 (0.59) 0.70, p=[0.651]
e-
Age 14 0.18 (0.12) 0.07 (0.23) -0.33 (0.95) -0.23 (0.54) -0.12 (0.19) -0.65 (0.63) 0.29, p=[0.944]
0-2 children 1.02 (0.53) 0.94 (1.71) 0.53 (1.16) 0.23 (0.62) 0.34 (0.57) -0.75 (1.06) 1.69, p=[0.121]
Pr
3-4 children 0.35 (0.13) -0.29 (0.98) -0.04 (0.11) -0.91 (1.69) -0.84 (1.68) -0.98 (1.47) 0.60, p=[0.661]
5-11 children -0.29 (0.26) -0.06 (0.43) -0.15 (0.94) -0.23 (1.29) 0.09 (0.34) -0.34 (0.94) 0.56, p=[0.759]
12-15 children 0.44 (0.30) -0.04 (0.18) -0.22 (0.88) 0.13 (0.47) -0.31 (0.74) -1.07 (1.87) 0.78, p=[0.587]
al
Parent(s) own home -1.08 (0.49) 0.28 (1.11) -0.05 (0.18) 0.07 (0.23) 0.55 (1.28) -0.21 (0.34) 0.59, p=[0.741]
Parent(s) employed 3.46 (0.95) 0.13 (0.50) 0.00 (0.00) 0.73 (2.34) 0.24 (0.56) 0.44 (0.79) 1.23, p=[0.289]
Parent(s) have degree -1.72 (1.38) -0.13 (0.57) -0.22 (0.90) -0.23 (0.86) 0.04 (0.09) -0.03 (0.05) 0.58, p=[0.745]
n
Single parent household 2.04 (2.49) 0.21 (0.87) 0.32 (1.17) -0.00 (0.01) 0.25 (0.54) -0.69 (1.11) 1.60, p=[0.143]
ur
Log real equiv. hh income 0.80 (0.61) -0.36 (1.49) -0.28 (1.00) -0.09 (0.31) -0.18 (0.34) -1.21 (1.94) 1.11, p=[0.356]
Log unemployment rate -1.99 (0.70) -0.98 (1.43) 0.02 (0.02) -0.83 (1.41) -2.57 (2.28) -1.06 (1.05) 1.13, p=[0.341]
Log GVA per capita -8.60 (1.99) -7.01 (1.40) -5.87 (1.01) -6.55 (2.22) -7.62 (1.56) -7.37 (1.45) 2.53, p=[0.079]
Jo
Log share of females -3.61 (1.09) -1.43 (0.43) 1.21 (0.76) 0.57 (0.04) -3.66 (0.93) -2.68 (0.73) 1.76, p=[0.103]
Log share of population 65+ -1.25 (1.06) -6.15 (1.54) -6.16 (1.56) -7.41 (1.93) -8.78 (1.97) -8.52 (1.63) 0.95, p=[0.461]
Log share of population 16-65 2.00 (0.71) 1.72 (1.20) 2.36 (1.66) 2.39 (1.70) 2.55 (1.81) 0.28 (0.19) 1.43; p=[0.200]
F-statistic 𝐻0 : 𝛿𝑤 = 𝜓𝑤 = 0 1.02; p=0.4159
Number of LSOAs (N) 2,368
Observations (NT) 4,766
This table considers the relationship between improving BB performance and the control variables based upon their mean value in 2012, 𝑿̅𝑗,2012, including the area level controls to ascertain whether
service expansion was concentrated in those locations where the economy was growing fastest. Figures reported are coefficients and those in parenthesis are t-statistics. The final column shows an F-
statistic, which tests the null hypothesis that the parameters are jointly equal to zero across time for each covariate in turn.
 f
oo
TABLE 3: The role of selection on observables – full sample and heterogeneity
FULL SAMPLE GENDER AGE URBAN-RURAL
pr
Boys Girls Aged ≥13 Aged <13 Urban Rural
SCHOOL WORK -0.00095 -0.00188 -0.00136 -0.00089 -0.00117 -0.00017 -0.00050
e-
APPEARANCE 0.00139 0.00275 -0.00073 0.00372 0.00038 0.00013 0.00018
FAMILY -0.00097 -0.00057 -0.00152 -0.00748 0.00384 0.00005 -0.00015
Pr
FRIENDS -0.00086 -0.00132 -0.00134 0.00016 0.00073 0.00013 0.00054
SCHOOL -0.00171 -0.00031 -0.00362 0.00035 -0.00149 -0.00001 0.00041
LIFE OVERALL -0.00093 -0.00119 -0.00160 0.00093 0.00055 0.00016 0.00017
al
Number of children (N) 6,310 3,179 3,131 4,082 4,154 5,346 993
Observations (NT) 13,938 6,964 6,974 7,073 6,865 11,637 2,301
n
This table considers whether BB speed is uncorrelated with time-varying factors which are associated with psychological wellbeing, by considering selection on observables as an indication
′
ur
of selection on unobservable effects, where 𝑐𝑜𝑣(𝑿𝑖𝑗𝑤 𝜷, 𝑍𝑗𝑤 |𝛼𝑖 )/𝑣𝑎𝑟(𝑍𝑗𝑤 |𝛼𝑖 ) is shown for each outcome.
Jo
 TABLE 4: Coefficients for the reduced form relationship between wellbeing domains and broadband speed
f
PANEL A SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
oo
BB speed -0.0095 (9.70) -0.0156 (15.80) -0.0162 (15.84) -0.0114 (10.11) -0.0174 (16.37) -0.0139 (13.76)
Controls × × × × × ×
Area fixed effects × × × × × ×
Year fixed effects × × × × × ×
pr
Child fixed effects
R-squared 0.0140 0.0424 0.0417 0.0169 0.0446 0.0301
e-
PANEL B SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB speed -0.0061 (4.07) -0.0080 (5.50) -0.0076 (5.04) -0.0059 (3.54) -0.0073 (4.61) -0.0054 (3.57)
Controls
Pr
Area fixed effects × × × × × ×
Year fixed effects × × × × × ×
Child fixed effects
R-squared 0.0180 0.0527 0.0518 0.0225 0.0586 0.0419
al
PANEL C SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB speed -0.0045 (2.70) -0.0061 (3.53) -0.0026 (1.50) -0.0037 (2.92) -0.0037 (2.93) -0.0025 (1.50)
Controls
n
Area fixed effects × × × × × ×
ur
Year fixed effects
Child fixed effects
R-squared 0.0191 0.0562 0.0582 0.0236 0.0623 0.0440
PANEL D SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Jo
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB speed -0.0041 (2.40) -0.0062 (3.60) -0.0024 (1.36) -0.0037 (2.91) -0.0035 (2.89) -0.0024 (1.43)
Controls
Area fixed effects
Year fixed effects
Child fixed effects
R-squared 0.0283 0.0612 0.0609 0.0314 0.0659 0.0494
Number of children (N) 6,310
Observations (NT) 13,938
Controls include: age of child; number of children in household aged 0-2; number of children in household aged 3-4; number of children in household aged 5-11; number of children in
household aged 12-15; whether live in a single parent household; whether parent(s) own home; whether either parent has a degree; log equivalized real household income and the following
LAD level covariates, the natural logarithm of: the unemployment rate; GVA per capita; share of females; share of population over 65; and share of population aged 16-65. In Panel C we
also include year fixed effects and in Panel D 320 Local Authority indicators, i.e. area fixed effects, are also included.
 f
TABLE 5: Coefficients for all covariates (from Table 4 panel C)
oo
SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB speed -0.0045 (2.70) -0.0061 (3.53) -0.0026 (1.50) -0.0037 (2.92) -0.0037 (2.93) -0.0025 (1.50)
pr
Age 10 0.1927 (1.34) 0.3014 (2.17) 0.4255 (3.13) 0.4556 (2.99) 0.4965 (3.40) 0.3111 (2.17)
Age 11 0.2237 (1.92) 0.2638 (2.33) 0.3618 (3.31) 0.3495 (2.81) 0.4205 (3.52) 0.3009 (2.55)
Age 12 0.1480 (1.73) 0.1116 (1.36) 0.2014 (2.44) 0.2985 (3.29) 0.2955 (3.34) 0.2034 (2.37)
e-
Age 13 0.1198 (1.99) 0.0086 (0.15) 0.1234 (2.14) 0.1858 (2.91) 0.1417 (2.27) 0.1299 (2.15)
Age 14 0.0732 (1.94) -0.0274 (0.79) 0.0351 (0.93) 0.1049 (2.58) 0.0354 (0.92) 0.0494 (1.29)
0-2 children -0.0688 (1.15) -0.0138 (0.24) -0.0477 (0.87) 0.0279 (0.42) -0.0161 (0.25) -0.0874 (1.57)
3-4 children 0.0407 0.0131 -0.0589 (0.99) 0.0217 (0.33) -0.0268 (0.45) 0.0038
Pr
(0.76) (0.24) (0.07)
5-11 children -0.0477 (1.13) -0.0155 (0.39) -0.0885 (1.94) -0.0305 (0.65) -0.0236 (0.55) -0.0262 (0.62)
12-15 children -0.0283 (0.95) 0.0292 (1.06) -0.0024 (0.07) 0.0195 (0.60) -0.0117 (0.39) 0.0076 (0.25)
Parent(s) own home 0.1076 (1.02) -0.0848 (0.96) 0.0433 (0.44) 0.0157 (0.18) 0.0458 (0.49) 0.0277 (0.30)
Parent(s) employed -0.0729 (1.59) -0.0456 (1.20) -0.0699 (1.72) -0.0274 (0.64) -0.1028 (2.17) -0.1042 (2.54)
al
Parent(s) have degree -0.1787 (1.97) -0.0512 (0.46) -0.0222 (0.21) 0.1375 (1.26) -0.0992 (1.00) 0.0520 (0.45)
Single parent household -0.0790 (0.91) -0.0606 (2.04) -0.1039 (1.03) 0.1671 (1.52) -0.1366 (1.43) -0.2203 (2.67)
Log real equiv. hh income 0.0024 (0.08) -0.0318 (1.18) 0.0288 (0.93) -0.0513 (1.52) -0.0200 (0.68) -0.0285 (0.95)
Log unemp. rate 0.0529 (1.46) 0.0236 (0.69) -0.0296 (0.83) 0.0537 (1.31) 0.0241 (0.66) -0.0928 (2.45)
n
Log GVA per capita -0.1815 (0.55) 0.1639 (0.55) -0.4040 (1.25) -0.4028 (1.14) -0.0287 (0.08) -0.2203 (0.73)
ur
Log share females 0.4910 (0.49) -0.1815 (0.19) 0.5060 (0.49) -0.1570 (0.14) 1.6044 (1.64) 0.3306 (0.34)
Log share pop. 16-65 0.1618 (0.59) 0.4988 (2.00) -0.2863 (1.07) -0.0083 (0.03) 0.2458 (0.90) 0.0394 (0.15)
Log share pop. 65+ -0.0925 (0.85) -0.2073 (2.11) -0.0139 (0.14) -0.0550 (0.47) -0.1504 (1.42) -0.0845 (0.77)
R-squared 0.0191 0.0562 0.0582 0.0236 0.0623 0.0440
Jo
Number of children (N) 6,310
Observations (NT) 13,938
Controls as per Table 4 Panel C.
 f
TABLE 6: Coefficients for the reduced form relationship between wellbeing domains and broadband speed – non-linearity
oo
SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB 10 and <20 mbps -0.0598 (1.19) -0.0693 (1.43) -0.0546 (1.06) -0.0378 (0.76) -0.0331 (0.66) -0.0794 (1.65)
pr
BB 20 and <30 mbps -0.0961 (1.56) -0.1135 (1.95) -0.0388 (0.63) -0.0318 (0.50) -0.0335 (0.06) -0.1091 (1.80)
BB 30 and <40 mbps -0.1605 (2.18) -0.1327 (1.94) -0.0161 (0.22) -0.0222 (0.29) -0.0290 (0.39) -0.1189 (1.73)
BB 40 and <50 mbps -0.1906 (2.28) -0.2238 (2.87) -0.1136 (1.36) -0.1261 (1.46) -0.0705 (0.84) -0.1876 (2.28)
e-
BB 50 and <60 mbps -0.2483 (2.65) -0.2938 (3.30) -0.0948 (1.02) -0.1327 (1.34) -0.1029 (1.08) -0.2240 (2.44)
BB 60 mbps -0.2635 (2.41) -0.3671 (3.35) -0.0562 (0.50) -0.2103 (1.76) -0.2033 (1.78) -0.2300 (2.18)
Controls
Area fixed effects
Pr
Wave fixed effects
Child fixed effects
R-squared 0.0288 0.0619 0.0618 0.0323 0.0667 0.0503
Number of children (N) 6,310
al
Observations (NT) 13,938
Notes: 1 BB refers to broadband speed which is defined as a set of binary indicators for whether the average synchronization speed is between: 10 and <20 megabits per second (mbps);
20 and <30 mbps; 30 and <40 mbps; 40 and <50 mbps; 50 and <60 mbps; and 60 mbps, respectively. 2 Controls include: age; number of children in household aged 0-2; number of
n
children in household aged 3-4; number of children in household aged 5-11; number of children in household aged 12-15; whether live in a single parent household; whether parent(s) own
ur
home; whether either parent has a degree; log equivalized real household income; and year of interview. We also include 320 Local Authority indicators, i.e. area fixed effects, and the
following LAD level covariates, the natural logarithm of: the unemployment rate; GVA per capita; share of females; share of population over 65; and share of population aged 16-65.
Jo
 TABLE 7: Coefficients for the reduced form relationship between wellbeing domains and broadband speed – heterogeneity
f
PANEL A: Males SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
oo
BB speed -0.0017 (0.74) -0.0051 (2.29) -0.0013 (0.53) -0.0015 (0.60) -0.0041 (1.63) -0.0001 (0.05)
R-squared 0.0185 0.0339 0.0446 0.0247 0.0428 0.0241
Observations (NT) 6,964
pr
PANEL B: Females SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB speed -0.0057 (2.27) -0.0063 (2.43) -0.0031 (1.19) -0.0052 (1.90) -0.0024 (0.87) -0.0040 (1.50)
e-
R-squared 0.0514 0.1047 0.0890 0.0503 0.0966 0.0941
Observations (NT) 6,974
PANEL C: Aged <13 SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Pr
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB speed -0.0026 (0.75) -0.0068 (2.02) -0.0060 (1.96) -0.0060 (1.53) -0.0006 (0.20) -0.0020 (0.58)
R-squared 0.0164 0.0494 0.0533 0.0196 0.0344 0.0293
Observations (NT) 6,865
al
PANEL D: Aged 13 SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB speed -0.0137 (4.26) -0.0090 (3.10) -0.0029 (0.89) -0.0072 (2.86) -0.0006 (0.17) -0.0061 (1.92)
R-squared 0.0292 0.0183 0.0281 0.0226 0.0294 0.0307
Observations (NT) 7,073
n
ur
PANEL E: Rural SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB speed -0.0052 (1.12) -0.0017 (0.35) 0.0042 (0.94) -0.0047 (0.83) 0.0006 (0.15) -0.0011 (0.21)
Jo
R-squared 0.0244 0.0813 0.0581 0.0350 0.0802 0.0603
Observations (NT) 2,301
PANEL F: Urban SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
BB speed -0.0039 (2.06) -0.0085 (4.41) -0.0040 (2.69) -0.0039 (1.72) -0.0052 (2.41) -0.0036 (1.95)
R-squared 0.0274 0.0580 0.0637 0.0318 0.0634 0.0471
Observations (NT) 11,637
Controls include: age of child; number of children in household aged 0-2; number of children in household aged 3-4; number of children in household aged 5-11; number of children in
household aged 12-15; whether live in a single parent household; whether parent(s) own home; whether either parent has a degree; log equivalized real household income; and year of
interview. We also include 320 Local Authority indicators, i.e. area fixed effects, and the following LAD level covariates, the natural logarithm of: the unemployment rate; GVA per capita;
share of females; share of population over 65; and share of population aged 16-65.
 TABLE 8: Potential mechanisms and children’s wellbeing
PANEL A: Transitions 1 SCHOOL WORK
Coef t-stat
APPEARANCE
Coef t-stat
FAMILY
Coef t-stat
f FRIENDS
Coef t-stat
SCHOOL
Coef t-stat
LIFE
Coef t-stat
oo
UE to employed -0.0374 (0.82) 0.0130 (0.31) -0.0580 (1.20) -0.0214 (0.41) -0.0194 (0.40) 0.0092 (0.21)
R-squared 0.0335 0.0510 0.0461 0.0405 0.0561 0.0454
Observations (NT) 5,668
pr
PANEL B: Duration of UE spell 1 SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
UE duration -0.0459 (2.63) -0.0234 (1.49) -0.0270 (1.52) -0.0189 (1.09) -0.0452 (2.61) -0.0326 (0.68)
e-
R-squared 0.0365 0.0479 0.0520 0.0448 0.0661 0.0513
Observations (NT) 5,047
PANEL C: Parental wellbeing 1 SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Pr
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
GHQ12 -0.0110 (2.32) -0.0110 (2.35) -0.0067 (1.41) 0.0032 (0.58) -0.0042 (0.91) -0.0120 (2.54)
R-squared 0.0302 0.0613 0.0646 0.0319 0.0677 0.0519
Observations (NT) 13,227
PANEL D: Key stage 2 2 SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
al
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
GCSE attainment 0.0100 (0.38) 0.0227 (0.78) -0.0097 (0.32) -0.0103 (0.39) -0.0062 (0.22) -0.0035 (0.13)
R-squared 0.5285 0.5281 0.4879 0.5186 0.4881 0.5221
Observations (NT) 778
PANEL E: Key stage 4 2
n
SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
ur
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
Hours spent online -0.0001 (0.01) 0.0043 (0.32) 0.0026 (0.20) -0.0016 (0.13) 0.0002 (0.02) 0.0107 (0.79)
R-squared 0.3376 0.3484 0.3497 0.3269 0.3210 0.3379
Observations (NT) 1,096
Jo
PANEL F: Activities 3 SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
Number of activities 0.0542 (3.58) 0.0447 (3.26) 0.0036 (0.25) 0.0557 (3.58) 0.0221 (1.48) 0.0229 (1.79)
R-squared 0.0426 0.0979 0.0632 0.0513 0.0928 0.0754
Observations (NT) 7,389
Notes: 1 controls include: age of parent(s); age of child; number of children in household aged 0-2; 3-4; 5-11; 12-15; whether live in a single parent household; whether parent(s) own home; whether
either parent has a degree; log equivalized real household income; and year of interview. Local Authority indicators, i.e. area fixed effects, are also included as well as the following LAD level
covariates, the natural logarithm of: the unemployment rate; GVA per capita; share of females; share of population over 65; and share of population aged 16-65. 2 Regressions are at LAD level based
on mean characteristics including area fixed effects, but excluding the age of the child and parent(s). 3 controls include: age of child; number of children in household aged 0-2; 3-4; 5-11; 12-15;
whether live in a single parent household; whether parent(s) own home; whether either parent has a degree; log equivalized real household income; and year of interview. Local Authority indicators,
i.e. area fixed effects, are also included as well as the following LAD level covariates, the natural logarithm of: the unemployment rate; GVA per capita; share of females; share of population over
65; and share of population aged 16-65. Each model includes child fixed effects, with the exception of Panels D and E.
 f
TABLE 8 (cont.): Potential mechanisms and children’s wellbeing
PANEL G: Social media 3 SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
oo
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
Hours spent online -0.0331 (3.95) -0.0284 (3.77) -0.0232 (2.87) -0.0038 (0.42) -0.0368 (3.94) -0.0374 (4.54)
R-squared 0.0372 0.0586 0.0620 0.0386 0.0733 0.0624
pr
Observations (NT) 9,948
PANEL H: Social media 3 SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
e-
5 hours spent online -0.1646 (4.24) -0.1348 (3.81) -0.1310 (3.59) -0.0343 (0.84) -0.1697 (4.07) -0.1593 (4.31)
R-squared 0.0302 0.0607 0.0624 0.0309 0.0679 0.0517
Observations (NT) 9,948
Pr
PANEL I: Social media 3 SCHOOL WORK APPEARANCE FAMILY FRIENDS SCHOOL LIFE
Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat Coef t-stat
Belong to social media site -0.0812 (3.29) -0.0653 (2.90) -0.0386 (1.65) 0.0538 (2.10) -0.0131 (0.54) -0.0163 (0.66)
R-squared 0.0288 0.0596 0.0609 0.0313 0.0650 0.0491
Observations (NT) 13,938
n al
ur
Jo
 f
TABLE 9: How broadband influences potential mechanisms
oo
EFFECT OF BROADBAND SPEED ON POTENTIAL MECHANISMS
Coef t-stat R-squared Observations (NT)
PANEL A: Transitions 1 -0.0011 (0.83) 0.0875 5,668
pr
PANEL B: Duration of UE spell 1 0.0091 (1.60) 0.2400 5,047
e-
PANEL C: Parental wellbeing 1 -0.0044 (0.99) 0.0191 13,227
Pr
PANEL D: Key stage 2 – SATs 2 0.1832 (9.52) 0.8742 778
PANEL E: Key stage 4 – GCSEs 2 -0.3761 (8.25) 0.5534 1,096
al
PANEL F: Number of activities -0.0100 (3.01) 0.0810 7,389
PANEL G: Social media – hours spent online 3 0.0088 (2.73) 0.1298 9,948
n
ur
PANEL H: Social media – 5 hours spent online 3 0.0016 (2.42) 0.0650 9,948
PANEL I: Social media – Belong to social media site 3 0.0045 (5.37) 0.0184 13,938
Jo
Each row represents a separate regression model; notes as per Table 8.
 f
APPENDIX: TABLE A1 – Variable definitions
VARIABLE NAME DEFINITION
oo
DEPENDENT VARIABLES, 𝑦 The next few questions are about how you feel about different aspects of your life. The faces
express various types of feelings. Below each face is a number where ‘1’ is completely happy and
pr
‘7’ is not at all happy. Please tick the box which comes closest to expressing how you feel about
each of the following things...
e-
Pr
We reorder this variable so that it is increasing in happiness, i.e. ‘1=not at all happy’ and
al
‘7=completely happy’
YPHSW (School work) Your school work?
YPHAP (Appearance) Your appearance?
YPHFM (Family)
n Your family?
ur
YPHFR (Friends) Your friends?
YPHSC (School) The school you go to?
YPHLF (Life) Which best describes how you feel about your life as a whole?
BROADBAND SPEED
𝑍𝑗𝑤 Jo Defined at the neighbourhood level via Lower Layer Super Output Area (LSOA)
The average synchronisation speed of existing broadband connections based on official Ofcom
figures, defined as mega-bits per second (mbps).
APPENDIX: TABLE A1 – Variable definitions (cont).
VARIABLE NAME DEFINITION
f
oo
EXPLANATORY VARIABLES, 𝑿
Male 1=child is male; 0=female
pr
Age10 1=child aged 10; 0=otherwise
Age 11 1=child aged 11; 0=otherwise
Age 12 1=child aged 12; 0=otherwise
e-
Age 13 1=child aged 13; 0=otherwise
Age 14 1=child aged 14; 0=otherwise
Pr
Parent(s) employed 1= mother and/or father in paid employment or self-employed; 0=other labour market state
Parent(s) degree 1= mother or father has a degree qualification or equivalent; 0=other qualification or none.
Single parent household 1=child is in a single parent household; 0=otherwise
Real equiv. hh income Natural logarithm of real equivalized net household monthly income in 2009 prices
al
HH child 0-2 Number of children in household aged 0-2
HH child 3-4 Number of children in household aged 3-4
HH child 5-11 Number of children in household aged 5-11 (excluding respondent)
n
HH child 12-15 Number of children in household aged 12-15 (excluding respondent)
ur
Parent(s) own home 1=parent owns home outright or on a mortgage; 0=other housing tenure state
Unemployment rate Natural logarithm of the local area district (LAD) unemployment rate
Gross value added (GVA) per capita Natural logarithm of the local area district (LAD) gross value added per head
Jo
Share of females Natural logarithm of the local area district (LAD) female share of population
Share of population over 65 Natural logarithm of the local area district (LAD) share of population aged over 65
Share of population 16-65 Natural logarithm of the local area district (LAD) share of working age population