Are slow internet connections limiting home working opportunities?

The outbreak of COVID-19 in early 2020 resulted in many governments imposing restrictions on people ’ s mobility and ordering workers to work from home where possible. As the pandemic has progressed, these restrictions have eased but many hope to maintain a higher frequency of working from home than they had before the pandemic. However, home working often requires a good internet connection. Those without good connections may have limited options for home working and may therefore have to travel to the office more often. In this paper, we examine whether people living in areas with faster internet connections are more likely to have increased their frequency of telecommuting compared to those in areas with slower connections. We also examine whether home working reduces overall travel demand, or whether there is a rebound effect. We find that faster connections are associated with an increase in the frequency of home working. As expected, these workers do not travel as much for work purposes. There is no evidence of a rebound effect for these workers.


Introduction
Advances in Information and Communication Technologies (ICTs) over the past decades have reshaped the world.The rapid fall in the cost of communicating over vast distances led economist Frances Cairncross to declare The Death of Distance (Cairncross, 1997).Part of this was the idea that jobs would disperse from cities and workers would be able to work from a variety of different locations, a practice sometimes referred to as teleworking or telecommuting.
The term telecommuting is thought to have been introduced by Jack Nilles in 1973, who believed that taking work to the workers rather than the other way around could be a way to reduce traffic congestion.Results about the impact of telecommuting on travel demand have been mixed.Hook et al. (2020) present a systematic review of the literature on the impact of telecommuting on energy use (of which travel demand is a substantial component).Most of the studies they consider show a reduction in energy use, though a significant minority show neutral effects or even an increase in energy use.
There are, however, other potential benefits from teleworking.Fonner and Roloff (2010) found that high intensity teleworkers are more satisfied than their office-based counterparts.For instance, teleworking may help to achieve a better work-life balance.Baard and Thomas (2010) reported improved job satisfaction as well as productivity gains.In China, Bloom et al. (2015) found that working from home could increase productivity.In addition, there is the potential to improve access to employment for people with disabilities (Schur et al., 2020).Anderson et al. (2015) found that there can be substantial benefits from home working, but note that the effects may vary between different types of people.Lara-Pulido and Martinez-Cruz (2022) show that workers have a substantial willingness to pay in order to be able to work at least some of the time from home.
Measuring the prevalence of teleworking can be difficult due to a lack of agreement on the definition of the term (Mokhtarian et al., 2005).Some studies focus on workers who work mainly from home.For example, using the Sixth European Working Conditions Survey (EWCS) Ojala and Pyöriä (2018) found that only around 3% of wage/salary earners in the EU teleworked mainly from home.Anderson et al. (2015) estimated the share of workers in the US who worked at least some hours from home, and found that in 2011 around one quarter of Americans worked at least some hours from home each month.Similarly high shares were reported in Sweden and Finland.
There is renewed interest in teleworking in light of the COVID-19 pandemic.The spread of the virus in early 2020 resulted in many governments around the world imposing various restrictions on their citizens.Over 80% of the global workforce lived in countries with full or partial lockdowns (ILO, 2020).Among these restrictions was a requirement for workers to work from home where possible.This led to an unprecedented rise in working from home.Okuyan et al. (2022) noted that as of February 2020, 46% of businesses in the US had implemented teleworking policies.This creates an interesting and useful natural experiment.
Many of the requirements to work from home were subsequently relaxed.However, there is evidence that many wish to incorporate more frequent home working beyond the pandemic.In a study in Melbourne, Australia, Jain et al. (2022) found that post-pandemic working from home could increase by 75%.Currie et al. (2021) also suggested increased home working in the future.Barrero et al. (2021) suggested that 20% of workdays will be supplied from home post-pandemic compared to around 5% before.Sanchez et al. (2021) estimated that globally, around 20% of jobs could be performed from home.
However, the choice to work from home is not evenly distributed across workers.While it is estimated that 37% of US jobs could be done entirely from home (Dingel and Neiman, 2020), certain jobs are better suited to this than others (Okuyan et al., 2022;Budnitz and Tranos, 2022).Access to appropriate equipment is also important in determining whether a person works from home (Jain et al., 2022).During the pandemic, it became clear that access to a high-speed internet connection was essential for most people to be able to work effectively from home (Lai and Widmar, 2021;Ford et al., 2021).
There is already evidence that access to a high-speed internet connection increases the likelihood of people working from home (Han, 2021).However, there is no generally agreed upon definition of highspeed internet.We extend the current literature by posing two research questions: 1) Are people living in areas with faster internet connections more likely to have increased their telecommuting frequency relative to its pre-pandemic level?2) Do people who have increased their telecommuting frequency travel less than people who have not?It is particularly important to better understand these relationships at the current juncture.Levels of home working have increased, but some people may be excluded from these benefits if they live in areas with inadequate digital infrastructure or if internet connections are not affordable.
We make use of a 2021 travel survey from the Puget Sound region of Washington state, US and a series of regression models.We find evidence that higher speed connections in a person's neighbourhood are positively associated with increases in telecommuting frequency relative to pre-pandemic telecommuting frequencies.Furthermore, we find that people who have increased their telecommuting frequency tended to travel less far for work and work-related purposes.We do not find any increase in non-work travel for this group.
The paper is structured as follows.Section 2 provides a review of studies examining impacts of access to the internet, variations in the speed of access, and impacts of telecommuting on travel demand.In Section 3 we outline the data, methods and models used in the paper.Results are presented in Section 4 with some concluding remarks made in Section 5. Athanasiadou and Theriou (2021) noted that several different terms are often used interchangeably in the literature to discuss the same phenomenon e.g., homeworking, telehomeworking, telecommuting, remote working, virtual work, electronic home working, e-work, distributed work, home-anchored work, and telework.During the pandemic it became common to refer to work from home, often abbreviated to WFH (Loo and Huang, 2022;Mehta, 2021;Galanti et al., 2021).We consider studies naming any of these topics to be potentially relevant to our study.

Impacts of access to the internet
Access to a high-speed internet connection has been shown to have many benefits.It can improve student performance (Sanchis-Guarner et al., 2021;Hampton et al., 2020), stimulate the birth of new enterprises (Conroy and Low, 2022), improve employment outcomes (Zuo, 2021;Lobo et al., 2020;Dettling, 2017;Bai, 2017), improve public health and racial justice (Early and Hernandez, 2021), and even improve fertility (Billari et al., 2019).The COVID-19 pandemic and the subsequent response to it has generated renewed interest in the relationship between internet access and telecommuting.
Telecommuting and its impacts have long been of interest to researchers.Towards the end of the 20th Century, Handy and Mokhtarian (1996) discussed the potential for telecommuting.At that time, they noted that the telephone was one of the most vital pieces of equipment.In 1992, only 30% of telecommuters made use of a modem to connect to the internet.In the time since then, access to a high-speed internet connection has become essential for home working (Lai and Widmar, 2021;Ford et al., 2021;Budnitz and Tranos, 2022;Kaushik and Guleria, 2020;Cotterill et al., 2020).
During the pandemic, Ford et al. (2021) found that slow connection speeds and unreliable connections were among the most frequently mentioned challenges reported by home workers.In an application of the theory of planned behaviour, Jain et al. (2022) found that access to relevant materials (e.g., ICT) was crucial in determining people's intention of working from home post-pandemic.Lai and Widmar (2021) noted that while inadequate access to the internet had long been considered an irritation, during the pandemic it became a crisis for many.
Inadequate internet access is not evenly distributed across space or the population.This contributes to what is often referred to as a digital divide (Scheerder et al., 2017).Internet access is usually subject to an urban/rural divide, with rural areas having lower quality connections and infrastructure (Phillipson et al., 2020;Ali, 2020;Salemink et al., 2017;Whitacre and Mills, 2007).Sarrasanti et al. (2020) noted that women typically have less exposure to digital technologies, thus creating a potential gender gap.This gap is found to be worse in developing countries and low-income countries (Kaushik and Guleria, 2020;Singh, 2017;Fuchs, 2009).Digital divides have also been found to exist with respect to age and education (Elena-Bucea et al., 2021;Hong et al., 2020) as well as race and income (Gao and Hayes, 2021).Sanders and Scanlon (2021) noted that the digital divide is a social justice as well as a human rights issue, and that minorities are particularly likely to be excluded from reliable high-speed internet connections.

Internet speed
Much of the literature discusses internet access, or high-speed internet access.The term is often ill defined and sometimes is used only to indicate a difference between a broadband connection and a slower dial-up connection (Dettling, 2017;Savage and Waldman, 2005).As technology has advanced, the distinction between dial-up and broadband has become less important.Martins and Wernick (2021) discussed current and future requirements for broadband connections.They noted that it is not only important to consider the download speed, but also the upload speed, latency and packet loss.The relative importance of these characteristics will vary according to the application.For home working, they suggested that an upstream and downstream speed of around 250 Mbps will be required by 2025 and that low latency and packet loss will be very important.Kaushik and Guleria (2020) recommended that a reliable and fast internet connection is essential for effective home working.Good connections can be particularly important for applications such as group video calls (Amankwah-Amoah et al., 2021).Cotterill et al. (2020) noted that during the pandemic employees experienced difficulties with slow internet connections.Inadequate connections also caused problems for the academic sector as it moved many of its activities online during the pandemic (Corbera et al., 2020).
The results with respect to the role played by speed are somewhat mixed.For example, Bai (2017) studied employment impacts of broadband access at the county level.They found that access to D.P. McArthur and J. Hong broadband is important, but that the speed didn't seem to have an effect.Lai et al. (2020) studied the willingness to pay for internet connections in Indiana, US.They found a positive willingness to pay (WTP) for speed, though don't differentiate between upstream and downstream speed.The reliability of the connection was found to be an important attribute.Liu et al. (2018) looked at willingness to pay for different aspects of broadband i.e., upload speed, download speed, data caps, and latency.They found a highly concave relationship for bandwidth.The WTP for different aspects depends on the usage type.For instance, people who transfer files have higher WTP for upload speeds.Gamers were more sensitive to latency than other groups.
Differentiating between different aspects of the internet connection may be important.For example, Dahiya et al. (2021) studied the performance of internet connections during the pandemic.They note that many ISPs reported the internet performed well during the pandemic despite the increase in demand.However, there was a large rise in the number of complaints.The authors suggested that part of the reason for this may be that ISPs focus on downstream speeds.These performed well during the pandemic.However, upstream speeds suffered far more.They suggested that during the pandemic, upstream speeds may have become more important and that it may be important to focus more on them.
It is worth noting that measuring the different aspects of internet connectivity can be challenging, and that even measuring speed is not entirely straightforward (Feamster and Livingood, 2020).Bronzino et al. (2021) discussed some of the limitations of different data sources for measuring speed.They noted, for instance, that survey data often gives very crude measures of access often only recording whether someone has broadband or not.Some studies adopt a different approach.For instance, Barrero et al. (2021) asked people to rate their internet connection quality based on a five-point scale detailing what percentage of the time the person can work from home without interruption caused by their connection.They found that people who rated their internet connection quality as higher indended to work from home on more days post-pandemic.

Work from home and travel demand
Our second research question asks whether people who work form home more often have reduced travel demand.Previous results on this topic are mixed, and there is substantial variability in the robustness of the methodological approaches taken (Hook et al., 2020).In a study in India, Nayak and Pandit (2021) found that work related travel decreased during the pandemic and that there was no compensatory increase in other travel.Shabanpour et al. (2018) studied the Chicago area and found that increased home working can reduce congestion, vehicle miles travelled, and pollution.Jaff and Hamsa (2018) found that encouraging telecommuting among women in Kuala Lumpur can reduce the amount of travel they undertake.
Several other studies found that telecommuters tend to travel more.In the UK, De and Melo (2018) found that teleworking increases travel demand, and particularly travel by car.Similarly, Silva and Melo (2017) found that teleworking increases travel demand in single worker households.Several other studies also found a positive relationship between telecommuting and travel demand, sometimes caused by teleworkers having longer commute distances (Zhu et al., 2018;Zhu and Mason, 2014;Zhu, 2012).
In South Korea, Kim (2017) concluded that telecommuting can likely play only a limited role in reducing travel demand.They noted that other members of a household with a telecommuter may also have their travel patterns impacted.They did note that the departure times of telecommuters tended to be more flexible, and that telecommuting may therefore help to reduce congestion.Chakrabarti (2018) highlighted the fact that while telecommuting may help reduce travel demand, that the effect will not be automatic and that other policies may have to be implemented e.g., pedestrian infrastructure, suitable housing close to workplaces, and mixed land-use.

Summary
Previous studies have generally indicated that access to high-speed internet is an important determinant of home working.The responses to the COVID-19 pandemic saw a substantial increase in the volume of home working, and there is evidence that the level is unlikely to return to pre-pandemic levels.There is also evidence that poor internet connections posed challenges to those working or studying from home.However, there is no evidence about how the speed of available internet connections is associated with increases in telecommuting frequency.
Results about the impact of telecommuting on travel demand are mixed, with studies showing positive, negative and no effects.Given that the pandemic has increased home working so drastically, it is not clear whether the results which were relevant for pre-pandemic telecommuters are applicable to the current, expanded cohort of telecommuters.

Data and methods
Our first question is to establish whether there is an association between the speed of internet connections offered in an area and the likelihood of a person increasing the frequency of their telecommuting relative to pre-pandemic levels.Binomial logistic regression models will be used to test this hypothesis.For our second question, we wish to test whether travel demand is associated with changes in telecommuting frequency.This will be tested using Poisson regression models.All data processing and analysis is carried out using R [R Core Team (2022)]1 .

Data sources
We make use of the 2021 travel survey conducted by the Puget Sound Regional Council 2 (PSRC).The council is a planning organisation for the metropolitan area of the city of Seattle, WA.The area has a population of around 4.3 million inhabitants.Data are collected from households across the King, Kitsap, Pierce and Snohomish counties.The survey is conducted every two years.The most recent edition, from 2021, was collected though an online survey.Respondents reported one day's travel designed to capture travel on a typical weekday.Address-based sampling was used with people being invited via mail to participate in the survey.A total of 6482 people from 2793 households were included in 2021.See RSG (2019) for more details on the survey.
Data on available internet speeds is taken from the Federal Communications Commission (FCC) 3 .The data refers to December 2020.It is derived from information filed by Internet Service Providers (ISPs) on the maximum advertised speeds (upstream and downstream) offered to different census blocks.

Data processing
We begin by processing the data held in the person file.We keep only data from the 2021 survey.We include only people who are of working age and who are: 1) employed (full-time), employed (part-time), or selfemployed4 .Our analysis requires two dependent variables: one to look at whether a person has changed telecommuting frequency and the other to look at the distances travelled.
Our second dependent variables are extracted from the trips file.We sum the total distance travelled depending on the type of trip.If a trip's origin or destination was recorded as "work", then we class it as a work trip.Otherwise, it is classed as a non-work trip.
Next, we process the household PSRC data.We filter the data to include only the 2021 survey.The household file gives the census tract of the respondent, and a variable indicating whether the household has access to broadband.The variable records whether the household has access to broadband, whether they have the option of having broadband but don't have it, they have neither access nor the option of access, or they don't know.We recode this as a binary variable where a value of one indicates that they have access to broadband and zero indicates that they do not.
The data on broadband speeds are then processed.We extract records for the state of Washington.Next, the data is restricted to only operators offering services to consumers.The census tract's geographic identifier is exacted from the census block code.Only records relating to tracts where survey respondents live are retained.The remaining records are aggregated at the census tract level.We take the average of the maximum advertised upload and download speeds across all providers for each tract measured in Megabits per second (Mbps).
The datasets are then merged.We derive three variables.The first two are adjusted upload and download speeds.We derive these by multiplying the average speeds offered in the census tract by the dummy variable indicating whether the household has access to broadband.Therefore, if a household indicates that they have access to broadband, then they are assigned the average upload and download speeds from their census tracts.If they indicate that they do not have access to broadband, they are assigned upload and download speeds of zero.
The first dependent variable we create indicates whether someone has increased their telecommuting frequency.This is done using two variables.The first asks on how many days in the previous week the person telecommuted.The following options were available: None/ never, 1-2 days, 3-4 days, 5 + days.They were also asked their telecommute frequency pre-COVID.Here, they could select: Not at all, Less than monthly, 1-3 times per month, 1-2 days a week, 3-4 days a week, 5 + days a week.Some were not asked the question, for example if they were not employed pre-COVID.If someone was not asked the question about pre-pandemic telecommuting, we counted them among the people who did not telecommute.Though their reasons for not telecommuting were different, they were nevertheless not telecommuting.Respondents were said to have increased their telecommuting frequency if they indicated that in the previous week they telecommuted more than their pre-COVID telecommuting frequency e.g., someone who telecommuted 1-2 days last week and indicated that they previously telecommuted less than weekly would be counted as having increased their telecommuting frequency.
Finally, we extract a number of control variables.We begin with the person file.Industry5 is included as opportunities for home working vary between industries (Bartik et al., 2020;Barrero et al., 2021).Gender identity is included, with previous studies generally showing that men are less interested in home working (Barrero et al., 2021;Nguyen, 2021).Age is often included as a determinant of home working though does not always have a significant effect (Bloom et al., 2015;Barrero et al., 2021;Loo and Wang, 2018;Nguyen, 2021).We also include the distance between a person's home and workplace.This has been shown to have a broadly positive correlation with home working (De and Melo, 2018;Loo and Wang, 2018;Nguyen, 2021).From the household file we extract household income.Income has been shown to have a positive relationship with home working interest and intensity (Barrero et al., 2021;Loo and Wang, 2018;Mannering and Mokhtarian, 1995;Singh et al., 2013).We also include whether the household has children aged under 5 years or children aged 5 to 17 years.The presence of children in a household has been shown to generally have a positive relationship with the level of home working though it is not always a statistically significant effect (Barrero et al., 2021;Nguyen, 2021;Mokhtarian and Salomon, 1994;Loo and Wang, 2018;Iscan and Naktiyok, 2005;Sener and Reeder, 2012;Singh et al., 2013;Yen, 2000).We exclude observations with data missing for these key variables.We also exclude respondents where the distance between their home and workplace is more than 1000 miles.

Summary of the data
In Table 1, we present summary statistics for our variables broken down by whether a person changed their telecommuting frequency.We see that around one third of the respondents reported an increase in telecommuting frequency compared to before the pandemic.We can also see that people who are now telecommuting more often live in areas where faster internet connections are advertised.This applies both to upload and download speeds.These people have less travel for work purposes as well as slightly less travel for non-work purposes.
One result from Table 1 is that income seems to have an important association with changes in telecommuting frequency.People living in households with the highest incomes are more likely to have increased their telecommuting frequency.This highlights the importance of controlling for income.It could be that people on higher incomes are more likely to have the kind of jobs which are amenable to home working.There are also reasons to believe that income may be correlated with internet speed and access.In our data, household income explains around 1.41% and 5.08% of the variation in upload and download speeds respectively.Recall that we multiply the connection speed by an indicator of whether the person has access to broadband.These measures thus account both for access to broadband and the available speed.If we consider only what speed is available in the area, then income is not a statistically significant predictor of either upload or download speed.For brevity, we do not include our industry variable in the summary statistics.It is worth noting, however, that industry explains around 3.55% of the variation in download speed and 3.95% of the variation in upload speed.
In Fig. 1, we consider the spatial distribution of upload and download speeds in the study area.Perhaps the most obvious pattern is that download speeds are on average higher than upload speeds.An urban/ rural divide is also visible with speeds tending to be higher in the most urban areas and lower elsewhere.However, even within the urban areas there is variation in speeds between tracts.

Models
Our first research question requires us to test for an association between average internet connection speeds in an area and change in telecommuting frequency.To do this, we use a binary logistic regression model.Our dependent variable is whether someone has increased their telecommuting frequency relative to pre-pandemic levels (y = 1) or whether there has been no change (or a decrease) (y = 0).Our key independent variables are the average upload and download speeds advertised by ISPs in the respondents' census tracts.We also control for household income, age, gender identity, industry, the presence of children, and distance (by car) from a person's home to their workplace.The logistic regression model is given by: Where y is the probability of increasing the frequency of telecommuting, X is a matrix of explanatory variables, and β is a vector of parameters to be estimated.We are particularly interested in the parameters attached to the internet speed variables.
Our second research question asks whether an increase in telecommuting frequency is associated with changes in the distance travelled for work and non-work travel.Distance variables are highly skewed, positive, and include zero values.This makes robust estimation with ordinary least squares problematic.To deal with this skew, we use a generalised linear model (GLM) with a log link and assuming a Poisson distribution.This modelling approach is typically deployed for count data (Cameron and Trivedi (1986)).However, consistent estimates of the parameters can be obtained even when the outcome variable does  (Gourieroux et al., 1984;Wooldridge, 2010).The assumption that the conditional variance and conditional mean are equal is unlikely to hold.While this does not affect the consistency of the parameter estimates, it does affect the standard errors.Using Eicker-Huber-White standard errors allows us to correct for this.For more on this see, for example, Motta (2019).See also Blackburn (2007) for a discussion of why quasi-maximum-likelihood methods may be more appropriate than more traditional log-transformed OLS approaches when dealing with skewed, continuous data.The distance travelled, G is given by: where λ = e Xγ , X is our matrix of explanatory variables, and γ is a vector of parameters to be estimated.We are interested in the γ estimates attached to the telecommuting frequency variable.We estimate this model for both work and non-work travel.Goodness of fit is measured using the pseudo-R 2 recommended by Heinzl and Mittlböck (2003) for Poisson models with over-or under-dispersion.

Results
In this section, we present the results of our econometric models.We first examine the association between change in telecommuting frequency and average advertised internet speed.We then turn our attention to the association between distance travelled and change in telecommute frequency.

Table 2
Logistic models of whether or not the person has increased their telecommuting frequency.

Telecommute frequency and broadband speed
We present six models in Table 2*****.Models 1, 2 and 3 differ in the way internet speed is included.Model 1 includes both the upload and download speeds.These variables are highly correlated (ρ = 0.82), so there is likely to be significant multicollinearity in this model.In Model 2 we include only download speed and in Model 3 we include only upload speed.
In Model 1, only the average upload speed is statistically significant.The average download speed is not statistically significant and is incorrectly signed.This may be due to the high correlation between them.In model two, where upload speed is excluded, the download speed variable is positive and statistically significant at the 5% level of significance.In Model 3, where download speed is excluded, the upload speed is positive and statistically significant.According to both the AIC and BIC, Model 3 fits the data best out of these three models.
As with many questions about travel behaviour, there are substantial risks of endogeneity, making causal inference difficult.For instance, while having access to a good internet connection may increase telecommuting frequency, it might also be that people who want to work from home choose to live in areas with good internet connections.This is more problematic if we look at the speed of the internet connections in a person's home.For example, someone who wants to work from home may opt to subscribe to a higher speed service.We do not work with the data at the individual level in this way.Instead, we look at the speeds being offered by different providers in the area the person lives in.Still, it could be the case that when people choose where to live, they consider the speed of the available internet connections.Higher speed connections don't seem to capitalise into house prices (Conley and Whitacre, 2020), but may still play a role in the decision making process.To eliminate this possibility, Models 4, 5, and 6 are estimated.These are the same as Models 1, 2, and 3 but are estimated using only people who have been at their current address for at least two years.This means that they have not moved since before the pandemic.If we think of the pandemic as an exogenous shock which impacted home working, then we can have more confidence that residential location choice does not fully explain the relationship between speed and change in telecommute frequency.
Among the workers in our sample, around 30% report that they have lived at their current address for 2 years of less.This is a high percentage of movers, particularly at a time when we might expect moving rates to be lower as a result of the pandemic.However, data from the US Census Bureau shows that for the Seattle-Tacoma-Bellevue metropolitan area in 2021, 16.8% of the population reported living at a different address in the previous year.Our data covers a two-year period, making a figure of  Reassuringly, the results from Models 4, 5, and 6 are similar to their counterparts estimated on the full sample.Out of the subsample models, the upload-speed-only model (Model 6) fits the data best.Here we can see a positive and statistically significant association between average advertised upload speed and change in telecommuting frequency.
Our preferred model is therefore Model 3.This model includes the average advertised upload speed and makes use of our full sample.According to this model, a 100 Mbps increase in upload speed is associated with 65% (e 0.539 ) higher odds of increasing telecommuting frequency relative to pre-pandemic levels.In line with previous findings in the literature, this model finds a positive association between commute distance and change in telecommuting frequency.The model also suggests people from higher income households are more likely to have increased their telecommuting frequency.Age does not appear to have a statistically significant effect on the increasing telecommuting frequency.Men seem to be less likely to have increased their telecommuting frequency compared to women.

Telecommute frequency and travel demand
We next turn our attention to the question of whether an increase in telecommuting is associated with travel demand.An important point to note here is that the dataset we use is designed to capture travel on a typical weekday.Our results, therefore, will not capture weekend travel.This may have important implications for our results on non-work travel.The results are presented in Table 3.We estimate models for work and work-related travel, as well as for non-work travel.Once again, we estimate the models for the full sample as well as for the subsample of people who have not moved in the past two years.

Table 5
Logistic models of whether or not the person has increased their telecommuting frequency for full-time workers.In both samples, there is the expected negative and statistically significant association between increasing telecommuting frequency and distance travelled for work.There are no statistically significant differences in demand for non-work travel between the people who have increased their telecommuting frequency and those who have not.People from households with a higher income tend to travel more for non-work purposes than lower income households.In general, older workers tend to travel more for work-related purposes.There are no statistically significant differences in distances travelled between men and women, though people identifying as non-binary appear to travel less for work than women.The model estimated on the full sample suggests that people who have increased their telecommuting frequency travel about 50% (e 0.698 ) less for work purposes.The model estimated on the subsample of non-movers suggests a reduction of around 40% (e 0.505 ).

Conclusion
Our paper set out to address two research questions.The first asks whether there was an association between the speed of internet connections in an area and a person's likelihood of increasing their telecommuting frequency relative to pre-pandemic levels.The hypothesis being tested was that people with faster connections would be more likely to have increased their frequency of telecommuting.The results supported this hypothesis.Faster connections were associated with increases in the frequency of home working.This result held even when using only respondents who had not moved since before the pandemic, thus eliminating the possibility that frequent telecommuters had chosen to live in areas with fast internet connections.In this way, the large increase in home working as a result of the pandemic can act as a natural experiment to help us better understand the complex relationships between travel demand and telecommuting.
The results also supported some discussion in the literature which suggested that for home working, it was not only the download speed which counts.Characteristics such as upload speed are identified as important for home working.Our results support this, with upload speeds being more strongly associated with increases in telecommuting frequency compared to download speeds.This may also reflect the fact that the upload speeds offered by ISPs tend to be substantially lower than download speeds.As a result, the marginal benefit from improvements in upload speed may be more important than marginal improvements in already high download speeds.
This finding has implications for ISPs as well as regulators and policy makers.For instance, in a study of willingness to pay for different aspects of an internet connection, Liu et al. (2018) suggest that the FCC punishes latency too much when issuing subsidies.The change in working and travel patterns brought about by the COVID19 pandemic may have changed the way people use internet connections and the value they place on different aspects of the connection.Our results support a hypothesis that upload speed facilitates more home working.
Our results also suggest that policy makers should be aware that the digital divide may be limiting options for home working in the postpandemic context.Areas with slow internet connections are less suited to home working.This may leave workers with little option but to commute to the office, thus forgoing the potential benefits of home working.Poor quality infrastructure may also reduce the resilience of the economy should further restrictions need to be introduced or if there are future virus outbreaks requiring similar action.
There has been a decades long discussion about whether telecommuting can be an effective policy measure to reduce travel demand.Several studies have found that it cannot and there is a tendency for any time saved commuting to be used to increase the commute length or on non-work travel.Our results do not find evidence of this rebound effect.People who had increased their frequency of telecommuting tended to travel less for work.There was no association with non-work travel.However, it is worth noting that this survey took place in the Spring of 2021 when people may still have been reducing their travel to avoid COVID19.An increase in non-work travel may also occur at weekends, but our dataset focussed on weekday travel.It may also be a short run effect.Once people settle into new working patterns, they may consider changing their residential location.Still, it is important to investigate whether what we know about home working and travel demand from pre-COVID19 times still applies post-COVID.If home working becomes more widespread it is not clear whether results which applied to the pre-COVID19 telecommuters would apply to the expanded population of home workers in a post-COVID context.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Table 1
Summary statistics by change in telecommuting frequency.

Table 3
Poisson models of total distance travelled by trip type.

Table 4
Summary statistics by change in telecommuting frequency for full-time workers.

Table 6
Poisson models of total distance travelled by trip type for full-time workers.
D.P. McArthur and J. Hong