Impact of Broadband Quality on Median Income and Unemployment: Evidence from Sweden

Based on a unique and exhaustive database, including micro-level cross-sectional data on 23 million observations over nine years, from 2009 to 2017, we assess whether broadband quality has an impact on income and unemployment reduction. Overall, the results do not show any significant effect of download speed on either income or the unemployment rate. However, after distinguishing between educational attainment and the city size, we obtained heterogeneous results. While we highlight a substitution effect between low-skilled workers and broadband in smaller cities, we also show that broadband quality has a positive impact on unemployment reduction for low-skilled workers in bigger cities. However, the model predicts a negative effect of broadband quality on both the median income and the unemployment rate in areas having a higher proportion of college graduates. This result tends to support the analyses showing that, with the progress made in machine learning, artificial intelligence and the increasing availability of big data, job computerization is expanding to the sphere of high-income cognitive jobs.

This project has been mainly funded by the European Investment Bank (EIB) in the frame of the STAREBEI (STAges de REcherche BEI-EIB research internships) programme. This is a programme that provides grants to universities in order to finance junior researchers carrying out research projects. I would like to express my very great appreciation to Anders Bohlin (EIB) for its valuable support on this project.
My grateful thanks are also extended to Internetstiftelsen, who has been partly financing this report and has been contributing with a large and rich dataset on broadband quality. I direct my special thanks to Anna Caracolias for a considerable help and to Pamela Davidsson and Göran Andersson for their help with getting the data from Bredbandskollen.
I would also like to express my gratitude to the Swedish Telecommunication Regulator, PTS (Post och Telestyrelsen), which contributed to this projects by purchasing the socio-economic data from the national statistics institute (SCB) and for the provision of data on broadband deployment in Sweden.
I would like to extent my thanks for Maria Elmqvist (head of department) and Per Lundin (head of division) from Chalmers University of Technology for letting me use Chalmers' facilities and IT services to conduct my research.
A special thanks to Erik Bohlin, professor at Chalmers University of Technology and project supervisor for his valuable support and great help.

Introduction
Very high-speed broadband networks are seen as a key enabler for socio-economic development.
Their roll-out comes along with the development of advanced digital services in all areas of the society. As an example, the progresses made in artificial intelligence (AI) are opening doors to an unprecedented level of progress in health engineering and transportation. The digitalization of the society also creates benefits for the citizens. It gives them greater opportunities to participate in the civil society, to enjoy a broad range of services at home, such as welfare and healthcare services. In addition, job searches are facilitated increasing opportunities to find a work in lines with the person's competences. Digitization also creates opportunities for companies and entrepreneurship. It bridges geographical distances enabling more communication, exchanges and commerce. Thus, digitalization is considered important for the economy and for democracy.
Sweden, which is a sparsely populated country characterized by a long coastline with archipelagos, extensive forests and numerous lakes, is though one of the leading countries in terms of broadband deployment. Sweden has been particularly active in the roll-out of next generation broadband technologies, both fixed and mobile. The government is aiming for a completely connected Sweden, by 2025, with 98% of the population which should have access to broadband at a minimum speed of 1 Gbit/s both at home and at the workplace. The remaining 1.9% should have access to connections of at least 100 Mbit/s, and 0.1% at speeds of at least 30 Mbit/s. In addition, by 2023 everyone should have access to reliable and high-quality mobile services. The growth of mobile services is also dependent on the existence of a very high-speed broadband infrastructure connecting the base stations together.
To reach these ambitious targets, Sweden is not only relying on private initiatives but also on the commitment of the public sector. As a forerunner, Sweden launched its ambitious and interventionist national ICT infrastructure program already in 2000. In contrast with other European countries, which defined broadband as an Internet connection of at least 256 Kbit/s, Sweden had already then a more stringent definition including only transmission capacities of at least 2 Mbit/s in both directions. In line with this interventionist vision, local collectivities have traditionally played a significant role in the roll-out of broadband technologies, which has had an impact on the digitalization level of rural areas. Swedish municipalities lead the way to a competitive and well-functioning broadband market by being facilitators, landowners or active operators acting in the market in competition with private actors.
Sweden's high performances are far above the European average. The share of broadband connections of at least 100 Mbit/s is the second highest of the European Union. Besides, Sweden ranks second in the use of Internet services by households since 2017 1 . Given the recent deployment of next generation access networks (NGN), little empirical research has investigated the economic impact of very high-speed Internet on the society. Considering that Sweden is a frontrunner in very high-capacity connectivity in Europe, it is also a good candidate to assess the effects of broadband on socio-economics variables related to the households.
In this report, we quantify to which extent broadband quality, measured in terms of download speed, impacts income and unemployment in both urban and rural areas. This will enable us to investigate whether the roll-out of NGN networks is a way to answer the challenges encountered in the cities and in rural areas. More precisely, one challenge which is mostly encountered in rural areas and smaller cities is the lack of companies, which leads to a lack of jobs, a higher unemployment rate and a lower income, i.e. a lower purchasing power for households. Companies tend to locate in bigger cities where they can benefit from the local infrastructures, find their workforce, especially highly educated experts. In these areas, they can also find other companies which could potentially become their suppliers, clients, intermediaries. Another challenge is the depopulation of rural areas and smaller cities as those living in economically deprived areas may decide to relocate to areas with better job prospects. People tend to locate in areas where they can receive an education, find a workplace and benefit from the local infrastructures and the cultural structures. They have more chances to find a job in an area with more companies and especially to find a job they are interested in. Therefore, non-attractive areas may see their population decreasing and ageing; their cultural life and sociability places reducing; their public services being maintained to a minimum level and their grocery stores disappearing. As such the digital divide is also triggering the territorial and social divides.
These challenges could potentially be answered by the deployment of very high-speed broadband infrastructures. The possibility to work from home, to benefit from a reliable and sufficient internet connection to start a business could be a way to give a new life to non-attractive areas 1 European Commission Digital Economy and Society Index (DESI 2017-DESI 2018-DESI 2019 and help others to remain attractive. Some studies have also been showing that two otherwise similar urban areas could see their attractiveness impacted by the absence of very high-speed internet, especially in terms of attractiveness for companies (See for example McCoy et al. (2018) and Hasbi (2020)).
This study relies on a unique and rich dataset including 23 million measurement tests realized by Internet users all over Sweden. The tests have been performed using the "Bredbandskollen" test managed by the Swedish Internet Foundation. It provides us with micro-level cross-sectional data covering approximately 700 Swedish localities, located in 287 municipalities, over 9 years, from 2009 to 2017, giving us a wide coverage of Sweden. 2 The results highlight a positive effect of broadband quality on unemployment reduction in localities with a lower proportion of highly educated inhabitants. However, improvement in broadband quality is shown to have a detrimental effect on both the median income and the unemployment rate in localities with a higher proportion of highly educated inhabitants. In addition, the model highlights heterogeneous effects while taking into account the municipality size.
To the best of my knowledge, this is one of the first papers to estimate, at a granular local level in both urban and rural areas, the impact of very high-speed broadband network on socioeconomic variables related to the demand, the median income and the unemployment rate. The results provide policy-makers with better insights on the role of very high-speed broadband for local economic growth and social development.
The remainder of the paper is organized as follows. Section 3 discusses the relevant literature on the effect of broadband on income and unemployment. Section 4 provides an overview of the state of the broadband market in Sweden. Section 5 presents the data. Section 6 introduces the econometric framework. Section 7 presents the estimation results. Finally, section 8 concludes.

Literature Review
As internet connectivity becomes ubiquitous and comprehensive datasets become available, many economists have tried to find answers to the 1987 Solow paradox: "you can see the computer age everywhere but in the productivity statistics". In the last decade, an extensive range of macro-level studies bring empirical evidence on the positive effect of broadband adoption on economic growth (see Holt et al. (2009), Greenstein et al. (2011 and Bertschek et al. (2013) for comprehensive literature reviews). It is widely accepted that there is a positive relation between broadband adoption, broadband availability and economic growth, especially as measured by the GDP, at the national and regional level. A positive relation has also been found among others by Gillett et al. (2006), Crandall et al. (2007) and Ford et al. (2006), between broadband availability and employment at the macro-level. They find that communities with broadband experienced a more rapid growth in employment and a faster firm growth, especially in ITintensive sectors than non-broadband communities. Gruber et al. (2014) show that the economic benefits that would derive from the achievement of the objectives of the 2020 Digital Agenda for Europe outweigh the costs of investment.
They show that the economic benefits mostly spill over to users and to the national economy highlighting the rationale for public subsidies in the roll-out of broadband networks. Another stream of literature has explored the relation between broadband and consumer surplus bringing positive evidence on the existence of a positive correlation. For example, Greenstein et al. (2012) point out that countries with large Internet economy have experienced increased consumer surplus because they simultaneously experienced large improvements in broadband quality and a decline in real prices.
The Great Recession of 2007-2009 has, however, led to challenge the common acceptance of the positive relationship between GDP and employment as digital technologies become more prevalent in all areas of the society. Analysing the situation in the US, Brynjolfsson et al. (2011) shed some light on the appearance of a weakening in this relation showing that, even though GDP is increasing in the US, there is no net employment creation. They highlight a decline in the quantity of labour demanded bringing back the 1800s concept of technological unemployment introduced in the economic theory by Ricardo and popularized by Keynes in the 1930s. "This means unemployment due to our discovery of means of economising the use of labour outrunning the pace at which we can find new uses for labour", Keynes, 1930. 3 The paradox of today is that we see everywhere that very high-speed broadband affects positively economic growth and social development at the local level, but the empirical evidences are very limited, divergent and ambiguous, especially as regards rural areas. A burgeoning literature has explored the relation between broadband and income. Whitacre et al. (2014) find that broadband adoption, availability and download speeds have an impact on economic growth in rural areas. They highlight a positive impact on unemployment reduction and on median household income. They also show that rural areas with high levels of download speeds tend to attract more creative class workers and to have a lower poverty level. Autor et al. (2013) and Akerman et al. (2015) focus on job polarization and wage inequality.
They show that broadband availability and adoption impacts positively the productivity of highskilled workers, acting as a skill complement and lowers the productivity of unskilled workers, acting as a substitute for routine tasks. In 2009, Goos et al. observe a trend towards the polarization of the labour market between high-income cognitive jobs and low-income manual jobs leading to a hollowing-out of middle-income routine jobs. In addition, Autor et al. (2013) confirms the existence of a job polarization effect as low-skilled workers being specialized in routine tasks reallocate into the service sector. Brynjolfsson et al. (2011) highlight that the polarization of the market between high-skilled and low-skilled workers leads to a stagnation of the median income with an increase in wage inequalities. Adding to that, Boland et al. (2015) show that broadband deployment in rural areas has stimulated wage increase and employment growth in the service industries, but not in the good industries. The complementary effect between broadband and high-skilled workers has also been highlighted by Hasbi (2020). She finds that on average areas covered by very high-speed broadband and having a high proportion of college graduates are more attractive for firms. However, and in line with the existence of a job polarization effect, she shows a higher positive effect of broadband on the creation of companies operating in the service sector in areas with a higher proportion of low-skilled workers. Unlike Autor et al. (2013) and Akerman al al. (2015), Houngbonon et al. (2019) show that the impact is larger at the bottom of the income distribution.
As regards broadband effect on employment, Czernich (2011 and for German municipalities and Jayakar et al. (2013) for eight States in the US, find no evidence that broadband availability reduces unemployment. On the contrary, Kolko (2012) highlights that broadband expansion is associated with population growth and employment growth. However, he also finds that the average wage and the employment rate are unaffected by broadband expansion.
Kolko underlines that the positive effects of broadband are stronger in industries that rely more on ICTs and in less densely populated areas. Atasoy (2013) confirms this result by showing that gaining access to broadband in a county is associated with an increase in the employment rate, especially in rural areas. She underlines also a complementary effect between broadband and skilled workers with college educated workers encountering larger effects. On the contrary, Canzian et al. (2015) find no impact on employment creation. But they find a positive effect of broadband diffusion on firm performance, especially in rural areas. in areas such as machine learning, machine vision and artificial intelligence and with the availability of increasing amount of big data, non-routine tasks, such as legal writing, truck driving, medical diagnoses, some bank and financial services are becoming computerized. They highlight that 47% of total US employment is at high-risk of being computerized. Jobs that requires high dexterity, creativity or social intelligence are the least susceptible to be computerized. In addition, they show that wages and educational attainment are strongly negatively correlated with the probability of computerization. In a similar study for Sweden, Fölster (2014) shows that it is 53% of total employment, which is susceptible to be computerized within 20 years in Sweden. Brynjolfsson et al. (2011) argue that advances realized in what digital technology can do will inevitably continue increasing the number and types of jobs susceptible to be computerized.
The use of speed to measure the economic impacts of broadband is considered as more and more important in the literature. Middleton et al. (2013) argue that along with broad- This could partially be explained by the lack of sufficient data.

Background on Broadband Roll-Out in Sweden
Very high-speed broadband should be accessible for everyone at the latest in 2025. This is the goal set by the Swedish government in its 2016 broadband strategy. More precisely, 98% of the population should have access to broadband at a minimum speed of 1 Gbit/s both at home and at the workplace. The remaining 1.9% of the population should have access to connections of at least 100 Mbit/s, and 0.1% at speeds of at least 30 Mbit/s. These ambitious targets should be put in context with the geographical reality of Sweden and its previous broadband strategies.
Sweden is a long and sparsely populated country with long coastlines and inhabited archipelagos. Forests cover 69% of the territory and only 3% of the land surface is inhabited. The population density is fairly low with small cities and villages scattered over the whole territory.
In 2019, the average density of population was 25.4 inhabitants per square kilometre with big differences within the country 4 . 87% of the population lives in urban areas, mostly in the south of Sweden, principally in the three biggest cities: Stockholm, Gothenburg and Malmö 5 . Figure   1 provides an overview of the population density in Sweden. Sweden's geographic characteristics could be seen as a hinder for an extensive broadband coverage having a detrimental effect on the deployment costs of next generation broadband networks. This problem of sparsity is especially visible in rural areas where the potential number of customers reachable is lower 6 .
To reach the government's goals both the private and the public sectors work together. Sweden has a long tradition of public involvement in the deployment of broadband infrastructures.
In 1999, Sweden was the first European country to define a national broadband strategy. In its vision of "an information society for all", the government emphases the importance of having a nationwide and reliable broadband infrastructure to ensure a healthy and well-functioning broadband market. Being a strategic infrastructure, it has been decided that it was the State which should bear the responsibility of its roll-out to ensure its availability throughout the whole territory. The broadband plan also provided a definition of broadband by including only transmission capacities of at least 2 Mbit/s in both directions, far above the European reference of 256 Kbit/s. The plan contained two main strategies.
The first one consisted in restricting Telia's monopoly position, the incumbent operator, by opening-up its network by the mean of local loop unbundling (LLU). LLU is a regulatory tool enabling alternative broadband operators to provide their services by using the network of the incumbent. Therefore, they do not need to deploy first their own infrastructure. LLU is a tool to ensure service-based competition (in opposition to infrastructure-based competition) and to create sustainable competition in the provision of broadband services that should then lead to lower prices for municipal residents. The second main strategy was to provide government funding of 4,150 million Swedish crowns (approximately 400 million Euro) to promote infrastructure deployment between 2000 and 2005. These funds were allocated to three different levels of network hierarchy: the backbone networks, the regional and local networks. The public investment in the base network aimed at building a competitive national backbone to Telia's network.
The municipal networks build at the regional and local levels were open and operator-neutral.
Municipal as well as private operators could provide broadband services to the end-users. The national broadband strategy resulted in the availability of a nationwide broadband network, with 99% of households having access to a DSL connection.
In Sweden, municipalities can promote competition in the broadband area in different ways: (1) by promoting the expansion of broadband networks; (2) by providing support in land planning and (3) by investing in network deployment. Municipalities involvement in broadband deployment is a way to meet the broadband targets defined by the government. Their role is to ensure that everyone gets access to the broadband infrastructure. This means that they are responsible for the definition of local broadband strategies and have at their disposal a broad range of tools. They can be a facilitator by easing the administrative processes, for example by making it easier to get specific authorizations. They can also be a landowner or play the role of a coordinator between different actors. They can also be a financial partner or the main investor. A municipality can also be an active operator acting in the market in competition with private actors. To ensure the realization of the government strategy and frame competition in the broadband market, municipalities rely on the deployment or upgrade of different technologies, often in combination, such as optical fibre, copper upgrade with xDSL technologies, cable upgrade, wireless technology, mobile broadband or satellite.
The strategies defined by the municipalities are based upon their specific characteristics and differ from one another. Some of them results from a municipal vision, some others are the product of civil initiatives. The need for broadband deployment can stem from a public consultation or from the civil society. 7 High-speed broadband is seen as a tool to preserve a dynamic economy and counteract the depopulation movement. In all cases, before public funding are granted, municipalities must first define a broadband project that is then submitted to the regional county administrative board for approval 8 .
Funding can be delivered to any legal person who intend to build a broadband network for more than just its own needs or its own business. Therefore, all types of associations, organizations, companies, municipalities and other authorities can receive financial support. Public funding is only directed towards the roll-out of broadband networks that meet the definition of a next generation access network, i.e. broadband networks with high transmission capacity that are of very high-speed, reliable and with symmetric signal transmission. Fibre optical networks are eligible to receive support, along with upgraded copper networks via VDSL, upgraded cable networks: DOCSIS 3.0 or wireless networks (under certain conditions). Other technologies could be eligible if it is possible to justify that they meet the requirements for NGA networks.
In line with the European and national regulation, a prerequisite for a project to be eligible for broadband funding is the existence of a market failure, i.e. the lack of private initiative. When a municipality builds a municipal network, the national regulator PTS recommends that the municipality grants passive access to the network by the use of dark fibre. The solution of an open net is a way to increase the customer's choice of an Internet service provider and enables more innovations. 9 There are more than 200 municipal networks in Sweden. Figure

Broadband market in Sweden
Sweden is one of the most connected country in the EU. In 2008, the first year in the dataset, broadband penetration in Sweden was already high, exceeding by far the European average.
The proportion of households with internet access was of 84%, the second highest of the EU27. and optical fibre. Figure 4 shows the evolution of fixed broadband subscriptions per technology between 2008 and 2018.
10 source: All the statistics come from Eurostat. 11 VDSL is a DSL technology providing faster transmission speeds than standard DSL for short copper lines.

Data
Data on broadband speed comes from Internetstiftelsen. The Swedish Internet Foundation, Internetstiftelsen, is an independent organization that works for the development and improvement of the internet in Sweden. One of the main focus is to raise end-users' awareness and knowledge about the Internet. To that purpose Internetstiftelsen is managing an online tool called Bredbandskollen aiming at measuring access connection speed. This tool is free of use and allows everyone across the country to test its internet connection in terms of download and upload speed in Mbit/s and latency in seconds. We got a dataset with broadband measurements over 9 years, from 2009 to 2017. These data have been collected at an established aggregation level in Sweden (tätort). Tätort is defined as urban areas with contiguous buildings with no more than 200 metres between houses and at least 200 residents. There are around 2,000 tätorter in Sweden comprising around 87% of the population.
Using a dataset on broadband speed instead of on the type of Internet technologies has a number of potential benefits. First, the importance of taking into account Internet speed to measure the impact of broadband on different socio-economic variables has been highlighted in the literature. With these data, we have information on the demand and therefore, on the actual Internet access speed of the households. This gives us an overview of the network quality and enables us to avoid biases related to the different physical characteristics of a technology. For example, the connection speed on a DSL line is dependent on the line length. Longer is the line, lower is the speed. As such, for the same technology, the broadband speed experienced by the end user can vary significantly. In addition, it avoids biases related to the use of the theoretical speed marketed by the operators, which is the maximal theoretical speed reachable under specific circumstances and not necessarily the actual speed experienced by the customers.
However, the dataset includes some sort of positive self-selection in which the user expresses an interest and demand in order to use the measurement test (Bredbandskollen). There are two main reasons to test an Internet connection. The first one is that the user is experiencing a problem, such as a slow download or upload speed or a high latency. To correct this bias, all observations with abnormally long latency or very slow download speed or upload speed are dropped out. The second main reason to use the test is when an Internet user has changed his or her subscription. However, because these behaviours repeat every year, it is reasonable to consider that it doesn't introduce a bias in the data. There will be each year tests on new (and most probably faster) Internet subscriptions. All in all, because it is the user expressing an interest in testing its Internet connection, our data is closer related to expressed demand rather than supply.
Finally, the data set is very extensive. Over the 9 years, the number of actual tests conducted through Bredbandskollen amounts to more than 300 million use cases, of which a third had usable location coordinates. After dropping observations related to unrealistic broadband speeds or latencies, only data on tests performed the second week of each month have been kept. The estimations have been performed on a representative sample of 23 million observations (tests).
Considering the large number of measurements, we get a wide coverage of Sweden.
Data on unemployment and income come from the Swedish statistical agency, SCB. The data on unemployment represents the number of openly unemployed aged 18 to 64 registered as job seekers. Data on income from employment and business are available for the population aged 20 and over. So we have, In model (1) Y it refers to the median income whereas in model (2)  In other words, operators are more likely to upgrade network quality or deploy a next generation access network in areas with a higher quality of demand. In the meantime, areas with better broadband infrastructures are more likely to attract households with a higher income.
As argued by Lobo et al. (2020), the use of lagged explanatory variables is a common approach in the literature to mitigate problems related to endogeneity in the social science.
Following the literature, I control for pre-existing location characteristics and pre-existing labour market characteristics by using one-year lagged variables. Robustness checks are performed with lags of 1 year and 3 years, which give similar qualitative results. Nevertheless, one can suspect that the estimation results might suffer from an upward bias.
Omitted variables may also be a potential source of endogeneity. For example, operators may have higher incentives to deploy a very high-speed broadband network or upgrade network quality in areas in which they can benefit from a more favourable tax regime or in which there is higher demand for faster broadband services. To mitigate this problem, I follow the econometric literature by using time-varying and time invariant fixed effects.
Using (1), the median income can be derived: With income it the yearly income from employment and business in locality i at time t, which is expected to be influenced by broadband quality hbb it−1 , by the locality size, i.e. lo-cal population denoted by ln pop it−1 and population density density it−1 . 13 The unemployment rate, unempl it−1 , is also likely to impact the median income as well as the education level of the population, measured by the proportion of the population having a diploma from superior education, ln uni diploma it−1 . It has also been shown for the US (Lobo et al. (2020)), that areas with a higher proportion of non-whites are more likely to experience a higher unemployment rate. This potential impact is captured by non EU it−1 , which measures the proportion of inhabitants born outside Scandinavia and outside the European Union. In addition, the model includes information on the number of new establishments establishment it−1 operating in the locality. All socio-demographic variables are estimated in locality i at time t − 1. One variable is measured at a different geographical level, as the model includes information on the population of the municipality to which the locality belongs, denoted by ln pop m jit−1 .
Using (1), the unemployment rate could be derived as follows: unempl it = α + δ hbb it−1 + β 1 ln pop it−1 + β 2 ln pop m jit−1 + β 3 density it−1 + β 4 ln income it−1 + β 5 non EU it−1 + γ 1 ln uni diploma it−1 With unempl it , the number of openly unemployed aged 18 to 64 registered as job seekers in locality i at time t, defined as a function of broadband quality, the local and municipal population, the local population density, the proportion of inhabitants born outside Scandinavia and outside the EU, the educational level, the number of new establishments. In addition, the yearly income, ln income it−1 , is expected to influence the unemployment rate.
It has been shown in the literature that some benefits of broadband were found to be complementary with high-skilled labour forces and highly educated inhabitants (Akerman et al. (2015) and Hasbi (2020)). To assess whether broadband quality affects differently areas with (1) a higher, (2) a lower proportion of highly educated inhabitants, an alternative specification to 1 taking into account educational attainment is defined.
With high skilled it−1 taking two values: 1 if the proportion of inhabitants having obtained a diploma from superior education is higher than the median and 0 otherwise.
The potential effects of broadband may also differ across municipality size. Political decisions are not made at the DeSo level but at a higher administrative level, such as the region or the city. Cities of different sizes have also different budgets or possibilities to grow and develop. This is why, it is important to compare localities belonging to the same type of cities. In our database, 50 % of the population lives in municipalities with less than 135,500 inhabitants, while 10% lives in municipalities with less than 19,000 inhabitants or more than 920,600 inhabitants. To analyse whether the effects of broadband quality materialize differently depending on the municipality size, another specification is defined: with municipality size jit the size of municipality j, to which locality i belongs, at time t − 1, which can take 5 values: -municipalities with a number of inhabitants equal or lower than 18,937 inhabitants (rank 1); -municipalities with a number of inhabitants comprised between 18,938 and 44,813 (rank 2); -municipalities with a number of inhabitants comprised between 44,814 and 132,536 (rank 3); -municipalities with a number of inhabitants comprised between 132,537 and 354,349 (rank 4); -municipalities with more than 354,349 inhabitants (rank 5).

Impacts of Broadband Quality on the Society
In this section are presented both the estimation results for the impact of broadband quality on income (6.1) and on unemployment ( for localities having a high-skilled population, i.e. the proportion of inhabitants having obtained a diploma from superior education is higher than the median. While tables 4 and 7, (low skilled), displays the estimation results for localities having a low-skilled population, i.e. the proportion of inhabitants having obtained a diploma from superior education is lower than the median.
The database is further divided in 5 sub-samples (ranks 1 to 5) based on the size of the city to which the locality belongs. Therefore, it enables us to make within-group estimations and see how broadband quality affects income and unemployment within a group. Column 1 of each table presents the estimation results for the whole sample and column 2 for localities located in smaller municipalities of rank 1. Columns 3 and 4 shows the results for localities located in medium-sized municipalities of ranks 2 and 3 respectively. Column 5 presents the results for localities located in bigger cities of rank 4. In addition, tables 4 and 7 column 6 shows the results for localities located in the biggest cities, those having a population higher than the average: rank 5.

Impacts of Broadband Quality on Income
The exhaustive estimation results are displayed in Annex B in tables 12 to 14. Table 2 do not show any significant effect of broadband quality on median income. This result holds irrespective of the size of the city to which the locality belongs. Therefore, we cannot conclude that broadband quality has an impact on income. The results, displayed in table 12, highlight that the median income is more likely to be higher in more populated localities as well as in localities with a lower unemployment rate.
Nevertheless, we also notice that more densely populated localities as well as localities in which the proportion of inhabitants born outside Scandinavia or outside the EU is high, are more likely to have a lower median income. In line with the literature, the results show a positive effect of educational attainment on income. The median income is more likely to be higher in localities having a higher proportion of college graduates.

High-skilled Localities
The second part of the results concerns localities characterized by a high proportion of highly educated inhabitants. Overall, the model predicts a negative effect from broadband on the median income of high-skilled workers. However, after distinguishing between the different municipality sizes, this result holds only for localities belonging to a medium-sized municipality.
Otherwise, in smaller medium-sized municipalities broadband quality is found to have a positive impact on the income of high-skilled workers.
More specifically, the first column of table 3 reveals a negative impact of broadband quality on median income for high-speed and very high-speed broadband (speeds over 30 Mbit/s and especially over 100 Mbit/s). The results show that the negative impact of broadband quality increases with download speed. On average, the median yearly income in localities with highspeed broadband (between 30 and 50 Mbit/s) is 430 Swedish crowns (about 43 euro) lower than in localities without broadband. This difference in median income increases to 1,060 Swedish crowns (about 106 euro) in localities with very high-speed broadband (speed over 100 Mbit/s).
However, this result hides differences depending on the size of the city to which the locality belongs. In that respect, column 2 shows that there is no significant impact of broadband quality on income in localities located in smaller cities (with less than 18,937 inhabitants). A positive effect, increasing with download speeds, is found in localities located in a medium-sized city Similarly, column 5 shows that localities located in bigger cities (rank 4) that have a low broadband quality (under 5 Mbit/s) are more likely to have a slightly higher yearly median income than other localities of the same rank without broadband: 138 Swedish crowns (about 14 euro) more per year. However, localities located in areas with higher broadband speed do not encounter any significant impact of a higher broadband quality.
In accordance with the general picture, localities located in larger medium-sized cities of rank 3 are more likely to encounter a negative impact of broadband quality on median income: higher are the download speeds, lower is the median income. On average, the median yearly income in localities with broadband speed between 5 Mbit/s and 50 Mbit/s is of about 1000 Swedish crowns (about 100 euro) lower compare to localities of rank 3 without broadband. This difference increases to around 1200 Swedish crowns (about 120 euro) in localities with very high-speed broadband (over 100 Mbit/s). A bit less than a third of the localities are located in a municipality of rank 3. This explains the general negative effect found for the estimation of the whole sample. 14 Table 3: Impact of broadband quality on income in localities characterized by a high proportion of highly educated inhabitants.

Low-skilled Localities
The third part of the results concerns localities characterized by a low proportion of highly educated inhabitants (inferior to the median). All in all, the model doesn't predict any effect of broadband quality on the income of low-skilled workers. However, after taking into account the city size, the model shows a negative effect of broadband on the income of low-skilled workers in smaller municipalities.  Overall, for both "high-skill" and "low-skill" groups, the results, displayed in Annex B, show that the median income is more likely to be higher in more populated localities as well as in localities with a lower unemployment rate. However, localities having a higher proportion of inhabitants born outside Scandinavia or outside the EU are more likely to have a lower median income.

Impacts of Broadband Quality on Unemployment
The exhaustive estimation results are displayed in Annex B in tables 15 to 17.
Similarly to the previous estimations for income and at one exception,  Overall, the results, displayed in table 15, highlight a positive effect of the size of the municipal population and of the median income on unemployment reduction. However, and in the same way as for the previous estimations, the proportion of inhabitants born outside Scandinavia or outside the EU has a negative impact on the unemployment rate. Having a look at the whole sample, we also observe a negative impact of population density. However, when the localities are divided in different categories in function of their city size, we no longer observe any significant effect of population density.
In addition, the results highlight a negative effect of the local population on the unemployment rate: more populated is a locality, more likely it is to have a higher unemployment rate.
However, after distinguishing between the different city sizes, we observe that this negative effect only holds for localities belonging to a small or medium-small municipality (ranks 1 and 2).
As previously, the models predict a positive effect of educational attainment on unemployment reduction.

High-skilled Localities
The second part of the results concerns localities characterized by a high proportion of highly educated inhabitants (superior to the median). Overall, the model predicts an increase in the unemployment rate of high-skilled workers in areas with better broadband quality. However, after distinguishing between the city size, the model highlights a complementary effect between broadband quality and high-skill jobs in smaller and in bigger municipalities.
More precisely, the first column of table 6 highlights a non-linear effect of broadband speed on the unemployment rate. Unlike what could have been expected, the results show only a positive effect on unemployment reduction for low-speed broadband. As such, we find an overall negative effect of broadband quality on the unemployment rate. In localities with basic broadband (under 5 Mbit/s), the unemployment rate is 0.77 percentage points lower than in localities without broadband.
On the contrary, localities with higher broadband quality are more likely to experience a higher unemployment rate of about 1 to 2.75 percentage points higher than in localities without broadband. This result is in line with the negative impact of broadband quality on income we found in areas with a high proportion of highly educated inhabitants. An explanation could be that a new substitution effect between certain types of high-skill jobs and new ICT technologies is appearing. As highlighted by Brynjolfsson et al. (2011) andFrey et al. (2017), the progress made in machine learning and artificial intelligent as well as the availability of big data has led to the development of more sophisticated programs in all kind of sectors using data. For example, one of them is the bank and insurance sector. The implementation of new applications for e-medicine has increased the use of distance monitoring and diagnose techniques enabling doctors to help more patient in a lower time frame. The considerable improvement made in smart systems enables to substitute for some cognitive (routine and non-routine) tasks realized by high-skilled workers.
After taking into account the city size, we find that broadband quality has heterogeneous effects on unemployment. Columns 2 and 5 show that localities located in smaller cities, with less than 18,937 inhabitants, and localities located in bigger cities of rank 4, benefit positively from a higher broadband quality. Higher is the download speed, lower is the unemployment rate.
As regards the localities located in smaller cities (rank 1), the results highlight an increasing positive effect of broadband quality on unemployment reduction. In localities with basic broadband (under 5 Mbit/s), the unemployment rate is 1.20 percentage points lower than in localities of rank 1 without broadband, while in localities with higher broadband speed (over 5 Mbit/s and especially over 100 Mbit/s), the unemployment rate is approximately 2 to 2.2 percentage points lower than in localities without broadband. These results are in line with Atasoy (2013), who highlighted the existence of a positive association between broadband and employment especially in rural areas along with a complementary effect between broadband and education.
With respect to bigger municipalities (rank 4), we observe only a significant effect of broadband quality on unemployment reduction in localities with basic broadband and in localities with high-broadband speed broadband over 50 Mbit/s. In these localities, the unemployment rate is approximately 0.1 percentage point lower than in the localities without broadband. The positive effect of broadband quality is however lower in areas with ultra-fast broadband, download speeds over 100 Mbit/s, which experienced an unemployment rate of about 0.04 percentage point lower than the other localities of the same group which don't have broadband.
In addition, the results highlight that localities with a higher proportion of highly educated inhabitants located in medium-sized municipalities (ranks 2 and 3) do not experience any effect of broadband quality on unemployment reduction. There is no significant difference in the unemployment rate in localities without broadband or in localities with high-speed broadband.

Low-skilled Localities
The third part of the results concerns localities characterized by a low proportion of highly educated inhabitants (inferior to the median). All in all, the model predicts a complementary effect between broadband quality and low-skilled jobs. However, this result holds only for medium-sized municipalities. A substitution effect is highlighted between better broadband quality and low-skill jobs in smaller municipalities.
More specifically, it stems from the results that, in general, localities which experience higher download speeds are more likely to have a lower unemployment rate. Except from localities with ultra-fast broadband, download speeds over 100 Mbit/s, where the results do not show any significant effect. Localities with broadband have on average an unemployment rate which is between 2 and 4 percentage points lower than in localities without broadband. This result confirms the job polarization effect highlighted by Autor et al. (2013). As broadband acts as a substitute to low-skilled workers to perform routine tasks, they reallocate into the service sector. A sector which is expected to benefit the most from ICT technologies (Hasbi (2020)).
Potentially, the absence of significant effect on income found earlier could also result from the fact that these new jobs created in the service sector are mainly low-paid jobs.
After distinguishing between the city size, the results show that localities located in smaller cities (of rank 1) tend to have a higher unemployment rate in areas with better broadband quality. This substitution effect had already been highlighted in the literature, especially by Akerman et al. (2015).
Similarly to the positive effect found for the whole sample, the results highlight a positive effect of broadband speed on unemployment reduction in localities located in a medium-sized city, but only for broadband speeds under 30 Mbit/s. Therefore, we do not observe any positive effect of high-speed or very high-speed broadband on unemployment reduction in these localities.
Localities located in a medium-sized city of rank 2 and having broadband speed over 5 Mbit/s and over 10 Mbit/s are more likely to have a lower unemployment rate than similar localities of rank 2 having no broadband. The difference in the unemployment rate is of about 2.4 and 1.8 percentage points respectively. The results also highlight a positive effect of basic broadband (under 5 Mbit/s) on unemployment reduction in medium-sized cities of ranks 3 and 4. The unemployment rate is more likely to be lower by roughly 2 percentage points in these localities compared to localities of the same category without broadband. Localities of rank 4 located in areas with broadband speed between 5 Mbit/s and 10 Mbit/s are also more likely to encounter a lower unemployment rate of about 2.3 percentage points. However, there is no significant effect of broadband quality on the unemployment rate of localities located in the largest cities. Overall, the results show, for both the "high-skill" and "low-skill" groups, that the unemployment rate is more likely to be higher in more populated localities. However, larger is the municipal population, more likely it is to have a lower unemployment rate. In the same way, localities with a higher median income are more likely to experience a lower unemployment rate.
Similarly, to the results for the median income, localities located in a city of rank 2 having a higher proportion of inhabitants born outside Europe are more likely to encounter a higher unemployment rate.

Conclusion
Based on a unique and rich dataset, provided by the Swedish Internet Foundation, we assess whether broadband quality has an impact on income and unemployment reduction. We exploit micro-level cross-sectional data covering approximately 700 localities over 9  The results underline a clear distinction between localities having a high and a low proportion of college graduates. Broadband quality is predicted to have a detrimental effect on both the median income and the unemployment rate in localities with highly educated inhabitants.
Whereas, it has no significant effect on the median income and a positive effect on unemployment reduction in localities with a lower proportion of highly educated inhabitants.
Interestingly, this result provides some evidence of the existence of a substitution effect between new ICT technologies and more complex tasks performed by high-skilled workforce.
The evolution of ICT technologies towards more intelligent systems capable of performing more complicated analytic tasks leads to a second wave of substitution to the detriment of highly educated workers. Therefore, the results are supporting the analyses from Brynjolfsson et al. (2011) andFrey et al. (2017), which show that the progress realized in digital technologies, especially in machine learning and AI combined with the increasing availability of big data, has led to computerize cognitive routine and non-routine tasks (that are usually performed by high-skilled workers).
After distinguishing between the different municipality sizes, the results tend to confirm the existence of a substitution effect between new broadband technologies and low-skill jobs but only in localities located in smaller municipalities. In medium-sized municipalities, the results tend to confirm the existence of a job polarization effect. As in a process of destruction creation, low-skilled workers, whose routine tasks have been first automated are now reallocating to the tertiary sector, where new jobs have been created. In addition, the model predicts that broadband quality has a complementary effect with high-skill jobs in smaller municipalities and to a lower extent in bigger municipalities.
With this report, we show that the relation between broadband quality and income or unemployment is not straightforward but rather is based on complex mechanisms. Both positive and negative effects are encountered. In accordance with our results, broadband quality can be seen as a tool to increase digital inclusion and reduce digital divide between people and territories, with a positive effect on unemployment reduction in medium-sized municipalities, where low-skilled jobs are predominant and in smaller municipalities, where high-skill jobs are predominant. In future research, when more detailed information becomes available on mobile data, it may become possible to estimate to which extent mobile broadband complement the effect of fixed broadband. Besides, it may also be interesting to compare how broadband quality affects company creation and entrepreneurship, especially in rural areas.