A systematic review of the energy and climate impacts of teleworking

Information and communication technologies (ICTs) increasingly enable employees to work from home and other locations (‘teleworking’). This study explores the extent to which teleworking reduces the need to travel to work and the consequent impacts on economy-wide energy consumption. The paper provides a systematic review of the current state of knowledge of the energy impacts of teleworking. This includes the energy savings from reduced commuter travel and the indirect impacts on energy consumption associated with changes in non-work travel and home energy consumption. The aim is to identify the conditions under which teleworking leads to a net reduction in economy-wide energy consumption, and the circumstances where benefits may be outweighed by unintended impacts. The paper synthesises the results of 39 empirical studies, identified through a comprehensive search of 9000 published articles. Twenty six of the 39 studies suggest that teleworking reduces energy use, and only eight studies suggest that teleworking increases, or has a neutral impact on energy use. However, differences in the methodology, scope and assumptions of the different studies make it difficult to estimate ‘average’ energy savings. The main source of savings is the reduced distance travelled for commuting, potentially with an additional contribution from lower office energy consumption. However, the more rigorous studies that include a wider range of impacts (e.g. non-work travel or home energy use) generally find smaller savings. Despite the generally positive verdict on teleworking as an energy-saving practice, there are numerous uncertainties and ambiguities about its actual or potential benefits. These relate to the extent to which teleworking may lead to unpredictable increases in non-work travel and home energy use that may outweigh the gains from reduced work travel. The available evidence suggests that economy-wide energy savings are typically modest, and in many circumstances could be negative or non-existent.


Introduction
Efforts to reduce greenhouse gas (GHG) emissions focus upon both technological innovation and behavioural change, while recognizing that these domains are interdependent (Bel and Joseph 2018, Creutzig et al 2018, Dubois et al 2019. One area that has received particular attention is encouraging technology-enabled changes in working patterns to reduce commuter travel and office-related energy consumption (Hopkins and Mckay 2019). Since the transport sector in the United States (US), for example, accounts for around 33% of final energy use, a reduction in commuter travel could have a significant impact (Zhu and Mason 2014).
One trend that could reduce energy consumption and thus carbon emissions from travel is teleworking, 1 where employees use information and communication technologies (ICTs) to work from home, in satellite telecentres or in other locations. Employees may telework part-time or, less commonly, full-time (Hynes 2016, Giovanis 2018. However, despite assumptions that teleworking would provide an important contribution to a 'lower energy future' , evidence of its impacts is inconclusive (Brand et al 2019). Indeed, while some studies suggest that teleworking can reduce energy consumption (primarily through avoided commuting) by as much as 77% (e.g. Koenig et al 1996), others find much smaller gains, with some studies suggesting a paradoxical increase in energy consumption (e.g. Rietveld 2011).
This lack of consensus on the energy and environmental benefits of teleworking has arguably contributed to the lack of coordinated promotion of teleworking by business or government, even in countries where multiple studies have been conducted-such as the US (Allen et al 2015). Indeed, despite the promise of energy savings and other social benefits, teleworking has not grown as rapidly as predicted, even in sectors and regions that appear well-suited to it-such as growing cities in developing countries (Ansong and Boateng 2018). For example, (Zhu et al 2018) estimates that only around 9% of the US working population teleworks more than once a week.
This uncertainty about environmental benefits is compounded by persistent scepticism about the social implications of teleworking. Many believe, for example, that practices such as 'face to face' meetings' are essential for building confidence between colleagues and clients and cannot be substituted by 'virtual meetings' enabled by ICT (Baruch 2001). Other studies have suggested that concerns about emotional isolation or future career advancement may hinder people's willingness to adopt teleworking (Golden et al 2008, Schulte 2015. For firms, concerns over accountability and productivity persist, despite evidence to the contrary (Pérez et al 2005).
In this context, this paper provides a systematic review of the current state of knowledge about the energy impacts of teleworking. This includes the energy savings from reduced commuter travel and the indirect impacts on energy consumption associated with changes in: (a) non-work travel by both the teleworker and other household members; (b) the size and occupancy of work premises; and (c) the location and occupancy of employees' homes. The aim is to identify the conditions under which teleworking can lead to a net reduction in overall energy consumption, and the circumstances where the benefits from teleworking are outweighed by the unintended impacts, such as greater private travel or increased non-work energy consumption. The latter are commonly referred to as 'rebound effects' (Berkout and Hertin 204, Horner et al 2016).
Our interest is the impacts of teleworking on economy-wide energy consumption, taking into account the full range of mechanisms through which those impacts occur. But many studies have a narrower scope, in that they focus upon a more limited range of impacts, such as the changes in commuter travel alone. These studies may nevertheless provide useful evidence, as they frequently capture the most important impacts. Similarly, many studies use different metrics to measure those impacts, such as changes in vehicle distance travelled. Again, these studies may provide useful evidence, as there is frequently an approximately linear relationship between those metrics and energy consumption. However, it is important to recognise that studies with a narrower scope will omit many important categories of impact, and studies with a different metric will provide rather inaccurate measures of the impact on energy consumption.
The article is structured as follows. Section 2 summarises the academic and policy debates about the energy and environmental benefits of teleworking. Section 3 outlines the systematic review methodology, while section 4 presents the key results of the 39 identified studies. Section 5 discusses these results in more detail, including the magnitude of the identified impacts, the determinants of those impacts, and the source and scale of associated rebound effects. It also assesses the scope of the studies in terms of the types of impact that are included, as well as their methodological quality. Section 6 summarises the overall findings and reflects upon their implications for research and policy.

Teleworking and the impacts on energy use and emissions
'Teleworking' means working either full-or part-time from home, from a 'telecentre' located close to home, or from other locations. The practice has grown in popularity as technology has improved, but definitional ambiguities and data limitations make it difficult to estimate the precise number of teleworkers at any one time . The concept of teleworking can be traced back to the 1960 s when ICTs such as telephones and fax machines were first mooted as offering the possibility of liberating workers from commuting to work every day (Mokhtarian 1997). At this time, however, teleworking was largely promoted as a social policy that would enable workers to spend more time with their families and less time travelling (Johnson 2003).
The advent of the internet in the mid-1990s and innovations such as teleconferencing coincided with a focus on the broader benefits of teleworking and a Note: In the case of teleworking, the substitution effect is normally considered to be the most significant.
shifting rationale for its expansion (e.g. Marvin 1997, Allenby andRichards 1999). The increasing prominence of climate change within popular discourse led teleworking to be seen as an environmental or energy strategy that could reduce air pollution related to peak-time traffic congestion (Niles 1994), along with energy use and emissions from travel to work  and energy consumption within workplaces (Matthews and Williams 2005). The main source of these benefits was that working from home (or from satellite telecentres that were closer to the home than the workplace) should reduce the energy expended in both travelling to work (typically by private car) and in heating, cooling and lighting large office spaces (Marcus 1995, Williams 2003. Appraising whether such changes in working practices have indeed had these benefits is difficult, however, since the enabling technology (ICTs) triggers a range of 'direct' and 'higher-order' effects that are very hard to measure. Frequently, these effects are both unexpected and unintended (Pohl et al 2019). 'Direct' effects relate to the energy used in the manufacture, operation and disposal of ICTs together with the associated network infrastructure, while 'higherorder' effects relate to the changes in energy consumption stimulated by ICTs, including changes in individual behaviour (e.g. commuting behaviour) and changes in prices, consumption, investment and other variables throughout the economy (Horner et al 2016). These higher-order effects take a number of forms that (both individually and collectively) may either increase or reduce energy consumption relative to a baseline scenario where those changes do not occur. Table 1 provides a classification of these different types of impact, and illustrates this with examples relevant to teleworking.
Whether the economic and behavioural changes stimulated by teleworking lead to an overall reduction in energy consumption therefore depends upon the sign and magnitude of these different categories of impact-the relative importance of which is likely to vary with context and change over time (De Graff 2004, Horner et al 2016. Since personal transport is significantly more energy intensive than ICT services, most studies of teleworking ignore the direct impacts altogether and focus solely upon the indirect impacts-and particularly those from reduced commuter travel (Horner et al 2016). However, factors such as the short lifetime and rapid replacement of ICTs, their increasingly complex supply chains (including dependence on a growing range of rare earth elements), and the advent of cloud storage and video streaming (which are relatively energy intensive) may be contributing to a growing energy footprint for ICTs. Hence, these direct impacts may become a more significant focus of teleworking studies in the future (Chapman 2007, Lachapelle et al 2018.
The focus of the majority of studies has been the nature and magnitude of the 'higher-order' impacts indicated in table 1 (Horner et al 2016, Pohl et al 2019. The most commonly cited benefit of teleworking is its 'substitution' effect, whereby commuter travel is substituted (or displaced) by less energy-intensive activities or behaviours that are enabled by ICTs (Salomon 1998). Historically, this has typically involved using ICTs to work from home or from a 'telecentre' located closer to the home than the workplace (Balepur et al 1998). More recently, there has been a rapid growth in mobile working from cafes, trains and other Wi-Fi-enabled locations, but the energy implications of these emerging practices have yet to be fully studied. Whether these substitution effects lead to a net reduction in energy consumption (at either the individual or societal level) depends, however, on the higher-order impacts (Mokhtarian 2009). Indeed, in some circumstances teleworking could encourage changes in behaviour that increase work and/or non-work travel, and thereby energy consumption (Pérez et al 2004, Williams 2011, Zhu 2012. In the case of work-travel, for example, the ability of teleworkers to live further away from their place of work could mean that the longer trips they make on non-teleworking days (where, as is the norm, they are only part-week teleworkers) wholly or partly outweigh the travel and energy savings they make on days that they work from home (Bailey and Kurland 2002). These impacts will also depend on the mode of transport they use to commute to work: in countries where public transport is a common mode of commuting, teleworking practices will have less impact on energy use than in countries (such as the US) where the private car is the dominant mode (Mokhtarian 2009, Van Lier et al 2014. The energy impacts will further depend upon the energy efficiency and level of occupancy of the relevant mode (e.g. one person in a Sport Utility Vehicle (SUV) versus several hundred in a crowded train), and the carbon impacts will additionally depend upon the carbon intensity of the relevant energy carriers (e.g. gasoline versus electricity).
In the case of non-work travel, there is evidence that gaining more time at home as a result of teleworking may induce extra trips by teleworkersfor leisure and social purposes, for example-that would not have been made had the teleworker been commuting to work every day (Lyons et al 2008). It may also enable greater use of the household car by other household members on days that the commuter works from home. This latter trend has been observed in countries where households have fewer cars, such as South Korea, where the car is more of a 'scarce'-and thus more desirable-commodity (Kim et al 2015). Such examples of additional, nonwork travel enabled by teleworking may be considered expressions of 'latent' travel demand (Mokhtarian et al 1995). Another induced travel effect could be where the feelings of isolation and sedentariness generated by teleworking stimulate a desire for movement and mobility (Gurstein 2002). This compensatory travel, which may involve habitual trips to libraries or cafes for work, may partly offset the travel and energy savings achieved by avoiding commuting (Rietveld 2011). Overall, these examples suggest that the travel demand displaced by teleworking may be partly offset by induced travel demand in other areas.
Home and office energy consumption is another area where the benefits of teleworking could potentially be offset (Pérez et al 2004). For example, teleworking may lead to more energy being used at home (e.g. for heating, cooking and lighting) without any compensating reductions in the energy used at work (e.g. offices may continue to be heated and lit as much as before). There could be an 'additive' impact of teleworking if businesses neither move to smaller offices (which have a smaller energy footprint) nor close their offices in response to increased teleworking. The net result could be an increase in building energy consumption, and possibly total energy consumption, as a result of greater teleworking (Kitou and Horvath 2008).
At the societal level, the aggregation of these and other trends may generate broader indirect and economy-wide rebound effects (Horner et al 2016). If households reduce car travel, they may spend the money they save on road fuel on other goods and services that also require energy and emissions to produce (Sorrell et al 2020). Alternatively, if teleworking boosts labour productivity and stimulates economic growth, it could encourage increased consumption, travel and energy use by both producers and consumers (Lachapelle et al 2018). (Berkhout and Hertin 2004) draw attention to the potentially small impact of teleworking on energy consumption relative to other driving forces such as population and income growth. A summary of the direct and higher-order effects of teleworking at the individual and societal level is given in table 2.
The identification of these higher-order effects suggests that, to accurately estimate the net energy impacts of teleworking, a study must have as broad a scope as possible: a narrow scope may mean that important impacts are overlooked (Berković et al 2013). For example, a study may estimate the reduction in commuter travel from teleworking but ignore the increase in other forms of travel. Alternatively, a study may overestimate the energy savings from teleworking by assuming that all commuting is by car, thereby neglecting any commuting by public transport. Similarly, a study may estimate the energy savings from reduced commuting and reduced office use but ignore the increase in home energy use. A limited scope could therefore lead to either an over-or under-estimate of the energy savings from teleworking depending on the context-specific interactions between a range of variables (Mokhtarian 2009).
While the range of possible interactions among different variables suggests that studies should have a wide scope, there are considerable methodological challenges in designing studies that capture the full range of impacts from teleworking. As a result, most studies focus upon a narrower range of impacts, such as commuter travel alone, whose measurement is more feasible. As Horner et al (2016, p. 14) observe, this is a more general problem when studying the impact of ICT on energy use: The inability to draw concrete conclusions reflects, in large part, uncertainty regarding the rebound effect for ICT and the inability to disentangle root causes of interrelated economic effects. The dynamics of these effects are hugely dependent upon human behavior, which is laden with uncertainty and confounds efforts to achieve the full technical potential of ICT interventions.

Research questions and approach
Our primary research question is as follows: • What are the determinants and magnitude of the impacts of teleworking on energy consumption or proxies for energy consumption such as distance travelled by car?
Our sub-questions were are follows: • What are the full range of impacts identified in the literature? • What are the key socio-technical determinants or drivers of those impacts?
To review the evidence on this topic, we employ the methodology of 'systematic reviews' (Petticrew and Roberts 2006). These offer a number of advantages over traditional literature reviews, including minimising unintentional bias (e.g. excessive selfcitations, or citations of colleagues) and avoiding the exclusion of studies that are frequently overlooked (Haddaway et al 2015). For these and other reasons, many authors have called for greater use of systematic reviews in the area of energy and climate research (Sorrell 2007, Sovacool et al 2018, Pereira and Slade 2019. The first stage of our systematic review involved choosing search terms that were relevant to the selected topic. These were combined to construct search queries that were used in the search engines of two scholarly databases. The process was iterative, since small changes in the search terms can have a large influence on the number of identified sources. As such, while reviewing the bibliographies of review articles in the area (e.g. Horner et al 2016), we repeatedly refined our search strings to ensure that they captured all of the identified studies.
The references generated by this search phase were then screened in order to remove irrelevant studies. This involved applying explicit inclusion and exclusion criteria to the title and abstract of the study, and if necessary, to the full text. Following this, information was extracted in a consistent way from each of the selected studies. Since the evidence was both quantitative and qualitative, as well as being highly variable and using a variety of metrics (e.g. energy use, distance travelled, carbon emissions), a narrative synthesis was considered most appropriate (Snilstveit et al 2012). To formulate our search and screening protocols, we followed the guidelines of the Collaboration for Environmental Evidence  and used the free online platform CADIMA to perform the screening phase (Kohl et al 2018).

Sources and databases
The evidence base includes peer-reviewed academic journals, conference proceedings, books, working papers, doctoral theses, and technical reports. We gave priority to studies that provided quantitative estimates, but also examined qualitative evidence to obtain a deeper understanding of the relevant mechanisms and determinants. Given the pace of technical change in this area, we considered that older studies were unlikely to be of much value. Hence, we confined the review to studies published after 1995, approximately the start of the 'internet age' (Huws 2013). We also confined the review to English language studies, since this was the language of the research team. We applied our search protocol to Scopus and Web of Science, which are the most widely used scientific literature databases. We also searched for relevant grey literature (technical reports, doctoral theses, working papers) through a combination of internet searches and checking the profiles of key researchers in the field and the bibliographies of the identified studies.

Search terms and combinations
We combined three types of keywords in our search query, namely: a synonym for 'teleworking'; a second for 'energy' (including various proxies for energy such as distance travelled); and a third that referred to the relationship or interaction between these two. We investigated exhaustive variations around these terms using the Boolean OR operator, and combinations of them using the Boolean AND operator, and made sure that studies identified by other authors (e.g. Horner et al 2016) were caught. This led to an extensive search string for each database (see supplementary material 1 (available online at stacks.iop.org/ERL/15/093003/mmedia)).

Inclusion and exclusion criteria
The search results were merged, and duplicates removed to obtain our initial sample. We then applied the inclusion/exclusion criteria in table 3 to select only those studies that appeared relevant to our research question. Analysis of this preliminary sample led to the exclusion of further studies in which results or data were duplicated or where, on closer inspection, relevant data were not present. Once the final set of studies had been defined, we extracted the data into an Excel file (see supplementary material 2). The key results are summarised in section 4.

Search and screening phases
As indicated in figure 1, the search phase generated an initial sample of 7041 references from Scopus and 4585 from Web of Science, making a total of 11 626 references. This is a very large number compared to Table 3. Inclusion and exclusion criteria used to screen the identified studies.

Inclusion criteria (IC) IC1 Refers to an analysis of ICT-enabled teleworking IC2 Refers specifically to an energy-related effect of teleworking IC3 Contains primary research results
Exclusion criteria (EC) EC1 The main topic does not relate to teleworking or energy EC2 The study contains no quantitative analysis of the effects of teleworking on energy demand EC3 The study is not accessible at the time of review (e.g. due to it being unpublished or behind a paywall) other systematic reviews because we were exhaustive when designing our search query. Adopting such a 'large nest' approach minimises the risk of missing relevant studies but leads to the inclusion of a large number of irrelevant studies that need to be screened out. After removing 2165 duplicates, our initial sample comprised 9461 references. Screening the titles and abstracts led to the removal of 9042 irrelevant references, while full text screening led to the removal of an additional 63 studies. Our preliminary sample therefore consisted of only 56 studies, which was further reduced to 39 by removing studies with data that was duplicated in other studies or those which had no relevant primary data. Table 4 summarises the extracted data from the 39 studies in our final sample, presenting the studies alphabetically (a more detailed table is provided in supplementary material 2). For each study, we include:

Data extraction
(a) the study's number in the list; (b) the main author's name and the year of publication; (c) the country location; (d) the methodology (distinguishing between analysis of survey data, evaluation of pilot schemes, and scenario modelling) (e) the most relevant metric (e.g. commuting distance travel); (f) the scope of the study (i.e. coverage of: (i) commuter travel; (ii) non-commuter travel; (iii) home energy use; and/or (iv) office energy use); (g) the estimated impact on the relevant metric ('increase, 'neutral' , 'reduce' , or 'unclear'); (h) the nature and scale of that impact, including quantitative estimates; and (i) our appraisal 2 of the methodological robustness of the study ('good' , 'average' , or 'poor').

Discussion: impacts and rebounds of teleworking
This section discusses the results of the systematic review that are presented in table 4 and in greater detail in supplementary material 2. It first provides an overview of the results, before discussing the sources and conditions of impacts on the relevant metrics, the potential rebound effects from teleworking, the scope of the studies and the methodological quality of the evidence base.

Overview of the studies
The 39 studies in the final sample examine a range of configurations and scales of teleworking in a variety of contexts. The studies examined two main types of teleworking, home-based (35 studies) and telecentrebased (four studies). As table 4 shows, most studies are from the US (19 studies) and Europe (11 studies), with only three from the Global South (Thailand, Malaysia, and Iran). The dominance of US studies may influence the overall findings, since most US commuters travel by private car rather than public transport, and vehicles and buildings in the US tend to be larger and less energy efficient than those in other OECD countries.
As figure 2 indicates, there is a fairly even distribution of studies across the time range . While this suggests there has been no slackening of interest in teleworking over this period, the character of teleworking has changed as ICTs have evolved. In particular, telecentre-based teleworking is now largely obsolete, and the three studies that involved the collection of data on telecentre pilot schemes were all published before 1998.
The studies employ a variety of methods that are described in detail in supplementary material 2. The studies also vary in methodological quality and include both ex post estimates and ex ante projections of impacts on a number of different metrics (e.g. commuting trips, commuting distance, energy consumption) which makes them challenging to compare. These methods can be grouped into three broad categories: • Scenario modelling: using simulation models or other types of model to project future impacts from teleworking (often using rather sparse datasets) (e.g. Larson and Zhao 2017

Increase
For single-worker households where the person teleworks once a week, there is an average increase of 9.7 miles travelled by all modes (9.0 by car, or 3.9%, 0.5 by public transport, or 2.4%, and 0.2 by nonmotorized modes, or 3.8%).
For two-worker households one day a week teleworking increases miles travelled by car by 1.6 miles (or 0.4%). This is lower than above because workers share trips.

Reduce
On non-teleworking days, telecentre-based teleworkers have 91% higher VMT than non-teleworkers, while home-based teleworkers have 54% higher VMTsuggesting they live further from work than regular commuters.
On teleworking days, home-based teleworkers have 67% less travel distance than commuters and telecentre workers have 54% less distance.

Reduce
Average quarterly per capita total commute distances are generally 15% lower for teleworkers than for nonteleworkers, indicating that they telework often enough to more than compensate for their longer one-way commutes (which tend to be on average 16 miles compared with 11 miles for non-teleworkers).

Reduce
Provided the floor area of office buildings being utilized decreases as the rate of teleworking increases, 60% of the population could teleworking lead energy consumption to decrease by 0.6% of the total energy consumption of the residential and non-residential sectors in Osaka City.   Table 6 classifies the studies by their scope, or the 'number of impact categories' included. We distinguish four categories of impact, namely the energy used in: (a) commuting; (b) non-work travel; (c) the home; and (d) the office. Most studies do not estimate energy consumption directly, but use other metrics such as distance travelled that serve as proxies for energy consumption. While there may be additional impact categories, such as economy-wide rebound effects, these are not included in any of the reviewed studies. The scope of a study depends in part upon the research questions employed: for example, if the primary interest is the impact of teleworking on congestion, a narrow scope is appropriate. Conversely, if the primary interest is the impact on energy consumption, a wide scope is appropriate. While our interest lies with the latter, studies with a narrow scope nevertheless provide useful evidence on the impacts on energy consumption within a particular area.

Average
While approximately half the studies (19) only consider the impact of teleworking on commuter travel, the remainder estimate a wider range of impacts. For example: 12 studies also estimate the impact on non-commuting travel by either the commuter or other household members; five studies estimate the impacts on home and/or office energy use as well as commuting travel; and two studies estimate the impacts on commuting travel and home energy use (but not on non-work travel). An exception is (Shimoda et al 2007), who ignore the impact on travel altogether and only consider the impact on home and office energy consumption. It is notable, however, that none of these studies encompass all four of our impact categories. Table 7 summarises our assessment of the methodological quality of each study. We ranked 14 of the studies as methodologically 'good' , 11 as 'average' and 14 as 'poor' . Some justification for these rankings can be found in supplementary material 2. Section 5.5 discusses the relevance of methodological quality to the estimated impacts on energy consumption. Table 8 shows that the majority of the studies (26 out of 39) suggest that teleworking (both from home and telecentres) leads to a net reduction in energy use and/or emissions, with only five studies finding a net increase. These benefits largely result from the elimination of the commute, reductions in congestion, concomitant reductions in vehicle emissions, and reductions in office-based energy consumption.

A summary of the energy, climate and environmental impacts of teleworking
While most studies compare the net energy/environmental impacts of a teleworking and non-teleworking mode of working, a few studies (e.g. Atkyns et al 2002) provide only absolute estimates of changes in key variables, such as gallons of gasoline saved. These studies are less useful than those providing relative figures expressed in terms of a percentage gain or loss. Only the latter are included in table 9, which displays the full range of estimates found in our sample of the net impact of teleworking on different metrics. As with scope, the diversity of metrics used by the different studies reflects their different research questions.
While all of the metrics in table 9 are relevant to the impact of teleworking on energy consumption (our research question), some are more useful than others. It is important to stress, furthermore, that the estimates in table 9 are a mix of relative and absolute figures, reflecting the diversity of the studies. So, while some studies estimate the impact of telecommuting versus not telecommuting for single journeys (a relative figure), other studies estimate the total impact of a specific proportion of the population telecommuting a certain number of times per week or month (an absolute figure). This makes it difficult to extract any normalized estimates of relative energy savings from across contexts. Table 9 indicates that the most commonly used metric (used by 26 of the 39 studies) is vehicle distance travelled, which is a proxy for the energy consumed by motorized travel. Studies using this indicator give the widest range of estimates, ranging from a 20% reduction in distance travelled (Balepur et al 1998), to a 3.9% increase (De Abreu E Silva and Melo 2018). These studies either measure or assume different proportions of teleworkers and/or differing frequencies of telework, making comparison between them difficult. In addition, most of the studies do not disaggregate 'avoided travel' by mode and instead (implicitly) assume that it relates to travel by private car. In fact, (De Abreu E Silva and Melo 2018) is the only study to recognise that the (avoided) commuter travel may be by other modes such as public transport. This bias partly reflects the dominance of US studies, but the assumption that avoided travel necessarily take the form of avoided car travel may lead to an overestimate of energy and travel savings (see section 5.6).
Seven studies measure impacts in terms of the number of commute trips and find reductions of between 2.3% and 30% per week. This metric tells us less about energy savings, however, as it does not indicate the distance travelled. (Mitomo and Jitsuzumi 1999) measure impacts in terms of time savings from reduced traffic congestion and estimate that these range from 1.9% to 28%, with implications for energy use and emissions from stationary traffic.
Seven studies estimate the impact of teleworking on energy consumption and estimate reductions of between 0.01% and 14%. Several of these take into account the impacts on both commuting energy use and home or office energy. For example, (Matthews and Williams 2005) estimate that, if half of the 'information workers' in the US and Japan telework four days per week, this would reduce primary energy consumption by~1%. Finally, ten studies suggest that teleworking could reduce carbon emissions by between 0.1% and 80%, with this higher estimate assuming a five-day teleworking routine by the whole population (Kitou and Horvath 2003).

Sources and estimates of environmental benefits from teleworking
The majority of the 39 studies suggest there are energy savings and other environmental benefits from teleworking. This section examines the main sources and estimates of these savings in more detail and contextualizes these results in terms of the broader literature.

Elimination or reduction of commutes
The main source of energy savings is the reduction in commuter travel to and from work. This is a substitution effect, whereby ICT facilitates remote working and removes the need to commute for at least part of the week. Overall, the studies suggest varying reductions in weekly, monthly or annual vehicle distance travelled, up to a maximum of 20%. They also suggest corresponding benefits, including reductions in the number of trips by up to 30%, time savings from reduced congestion of up to 28% (which in turn could lead to significant energy savings since slow-moving traffic is inefficient), and associated reductions in carbon emissions. It should be stressed, however, that the majority of the studies finding reductions in vehicle distance travelled neglect potential rebound effectssuch as increased non-work travel (see section 5.4).
Studies of telecentre workers find significant reductions in commuting distance travelled. For example, (Balepur et al 1998) show how participants in the Puget Sound pilot who teleworked once a week reduced their total weekly commuting vehicle travel by 19% (10 miles). However, different studies make different estimates of, or assumptions for, the number of households that are teleworking and the frequency with which they are teleworking. They also estimate both relative and absolute figures, making it difficult to compare their estimates of travel/energy savings and to generalise their findings. For example,  estimate that teleworking is practised by 12% of the US workforce once a week and estimate a resulting 0.8% reduction in the total distance travelled by private cars. In contrast, (Martens and Korver 2000) assume a teleworking rate of 'between 20% and 60%' of the US working population and estimate a resulting 5% reduction in vehicle distance travelled. But (Martens and Korver 2000) do not state the assumed frequency of teleworking (i.e. how many times per week these 20%-60% of the population will telework). They moreover make arguably unrealistic assumptions about the potential future uptake of teleworking considering that the current proportion of teleworkers is only 9% in the US and 5% in the UK. Other studies (e.g. Röder and Nagel 2014) fail to state either the proportion of the population teleworking or their frequency of teleworking, making it impossible to extrapolate useful lessons from their results.
The studies also relate to very different geographical contexts, where differences in the patterns and modes of commuting differ have important implications for the potential energy savings from teleworking. For example, (Helminen and Ristimäki 2007) estimate that teleworking by 4.7% of the Finnish labour force once a week would reduce commuting distance travelled by 0.7%. (Larson and Zhao 2017) meanwhile estimate that if 20% of US workers telework once a week, commuting energy consumption would decrease by 20%. However, Finland and the US differ significantly in terms of the average distance between home and work, the modal mix for commuter travel and the relative energy efficiency of different modes; with the result that the energy savings from teleworking in Finland may be substantially lower than in the US. For example, while (Helminen and Ristimäki 2007) state that 70% of the Finnish population commute by car or motorbike, (Larson and Zhao 2017) assume that all US commuting is by car. This means that the energy savings from telecommuting will be higher in the US, where the forgone travel is in the form of avoided car trips, compared with Europe, where a large proportion of commuting is by other modes (Van Lier et al 2014).
For our purposes, however, the most fundamental problem with many of the studies is their limited scope. Indeed, whether teleworking reduces economywide energy consumption depends upon the impacts on commuting travel, non-work travel, home energy use and office energy use, together with the relative energy efficiency of transport modes, homes and office buildings. Most studies only provide a partial coverage of these different variables. While some studies examining both work and non-work travel find that increases in non-work travel as a result of teleworking do not lead to increases in overall travel (e.g. Mokhtarian and Varma 1998), others find evidence to the contrary (e.g. Zhu 2012). Capturing these nuances in order to appraise the impact of teleworking on overall energy use is difficult but essential-an issue that will be returned to in section 5.5.

Reductions in office energy consumption
Some of the literature on ICT and energy suggests that more remote working may lead to higher energy consumption at home (e.g. Chapman 2007). However, several studies show how, even allowing for increases in home energy consumption, teleworking could achieve overall energy savings since it enables reductions in per capita office space (e.g. through hotdesking) and potentially means that offices no longer need to be heated or cooled to the same level or for the same period of time. (Williams 2003), for example, estimates that the adoption of 4-day per week teleworking by the specialist/technical workforce in Japan (~14% of the total) could reduce national energy consumption by 1.0% by eliminating the need for office heating and cooling on non-working days. Similarly, (Matthews and Williams 2005) estimate that the potential energy savings from reducing office space are comparable to those from reduced commuting. In countries such as Japan, where there is a lower level of office space per worker, the energy savings from reduced office use may be smaller than in the USA, where offices tend to be larger (Matthews and Williams 2005). The gains may also be smaller in more temperate regions, since less energy is required to heat and cool office buildings (Kitou and Horvath 2003) and may also be partly offset by the embodied energy associated with duplicated equipment such as printers. The latter forms part of the direct impacts of ICT on energy consumption (table 1), but this is ignored in all of the reviewed studies.
As with gains from reduced commuting, these potential gains also depend upon a range of factors, including the extent to which firms downsize or close their offices as the number of teleworkers increase. (Shimoda et al 2007) estimate that, if utilised office space decreases in proportion to the rate of teleworking, full-time teleworking by 60% of workers in Osaka City (Japan) would reduce total energy consumption for residential and non-residential buildings by 0.6%. (Shimoda et al 2007) stress, however, that if teleworkers are only part-time, companies may not downsize their offices or reduce energy consumption since they will need to retain the same sized premises for the days that teleworkers join non-teleworkers in the office. Since part-time teleworking is more common than full time teleworking, the latter appears a more likely outcome. Thus, the potential gains in terms of reduced office energy consumption may not be realised.
More generally, (Shimoda et al 2007) demonstrate that, even assuming office energy use falls in proportion to the rate of teleworking, very high levels of teleworking may achieve only modest reductions in aggregate energy consumption. Similar conclusions are reached by (Matthews and Williams 2005), who estimate that if all US 'information workers' teleworked four days a week, US energy consumption would fall by only~2%. This is partly because teleworking is expected to be suitable for less than half of the US workforce. For comparison, (Matthews and Williams 2005) estimate that a 20% improvement in average car fuel efficiency in the US would reduce aggregate energy use by~5. 4%. 4 Although (Shimoda et al 2007) provide some useful evidence about the potential impacts of teleworking on home and office energy consumption, their study provides no analysis of the impacts of teleworking on work or non-work travel. Hence, it still provides only a partial picture of the net impacts of teleworking on energy consumption.

Rebound effects from teleworking
While teleworking is framed by some studies as a promising way to reduce energy consumption, particularly from commuting travel, other studies draw attention to potential unintended impacts that could increase energy consumption and negatively affect various environmental indicators. They also highlight the uncertainty about the impacts of teleworking, owing to the complexity of impact pathways and the unpredictably of human behaviour.

Dispersion of residential location and longer commutes
Although 70% of the studies in our review suggest that teleworking reduces energy use, five studieswhich we also consider to be methodologically rigorous-suggest that the gains from eliminating commutes on teleworking days may partly or wholly offset by longer commutes on non-teleworking days (Balepur et al 1998, Chakrabarti 2018. For example, (Helminen and Ristimäki 2007) find that Finnish teleworkers have a 3.7 km longer commute than nonteleworkers. This concurs with (De Abreu e Silva and Melo's 2017) finding that, controlling for a wide range of sociodemographic variables, UK teleworkers (in one-worker households) have a 10.7 mile longer commute than non-teleworkers. Several studies moreover find that some teleworkers also travel further than regular commuters on days that they are not teleworking. For example,  find that home-based teleworkers in the US travel 67% less than regular commuters on teleworking days, but 54% more on non-teleworking days. Thus, over the course of a week-and given a part-week teleworking lifestyle-teleworkers may potentially travel further than regular commuters.
However, such studies do not establish the direction of causality, i.e. do people telework to avoid a long (and/or a slow or difficult) commute, or do they choose to live further away from the workplace because their job enables them to telework? One approach to identifying whether teleworking has a causal influence on commuting distance is to use instrumental variables. In his analysis of US national household survey data, (Zhu 2012) used the frequency of internet use as an instrument for teleworking since this should be correlated with the latter while not affecting commuting distance. (Zhu 2012) finds that teleworking has a positive influence on commuting distance that has increased over time. In 2009, US teleworkers' work trips were 43% longer in distance than those of non-teleworkerscompared to 34% in 2001.
An alternative approach to addressing endogeneity is to use panel data, since this allows the changes in teleworking and commuting distance over time to be identified whilst controlling for time-invariant fixed effects. Using this approach, (de Vos et al 2018) estimate that Dutch teleworkers have 5% longer commuting times on average, with every additional day of home working being associated with a 3.5% longer duration commute. In a more recent study using a different data set, (de Vos et al 2019) obtain larger estimates of 12% and 16% respectively. Both studies use commuting duration rather than commuting distance as the dependent variable, but these two variables should be correlated-and (Zhu's 2012) results suggests that the impact of teleworking on distance travelled could be larger than the impact on commuting duration.
Overall, evidence from both the US and Europe suggests that the adoption of teleworking may induce long-term changes in residential location that offset some of the environmental benefits. The size of this effect may be expected to vary with contextual factors, such as the differential in property prices between urban and peri-urban regions and the financial and temporal cost of the commuting journey. However, it seems clear that, in some circumstances, the increased adoption of part-time teleworking could increase weekly, monthly, or annual commuter travel. More generally, the environmental benefits of teleworking will depend upon both the frequency of teleworking and how far teleworkers live from their workplace (Lachapelle et al 2018).

Non-work travel
Another potential unintended effect of teleworking is that it may encourage more non-work travel. In this case, the travel avoided by the daily commute is partly or wholly offset by additional travel by the teleworker for other reasons. This is sometimes termed a 'complementary' effect of teleworking (Mokhtarian 2002(Mokhtarian , 2009. Several studies find such effects, though it is important to underline that they only do so because their wider scope enables the interactions between teleworking behaviour and non-work travel to be explored. For example, (Elldér 2017) finds that teleworkers travel further than non-teleworkers on both teleworking and non-teleworking days. While nonteleworkers travelled an average of 46 km per day, teleworkers travelled 54 km on teleworking days and 64 km on non-teleworking days. Similarly, (Zhu 2012) find that, according to US National Household Travel Surveys (NHTS), teleworkers took 10.8% more non-work trips per day than non-teleworkers (4.18 versus 3.77) and that these were, on average, 15.7% longer (36 km versus 32 km). Again, using instrumental variables, (Zhu 2012) finds that teleworking has a significant impact on non-work travel.
The reasons for greater non-work travel on teleworking days are complex and are not explored by most of the studies in the sample. Of the studies that did attempt to explain causality, (Zhu 2012) suggests that non-commuting workers are less able to 'daisy chain' (or 'link') trips together in an efficient way, and thus have to make specific trips for non-work activities. This effect may be particularly pronounced where there is one household member who works: with that member no longer commuting, other household members may have to make separate trips out to carry out specific non-work duties (Kim et al 2015, De Abreu E Silva andMelo 2018). The distance travelled for non-work activities will also vary with geographical context, including the proximity of the home to schools, retail outlets and other destinations-which again suggests that the results from US studies may not necessarily apply to other contexts.
Teleworking could also increase daily/weekly travel among those who telework by creating a displacement effect, whereby commuting trips are replaced with other forms of non-work travel, such as leisure trips (Lachapelle et al 2018). These trips could be due to boredom or could merely be opportunistic where teleworkers take advantage of their free time to travel more or to engage in more social activities (Rietveld 2011). This type of induced travel is consistent with the broader evidence on the stability of daily travel time in widely different contexts-at slightly over one hour a day (Schäfer andVictor 1997, Stopher et al 2017).
The evidence for a definitive, non-work travel rebound is, however, inconclusive. For example, , in their analysis of travel diary data in a teleworking pilot in California, find no evidence of increased non-work travel on teleworking days. However, this lack of evidence is partly because most studies neglect non-work travel altogether, and therefore fail to detect these effects. For example, although only 15 of the 39 studies in our sample examine non-work travel, five of these find complementary travel effects. As most studies focus more narrowly upon commuter travel and ignore interactions between teleworking practices and nonwork travel, it seems likely that they overestimate the energy savings from teleworking.

Intra-household dynamics and non-work travel
The potential rebound effects discussed above may be further amplified by intra-household travel dynamics. Indeed, two studies examined the ways in which the travel behaviour of all household members is affected by one or more members' teleworking. (De Abreu E Silva and Melo 2018), for example, find that the travel effects of teleworking by one household member were different when there were two household members working. Using UK National Travel Survey data, they find that higher teleworking frequencies in one-worker households were associated with more travel by all modes, particularly by car. But in two-worker households, the estimated increase in travel was much smaller and not statistically significant. They claim that this lower increase in travel in two-worker households is due to a greater degree of sharing of household-related travel tasks between workers.
In South Korea, an additional effect was discovered, whereby home-based working by the 'head of household' led the level of household vehicle usage to increase. Using cross-sectional data for Seoul, South Korea, (Kim et al 2015) find that teleworkers' non-work trips as well as his/her household members' non-work trips were greater than those of nonteleworkers and their household members. While the daily distance travelled by the teleworking head of household fell by 7.8 km per day, this was offset by increases in the teleworker's non-commute travel (+24.2 km per day), as well as by increased non-work travel by other household members' (+1.5 km per day). But these differences were only significant in households with less than one vehicle per employed member. Car ownership is lower in South Korea than the USA (0.91 per household compared to 1.79), so the car is more of a scarce commodity. More generally, the focus of the teleworking literature on the US (where per capita car ownership is exceptionally high) may have led researchers to pay insufficient attention to the induced impact on travel by other household members.

Reflections on the types of teleworking and teleworkers
The studies examined two types of teleworking: home-based and telecentre-based. It is however difficult to assess the merits of one type over the other due to the highly specific conditions examined by different studies.
In terms of types of teleworkers, most studies examined office-based and computer-dependent workers, recognizing that these professions have the greatest potential for teleworking. For example, Matthews and Williams (2005) estimated that approximately 40% of jobs in the US and Japan would be suitable for teleworking. Within this group, studies emphasise that, above all, it is the frequency of teleworking over the course of a week that is the crucial factor in determining impacts-especially among those who live far from their place of work. Thus, from this perspective, it is full-time (or near full-time) telework that has the greatest potential for energy savings. Many of the studies examine schemes within larger companies (e.g. Atkyns et al 2002) and suggest that mass teleworking may be more realistic within large firms that can still retain a few officebased workers. In contrast, small firms whose workers take on multiple roles may be less able to encourage teleworking (Aguilera et al 2016).
While most studies investigate the impacts of a single teleworker, others examine the impacts of intra-household travel dynamics where more than one household member works, suggesting that teleworking impacts may be conditional on households being able to reconfigure non-work duties (e.g. De Abreu E Silva and Melo 2018). This would depend on economic and social capacity, as daily commuting may be an important part of households' economic strategy, with commuting travel being combined with other non-work duties, such as childcare and shopping.
Beyond the fairly unsophisticated analysis of the differing temporal frequencies of teleworking and certain intra-household work and travel dynamics, there is relatively little exploration of social differentiation among teleworkers and its impact on energy, suggesting that further research would be useful in this area. For example, none of the studies examine the gender dimensions of teleworking, although some studies in the preliminary sample of studies (e.g. Jaff and Hamsa 2018) consider such dynamics. Nor do any of the final studies examine other demographic dimensions of teleworking, such as ethnicity or political affiliation. However, many studies note the importance of household income, and observe that wealthier households may have longer commute distances on non-teleworking days (e.g. Fu et al 2012, Kim et al 2015.

Methodological assessment: a question of robustness and scope
As noted, the studies vary widely in both methodological quality and scope-raising the question of whether there is any correlation between these variables and the estimated impacts of teleworking. Table  10 maps our assessment of methodological quality against the sign of the estimated impact. This suggests that the more methodologically rigorous studies are less likely to estimate energy savings from teleworking. Specifically, 19 out of the 27 studies judged to be methodologically 'poor' or 'average' found reductions in energy use, while all six of the studies that found that teleworking led to negligible reductions or an increase in energy use were judged to be methodologically 'good' .
In terms of methods, table 11 shows that the strongest studies tended to be those analysing survey  The impact on commuting travel and one other variable The impact on only home and office energy demand (and not travel) The impact on commuting travel and two other variables data, especially those using large-scale national transport surveys and using panel and time-series data on work and travel behaviour (e.g. Kim et al 2015, Chakrabarti 2018. Although based on much smaller data sets, the studies examining specific teleworking pilot schemes-either within firms or within bounded regions (e.g. Henderson et al 1996)-also contain rich data on travel behaviour in response to teleworking. The weaker studies meanwhile projected future impacts from teleworking using scenario modelling rather than estimating historical impacts. These studies frequently relied upon limited datasets and/or unrealistic assumptions (e.g. Dissanayake and Morikawa 2008, Mamdoohi and Ardeshiri 2011). More fundamentally, as they are projecting impacts rather than measuring them, the estimated impacts rely upon modelling assumptions rather than empirical data. They are therefore a much weaker form of evidence. Table 12 maps the scope of the studies against the sign of the estimated impacts. This suggests that studies with a wider scope are also more likely to find that teleworking leads to an increase in energy use, or else has a negligible impact on energy use. Indeed, table 11 shows that all five of the studies finding that teleworking lead to an increase in energy use examined at least two variablestypically the impact on commuting travel and noncommuting travel. By contrast, 15 of the 27 studies finding that teleworking causes a reduction in energy use examined the impact on commuting travel alone.
Finally, table 13 shows the relationship between methodological quality and scope, with the studies having a wider scope (considering impact variables beyond just commuter travel) tending to be judged of higher methodological quality. Conversely, most of the studies with a narrower scope (considering the impact on commuter travel alone) are judged of lower methodological quality. Specifically, we can see that 17 out of the 19 studies with a narrow scope were rated methodologically poor or average, while 13 out of the 20 studies with a wide scope were rated methodologically good.
Overall, this analysis suggests that researchers should be wary of drawing conclusions from methodologically weaker studies that also have a narrow scope.

Conclusion and implications
This article has conducted a systematic review of the evidence on the impacts of teleworking on energy consumption. It reduced an initial sample of over 9000 academic articles to a final sample of 39 relevant studies by using specific inclusion and exclusion criteria. The final sample contained studies which investigated teleworking in a variety of contexts and which employed a range of different research methods-including scenario analysis and the quantitative analyses of survey data. The studies were predominantly focused on the US, with fewer from the EU and only three from the Global South. The studies mainly examined home-based teleworking, with three older studies examining experience with telecentres.
Overall, 26 out of 39 studies found that teleworking reduced energy use via a substitution effect, with only eight studies finding that teleworking led to higher-or else had a negligible impact on-energy use. This suggests that teleworking has some potential to reduce energy consumption and associated emissions-both through reducing commuter travel and displacing office-related energy consumption. However, a major difficulty in establishing whether teleworking does lead to a consistent relative reduction in energy use is the fact that every study provides estimates of energy savings based on a different set of conditions. Indeed, while some studies estimate the energy savings from telecommuting versus not telecommuting for single journeys (a relative figure), other studies present estimations of the total energy savings based on a specific proportion of the population telecommuting a certain number of times per week or month (an absolute figure). Some studies fail to specify either the frequency of teleworking or the proportion of teleworkers within the population that their estimates are based on. This makes it difficult to extract any normalized estimates of relative energy savings from across contexts based on identical (or similar) proportions and frequencies of teleworking. It also demands that researchers examine closely the specific configurations of conditions within particular studies that have led to particular estimates to be made for specific time periods.
While most studies conclude that teleworking can contribute energy savings, the more rigorous studies and/or those with a broader scope present more ambiguous findings. Indeed, where studies include additional impacts, such as non-work travel or office and home energy use, the potential energy savings appear more limited-with some studies suggesting that, in the context of growing distances between the workplace and home, part-week teleworking could lead to a net increase in energy consumption. In short: it is likely that many studies in the sample may have concluded that teleworking reduces energy use because their scope was too narrow-a problem identified by (Mokhtarian 2009, p. 43): Although direct, short-term studies focusing on a single application (such as teleworking) have often found substitution effects, such studies are likely to miss the more subtle, indirect, and longer-term complementarity effects that are typically observed in more comprehensive analyses.
These uncertainties and complexities suggest that, despite the positive evidence for energy savings that was found across the sample of studies, we should be cautious in drawing conclusions about the scale and consistency of energy savings from teleworking. Context matters, and in many circumstances the savings could be negative or non-existent. Moreover, the associated carbon savings will depend upon additional factors such as the carbon intensity of the energy used for transport (e.g. conventional versus electric vehicles), as well as that used for heating and cooling buildings (Moradi and Vagnoni 2018) (Giovanis 2018). Both of these are undergoing rapid change.
Furthermore, while 'teleworking' or 'telecommuting' , as terms, predate the internet itself, they also arguably refer to practices that do not reflect the dynamic new realities of working practices. Indeed, the technological basis of the working environment has changed dramatically since the 1990 s, driven by the panoply of new innovations, such as 'cloud' storage, ubiquitous high-bandwidth Wi-Fi, video streaming, and '5 G' mobile services (Appio et al 2018). So too has the range of social forms of work, with stable, single-location jobs yielding to 'zero hours' contracts and flexi-time arrangements (Akbari and Hopkins 2019). In short, modern modes of flexible or mobile work have become so non-linear and fluid (but also increasingly energy intensive in places) that it has become increasingly difficult to track their energy footprint, or to compare it with a dissolving notion of 'regular' work (Hopkins and McKay 2019).
Studies interested in appraising the potential of more flexible, ICT-enabled work practices should therefore aim to combine a range of methods capable of capturing the dynamic new configurations of working conditions. As well as accounting for change in commuting travel, non-commuting travel, distance between home and office, and home and office energy consumption, these studies must also consider other factors, such as the mode of commuting transport in the region being studied and the ways that people choose to use their time when they no longer have to commute to and from work. As many of these realities can only be established through qualitative methods, modellers must work together with other social scientists in order to build a better picture of the changing patterns of work and the energy saving potential of new working practices (e.g. Hampton 2017).
Finally, as 'flexible work' has become increasingly dependent on new energy-intensive forms of digital technologies (not to mention the reliance on rare earth metals and minerals (e.g. Sovacool et al 2019)), researchers should examine critically whether indeed new, flexible ways of working are indeed 'sustainable' , in the broadest sense (Mattila et al 2014, Priest et al 2016. Future studies in this area should therefore aim to combine a range of methods, types of work, and work arrangements in order to attempt to capture the dynamic configurations of conditions that could potentially support teleworking as a socially, economically, and environmentally constructive practice for the future.