key: cord-0063007-9foeam8h authors: Garrote Sanchez, Daniel; Gomez Parra, Nicolas; Ozden, Caglar; Rijkers, Bob; Viollaz, Mariana; Winkler, Hernan title: Who on Earth Can Work from Home? date: 2021-03-03 journal: World Bank Res Obs DOI: 10.1093/wbro/lkab002 sha: 4423318073cc476c9f143af60866cd27ad905b76 doc_id: 63007 cord_uid: 9foeam8h This paper reviews the emerging literature on which jobs can be performed from home and presents new estimates of the prevalence of such jobs based on the task content of occupations, their technology requirements and the availability of internet access by country and income groupings. Globally, one of every five jobs can be performed from home. In low-income countries, this ratio drops to one of every 26 jobs. Failing to account for internet access yields upward biased estimates of the resilience of poor countries, lagging regions, and poor workers. Since better paid workers are more likely to be able to work from home, COVID-19 is likely to exacerbate inequality, especially in richer countries where better paid and educated workers are insulated from the shock. The overall labor market burden of COVID-19 is bound to be larger in poor countries, where only a small share of workers can work from home and social protection systems are weaker. Across the globe, young, poorly educated workers and those on temporary contracts are least likely to be able to work from home and more vulnerable to the labor market shocks from COVID-19. The widespread availability of home-based work in high-income countries is in part due to changes in the nature and task content of jobs that have accompanied advances in communication and information technologies (ICT) over the last decades. The ICT revolution has fueled an expansion of high-skilled jobs that are intensive in tasks requiring cognitive skills and can feasibly be performed at home (World Bank 2019). By contrast, the demand for relatively lower-skilled workers executing routine tasks has fallen, resulting in widening labor market inequality. Since workers in jobs with routine tasks are less likely to be able to WFH, the COVID-19 pandemic will most likely reinforce longer-term trends towards job polarization and increased income inequality. This paper reviews the rapidly growing literature on the share and type of jobs that can be performed from home, various constraints on home-based work as well as their implications for the labor market and income inequality across the globe. While the academic and policy literature on this topic is expanding as rapidly as the pandemic itself with new papers being circulated each week, the ability to work from home will continue to occupy economists and policymakers long after the pandemic is over. The type and distribution of jobs which can be performed at home will stay as crucial determinants of the spatial division of economic activity, labor market competition, and income distribution in all countries regardless of their level of economic development. The remainder of this paper proceeds as follows. The next section discusses the evolving task content of jobs and how it varies across countries. The third section reviews the recent literature on the vulnerability of the COVID-19 shock across occupations depending on their task content, particularly focusing on measures of homebased work and face-to-face interactions. In the fourth section, we provide an original analysis of the importance of accounting for internet access to assess the feasibility of working from home, in particular in developing countries. It presents estimates of the prevalence of jobs amenable to working from home across countries and shows that poorer countries have more jobs at risk. The following section expands the analysis using different measures and sources of data. The sixth section explores implications of COVID-19 for inequality within countries and the seventh demonstrates that labor market risk is inversely correlated with age, job tenure, and, crucially, education levels. The subsequent section describes recent labor market developments across the globe and assesses the predictive capacity of different ex-ante measures of job vulnerability in explaining recent employment trends and the next highlights the main gaps in the literature as well as current data constraints. A final section concludes. The vast majority of studies of the labor market impacts of COVID-19 have focused on the type and nature of tasks performed in different jobs in order to determine their holders' ability to work from home. Before delving into the literature on WFH under COVID-19, we briefly discuss its genesis-the literature on the analysis of the task content of jobs. The strong correlation between expanding use of computers and demand for high-skilled labor was the main empirical observation that motivated the task content literature. Emergence of this correlation, interpreted as evidence of skill-biased technical change, can be traced back to the 1960s in the United States. Autor, Levy, and Murnane (2003) -ALM2003 from now on-argue that computers substitute for workers "carrying out a limited and well-defined set of cognitive and manual activities that can be accomplished by following explicit rules." They refer to these as "routine tasks." In contrast, computers complement workers who are "carrying out problem-solving and complex communication activities" which they label as "nonroutine tasks." They hypothesize that occupations where routine tasks are more prevalent would experience replacement by computers at a faster rate and to a greater degree as well as a parallel decline in their employment levels. The outcome is a steady increase in the relative demand for highly educated workers who have comparative advantage in nonroutine tasks. ALM2003 use data on the task requirements of jobs from the Dictionary of Occupational Titles (DOT) to characterize the occupations of workers in the Census and the Current Population Survey (CPS). 1 They find that the share of workers employed in occupations with high levels of nonroutine analytical and interpersonal tasks increased substantially between 1960 and 2000. In contrast, the share of workers in occupations with lower levels of routine cognitive and manual tasks declined. 2 Qualitatively similar trends have been documented in Europe (Goos, Manning, and Salomon 2009, Gorka et al. 2017) , and in several developing economies (World Bank 2016) . 3 This labor market polarization, in conjunction with offshoring and the rapid adoption of labor-displacing technologies stoked concerns about the future of jobs with high levels of routine tasks and the economic welfare of relatively low-skilled workers who hold these jobs. The roll-out of broadband internet access induced firms to substitute workers performing routine tasks in Norway (Akerman, Gaarder, and Mogstad, 2015) . Similarly, the adoption of industrial robots led to a significant decline in the employment of especially lower-skilled workers in several OECD countries as well as Mexico and China (Artuc, Christiaensen, and Winkler 2019; Giuntella and Wang 2019; Graetz and Michaels 2018; Acemoglu and Restrepo 2020) . 4 Autor (2015 Autor ( , 2019 argues that the technological progress "pushed the demand for skilled labor over many decades and will continue to do so." Analysis of the task content of jobs also attests to differential labor market trends in developed and developing countries as the pace of change seems to be correlated with income levels. Jobs with high levels of cognitively demanding tasks are more common in countries with higher levels of GDP per capita and technology use (Lo Bello, Sanchez-Puerta, and Winkler 2019; Lewandowski, Park, and Schotte 2020) . In contrast, routine tasks are more prevalent in jobs in developing countries. Despite the secular decline in the price of digital technologies and computing power, ALM2003's prediction-i.e., that labor performing routine tasks would be replaced by computers-has not materialized at the same pace in developing countries. Lower labor costs and barriers to the adoption of new technologies seem to have prevented poorer countries from experiencing a more rapid de-routinization of labor and increases in labor productivity (World Bank 2016; Artuc, Bastos, and Rijkers 2018) . These patterns help us to explain why the fraction of jobs amenable to WFH is lower in the developing world and why its workers may be more vulnerable to COVID-19 induced job losses. The COVID-19 pandemic has accelerated long-term trends in the task content of jobs, their relationship with communication technologies as well as their social and economic impacts. As soon as lockdown measures were in place, especially in highincome countries, workers with jobs involving high levels of nonroutine analytical and interpersonal tasks could make arrangements to work from home and thereby to keep their jobs. Their jobs can be carried out almost anywhere if there is reliable internet access. This is the case even for people with managerial responsibilities that require intensive interpersonal interactions, since most of their tasks can be carried out via online communication. In contrast, those with jobs involving intensively routine and manual tasks requiring low or mid-leves of skill were not able to work remotely. As a result, they were more likely to face income and possibly job losses. In most cases, their jobs require them to be present at a specific location or interact with clients and/or co-workers in person. Following the work of Dingel and Neiman (2020)-DN2020 from now on-most recent studies start by identifying the types and nature of tasks required by different occupations in order to assess workers' ability to work from home. To assess the feasibility to telework, DN2020 use information from 17 characteristics of more than 900 occupations based on two surveys from the US Department of Labor, Employment and Training Administration's Occupational Information Network (O * NET). They categorize jobs as "not amenable to telework" if an occupation requires daily work outdoors, physical activity, frequent contact with the public, or operating vehicles, mechanized devices, or equipment, or if its holders use email less than once a month, interact with violent people on a weekly basis, spend the majority of their time walking or running, sustain minor burns or cuts each week, or have to routinely wear specialized protective equipment. They assume that a job cannot be performed entirely from home if it meets at least one of these conditions. The DN2020 results are highly correlated with the share of time working at home according to the American Time Use Survey. DN2020 expand the task content analysis in the United States to 85 countries across the globe using ILO 2-digit ISCO occupations data. They show that there is a strong positive correlation between the GDP per capita level of a country and the share of jobs that are amenable to working from home. One key assumption of this analysis is that the task content of each occupation is the same across countries as they are all based on the O * NET data. Several articles have implemented different adaptations of DN2020. Mongey, Pilossoph, and Weingerg (2020) use O * NET questions to estimate which jobs can be done from home but by constructing a continuous index of home-based work instead of categorizing jobs as either fully "teleworkable" or not. Gottlieb, Grobovsek, and Poschke (2020) use labor force surveys from 57 countries and point out that low-income countries have a very high share of self-employed agricultural workers. Their ability to work from home impacts the overall labor market effect of COVID-19 in lower-income countries. The DN2020 measure is based on data from the United States, where farms are typically large and more reliant on hired labor. As a result, DN2020 assumes that only 8.3 percent of all agricultural workers can work from home. In poor countries farms are much smaller and a large share of agricultural workers is self-employed, which could imply that farming may be possible from home if plots are located very close to home (or perhaps, more relevant, may be feasible while respecting social distancing guidelines). Assuming that all these self-employed agricultural workers can do their jobs from home leads to a negative association between home-based work and GDP per capita. (See the analysis in the appendix based on alternative assumptions, especially on agricultural workers' ability to work from home.) Several articles use skills surveys from different countries to test if US-based measures such as O * NET lead to biased results. Saltiel (2020) follows an approach similar to DN2020 but uses worker-level data from the Skills Toward Employability and Productivity survey (STEP) from ten developing countries. Hatayama, Viollaz, and Winkler (2020) use skills surveys from 53 countries at varying levels of economic development to identify the share of jobs that can be done at home. 5 They find that there is variation in the task content for each occupation across countries, mostly due to differences in technology adoption and organization of production. Still, both papers' measures of working-from-home amenability are positively correlated with those of DN2020 and both document that the ability to telework is higher in richer countries and among more educated and formal workers. When deciding which data sources to use to assess the ability to work from home, there is a trade-off between capturing country-specific variation in task content of otherwise similar occupations and the breadth of countries for which data are available. Garrote Sanchez et al. Other papers have followed a different approach to DN2020 to identify the jobs that can be done at home, relying either on actual measures of home-based work or answers to questions in surveys about the ability to work from home. Using the American Time Use Survey (ATUS), Hensvik, Le Barbanchon, and Rathelot (2020) show that approximately 15 percent of working hours were performed from home between 2011 and 2018 in the United States. They argue that this is a likely to be a lower bound on the share of jobs that can be supplied from home. Alipour, Falck, and Schüller (2020) use data from a 2018 employment survey for Germany that includes a question on whether the worker would accept an offer from his or her employer to work from home temporarily. Adams-Prassl et al. (2020b) collected new data from late March to early April 2020 in the United States and the United Kingdom, including a question about the share of tasks that they could do at home in their current (or last) job. They demonstrate that workers who are least able to work from home are most likely to lose their jobs. Bonacini, Gallo, and Scicchitano (2021) use data from the Italian Survey of Professions (ICP), which is the Italian equivalent of O * NET, to construct an indicator of attitudes toward working from home. In general, all these articles find that amenability to work from home increases with workers' earnings and education levels. Their indices are positively correlated with those from DN2020, but since they are based on developed countries the extent to which their findings generalize to developing countries remains unspecified, given that the task content of jobs may vary with development (Lo Bello, Sanchez Puerta, and Winkler 2019). The possibility of working remotely does not always translate into efficiently working from home. Bloom et al. (2014) provided the first causal evidence on the impact of working from home on workers' productivity in a randomized control trial of workers of the Chinese company Ctrip. Workers who worked from home experienced a 13 percent increase in performance due to longer effective working hours and more efficiency given the quieter work environment. However, the current setting during the COVID pandemic is different, where several family members might need to work from home alongside children (Bloom 2020a) . In a recent survey, Barreo et al, (2020) highlights the importance of having a private space to work at home, as was the case for all the workers in the Bloom et al. (2014) study. However, less than half of respondents in the United States currently report being able to work privately in a room other than their bedroom, which can negatively affect workers' productivity. Another important precondition for being able to work from home is internet access. Barreo et al. (2020) shows that in the United States, only two-thirds of respondents had an internet connection that was strong enough to sustain video 6 The World Bank Research Observer, vol. 0, no. 0 (2021) calls while the other third had poor or no internet connectivity that hindered their ability to work from home. Beyond the feasibility of working from home, a number of studies have used other characteristics of jobs that shield them from the COVID-19 supply shock, in particular the physical proximity to other people and face-to-face interactions. Even if a job cannot be performed from home, the worker may still have low exposure to health risks and can satisfy social distancing requirements if the role does not entail frequent close interactions with coworkers, customers, or suppliers. Avdiu and Nayyar (2020) , following Blinder (2009) and using O * NET surveys for the United States, create an index of face-to-face interactions based on the intensity of particular tasks that involve: (1) establishing and maintaining personal relationships; (2) assisting and caring for others; (3) performing for or working directly with the public; and/or (4) selling to or influencing others. Similarly, Leibovici, Santacreu, and Famiglietti (2020) construct an index of contact-intensity for the United States, based on an O * NET question that asks about the extent to which the job requires the worker to perform tasks in close physical proximity to others. Overall, measures of home-based work and close face-to-face interactions are negatively correlated. This is true partly by design, as the DN2020 measure includes "performing for or working directly with the public is very important" as one of the reasons for a job not being amenable to telework. For example, information, communication, and technology (ICT) and professional and scientific jobs can more easily be performed at home and require little face-to-face interaction. On the other hand, most jobs in hospitality, food services, and health and social services are not amenable to home-based work and require extensive face-to-face interactions. Measures of physical proximity and working from home may, however, diverge for other occupations. The majority of manufacturing jobs require a physical presence in the place of work but are not always associated with extensive face-to-face interaction, thus allowing social distancing. On the other end of the spectrum, education requires significant face-to-face interaction, but those occupations are still amenable to working from home when internet access is available. When mobility restrictions are imposed, all non-WFH jobs are vulnerable to disruptions. However, once mobility is restored, activities that require more face-to-face interactions can demonstrate a slower recovery and be more affected by new norms of social distancing and attendant changes in consumer demand(s) (Avdiu and Nayyar 2020) . Therefore, jobs that require tasks that involve physical proximity are still more likely to be hit in the medium term even if mobility restrictions are removed, particularly if they cannot be carried out from home (Leibovici, Santacreu, and Famiglietti 2020) . Another aspect of job vulnerability does not relate to the type of tasks required in an occupation but on government mandates on mobility. The vast majority of governments have introduced restrictive measures in order to contain the spread of COVID-19-affecting the ability to commute to work. Although restrictions were temporarily eased in many countries during the summer of 2020, the increase in the number of COVID-19 cases pushed government officials to reintroduce these measures in many parts of the world during the subsequent fall and winter. At the same time, governments have deemed certain occupations as essential, excluding them from mobility restrictions. A number of studies have quantified what jobs are essential. Fasani and Mazza (2020b) use the European Commission's guidelines concerning the exercise of the free movement of workers during COVID-19 outbreak 6 and supplement it with the Dutch government's definition of key workers. Based on these guidelines they identify essential workers based on ISCO-08 occupations at three digits and find that about 33 percent of the working-age population in the EU have jobs that are deemed essential. Garrote- use different lists of "essential sectors" based on the decisions of Italy (EU) and the US states of Delaware, Minnesota, and Oklahoma, and mapped them to broad NACE 1-digit sectors of activity. Their results show that slightly more than half of jobs in the EU are deemed essential. One of the challenges for researchers is that the definition of essential workers varies from country to country-and even across regions within countries-(Garrote- ). However, these lists overlap considerably and typically include the delivery of critical healthcare services (such as medical professionals) and the provision of basic goods and services (e.g., food, utilities, security, ICT). Overall, occupations most affected by the COVID-19 supply shock in the short term are those that are not categorized as essential by authorities and cannot be effectively performed from home. That is, it is not just that jobs have to be amenable to telework but workers in practice need the appropriate environment and resources to work from home. In the medium-term, after mobility restrictions are removed, the intensity of face-to-face interactions can determine the extent of job vulnerability in the context of changing social norms on social distancing and the subsequent sectoral shifts in demand. Even in the case of occupations that require frequent face-to-face interactions, new technologies can reduce physical exposure. For example, education services, which traditionally involve face-to-face interactions, have shifted in many contexts to online platforms after COVID-19. Technology and internet access have, thus, a central role in shielding workers' employment from the COVID-19 pandemic, which is analyzed in more depth in the following section. This section presents new estimates of the share of jobs that can be done from home across the globe comparing various approaches discussed in the previous section, with special attention paid to the constraints imposed by internet availability. As mentioned earlier, there are several different approaches used in the literature. DN2020 use information from characteristics of more than 900 occupations based on two surveys from O * NET, US Department of Labor/Employment and Training Administration's Occupational Information Network. When answers reveal that an occupation requires daily activities such as "working outdoors" or "operating vehicles, mechanized devices," or "contact with the public," they determine that the occupation cannot be performed entirely from home. DN2020's measure, which is based on the Standard Occupational Classification (SOC) system used in the United States, needs to be concorded to the International Standard Classification of Occupations (ISCO-08) that is widely used globally at the 2-(or 3)-digit level of granularity (depending on the country). As DN2020 acknowledge, their Home-Based Work (HBW) index is likely to present an "upper bound" on the number of jobs that could feasibly be performed entirely from home, as it "neglects many characteristics that would make working from home difficult." 7 For many jobs, one of the principal constraints on performing them from home is internet access. Even when a job is in principle suitable for working from home (teleworking), that option may not be available in practice if the worker does not have internet access at home. To properly measure this constraint and account for the importance of ICT, we first need to split the telework jobs identified by the DN2020 index into two categories-jobs that require internet and those that do not require it. Then, for those telework jobs that require internet, we must identify which workers actually have access to the internet and for which workers a lack of access constitutes a constraint. Our final objective is to classify jobs as amenable to being performed from home only if they do not require internet or if they require internet and are held by workers who have internet access. Our first step is identifying telecommutable jobs that require internet access using detailed information on occupation characteristics from the O * NET surveys. We use two specific questions on the importance and frequency of computer and email use in the performance of the tasks. The answers to these questions are scored on a 5-point scale with higher numbers indicating greater dependence on computers and email use. We consider an occupation as requiring internet access if the combined average score exceeds 8 (of a total of 10). This leads to 55 percent of all SOC 8-digit occupations in O * NET being classified as requiring internet. By combining this measure with the DN2020 index, we can now distinguish four different types of occupations: (a) those that can be performed from home and require internet; (b) those that can be performed from home without the use of internet; (c) those that cannot be performed from home and do not require internet; and (d) those that cannot be performed from home but do require internet. In the United States, 33.3 percent of all jobs can be done from home and require internet-e.g., fall into group (a), while a further 3.3 percent can be performed from home without internet usage-e.g., fall into group (b). To apply our occupation-level measures to other countries, we aggregate our SOC8-digit measures to the SOC2-digit level using US employment weights from the Current Population Survey (CPS). The second step is to assess the actual availability of internet services by occupation and country. For this step, we combine information on the share of internet users by country and income level from the Gallup World Poll 2019 with data on average wages by occupation (at the two-digit disaggregation) from ILOSTAT. Gallup survey data provide the share of internet access at home among the top 60 percent and the bottom 40 percent of the income distribution in each country. This distinction enables us to account for the fact that internet use is positively correlated with income levels. The ILOSTAT data allow us to rank occupations by their average wages and assign them to either the top 60 percent or the bottom 40 percent of the income distribution in their corresponding country, which in turn allows us to concord them with the Gallup data to construct country-specific measures of internet penetration by occupation. Once we have both the share of DN2020 jobs that require internet and the internet penetration across occupations, we calculate the share of jobs that can be done from home by summing the share of jobs that can be done from home and do not require internet with the share of jobs that can be done at home and need internet multiplied by the relevant internet access rates at home from Gallup. Most of our global analysis relies on the ILOSTAT database (see ILO 2020, for details) that provides information on wages and employment numbers per occupation for over 180 countries. We restrict the country coverage to 107 countries for which 2-digit occupations are available. We also use individual data from the most recent (2018) European Labor Force Survey (EU LFS) as well as the labor force surveys from several large countries-Brazil (2017), India (2012), Mexico (2018), and Turkey (2018). These data sets are the most recent surveys and include the education level, formality status, age, wages, and occupational category of a large and representative sample of the working population and enable us to validate the conclusions derived from the analysis based on ILOSTAT data. Our first exercise is to calculate the share of jobs that can feasibly be performed at home for all of the 107 countries for which data are available in the ILOSTAT database. On average, 23.9 percent of all jobs can be done from home based on the standard DN2020 measure. However, once we account for internet access, this share drops to 18.7 percent. Put differently, failing to account for internet access would cause us, on average, to overestimate the share of jobs that can be performed from home by almost 30 percent. Measures of the feasibility of home-based work that do not consider internet access are thus upwardly biased, and this bias is especially large in low-income countries. Figure 1a plots both the DN2020 and our modified home-based work (HBW) measure against GDP per capita. The distance between the two fitted lines is a measure of the magnitude of the bias associated with ignoring ICT and internet access constraints. The bias is largest in the poorest countries. For example, 5.5 percent of all jobs in Ethiopia can be performed from home according to the DN2020 measure, while accounting for internet access reduces the prevalence to 2.1 percent. Even more strikingly, in Nepal the number of jobs that can be performed from home drops from 14.7 to 6.3 percent once internet constraints are accounted for. By contrast, in rich countries such as Switzerland, Sweden, the United Kingdom, the Netherlands, and Luxembourg-where the DN2020 measure would be between 40 and 55 percent-internet access constraints hardly matter. 8 How large is this bias in different parts of the global income distribution? figure 1a does not provide a direct answer since it does not consider country size. Figure 1b presents the share of telecommutable jobs by level of income of the countries, weighted by the size of their employed population and separating between types of home-based work. The proportion of jobs requiring internet, but lacking access is the bias of the DN2020 measure. It shows that the share of telecommutable jobs that do not require internet access is consistently very low. The share is slightly over 3 percent on average and is no more than 5 percent in any country. In other words, few jobs can be effectively done from home without internet access. Developing countries are doubly disadvantaged: First, they have fewer telecommutable jobs; second, internet access is far more binding when compared to richer economies. In low-income countries, 11 percent of all jobs are telecommutable. However, only 3.8 percent of those jobs can be effectively performed from home. The DN2020 measure thus overestimates the number of telecommutable jobs by a factor of almost 3 in low-income countries. In contrast, internet access constraints in high-income countries only prevent 1 of every 12 telecommutable jobs (3.3 percent of 37.1 percent) from being performed from home. Upper-and lower-middle-income countries are intermediate cases where internet access limitations reduce the number of telecommutable jobs by approximately 22 and 41 percent, respectively. Accounting for residential internet access limitations thus leads to the largest reductions in the number of jobs that can feasibly be performed from home in countries where telecommutable jobs are relatively scarce to start with. As a result, the correlation between GDP per capita and the feasibility of home-based work strengthens when internet connectivity is taken into consideration. This in turn implies that conventional measures of labor market exposure to COVID-19 may underestimate its impact on inequalities in job vulnerability across countries, a theme we will return to below. The construction of the home-based work indices based on O * NET surveys and the use of the Gallup World Poll data to determine internet access requires certain assumptions which might bias our results. The goal of this section is to use alternative assumptions and assess the robustness of our main findings as represented in figures 1a and 1b. Figure 2 displays the share of jobs amenable to telework using alternative assumptions. As was the case in figure 1b, countries are grouped according to their average income levels. The DN2020 index is, in essence, a weighted average of different types of tasks and embodies ICT requirements via inclusion of email dependence as a criterion in its construction. We first remove this condition to expand the set of jobs that can feasibly be performed from home without internet usage and recalculate the DN2020 index. This modification guarantees that our results are not an artifact of constraining the set of jobs that can theoretically be performed from home to be ICT dependent by assumption. As expected, removing the constraint that jobs must require frequent email use in order to be performed from home indeed raises the share of jobs that can be done at home, but only marginally. This increase is solely driven by jobs that can be done at home and do not require internet whose share increases from 3.1 to 3.6 percent in high-income countries, and from 1.7 to 2.0 percent in low-income countries ( figure 2a) . Next, we use PIAAC (Programme for the International Assessment of Adult Competencies) surveys rather than O * NET data to identify the types and extent of occupations requiring internet access. The main shortcoming of O * NET data is that they are based on the task content of occupations as performed in the United States. The PIAAC surveys, in contrast, include rich information on jobs' characteristics for 35 countries. We restrict the sample to 29 high-income countries where internet coverage is near universal to avoid our measures of ICT usage being biased downward by limited internet availability. Following Hatayama, Viollaz, and Winkler (2020), we use several questions related to internet use at work such as frequency of computer and email use, frequency of ICT usge, programming, and participating in video calls. We construct a continuous index of ICT use and we calculate the share of jobs within each ISCO 2-digit occupation that are above the 50th percentile of this index. Occupations above the median in ICT usage are determined to require internet access. We then combine this occupation-level measure of ICT requirements with the DN2020 index to identify the share of jobs that are telecommutable and do not require the internet versus the shares of jobs that are telecommutable conditional on internet access. We identify the proportion of jobs that are telecommutable and require the internet as the minimum of the share of jobs that can be performed from home according to Dingel and Neiman and the share of jobs requiring internet in each ISCO 2-digit occupation. Telecommutable jobs that do not require internet are obtained by subtracting telecommutable jobs that require internet from all telecommutable jobs. This alternative index does not change the total share of jobs that can be done from home with respect to our baseline results significantly. It increases, marginally, the share of jobs that can be done at home and require internet access but lack connectivity, particularly for lower-income countries. The results of using PIAAC to identify internet dependence are presented in figure 2b. In this case, the share of jobs that can be performed from home without internet declines, from 3.1 to 2.6 percent in high-income countries and from 1.7 to 1.0 percent in low-income countries. Our next robustness check uses a different method to allocate internet connectivity along the income distribution. Recall that the Gallup data contain the share of people with internet access only among the richest 60 percent and poorest 40 percent of households; we do not have internet access information for every income level. As a result, our measure could be underestimating connectivity for the relatively richer, and overestimating connectivity for the relatively poorer households. To address this concern, we linearly interpolate internet access by income level based on those two estimates for each country to allocate internet. As seen in figure 2c, this assumption does not change the results significantly either. The proportion of jobs that can be done in high-income countries is identical due to widespread internet access. Similarly, the change in low-income countries is also minimal, but due to an overall lack of internet access. The biggest change occurs in middle-income countries. In both upper-and lower-middle-income countries, the share of jobs that require but lack internet access declines by around 0.5 percent, increasing the share of teleworkable jobs by the same amount. Our final exercise uses data from the World Development Indicators (WDI) to allocate internet access. Since the data are not disaggregated by income level, we match the share of internet users by country by allocating connectivity to the highest earners. As an example, if 37 percent of the population has internet access, we assume that everyone in the top 37 percent of the income distribution has access and nobody in the bottom 63 percent has access. This procedure tends to increase the share of occupations amenable to teleworking across all country groups (figure 2d). Higher wage occupations are more likely to be amenable to teleworking, and this alternative allocation rule gives them preferential access to internet. The increase is about 2.3 percentage points in low-income countries, 5 percentage points in middle-income countries, and 3.1 percentage points in high-income countries. We also recalculated figure 1a under each scenario to see how the results change for each country and the results are available upon request. We see that the different assumptions explored in each of the scenarios (as presented in figures 2a-d) do not lead to dramatic changes in our estimates, increasing our confidence in the results and further analysis. The appendix presents additional robustness tests in which we explore how our results change with alternative assumptions about the ability of the self-employed, especially agricultural workers, to work from home. If anything, data from PIAAC surveys suggest agricultural employment in lower-income countries may be less suitable for working from home than in high-income countries. Based on these patterns, the DN2020 index (and our adaptation of it) does not appear to dramatically underestimate the ability of agricultural workers to work from home in developing countries. However, more research is needed to test how well these measures capture the ability of agricultural workers in developing countries to work from home. The correlation between underdevelopment and labor market vulnerability to COVID-19 is not limited to cross-country analysis; a strong correlation also exists within countries. This is shown in figure 3 , which plots the correlation between the share of jobs that can be performed from home in 280 NUTS2 subregions of Europe against local GDP per capita. Richer regions have higher endowments of jobs that can be performed from home. The share of home-based jobs tends to be higher in more developed regions of Northern European countries, compared to relatively poorer Southern European countries and EU new member states in Eastern Europe. Yet, there is significant heterogeneity within countries. Systematically, jobs performed in metropolitan areas such as Madrid, Paris, Lisbon, or Warsaw are more likely to be suitable for home-based work relative to more rural areas in the respective countries. Figure 4 presents similar patterns for Brazil, India, Mexico, and Turkey. The home-based work-income gradient is much steeper in Brazil, Mexico, and Turkey than in India, reflecting the fact that they are not only richer but also have higher levels of internet penetration. These graphs suggest that the labor market impacts of COVID-19 are not only likely to increase inequality across countries but will also exacerbate spatial inequalities within countries as the lagging and poorer regions tend to have the highest share of vulnerable jobs. 9 The next step is to analyze inequality between individuals by assessing how the prevalence of jobs that can be performed from home varies with income. Figure 5 shows how the DN2020 and our internet-adjusted measure vary with income across different groups of countries, based on the ILOSTAT database of average wages across occupations. The home-based work-individual income gradient is steepest in high-income countries which have the highest prevalence of telecommutable jobs. In the top income decile in high-income countries, almost 80 percent of jobs can be performed from home. Accounting for internet access hardly impacts the Labor Force Survey 2017-18; and regional GDP per capita from Eurostats; the System of Regional Accounts from the Instituto Brasileiro de Geografía e Estatísitca (IBGE); Instituto Nacional de Estadística, Geografía e Informática (INEGI) of Mexico; the 2012 regional statistics from the Reserve Bank of India (RSI); and the 2018 regional statistics from the Turkish Statistical Institute (TSI). Notes: All regional GDP per capita are adjusted using conversion factors (PPP 2011 international dollars, CPI, and exchange rates) from the World Development Indicators (WDI). gradient in rich countries, where internet access is much less likely to be a binding constraint. By contrast, the gradient is the least steep in low-income countries, which have few telecommutable jobs to start with. Accounting for internet access flattens the gradient the most in poor countries with limited internet access. The impact is strongest for the higher-income people in the least developed countries since they are more likely to have jobs that can be done from home but face internet access constraints. Workers in poorer households, in contrast, typically do not have jobs amenable to home-based work and are therefore not affected by internet restrictions. In short, COVID-19's impact is quite different in rich and poor countries. Absent interventions, inequality is likely to rise the most in rich countries, but poor countries have higher shares of jobs at risk. Figure 6 illustrates these patterns using representative data at the individual level from India, Brazil, Mexico, and Turkey. The chance of having a job that can be performed from home increases with individual income in all countries but less so in India, where labor market vulnerability is most widespread and internet access most limited. A worker in the top earnings decile in India has a 19 percent chance of having a job that can be done from home, whereas a worker in the bottom percentile has less than 1 percent chance of having such a job. By contrast, a worker in the top earnings decile in Turkey (which has the highest PPP adjusted GDP per capita in this group) has a 55 percent probability of having a job that can be performed from home, whereas a worker in the bottom decile only has a 7 percent chance of having such a job. COVID-19 is thus likely to exacerbate inequality within almost every country, but more so in higher-income countries. This conclusion is supported by figure 7, which presents estimates of the impact of COVID-19 on earnings inequality-measured by the Gini coefficient-under alternative assumptions about its impact on incomes. We assume COVID-19 leaves the incomes of those working in jobs that can be performed from home unaffected while all other workers lose, respectively, 30 percent and 50 percent of their incomes. The figure shows the changes in the Gini coefficient under these two scenarios. Since richer individuals tend to be insulated from such shocks because of their ability to work from home, inequality is exacerbated, especially in rich countries, where a larger share of the higher income workforce can work from home. The increase in inequality varies with the magnitude of the income shocks; larger losses incurred by those who cannot work from home are associated with sharper increases in inequality. In poorer countries, because of lack of telecommutable jobs as well as Our analysis up to this point captured averages across countries by occupation groups, income deciles, or sub-national geographic areas. The data indicate that labor market shocks associated with the COVID-19 pandemic impact poor countries, poor regions, and poor people more negatively. We now assess whether there are personal characteristics of workers that can explain these patterns. More specifically, we assess which workers are most at risk by running individual-level regressions in which the dependent variable is having a job that can be performed from home. We estimate separate regressions for European Union Countries based on the EU Labor Force Survey of 2018. We control for age, gender, and education, first separately and then jointly. The results for the European Union countries are reported in table 1. Next, we perform the same analysis for Brazil, India, Mexico, and Turkey. Table 2 presents the results for these countries. Several common patterns emerge as seen in tables 1 and 2. Young workers (i.e., those between 15 and 24 years of age), who comprise the omitted age category in our regressions, are significantly less likely to have a job suitable for home-based work than older people across all countries. Unlike the health risks of COVID-19, which are disproportionately borne by the elderly, the economic risk is thus concentrated among the youth. Second, and most important, labor market vulnerability is inversely correlated with educational attainment. Workers with tertiary education are much more likely to be able to work from home in all countries and regions. Education is the strongest predictor of who has a relatively safe job among the set of explanatory variables we consider here. While education offers protection in all countries, the probability of having a job suitable for working from home increases least with additional education in India, which is not surprising given that India has fewer jobs that can be performed from home to start with. Interestingly, when using the DN2020 telework variable instead of our home-based work measure, the coefficient of education level Garrote Sanchez et al. Note: Standard errors are clustered at the region and occupation (ISCO 2-digit) level and presented in parentheses. *** , ** , * denote significance at the 10 percent, 5 percent, and 1 percent significance level respectively. The omitted category comprises 15-24 year old workers who completed primary education. in India becomes similar to the one in the other studied countries, which attests to a lack of internet access being a binding constraint on highly educated Indians' ability to work from home. Third, workers in temporary jobs are less likely to have jobs that can be performed from home. This is worrisome, as they are more susceptible to losing their jobs, and reinforces the conclusion that COVID-19 is likely to exacerbate labor market inequality and will disproportionately impact those least protected. Including all the explanatory variables, together with the regional fixed effects, does not change the significance of specific variables. Education level, age, and job security are still highly important for the ability to perform a job from home even when regional heterogeneity is taken into account in each country. Note: Standard errors are clustered at the region and 2-digit occupation (isco2d level) and presented in parentheses. *** , ** , * denote significance at the 10 percent, 5 percent, and 1 percent significance level respectively. Region fixed effects are included in all the estimations. The omitted category comprises 15-24 year-old workers who completed primary education. For India, the "Temporary" variable is based on an indicator of informality indicator instead of the type of contract due to data availability. These findings dovetail with evidence from Latin America (Bottan, Hoffman, and Vera 2020), South Asia (ILO 2020b; World Bank 2020), and sub-Saharan Africa (Balde, Boly, and Avenyo 2020) that workers in informal jobs are more likely to have lost their jobs and income, in part because informal jobs tend to be more contactintensive. Evidence for Italy shows that workers on a temporary contact are at higher risk of losing their job due to the pandemic (Casarico and Lattanzio 2020). In the same vein, employment in large firms has been more resilient to the COVID-19 pandemic than employment in small firms (IMF 2020). In the United States, Campello, Kankanhalli, and Muthukrishnan (2020) find a steeper decline in job postings by firms of smaller size, with higher levels of unionization, and in non-tradable sectors. Garrote Sanchez et al. One group that is particularly vulnerable to the economic consequences of the COVID shock are immigrants. Migrant workers systematically differ from the native labor force on a number of socio-economic characteristics and they are disproportionately concentrated in certain occupations and sectors where they have comparative advantage. While natives specialize in occupations requiring more intensive communication and language skills, immigrants pursue jobs needing manual or quantitative skills (Peri and Sparber 2009, 2011) . The self-selection of migrants into occupations has led to an asymmetric effect of COVID, with migrants concentrated in more vulnerable jobs. Yasenov (2020) follows DN2020 and finds that only one in three migrants in the United States have jobs suitable for telework, compared to 45 percent of natives. Fasani and Mazza (2020b) find similar results for migrants in the European Union, in particular those coming from extra-EU developing countries. Given that migrants were more likely to have more vulnerable working conditions, with lower income and less job security, the COVID shock exacerbates pre-existing inequalities vis-à-vis natives. Borjas and Cassidy (2020) analyze employment trends since the COVID outbreak and find that migrant males had higher rates of losing employment. Migrants out of work at the onset of the crisis were also less likely to find new employment. The authors attribute the larger impact on employment to the higher concentration of migrants in jobs with lower potential to being performed remotely. On the other hand, the presence of migrant workers pushes natives, in particular those with higher education, towards occupations that are more suitable for working from home, thus, partially shielding them from the COVID shock (Bossavie et al. 2020) . How well do these vulnerability measures predict actual outcomes? The aggregate labor market impact of COVID-19 is unprecedented; workplace closures, working hour losses, and labor income losses have been higher than in any previous recession or crisis according to the ILO (2020). Between the fourth quarter of 2019 and the second quarter of 2020, global working hours declined by the equivalent of 495 million full-time jobs. Global labor income is projected to decline by 10.7 percent in 2020. Consistent with the cross-country patterns of vulnerability identified using measures of home-based work, the labor market burden of COVID-19 is negatively correlated with GDP. Lower-middle-income countries suffered a loss in working hours of 15.6 percent in the third quarter of 2020 with respect to the last quarter of 2019, compared to an estimated reduction of 9.4 percent in high-income countries for the same period (ILO 2020). The vulnerability metrics also predict differences in labor market outcomes across different groups where disadvantaged socio-economic groups within countries seem to be impacted more negatively. Based on this comparison of the labor market outcomes of the second quarters of 2019 and 2020, workers with low incomes, with low levels of education, younger adults, ethnic minorities, and immigrants are concentrated in occupations that are more likely to be affected by the lockdowns. The employment levels of women, of younger people, and of less educated workers have declined the fastest. For example, while the number of employed college-educated workers declined by only 0.7 percent year on year, the employment levels of noncollege educated people decreased by over 10 percent. People between 15-24 years of age lost 22 percent of their jobs while the same number is only 5.6 percent for 55-64-year-olds. Employment in various services (trade, transportation, accommodation) declined by almost 10 percent whereas manufacturing employment increased marginally, and construction employment decreased only slightly. Differences in employment vulnerability by education levels are similar across sectors. While highly-skilled occupations that can be performed from home (such as for managers and professionals) experienced employment declines of between 3-6 percent, lower-skilled workers (such as those in sales) lost over 15 percent of their jobs. To assess which of the different ex-ante measures of job vulnerability has the most predictive power we conduct a validation exercise using data from the European Union. Specifically, table 3 reports their correlations with actual employment losses in the 2nd quarter of 2020 (vis-à-vis employment in the 2nd quarter of 2019) for a given 1-digit occupation in a given 1-digit sector in every EU country for which Eurostat data is available. As shown in the table, measures of ability to work from home (DN2020), face-to-face jobs (Avdiu and Nayyar 2020), or essential jobs (Fasani and Mazza 2020a) are strongly correlated with actual COVID-induced job losses. Among them the DN2020 measure adjusted for internet access has the highest predictive power. A 10 percentage point increase in the share of workers who can work from home reduces initial COVID-induced employment losses by 1.1 percentage points. When splitting the DN2020 measure into (i) teleworkable jobs that do not require internet, (ii) jobs that require internet and have access, and (iii) those that require internet but do not have access, we find that only the shares of jobs in the former two categories are correlated with employment outcomes. Jobs for which internet constraints are binding thus seem to be spuriously included in the original DN2020 measure. Put differently, taking internet constraints into consideration helps to predict who is most at risk. The results presented in table 3 are consistent with other very recent papers. Montenovo et al. (2020) and Liu and Mai (2020) use CPS monthly employment data in the United States up to May and April 2020 respectively and show that job losses were greater in occupations that required more physical proximity and those that were less suitable for teleworking. Using weekly administrative payroll data from the largest US payroll processing company up to June 2020, Cajner et al. (2020) also observe a strong positive correlation between changes in employment and measures of feasibility to work from home per 3-digit industry. In the EU, Fasani and Mazza (2020b) categorize occupations by level of ex-ante vulnerability based among other (Avdiu and Nayyar 2020 in listing.) 0.030 *** 0.031 *** 0.032 *** (0.006) (0.008) (0.008) Face-to-face ( Note: Standard errors clustered at the country level are presented in parentheses. *** , ** , * denote significance at the 10 percent, 5 percent, and 1 percent significance level respectively. Each observation represents one ISCO occupation (1-digit) * NACE sector (1-digit) * country. The dependent variable is year to year change in employment between the 2nd quarter of 2019 and 2020. Telework jobs are jobs that are suitable for working from home according to Dingel and Neiman (2020, DN 2020) . Home-Based Work (HBW) jobs are suitable working from home taking into consideration internet access constraints as explained in the article's fourth section "Working from Home and Internet Access". They can be divided into three groups: (i) Teleworkable without internet: teleworkable jobs that do not require internet, (ii) Teleworkable with internet, have access: jobs that require internet and have access, and (iii) Teleworkable with internet, no access: those that require internet but do not have access. Essential jobs are jobs classified as essential based on guidelines from the EU and the Dutch government by Fasani and Mazza (2020b, FM 2020b) . Face-to-face jobs are ones that require extensive in-person interaction as measured by the index constructed by Avdiu and Nayyar (2020, AN 2020) . factors on the contractual protection workers enjoy and the potential to telework, and they find that these measures are closely correlated with actual employment losses in European countries after COVID with data up to the second trimester of 2020. Nonetheless, these measures only capture one dimension of labor market vulnerability, as a preliminary review of the differential impact of COVID-19 by gender reveals. Even though women are more likely to hold jobs that can be performed from home in many countries, they appear more likely to have lost their jobs during the COVID-19 crisis (Adams-Prassl et al. 2020a,b; Alon et al. 2020; Andrew et al. 2020; Bèland et al. 2020; Farré et al. 2020; Mongey, Pilossoph, and Weingerg 2020; Montenovo et al. 2020; Qian and Fuller 2020; Sevilla and Smith 2020) . Their hours of work have declined more (Andrew et al. 2020; Collins et al. 2020) , and, in the United States, the partial recovery in employment over the last months has benefited men more than women (Bick, Baldwin, and Mertens 2020) . The available evidence also points to a positive correlation between spouses' home-based work opportunities, suggesting that the COVID-19 crisis is likely to impose a constraint on intra-household insurance through adjustments in spouses' (typically female) labor supply (Malkov 2020; Peluffo and Viollaz forthcoming) . The larger impact of COVID-19 on female unemployment exhibits a sharp contrast to the patterns from the previous recessions, in which men were harder hit (Alon et al. 2020) . Two main reasons can explain this difference. First, in regular recessions, sectors with higher shares of male employment, such as construction and manufacturing, suffer larger employment reductions. In the current recession, the sectors more affected have been the contact-intensive service sectors, such as travel, hospitality, and retail which also employ larger shares of women (Albanesi et al. 2020; Alon et al. 2020; Hupkau and Petrongolo 2020; Mongey, Pilossoph, and Weingerg 2020; Queisser, Adema, and Clarke 2020) . The second reason is related to the increased housework and childcare responsibilities resulting from the closing of schools and nurseries. The evidence shows that, in general, both women and men increased the amount of time allocated to childcare and housework, but the extra time was greater among women (Adams-Prassl et al. 2020b; Del Boca et al. 2020; Lyttelton, Zang, and Musick 2020; Sevilla and Smith 2020) . These findings are consistent with existing studies documenting that women are more likely to miss work to shoulder caregiving responsibilities resulting from illness shocks to family members (Heath, Mansuri, and Rijkers forthcoming) . This result is observed even when comparing women and men in the same employment situation (Adams-Prassl et al. 2020b) . More generally, how vulnerable workers are to crises induced by labor market strain also depends critically on both (i) the quality of pre-existing social protection systems (Paci, Revenga, and Rijkers 2012) , which tend to be weaker in developing countries, and (ii) remedial policy responses. Countries that acted swiftly and spent more have experienced smaller economic losses during the COVID-19 pandemic (Deb et al. 2020; Demirgüç-Kunt, Lokshin, and Torre 2020) . Simulation evidence for Latin America shows that losses may be greatest in the middle of the ex-ante income distribution rather than among the poorest, because social assistance policies put a "floor" on the incomes of the poorest (Lustig et al. 2020) . Similarly Clark, D'Ambrosio, and Lepinteur (2020) argue that policy responses to COVID-19 may have offset its disequalizing impact in Europe. By September, relative income inequality in France, Germany, Italy, Spain and Sweden had fallen compared to its pre-pandemic levels, despite an initial increase in income inequality in the early stages of the pandemic. In the previous sections we have discussed the data used in other papers as well in our analysis. Naturally, more detailed data can only improve the quality of the analysis and this section lists several suggestions that would enhance future analysis. First, the ability to work-from-home depends on other infrastructure characteristics of the dwelling where a worker lives, such as space restrictions. For instance, an office worker might lack adequate physical space to use as a home office or the proper equipment required to work from home. Space restrictions interact with family size and composition, hindering productivity. Similarly, data on family composition would be crucial to identify the constraints that care for dependants (children and the elderly) may impose on the potential to work from home. As the previous section points out, these family care constraints are among the reasons why women are more likely to be negatively impacted even though they are more likely to hold jobs that can be performed from home. The reason for not including these determinants in our HBW index is the absence of microdata for a large number of countries. Second, data on internet access at the worker (or household) level is necessary to obtain a more precise HBW index. Internet quality may impose a restriction on HBW possibilities as well, but comparable micro data on internet access and internet quality for a wide range of countries are not available. Third, using skills surveys to determine the precise task content of each country's jobs (as in Hatayama, Viollaz, and Winkler 2020) will improve the accuracy of HBW measures, particularly among less developed countries. For example, workers are more likely to use nonroutine skills and higher levels of technology to perform the same job in higher-income countries. The rapid adoption of skills surveys in countries at different levels of income is an encouraging development for this research agenda. In the absence of skills surveys, a possible approach would be to use questions such as "what fraction of the tasks in your current job can be done from home?" (as in Adams-Prassl et al. 2020b) . However, these types of questions are not available in most countries' labor force surveys. Fourth, and related to the previous point, existing data are silent regarding the suitability of jobs to WFH in the rural areas of low-income countries. As mentioned above, while farmers in subsistence agriculture may be able to WFH, this is not necessarily the case if they sell part of their production outside the home. Finally, while WFH has been a key factor in protecting jobs during the pandemic, it is still not clear to what extent it is an efficient organization of work in the longer term. While Bloom et al. (2014) find that WFH has positive impacts on productivity, it is not obvious if this would also be the case for workers who, for example, do not have the right infrastructure at home or live in a crowded dwelling. At the same time, more research is needed to understand for what types of workers, firms, or sectors, WFH provides an efficient solution. For example, empirical evidence from the Business and Management literature suggests that physical proximity and face-toface interactions can have significant impacts on innovation (Choudhury 2017) , but that physical distance between managers and employees can give rise to separation costs in certain situations (Bonet and Salvador 2017) . At the same time, understanding the productivity implications of hybrid models where WFH is not a binary outcome but a continuous one (e.g., WFH on certain days of the week or hours of the day) could be an important area of future research. The COVID-19 pandemic will continue to cause severe labor market pain across the globe in the foreseeable future. After reviewing the literature, we create a new measure of which jobs can be performed from home by combining information on the task content of jobs with information on internet access by country and income groupings. On average, one in five jobs across the globe can be performed from home, but this number masks enormous heterogeneity across countries because the ability to telework is correlated with income. In high-income countries one in every three jobs is suitable for home-based work, while in low-income countries only one in every 26 jobs can be done at home. Failing to account for internet access would cause one to overestimate the prevalence of jobs amthat can be home-based in low-income countries by a factor of 4, and hence cause one to underestimate the vulnerability of poor countries. The latter suffer two disadvantages: they have fewer jobs that are theoretically telecommutable to start with, and limited internet access is a bigger handicap. Telecommutable jobs are highly unequally distributed across space, not only across but also within countries. They are less prevalent in lagging regions. The COVID-19 pandemic is thus likely to exacerbate spatial inequality, especially when one considers that local governments in lagging regions may have less fiscal capacity to cushion the COVID-19 shock. Across all countries, jobs that can be performed from home tend to be much better paid. Absent remedial action, the COVID-19 pandemic is thus likely to exacerbate inequality, and especially so in relatively richer countries given the higher prevalence of jobs suitable for home-based work. Yet, the bulk of the labor market pain will be shouldered by workers in developing countries given the very limited feasibility of working from home and their limited recourse to social safety nets. Across the globe, young, poorly educated workers and those with temporary contracts are especially exposed to labor market pain induced y COVID-19, which is worrying since they are more vulnerable at the outset. The COVID-19 crisis is therefore bound to exacerbate domestic as well as global labor market inequality. We are grateful to Ana Fernandes, Chisako Fukuda, John Giles, Gaurav Khanna, Aart Kraay, Peter Lanjouw, Norman Loayza, William Maloney, David McKenzie, Harry Moroz, Nina Rahman, Achim Schmillen, Joana Silva, and three anonymous members of the editorial board of the World Bank Research Observer for useful comments and discussions, to Leora Klapper for her help with the Gallup data on internet access and to Efsan Nas Ozen for the Turkish Labor Force survey. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank of Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the countries they represent. We are also thankful to the Knowledge for Change Program, the World Bank Research Support Budget, and the Multi-Donor Trust Fund on Trade for financial support. All errors are our responsibility. 1. The DOT was succeeded by the Occupational Information Network (O * NET), which is the source of information used by Dingel and Neiman (2020) to categorize occupations by their amenability to WFH. 2. This is supported by other studies such as Autor and Dorn (2013) and Autor, Dorn, and Hanson (2015) , who find that local labor markets more exposed to computerization experienced job polarization. 3. Polarization is more muted in poorer countries. 4. Graetz and Michaels (2018) use data from 17 high-income countries and find that robots do not affect total hours worked, but they reduce the hours of low-and mid-skilled workers. 5. They use the Survey of Adult Skills from the Programme for the International Assessment of Adult Competencies (PIAAC), surveys from the Skills Toward Employment and Productivity (STEP) and the Labor Market Panel Surveys for Middle East countries. 6. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52020XC0330(03) 7. In addition, it is not clear how the tasks required to perform an occupation are the same across economies at different levels of development. Using PIAAC (Programme for the International Assessment of Adult Competencies) data at the country level, Hatayama, Viollaz, and Winkler (2020) observe changes in the ranking of countries in terms of their jobs' suitability to WFH when comparing the measure based on PIAAC with DN2020's measure, which are highly correlated. Since most of the lower income developing countries do not have data on the task content of their jobs (the PIAAC survey covers mostly OECD countries), we apply the DN2020 index to all countries. The results from PIAAC countries using the Hatayama, Viollaz, and Winkler's (2020) methodology are very similar and available upon request. 8. Recall from an earlier discussion that the association between GDP per capita and the share of jobs amenable to working from home depends critically on what we assume about the ability of agricultural workers to do their jobs from home; assuming that the agricultural self-employed can work from home would result in a negative association between GDP and the prevalence of work suitable for working from home (Gottlieb, Grobovsek, and Poschke 2020) . 9. In contrast, there is less spatial variability in the share of jobs amenable to telework in India (not shown here but available from the authors upon request), reflecting the severely limited internet penetration in the country at the residential level for employment purposes. Secular Stagnation? The Effect of Aging on Economic Growth in the Age of Automation Robots and Jobs: Evidence from US Labor Markets Inequality in the Impact of the Coronavirus Shock: Evidence from Real Time Surveys Work Tasks That Can Be Done from Home: Evidence of the Variation Within and Across Occupations and Industries The Skill Complementarity of Broadband Internet Household Insurance and the Macroeconomic Impact of the Novel Coronavirus Germany's Capacities to Work from Home The Impact of COVID-19 on Gender Equality 2020. Parents, Especially Mothers, Paying Heavy Price for Lockdown. London: Institute for Fiscal Studies Robots, Tasks and Trade Does Automation in Rich Countries Hurt Developing Ones? Evidence from the US and Mexico Why Are There Still So Many Jobs? The History and Future of Workplace Automation Work of the Past, Work of the Future The Growth of Low-skill Service Jobs and the Polarization of the US Labor Market Untangling Trade and Technology: Evidence from Local Labour Markets The Skill Content of Recent Technological Change: An Empirical Exploration When Face-to-face Interactions Become an Occupational Hazard: Jobs in the Time of COVID-19 Labour Market Effects of COVID-19 in Sub-Saharan Africa: An Informality Lens from Burkina Faso, Mali and Senegal Why Working From Home Will Stick Covid-19, Family Stress and Domestic Violence: Remote Work, Isolation and Bargaining Power Work from Home after the COVID-19 Outbreak How Many US Jobs Might be Offshorable? How Working from Home Works Out. Institute for Economic Policy Research (SIEPR) The New Empirical Economics of Management Working from Home and Income Inequality: Risks of a 'New Normal' with COVID-19 When the Boss Is Away: Manager-Worker Separation and Worker Performance in a Multisite Software Maintenance Organization The Adverse Effect of the COVID-19 Labor Market Shock on Immigrant Employment Do Immigrants Push Natives towards Safer Jobs? Exposure to COVID-19 in the European Union The Unequal Burden of the Coronavirus Pandemic: Evidence from Latin America and the Caribbean The U.S. Labor Market During the Beginning of the Pandemic Recession Corporate Hiring under COVID-19: Worker Downskilling, Job Flexibility, and Income Inequality The Heterogenous Effects of COVID-19 on Labor Market Flows: Evidence from Administrative Data Innovation Outcomes in a Distributed Organization: Intra-Firm Mobility and Access to Resources The Fall in Income Inequality during COVID-19 in Five European Countries COVID-19 and the Gender Gap In Work Hours The Economic Effects of COVID-19 Containment Measures Women's and Men's Work, Housework and Childcare, before and during COVID-19 The Sooner, the Better: The Early Economic Impact of Non-Pharmaceutical Interventions during the COVID-19 Pandemic How Many Jobs Can Be Done at Home? How the COVID-19 Lockdown Affected Gender Inequality in Paid and Unpaid Work in Spain Immigrant Key Workers: Their Contribution to Europe's COVID-19 Response A Vulnerable Workforce: Migrant Workers in the COVID-19 Pandemic World Poll Survey Data Which Jobs Are Most Vulnerable to COVID-19? Analysis of the European Union Job Polarization in Europe Tasks and Skills in European Labor Markets. Background Paper for Growing United Working from Home across Countries Robots at Work Is an Army of Robots Marching on Chinese Jobs Short term Labour Market Effects of COVID-19 and the Associated National Lockdown in Australia: Evidence from Longitudinal Labour Force Survey Jobs' Amenability to Working from Home: Evidence from Skills Surveys for 53 Countries Labor Supply Responses to Health Shocks: Evidence from High-Frequency Labor Market Data from Urban Ghana Which Jobs Are Done from Home? Evidence from the American Time Use Survey Work, Care and Gender during the COVID-19 Crisis Annual Report: A Year Like No Other International Labour Organization Database (ILOSTAT)-Employment by Occupation-ISCO 08 2 digit Social Distancing and Contact-intensive Occupations The Global Distribution of Routine and Nonroutine Work Employment during the COVID-19 Pandemic: Collapse and Early Recovery From Ghana to America: The Skill Content of Jobs and Economic Development The Impact of Covid-19 Lockdowns and Expanded Social Assistance on Inequality, Poverty and Mobility in Argentina, Brazil, Colombia and Mexico Gender Differences in Telecommuting and Implications for Inequality at Home and Work Nature of Work and Distribution of Risk: Evidence from Occupational Sorting, Skills, and Tasks Government of India: National Sample Survey Organization-NSSO. 2012. National Sample Survey 2011-12 61th round Which Workers Bear the Burden of Social Distancing Policies Determinants of Disparities in Covid-19 Job Losses Intra-Household Insurance in the Time of Covid-19: Lessons from Mexico Coping with Crises: Policies to Protect Employment and Earnings Task Specialization, Immigration, and Wages Highly-Educated Immigrants and Native Occupational Choice COVID-19 and the Gender Employment Gap among Parents of Young Children COVID-19, Employment and Women in OECD Countries Supply and Demand Shocks in the COVID-19 Pandemic: An Industry and Occupation Perspective Who Can Work from Home in Developing Countries?" Mimeo Baby Steps: The Gender Division of Childcare during the Covid-19 Pandemic Regional Accounts: Gross Domestic Product Per Capita World Development Report 2019: The Changing Nature of Word IZA Institute of Labor Economics Discussion Paper No Given the widespread prevalence of agricultural (self-)employment in poorer countries, assumptions about the ability of agricultural jobs to be performed from home are an important determinant of labor market vulnerability to COVID-19. To assess their importance, we recompute our measures of the ability to work from home by making an extreme assumption: All agricultural jobs can be performed from home. This assumption is similar to the one made by Gottlieb, Grobovsek, and Poschke (2020) . The results in figure A1 shows that this assumption (presented as the green line) leads to a U-shaped relationship between the share of jobs that can be done from home and GDP per capita, instead of the earlier positive relationship. We should note that the GDP per capita is presented using the logarithmic scale (x-axis) and we would have a more positive relationship if we used a linear scale. This is due to the fact that both the ICT constraints are more binding and the share of agricultural self-employment is higher in lower-income countries.We also explore whether the share of agricultural jobs that can be performed from home varies with development using the PIAAC surveys. Specifically, we construct a crude Work From Home Index similar to DN2020. We consider a job as not being suitable for working from home if (i) the job requires physical work for an extended period at least once a week, (ii) the frequency of email use is less than once a month, or (iii) the job involves selling products or services at least once a week. The results, which are presented in figure A2a, show that very few agricultural jobs can be performed from home across all countries according to these criteria. More importantly, the ability to work from home as proxied by this metric is not correlated with GDP per capita, but exhibits an inverse-U relationship. In figure A2b we recompute the share of agricultural jobs that can be performed from home when we eliminate the frequency of email use condition. This leads to an increase in the prevalence of jobs suitable for working from home. But home-based work remains the exception rather than the norm. Using this alternative proxy, the share of jobs suitable for working from home is negatively correlated with income per capita.These results illustrate that conclusions about the ability of agricultural jobs to be performed from home are very important for assessing the aggregate labor market vulnerability to COVID-19. They also suggest that the assumption that all agricultural jobs can be performed from home might not be too realistic, even in developing countries where farms are smaller, and agriculture is informal. As the share of agricultural jobs is significantly higher in the poorest countries, investigating how many agricultural jobs can in fact be performed from home remains an important question for further research. The World Bank Research Observer, vol. 0, no. 0 (2021) Garrote Sanchez et al.