Access the full text.
Sign up today, get DeepDyve free for 14 days.
The Covid-19 crisis has forced great societal changes, including forcing many to work from home ( WFH) in an effort to limit the spread of the disease. The ability to work from home has long been considered a perk, but we have few esti- mates of how many jobs are actually possible to be performed from home. This paper proposes a method to estimate the share of these jobs. For each occupation, we obtain a WFH friendly measure by asking respondents from Amazon Mechanical Turk (MTurk) to evaluate whether the corresponding tasks can be performed from home based on the descriptions from the International Standard Classification of Occupations 2008 (ISCO-08) standard. The share of WFH friendly jobs in an economy can then be estimated by combining these measures with the labor statistics on occupa- tional employments. Using Norway as an illustrating example, we find that approximately 38% of Norwegian jobs can be performed from home. The Norwegian results also suggest that the pandemic and the government’s attempts to mitigate this crisis may have a quite uneven impact on the working population. Those who are already disadvantaged are often less likely to have jobs that can be performed from home. Keywords: COVID-19, Working from home, Job advertisements, Unconventional data JEL Classification: D24, J22, J61, O30, R12, R32 1 Introduction “social distancing” policy and the Covid-19 crisis could be Covid-19 pandemic hit the world hard and unprepared. In highly unequal. Identifying non-WFH workers would be a study of the Spanish flu, Hatchett et al.(2007) show that essential for designing effective economic austerity pack - non-pharmaceutical interventions known as “social distanc- ages. However, there is rather limited knowledge of the ing” during a pandemic can significantly reduce the dis - prevalence and characteristics of non-WFH workers. ease transmission and lower both the peak and cumulative In this paper, we propose a method to answer the ques- excess mortality. Learning from the historical lessons, many tion: who and how many can work from home in an countries implemented measures to limit physical contact economy. For this purpose, we evaluate the WFH fea- between people. Encouraging working from home is an sibility for the 426 occupations listed in the ISCO-08 important part of such measures. Not all jobs can be per- (International Labor Organization 2012). In particular, formed away from offices. Workers with non-WFH friendly respondents from an on-line labor marketplace, Amazon jobs will be hit harder by such policies, since they are Mechanical Turk (MTurk) (Amazon 2020), are asked to forced into a situation where they have to choose between evaluate whether they think occupations can likely be two unfortunate options: increased risk of infection or sub- performed from home using the detailed descriptions stantial economic loss due to lost work opportunities. Simi- of tasks to be performed. Based on the responses, we larly, firms and regions with few WFH workers may be more establish a WFH friendly measure for each occupation. Combing these measures with occupational employment severely impacted than others. Thus, the potential impact of *Correspondence: firstname.lastname@example.org We will refer to workers with WFH friendly jobs and non-WFH jobs as Statistics Norway, Post Box 2633, 0131 Oslo, Norway WFH workers and non-WFH workers respectively hereafter. © The Author(s) 2021. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat iveco mmons .org/licen ses/by/4.0/. 4 Page 2 of 13 H. Holgersen et al. statistics, we can obtain information on the prevalence factors, such as regulations, working cultures, attitudes of of jobs that can be worked from home. If in addition workers and managers and etc. Moreover, the surveys are employment information on individual level is available, often implemented some years back which may lead to we can link workers, jobs and occupations together and concerns of timeliness. Considering the rapid technologi- identify what type of workers are less likely to have WFH cal progress in information technology, this concern may be friendly jobs. This method can be easily applied to other particularly relevant. On the other hand, new surveys on a economies of interest. In current paper, we focus on Nor- representative sample of the working population are costly way and use it as illustrative example. The patterns we in term of both time and resource, which makes it not a find based on Norwegian data may be informative for practical alternative for our purpose. The method we used other industrial countries since the occupational struc- in this paper is similar to the manual evaluation method tures in these economies are often similar. applied by Dingel and Neiman (2020). However, we don’t do We find that approximately 38% of Norwegian jobs evaluations ourselves but rely on respondents from MTurk. can be performed at home. WFH-friendly jobs typically While we acknowledge that well-designed surveys are still pay better than non-WFH friendly jobs. The prevalence the most reliable source, our study suggests an unconven- of such jobs varies a lot across geographical areas. There tional data source where reliable information can be easily is a larger share of WFH friendly jobs in urban than in obtained timely with much less cost. rural areas. More importantly, as many worried, workers There are clear limitations with our method. For exam - who are already disadvantaged in the labor market, such ple, the respondents from MTurk are most likely not as young workers, workers with migrant background, low expert in the field of interest. Or the description of tasks educated and lone parents, are often less likely to have may not be entirely clear. These issues will lead to poten - WFH friendly jobs. We have also combined our WFH tial bias in our measure. Although we cannot directly test friendly measures with country specific employment the reliability of our WFH friendly measures, we have data from Eurostat and estimated the prevalence of WFH done several consistency checks and we did not find evi - friendly jobs in other European countries. We find that dence of large bias. We first compare our results with rich and more developed countries have larger shares of that of Dingel and Neiman (2020). We find that these two jobs that can be performed from home than poor and measures are very similar on the ISCO-08 major group less developed ones. occupation level (correlation 0.96), although they differ The rapid development of information and communica - somewhat on the unit group occupation level (correla- tion technologies (ICTs) has revolutionized the way that tion 0.65). Dingel and Neiman (2020)’s results also pre- work is organized. Many workers can now perform their dict a higher share of jobs which can be done from home tasks and connect with their colleagues from any place. The than ours, but the difference is small (43% vs 38%). In the implications of this new way to organize work have been second attempt to check the consistency of our measures, studied extensively. Among them, many studies try to esti- we compare our results with observed WFH incidence mate the share of workers that actually work from home. from previous surveys in Norway. There have been two They find that there are relative large differences across dif - surveys in Norway that include questions on whether ferent countries, sectors and firms. See for example, studies the respondents actually work from home: The Norwe - by Gschwind and Vargas (2019), Lister and Harnish (2019) gian labor force survey in 2017 and a recent survey by for an review. Almost all these studies, including the most the Norwegian Institute of Transport Economics (TØI). recent ones (Alipour et al. 2020; Barbieri et al. 2020; Mongey Although the actual WFH incidence is not the same as et al. 2020), are based on existing surveys. One exception is the potential capacity of WFH, their results and ours a study using US data by Dingel and Neiman (2020). While are are broadly similar. Finally, we use Norwegian job their main analysis is based on the Occupational Informa- advertisements data published by the Norwegian welfare tion Network (O*net) surveys with questions covering administration (NAV) between January 2012 and March “work context” and “generalized work activities, they have 2019. Some of these advertisements mentioned possibili- also tried to manually evaluate WFH feasibility for each ties of WFH to attract more candidates. We identify those occupation themselves. Existing surveys are often designed advertisements and construct the relative frequencies of for other purposes so that information obtained may not remote-friendly jobs across 9 major ISCO-08 occupa- directly answer our question of interest. What we are inter- tional groups. A comparisons between the observed fre- ested in this paper is whether a given job can be potentially quencies and those predicted using our results could be worked from home, which in principal depends on only a crude way for quality check. We find that the NAV job the nature of tasks that need to be performed. These sur - ads data provide supports to our WFH friendly measures. veys typically provide information on actual incidences of While none of these checks can directly prove the reli- home office, which in addition depend on many contextual ability of our measure, they do suggest that our measures Who and how many can work from home? Evidence from task descriptions Page 3 of 13 4 are consistent with several observable empirical pat- personal electronic devices”. However, there are rather terns and help relieve the concern on the quality of our few jobs that can be done remotely without requiring measure. the use of ICTs nowadays, which makes this distinc- tion less useful. In practice, the use of “Remote work” is often intended to stress the geographically detachment 2 Background between work and fixed places of work. U.S. Office of Even before the Covid-19 pandemic, there is already an Personnel Management (2013) goes even further and increasing number of workers who regularly perform considers only worker who “resides and works at a loca- their work away from the office. This practice is often tion beyond the local commuting area of the employ- considered as a potential solution to many social and ing organization’s worksite” as “Remote worker”. On the organizational problems (Bailey and Kurland 2002). other hand, “Working from home” emphasizes that it This new work mode represents a fundamental change is the location (worker’s own home) in which the work that has substantial impacts on workers, employers and is performed. It rules out any non-home-based forms society. Since the early contribution by Nilles (1975), a of “Remote work”. Traditionally, “Mobile work” is asso- large body of literature has been developed. However, ciated with work arrangements that require workers the discussions remain to be intense despite of increas- to spend most of their time out of the office (Siha and ing understanding of this new work mode. Big compa- Monroe 2006), such as door to door salesman. More nies such as Yahoo, Best Buy and HP ended their work recently, the terminology “Mobile work” or “Mobile from home programs in 2010s due to worries on possi- office” is used to highlight the fact that work could be ble negative impacts on performance. Hubert Joly, Best done not only at office or home, but various locations in Buy CEO at the time, stated that Best Buy’s working from between, such as cafes, hotels, airport lounges and etc. home program was “fundamentally flawed from a leader - In this paper, our goal is to evaluate the workers’ ship standpoint” (Allen et al. 2015). Many criticize that ability to continue to work while avoiding or minimiz- these companies make these programs the scapegoat for ing physical contacts with others. Among the terms their bad managements. This criticism might be unjusti - we discussed above, “Working from home” is prob- fied, since there is no consensus in the literature on how ably the best suited definition for this purpose. There work from home influence the worker’s productivity. The is a relative large literature that tries to measure how effect of WFH on productivity is theoretically ambigu - many actually work away from the office. However, ous, while empirical evidences are mixed. Bloom et al. the estimates often vary considerably. These estimates (2014) suggest that WFH increase the productivity of involve typically different definitions (as discussed workers based on a randomized controlled trial in a large above), have different reference populations (all work - Chinese travel agency. Monteiro et al. (2019) find the ers or only a particular group of workers) and are based opposite effects using a large panel of Portuguese firms. on different thresholds on the intensity of home work - Glenn Dutcher (2012) suggests that the effect of WFH ing (occasionally or regularly). More importantly, these may be heterogeneous: it may have positive implications estimates measure the actual observed incidences. Even on productivity of creative tasks but negative implica- if the tasks in principal can be performed from home, tions on productivity of dull tasks. A complete review whether a worker actually work from home depends of researches on this new work mode is beyond of the on many different factors. For example, employment scope. Interested readers are refer to those by Messenger regulations on working time and workplace flexibility (2019), Allen et al. (2015), Siha and Monroe (2006) and can play important roles (Gschwind and Vargas 2019). Bailey and Kurland (2002). Working culture, such as management’s trust of work- There are several closely related and often confused ers, is another important factor (Bailey and Kurland concepts in the literature which we need to distinguish: 2002). While this strand of literature is highly relevant, “Telework”, “Remote work”, “Mobile work” and “Work- it cannot directly answer our question of interest: what ing from home”. Originating from Nilles (1975), “Tel- jobs are technically possible to be performed at home. ework” refer to work arrangements where tasks are A note on nomenclature: For brevity, we sometimes performed away from the employer’s premises with refer to “WFH friendly” occupations rather than” occu- the help of ICTs (Messenger 2019). International Labor pations of which required tasks can be performed from Organization (2020) consider “Telework” as a sub- home”. We use the terms interchangeably. This does not category of the broader concept “Remote work”. They mean that such employees in actuality work from home claim that “What makes telework a unique category is either permanently or occasionally. that the work carried out remotely includes the use of 4 Page 4 of 13 H. Holgersen et al. 3 Method et al. 2014) also indicates a large quality improvement As mentioned above, the incidence estimates from previ- from by requiring MTurk workers to have had an high ous studies do not directly address our question of inter- acceptance rate from previous work. However, we real- est. A traditional survey that asks workers whether they ize that while making these changes will help to poten- consider their job tasks can be performed from home or tially improve the reliability of our answers. They did not not could provide the most reliable answer to our ques- resolve possible systematic biases that caused by the fact tion. However this can be quite costly, in terms of both that MTurk workers are not a representative sample of time and resources. In this paper, we propose a simple, the working population. cheap and timely method by noting that different jobs For our purpose, we create labelling jobs for each often have similar tasks and duties undertaken and can occupation. The labelling jobs are done in two consecu - be organized into a limited number of occupations. We tive rounds. In the first round, all 426 occupations are evaluate, not for every job but every occupation, the fea- presented to at least 5 MTurk workers for classification. sibility of WFH. While this greatly reduces the workload Only those occupations on which the MTurk workers needed, it ignores possible heterogeneity across jobs in largely disagree (less than 80% MTurk workers agree) the same occupation and may introduce bias, which we were included in the second round. In the following we discuss in more detail below. will label these occupations as the occupations that lack We group the jobs into occupations following the of consensus. In this round, we make three changes in ISCO-08 standard. The ISCO-08 contains 9 major occu - contrast to round one. First, we increase the number of pation groups and 426 occupations (Armed forces occu- responses to 15, so that totally we have at least 20 answers pations are excluded in this analysis) at the unit group for each of these occupations. Second, we refine and level. Detailed task descriptions for these occupations expand the descriptions of tasks in more detail. Finally, are listed in the ISCO-08 documentation. Using these only workers with high acceptance rates are allowed to descriptions, we try to evaluate whether an occupation is participate in the labelling job. We hope this could help likely to be performed from home. To do this, we make us to further improve the quality of the classification. use the online platform MTurk, which is “a marketplace for work that requires human intelligence” (Amazon 3.1 The first round 2020). Users of MTurk can generate different tasks that In the first round, each occupation was presented MTurk workers work on. It has gained increasing popu- together with a brief description. The exact question for - larity in social sciences. Researches have shown that mulation was “Can this type of job likely be performed MTurk can provide quick and reliable responses at rela- from a home office?”, and an example of a job description tively low costs (Buhrmester et al. 2011; Berinsky et al. could be: 2012). Electrical engineers conduct research and advise In order to increase the quality of our WFH friendly on, design, and direct the construction and opera- measure in a resource-effective manner, we adopt a sys - tion of electrical systems, components, motors and tem of evaluation loosely inspired by the Delphi method equipment, and advise on and direct their function- (Ziglio and Adler 1996), adopted for MTurk. The Del - ing, maintenance and repair, or study and advise on phi method was originally created for situations where technological aspects of electrical engineering mate- researchers were unsure of what questions to ask for rials, products and processes. a survey, and outlined a process by which a group of experts were consulted iteratively in order to reach a The respondent was asked to evaluate whether it was consensus and find possible areas of contention. The dif - likely that the job could be performed primarily from ferences to MTurk may seem stark—MTurk workers are a private home. The alternatives were “Yes”, “No” and not experts, nor can we consult the same MTurk work- “Unknown”, which were provided with the following ers repeatedly. Instead, we can view the occupations with description: a consensus as non-contentious issues, where the panel (MTurk workers) largely understood the question and agreed on the answer. The remainder are issues where either the question (occupation description) was mis- formed, there was real disagreement, or where there was We thank an anonymous referee for pointing out this potential source of excessive noise in the replies. We do not return not to the bias. same group of respondents but to a more reliable group Note that we haven’t include 9 “residual” occupations in the second round, since the descriptions of tasks are often not quite informative. For with more informative questions, and we can get more example, occupation 1219 is defined as “This unit group covers business MTurk workers to answer each question. Research (Peer services and administration managers not classified elsewhere…”. Who and how many can work from home? Evidence from task descriptions Page 5 of 13 4 Among these three factors, real differences in judge - ment may be informative and increase our measure accuracy by incorporating several experiences and views on remote feasibility. The two other reasons for lack of consensus can be combated by collecting more labels from”more responsible” MTurk workers and by providing a better description of a job. 3.2 The second round The occupations with a clear lack of consensus are anno - tated twice more, using two different MTurk panels. One panel of 10 workers who all are “masters” accord- ing to MTurk (a simple albeit vague checkbox indicating a high acceptance rate), and a panel of 5 workers who, in addition to being “masters”, have an task acceptance rate above 80% (meaning that at least 80% of their work has been accepted by other MTurk users). The thresh - Fig. 1 Agreement among the respondents from MTurk old of 80% was chosen mostly at random, but somewhat informed by comments on MTurk user boards. In order to provide more information to the workers, 1. Yes: This job can be performed primarily from an we use the the full description of the occupation, includ- office in a private home ing examples of task descriptions from the ISCO-08 2. No: Substantial parts of this job must be performed documentation. This information often more than dou - outside the employees home bles the volume of text a worker has to read to properly 3. Unknown: There is not enough information to decide. answer the task. Due to the increase in text to read, the monetary reward for workers was increased. We provided an “unknown” option in addition to the Neither of the MTurk annotations performed in the “yes/no” options in order to reduce arbitrary responses second round yielded any significant level of consensus to uninformative occupation descriptions. In order to among the workers. By carefully reviewing and annotat- reduce the serendipity in the labels, we acquired at least ing the selected occupation descriptions, some causes of five labels from different respondents independently for ambiguity stands out: each occupation. As expected, these respondents are not always agree with each other. Figure 1 shows how they • Some jobs may or may not be performed from home agree/disagree. For the majority of occupations (around depending on the home. Sewing, wood-working and 77% of occupations), there are at least 4 respondents similar types of craft can plausibly be done from agree on the same answer. Interestingly, the “unknown” home provided the home is spacious and properly label is seldom used: it is assigned 40 times, less than equipped. On a longer time-horizon this may be 2% of all labels assigned. Thus, the disagreement among plausible to many, while few may be able to bring this respondents is probably not an indication that some type of work home on short notice. occupations are difficult to evaluate for the respondents. • Some jobs can be done from home for a short while, We conclude that the occupations with a consensus of postponing in-person meetings until a later time. at least 4 are sufficiently certain that we can accept the • Some jobs can be done from home, but with a lower answer, while the remaining 23–24% are needs to be quality result. A perhaps poignant example are teach- redone. ers, who have shown an ability to teach via video There may be several reasons for this lack of consensus, even in lower elementary education, but few parents one welcome and two unwelcome: would agree that this is an acceptable long-term or even medium-term solution. • There are real differences in opinion as to whether a • Some jobs are very dependent on the technology in job can be done from home. use at an employer. Filing clerks may have to stay in • Turks are opportunistic, answering at random to the office if filing is mostly on paper, but as more and minimize time spent on each task. more documents are digital only, such jobs may be • The information provided was insufficient, leaving done from home. The same is true for certain types workers to fill in the blanks themselves. of systems administrators in Information Technol- 4 Page 6 of 13 H. Holgersen et al. ogy: If they are responsible for on-premise data cent- An interesting question is whether the changes made in ers they may have to be physically present, while if the second round, namely selecting only MTurk workers they are using cloud-computing there are no physical with high acceptance rate and presenting more detailed servers to access. task descriptions, is useful in terms of getting “better” • Some occupations are extremely specialized, and responses. One might consider to perform a two-sam- people without experience in those occupations can ple T-test on the hypothesis that the mean responses in not be expected to accurately assess WFH feasibility. round one and round two are the same. However, the T-test requires strong assumption on the underlying This implies that the disagreement in both rounds is distributions. We also cannot appeal to the asymptotic not mainly due to human labelling error, but rather rep- results given the small sample sizes we have. Instead, resents the possible heterogeneity in evaluations across we apply a permutation test which imposes no distribu- MTurk workers. We consider the arithmetic average as tional assumptions and is valid in small sample (Good a imperfect but good measure. Formally, for occupation j, 2005). For 76 of 90 occupations that were submitted to we have n different labels. Define MTurk in both rounds, we fail to reject the null hypothe- sis that the WFH measure from round one and that from 1 if answer "Yes" � � round two is the same at 5% level of significance. In other γ j = 0.5 if answer "Don t know" words, for the majority of “uncertain” occupations, the 0 if answer "No" changes we made in the second round did not have sig- nificant effects on the results. However, we should note and the WFH friendly measure γ( j) can then simply that increasing the sample size will nevertheless improve defined as γ( j) = ∑ γ ( j)/n. Note that to assign the value the estimates’ precision. 0.5 to the answer “Don’t know” is somewhat arbitrary, Some aspects of our approach are worth further discus- since “Don’t know” may not imply that half part of the sion. It is obvious that the workers performing the classi- jobs in this occupation can be done from home. An alter- fications are not labor market subject matter experts, and native is simply to drop these evaluations, which we have so the results are not authoritatively reflecting the inten - also tried and the main results do not change much. tion and original meaning of the creators of ISCO. In Using the average annotation may work well given the addition, respondents to our task on MTurk likely reside uncertainties outlined above: Annotations are likely to in different countries. However, we did not try to correct reflect the experience and expectations of the annota - for possible cultural/technological differences—some tor, and aggregating the knowledge of several annotators jobs that cannot be performed from home in one coun- can provide a more accurate picture of WFH feasibil- try may be possible be performed from home in other ity, reflecting the fact that occupations, employers and countries. These are clear limitations of our approach. So employees are not a homogeneous figure. For most of in a way, We should consider the evaluations as “interna- the occupations we revisited, we would be skeptical tional”, which is also true for the ISCO-08 standard itself. of any binary label. One may question our practice of However, as shown later in Sect. 5 where we run several using the average value as the remote-friendly measure. consistency checks, these limitations may not affect the The problem is most serious for the “lack of consensus” reliability of our measure very much. occupations where at most three respondents that agree with each other. We could, however, treat evaluations of 4 Results these occupations as missing and assign either all 1 or all Combining our occupational specific WFH measures 0 to those occupations. This way, we treat jobs of these and the labor statistics per occupation in Norway, we occupations as either all WFH friendly or all non-WFH find that in 2019 around 38 percent of the jobs in Norway friendly, thus establish the lower or upper bound of the can be performed from home. Applying the bounding prevalence estimate, respectively. Note that no restric- approach discussed in Sect. 3, we get the lower bound for tion/assumption is made for these occupations when constructing these bounds, this bounding practice is very similar to the so-called “worst case” bounds in the partial Statbank, Statistics Norway. Data can be obtained at https ://www.ssb.no/en/ identification literature, see for example Manski (2003). statb ank/table /12542 /. To be precise, this should be stated as “ the average of our occupational specific WFH friendly measures for all registered jobs in Norway in year 2019 is 0.38”. Given that our measure of WFH friendliness is an estimate of We consider most of the lack of consensus to stem from above mentioned the mean of a binary variable, we can interpret our measure as the probabil- factors, but noise is still a contributing factor: One may ponder how in all ity of a job of the given occupation can be worked from home. We should three rounds of MTurk annotation at least one worker submitted that a prison keep in mind that this simplification is obviously based on strong assump- guard could work from home. tions. Who and how many can work from home? Evidence from task descriptions Page 7 of 13 4 Table 1 Percentage of occupations are WFH-friendly across occupational group Percent WFH friendly Occupational group Estimate Lower bound Upper bound No. of jobs Managers 65.7 40.7 83.8 222,678 Professionals 57.4 40.2 56.4 652,356 Technicians and associate professionals 42.7 24.4 53.6 374,858 Clerical support workers 63.0 57.4 64.9 169,230 Service and sales workers 26.7 7.7 59.9 573,415 Skilled agricultural, forestry and fishery workers 17.0 16.7 18.2 21,631 Craft and related trades workers 12.0 2.0 27.9 219,843 Plant and machine operators and assemblers 7.0 6.6 7.2 163,197 Elementary occupations 1.7 1.7 1.7 134,400 All occupations 38.3 24.0 50.9 2,531,608 Table 2 Monthly wage: WFH and Non-WFH jobs WFH friendly measure No. of occupations No. of jobs Average earnings Median earnings High (> 0.8) 53 391,873 55,576 50,792 Medium (0.2—0.8) 202 1,396,139 44,987 43,440 Low (< 0.2) 136 665,653 43,948 42,800 this prevalence to be 24% and the upper bound to be 51%. are more likely to be full-time jobs, indicating that work- The range is somewhat wide, but still informative. ers with less WFH friendly jobs are also less likely to have Splitting into ISCO-08 major occupational groups, a stable job. we estimate what percentage of jobs in these groups Previous literature shows that workers with different are WFH friendly. The results are presented in Table 1. characteristics sort into different occupations. Some The share of jobs that can be performed remotely varies groups of workers face more challenges in labor market from 2 to 66 percent. “Managers”, ‘clerical support work- than others, such as lower-skilled young workers, lone ers” and “Professionals” are groups where many of the parents and workers with migrant background (Barrett employees can work from home. Only a small fraction 2010). Unfortunately, our results show that these work- of workers in occupations like “elementary occupation ers are also less likely to have WFH friendly jobs and workers” and “plant and machine operators assemblers” thus are impacted disproportionately by the shutdowns can work from home. Table 1 also reports the lower and and social distancing policies. We find a strong positive upper bound for the percentage of WFH friendly jobs in correlation between education levels and WFH pos- each occupation group. sibilities. Among those with at least a master’s degree, more than 55% have a WFH friendly job, while only 15% among those with only primary/elementary edu- 4.1 WFH f easibility: job and worker characteristics cation have such a job. Older workers are found to be We have also access to several administrative registers more likely to have a WFH friendly job. Workers over from SSB, which contains detailed information on work- 40 years old have a chance of 40% to be able to WFH, ers and their jobs. This enables us to find what types jobs while workers below 30 have only a 30% chance. These that can be worked from home. Jobs are characterized observations imply that low-skilled young people are by wages and working hours. In general, WFH friendly likely be impacted by the crisis particularly hard. Lone occupations also pay better, as shown in Table 2. The dif - parents often work less hours and earn less than oth- ference between WFH friendly and non-WFH friendly ers, partially due to the need of caring for dependent pay is much less pronounced when we split the data by children. They do not have similar ability to diversify occupational groups, and the pattern is not unequivocal income shocks and share caring responsibilities of chil- (Fig. 2). In general, there is a wage premium for WFH dren as couples. This makes them probably the most friendly jobs. However, the difference is minimal for the needed group of WFH. However, our results show that occupation “professional”. In addition, WFH friendly jobs 4 Page 8 of 13 H. Holgersen et al. Fig. 2 Earnings for WFH and non-WFH friendly jobs across different ISCO groups they are also least likely to have WFH friendly jobs To account for possible correlations between these among workers of different family statuses. characteristics, we estimate the average marginal effects We also find that workers with a migrant background for these variables on the probability of having a job that have much less chance (32%) to have WFH friendly jobs can be performed from home (represented by our WFH compared with native workers (40%). However, labor friendly measure). The patterns remain the same as dis - market qualifications and prospects vary widely among cussed above. To summing up, our results suggest that these workers. Table 3 shows the percent of work- those who are already disadvantaged in the labor market, ers with WFH friendly jobs by country background. such as workers with a migrant background, young work- Workers with migrant background from North Amer- ers and lone parents, are more likely to have non-WFH ica and Oceania top the table, while those from Africa friendly jobs. Thus, the pandemic and the government’s are found at the bottom of the rank. This finding is yet attempts to mitigate this crisis may have quite an uneven another example of the findings that immigrants from low-income countries face disadvantages in many con- text in the labor markets in developed countries (Careja Table 3 Percent of Workers with WFH friendly jobs: 2019). Country background There is also a clear difference across genders. Female World regions Percent WFH workers are more likely to have WFH friendly jobs than friendly (%) male workers, and thus might be less exposed to the Native workers 39.5 social distancing policy. However, there has been argued Workers with migrant backgrounds 32.2 that the possibility to work from home might actually not EU28/EEA 33.0 be beneficial for female workers as they often have to take Other European countries outside EU28/EEA 32.3 on additional housework in this situation (Collins et al. Africa 26.0 2020). Asia including Turkey 30.1 North America 43.6 By workers with a migrant background, we refer to both immigrants and South and Central America 32.6 Norwegian-born to immigrant parents. If we limit to only immigrants, the Oceania 42.2 chance is even lower, 26%. Who and how many can work from home? Evidence from task descriptions Page 9 of 13 4 cities have a higher share of WFH friendly jobs, which may be fortunate considering the greater need for social distancing in urban areas. The pattern looks clear, espe - cially in the area surrounding Oslo. Other major cities in Norway like Bergen, Trondheim, and Stavanger stand out on the map as well. By introducing a measure of urbanness, we can analyze the relationship more formally. We use population per square km as a proxy for urbanness. From Fig. 4, we see a clear correlation between “urbanness”, or population pr km , and the prevalence of remote-friendly jobs. Denser populated areas imply greater risks of COVID-19 spread, but this increased risk may be mitigated by better oppor- tunities for remote work. There is also significant variation in the distribution of WFH friendly jobs across different industries. Indus - tries such as Financial and Insurance (85%), Informa- tion and Communication (77%) have the highest share WFH friendly jobs, in contrast to primary and secondary industries which typically have much lower values rang- Fig. 3 Percentage of workers who can work from home, Norway ing from 20 to 30%. There are also few jobs that can be worked from home in accommodation and food service (14%), Transportation (25%) and Arts, entertainment impact on the working population. This hypothesis is and recreation (28%).Holgersen et al. (2020) studied the consistent with what actually happened during the first impact of Covid-19 crisis on the labor demand in Norway weeks of the crisis in Norway: these groups of work- using vacancy posting data and found that these indus- ers are more likely to be temporally laid off (Alstadsæter tries are among the industries that experienced largest et al. 2020). Policies aimed especially towards these par- drops in labor demand. ticular groups should have a high priority on the govern- ment’s list. Although the results are based on Norwegian 4.3 European results data, We believe that our findings can be informative The European statistical agency, Eurostat, publishes data for other countries as well, considering that workers of on employment by ISCO-08 major groups. Combining same occupations in different countries often share simi - the results presented in Table 1, we can use these data to lar characteristics. In fact, Mongey et al. (2020) studied estimate the prevalence of WFH friendly jobs in Europe. which workers are more likely to bear the burden of Note that we have used the Norwegian employment social distancing policy and found similar results as ours. sizes as the weights when aggregating from the ISCO-08 However, while the WFH friendliness is a strong positive unit groups to the major groups. Since the compositions predictor of the probability of job losses, institutional dif- of occupations can differ from country to country, the ferences across countries may play an important role on “true” weights can differ and lead to potential bias. its magnitude (Adams-Prassl et al. 2020). Figure 5 presents the geographic variation of pre- dicted share of jobs that can be worked from home across Europe. We observe a considerable difference. 4.2 V ariation of the prevalence of WFH friendly jobs The countries with the highest share of WFH friendly across different regions and industries jobs are Switzerland, Luxembourg, Norway and Sweden. The geographic location of jobs has been a point of inter - On the other end, the countries with the lowest share est for years, amid both pressure for workers to centralize of WFH friendly jobs are mostly less developed coun- and specialize, and fears of increased inequality between tries in southeast Europe, such as Turkey and Romania. cities and rural areas. Figure 3 shows the percentage of As Dingel and Neiman (2020) suggested, there seems to workers who can work from home in Norway. There is be a clear positive relationship between GDP per capita large heterogeneity across different regions. We estimate level and predicted share of WFH friendly jobs. Interest- that over 42 percent of the jobs in Oslo can be done from ing, the pattern on the prevalence of WFH friendly jobs home. On the other end of the spectrum, in some small we find above is very similar to the actual observed inci - rural municipalities in northern Norway just over a quar- dence of telework in Europe reported by Gschwind and ter of the jobs can be done from home. As we expected, 4 Page 10 of 13 H. Holgersen et al. Fig. 4 Shares of Remote feasible jobs and population density et al. (2020), and Hensvik et al. (2020) for the United States, Alipour et al. (2020) for Germany, and Barbieri et al. (2020) for Italy. Unlike our study, they rely on dif- ferent surveys, and the results are established on their national occupation classifications. Although their results are based on the OES/SOL occupation groups, Dingel and Neiman (2020) manage to use the crosswalk between the OES groups and the ISCO-08 groups. This crosswalk provides an opportu- nity to compare our results with theirs. As a robust- ness check, we have redone the above analyses using the US classification results. Given the many to many nature of the crosswalk, we do not expect that their results and ours agree with each other on the lower levels of occupation groups, but they should be simi- lar on a more aggregated level, such as the ISCO major group. Using the US classifications, the overall share Fig. 5 Percentage of workers who can work from home, Europe of remote-friendly jobs in Norway is estimated to be around 43%, slightly higher than our estimate 38%. Figure 6 presents the scatter plot of shares of jobs pre- Vargas (2019). They find also “a rough North/South and dicted using their measures against those using ours East/West divide in the incidence of telework”. for 9 major ISCO occupation groups together with the 45-degree line. The bubble size represents the employ- ment numbers in Norway. There is a strong positive 5 Validation and consistency check correlation between these results. The key patterns 5.1 C onsistency check: using alternative classification we found in Sect. 4 on earnings, worker’s characteris- results tics, geographic and industry variations remain to be There are several very recent analyses that study the the same when using the results by Dingel and Neiman prevalence of WFH jobs for different countries: Dingel (2020). and Neiman (2020), Brynjolfsson et al. (2020), Mongey Who and how many can work from home? Evidence from task descriptions Page 11 of 13 4 distribution of WFH jobs across different worker, occu - pation and industrial groups. Consistent with our results, the survey also finds that the share of workers who can work from home increases with worker’s education lev- els. The results on occupation and industry groups are broadly similar to what we obtained but with some dif- ferences. For example, around 29 percent of workers in “Clerical support workers” who participated the survey responded that they had the opportunity to work from home at times, much lower than the results we obtained. This is likely attributable to the distinction mentioned above: Not having the opportunity to work from home does not necessarily mean that the job can not be per- formed from home. Fig. 6 The US classification and ours, ISCO major groups 5.3 Validating results against job‑ads Another way to evaluate the results from Mechanical 5.2 C onsistency check: comparing with actual observed Turk is to use advertisements from the Norwegian wel- incidence of WFH fare administration (NAV). These job advertisements There was almost no surveys that particularly focus have been published as open data by NAV, and con- on the topic of home working/ “telework” in Norway. tain the text, title, employer information, and annota- The only exception is a recent survey by the Norwegian tions made by subject matter experts at NAV including Institute of Transport Economics (TØI) that is designed the occupational code of the job. Because the possibility especially for the COVID-19 situation. The main focus to work from home is a perk for many, some employ- is the effectiveness of working from home (Nordbakke ers mention it in their job ads to attract candidates. We 2020). In the survey, 950 respondents are asked about search the texts for mentions of “hjemmekontor” and their working situation for a given day, 19th March, 2020. “heimekontor”, two distinctive words unlikely to mean Among them, 32 already work from home already before anything other than the possibility of working from the Covid-19 crisis. 67 have the possibility to work from home. We find that there are quite few, only around 2.5 home, but still go to work on that day. And 401 have the among every 1000, announcements that actually include possibility and actually work from home on that day. these words. Obviously, far from all announcements of Summing up, we obtain an overall estimate of the preva- jobs that can be performed remotely include these words. lence of WFH friendly jobs (52%). This estimate is con - We cannot derive the total number of WFH friendly jobs siderably higher than our estimate. However, the survey from these job ads data alone. However, it says something is based on the job situation only for a given day and can important about the relative frequency of WFH-friendly potentially lead to a upward bias. It is unfortunate that jobs across the occupational groups. the survey contains very limited information: no char- Table 4 presents the actual observed relative frequen- acteristics of jobs or workers are collected, which makes cies from the job announcement data and those predicted further comparison of our results impossible. using our WFH friendly measure. Large discrepancies are Another survey that contains questions concerning found for three major groups “Professional”, “Technicians possibilities of working from home is the Norwegian and associate professionals” and “Clerical support work- labor force survey in 2017. The question was whether ers”. Our measure from the MTurk predicts more cases the respondent had the opportunity to work from home of the mentions of “home office” than what are actually when she/he wants to. This is not to say that the job could observed in the NAV data for the first group, and less be performed remotely in its entirety, and neither to say cases for the last two groups. It could be that our MTurk that those who weren’t given the opportunity all have results are biased. However, we think it is more likely jobs that cannot be performed from a home office. The that employers in these occupations have different per - results are discussed by Nergaard et al. (2018). According ceptions on the importance of the “working from home” to the labor force survey, around 35% have the possibility feature to attract potential candidates. Note that in the to work from home, which is very close to what we find job announcements, exact wage is seldom listed. From in the current analysis. However, as discussed earlier, the Fig. 2, we see that for the last two groups “Technicians magnitudes of these two estimates may not be directly and associate professionals” and “Clerical support work- comparable. What is really interesting to check is the ers”, wages of WFH friendly jobs are on average much 4 Page 12 of 13 H. Holgersen et al. Table 4 Relative frequency of WFH possibilities across ISCO groups Relative WFH frequency Occupational group MTurk (%) Job ads (%) Difference (%) Managers 3. 9 3. 1 0. 8 Professionals 67. 6 48. 0 19. 6 Technicians and associate professionals 13. 3 27. 2 − 13.9 Clerical support workers 3. 5 9. 9 − 6.4 Service and sales workers 5. 8 9. 1 3. 3 Skilled agricultural. forestry and fishery workers 0. 0 0. 0 0. 0 Craft and related trades workers 5. 2 1. 2 4. 0 Plant and machine operators and assemblers 0. 5 1. 4 1. 0 Elementary Occupations 0. 2 0.1 0. 2 higher than wages of non-WFH jobs. To some extent, several validation tests. We compare our results with the being able to work from home can be seen as proxy of classification results of Dingel and Neiman (2020), with high wages in this two occupations. So the employers the WFH incidence derived from two surveys in Nor- may have stronger incentives to include these words to way (Nordbakke 2020; Nergaard et al. 2018) and finally attract potential applicants. Overall, we think the correla- with the results generated from the job advertisements in tion we see in Table 4 is decent considering the spurious- Norway. We do not find evidence of large bias in our esti - ness of the data. mates. Although none of these checks can be considered as formal tests of reliability, the positive results enhanced our belief that our approach is a suitable alternative to 6 Conclusion the mainstream methods. In this paper, we propose a new method to evaluate This analysis is an attempt to combine conventional the prevalence of WFH friendly jobs in an economy. In (the administrative register and official statistics) and particular, we ask respondents from MTurk to evaluate unconventional (data from a web-based crowd-sourc- whether the main tasks of occupations can be performed ing platform) sources for statistical and research pur- from home and establish a measure of the feasibility of poses. The results we have found suggest that alternative WFH for each occupation. Compared with transitional approaches to collecting such information are feasible. approaches via experts or surveys, our approach is easier to implement, costs less and takes shorter time. Abbreviations The fact that the WFH feasibility is evaluated on the ISCO: International Standard Classification of Occupations; MTurk: Amazon occupation level but not on the job level may lead to Mechanical Turk; NAV: The Norwegian Welfare Administration; SSB: Statistics potential bias since it essentially ignored the heteroge- Norway; TØI: The Norwegian Institute of Transport Economics; WFH: Working From Home. neity across jobs within the same occupation. A related issue is that it might be difficult to assign a binary label to Acknowledgements some occupations. So far we have treated the lack of con- We thank the editor and two anonymous referees for comments and sugges- tions that greatly improved the paper. We also thank Terje Skjerpen, Simon sensus among the respondents as an indication of certain Bensnes, Oddbjørn Raaum, and seminar participants at the Research Depart- occupations being “problematic” and try to cope with this ment, Statistics Norway for inspiring comments and discussions. problem by aggregating the answers. However, in future Authors’ contributions practice, we may consider allowing for more detailed HH, ZJ and SS designed the study. HH collected the data via Amazon Mechan- labels or asking the respondents to provide own estimate ical Turk. SS prepared the restricted data owned by Statistics Norway. HH, ZJ of likelihood on a given scale directly. Another concern is and SS analyzed and interpreted the data. HH and ZJ wrote the manuscript. All authors read and approved the final manuscript. that the respondents from MTurk are not labor market subject matter experts and reside in different countries. Funding These could also lead to bias since the respondents are This paper is a part of the project “FL3017004: Analysis of the effects of COVID- 19 pandemic” funded by Statistics Norway. subject to possible cultural and technological differences and might misinterpret the task descriptions. Availability of data and materials These concerns highlight the need to check the reli - This paper uses both publicly available data and restricted data. The public available data and code for replication can be found at https ://githu b.com/ ability of our method. To do this, we have performed radbr t/worki ng from home. The restricted data that contains information on Who and how many can work from home? Evidence from task descriptions Page 13 of 13 4 individual workers are available from Statistics Norway but restrictions apply Good, P.: Permutation, parametric, and bootstrap tests of hypotheses: Springer to the availability of these data, which were used under license for the current series in statistics. Springer, New York (2005) study, and so are not publicly available. Data are however available from the Gschwind, L., & Vargas, O. (2019). Telework and its effects in Europe. In Telework authors upon reasonable request and with permission of Statistics Norway. in the 21st century. Cheltenham, UK: Edward Elgar Publishing. https ://www.elgar onlin e.com/view/edcol l/97817 89903 744/97817 89903 Ethics approval and consent to participate744.00007 .xml Not applicable. Hatchett, R.J., Mecher, C.E., Lipsitch, M.: Public health interventions and epi- demic intensity during the 1918 influenza pandemic. Proc. Natl. Acad. Sci. Consent for publication 104(18), 7582–7587 (2007) Not applicable. Hensvik, L., Barbanchon, T. L., & Rathelot, R. (2020). Which jobs are done from home? Evidence from the american time use survey. ( The University of Competing interests Warwick, Department of Economics, working paper 466) Not applicable. Holgersen, H., Jia, Z., & Svenkerud, S. (2020, May). Labor demand during the covid-19 crisis in norway: Evidence from vacancy posting data ( Working Received: 26 June 2020 Accepted: 27 January 2021 Paper No. 3663439). SSRN. Retrieved from https ://doi.org/10.2139/ ssrn.36634 79 International Labor Organization: International standard classification of occupations:Isco-08. International Labor Organization, Geneva (2012) International Labor Organization. (2020). Covid-19: Guidance for labour statistics data collection.International Labor Organization: Geneva. References Lister, K., & Harnish, T. (2019). Telework and its effects in the united states. In Adams-Prassl, A., Boneva, T., Golin, M., Rauh, C.: Inequality in the impact of the Telework in the 21st century. Cheltenham, UK: Edward Elgar Publishing. coronavirus shock: evidence from real time surveys. J. Public Econ. 189, Retrieved from https ://www.elgar onlin e.com/view/edcol l/97817 89903 104245 (2020). https ://doi.org/10.1016/j.jpube co.2020.10424 5 744/97817 89903 744.00009 .xml Alipour, J.-V., Falck, O., & Schu¨ller, S. (2020). Germany’s capacities to work from Manski, C. (2003). Partial identification of probability distributions: Springer series home. (IZA DP No. 13152) in statistics.Springer. Allen, T.D., Golden, T.D., Shockley, K.M.: How effective is Messenger, J. C. (2019). Introduction: Telework in the 21st century – an telecommuting?Assessing the status of our scientific findings. Psychol. evolutionary perspective. In Telework in the 21st century. Cheltenham, UK: Sci. Public Interest 16(2), 40–68 (2015). https ://doi.org/10.1177/15291 Edward Elgar Publish- ing. Retrieved from https ://www.elgar onlin e.com/ 00615 59327 3 view/edcol l/97817 89903 744/ 9781789903744.00005.xml Alstadsæter, A., Bratsberg, B., Eielsen, G., Kopczuk, W., Markussen, S., Raaum, O., Mongey, S., Pilossoph, L., & Weinberg, A. (2020, May). Which workers bear the & Røed, K. (2020, May). The first weeks of the coronavirus crisis: Who got hit, burden of social distancing policies? ( Working Paper No. 27085). National when and why? evidence from norway ( Working Paper No. 27131). National Bureau of Economic Research. http://www.nber.org/paper s/w2708 5 doi: Bureau of Economic Research. Retrieved from http://www.nber.org/ https ://doi.org/10.3386/w2708 5 paper s/w2713 1. https ://doi.org/10.3386/w2713 1 Monteiro, N. P., Straume, O. R., & Valente, M. (2019). Does Remote Work Improve Amazon. (2020). Amazon mechanical turk documentation. Retrieved from https or Impair Labour Productivity? Longitudinal Evidence from Portugal (CESifo ://docs.aws.amazo n.com/mturk /index .html Working Paper Series No. 7991). CESifo. https ://ideas .repec .org/p/ces/ Bailey, D.E., Kurland, N.B.: A review of telework research: findings, new direc- ceswp s/7991.html tions, and lessons for the study of modern work. J. Org. Behav. 23(4), Nergaard, K., Andersen, R. K., Alsos, K., & Oddervoll, J. (2018). Flexible working 383–400 (2002). https ://doi.org/10.1002/job.144 schedule: an analysis of arrangements in norwegian labor market. (In Nor- Barbieri, T., Basso, G., & Scicchitano, S. (2020). Italian workers at risk during the wegian, “Fleksibel arbeidstid. en analyseav ordninger i norsk arbeidsliv”, covid-19 epidemic ( Working Paper No. 3572065). SSRN. Retrieved from Fafo Institute for Labour and Social Research, working paper 2018:15) https ://ssrn.com/abstr act=35720 65 Nilles, J.: Telecommunications and organizational decentralization. IEEE Trans. Barrett, R.: Disadvantaged groups in the labour market. Econ. Labour Market Commun. 23(10), 1142–1147 (1975) Rev. 4(6), 18–24 (2010) Nordbakke, S. (2020). Reduced effectiveness with home office. Retrieved from Berinsky, A. J., Huber, G. A., & Lenz, G. S. (2012). Evaluating online labor markets https ://kanta r.no/kanta r-tns-innsi kt/redus ert-effek tivit et-med-hjemm for experimental research: Amazon.com’s mechanical turk. Political Analy- ekont or/ (in Norwegian, “Redusert effektivitet med hjemmekontor” sis, 20(3), 351–368. http://www.jstor .org/stabl e/23260 322 Kantar AS webarticle.) Bloom, N., Liang, J., Roberts, J., Ying, Z.J.: Does working from home work? Evi- Peer, E., Vosgerau, J., Acquisti, A.: Reputation as a sufficient condition for dence from a chinese experiment. Quart. J. Econ. 130(1), 165–218 (2014). data quality on amazon mechanical turk. Behav. Res. Methods 46(4), https ://doi.org/10.1093/qje/qju03 2 1023–1031 (2014) Brynjolfsson, E., Horton, J., Ozimek, A., Rock, D., Sharma, G., & Ye, H. Y. T. (2020). Siha, S., Monroe, R.: Telecommuting’s past and future: a literature review and Covid-19 and remote work: An early look at us data. https ://john-josep research agenda. Business Process Manag J 12(4), 455–482 (2006). https h-horto n.com/paper s/remot e work.pdf (MIT, manuscript) ://doi.org/10.1108/14637 15061 06780 78 Buhrmester, M., Kwang, T., & Gosling, S. D. (2011). Amazon’s mechanical turk: A U.S. Office of Personnel Management. (2013). 2013 status of telework in the new source of inexpensive, yet high-quality, data? Perspect. Psychol. Sci. federal government (Report to the Congress). U.S. Office of Personnel 6(1), 3–5. http://www.jstor .org/stabl e/41613 414 Management. https ://www.telew ork.gov/repor ts-studi es/repor ts-to- Careja, R. (2019). Immigrants’ labor market outcomes: Contributions from congr ess/2013r eport tocon gress .pdf multilevel studies. Ko¨ln Z Soziol, 71, 187-220. Ziglio, E., Adler, M.: Gazing into the oracle : the delphi method and its applica- Collins, C., Landivar, L.C., Ruppanner, L., Scarborough, W.J.: Covid-19 and the tion to social policy and public health. Jessica Kingsley Publishers, New gender gap in work hours. Gender Work Organ. (2020). https ://doi. York (1996) org/10.1111/gwao.12506 Dingel, J.I., Neiman, B.: How many jobs can be done at home? J. Public Econ. 189, 104235 (2020). https ://doi.org/10.1016/j.jpube co.2020.10423 5 Publisher’s Note Dutcher, E.: The effects of telecommuting on productivity: An experimental Springer Nature remains neutral with regard to jurisdictional claims in pub- examination. The role of dull and creative tasks. J. Econ. Behav. Organ. lished maps and institutional affiliations. 84(1), 355–363 (2012). https ://doi.org/10.1016/j.jebo.2012.04.009
Journal for Labour Market Research – Springer Journals
Published: Feb 28, 2021
Access the full text.
Sign up today, get DeepDyve free for 14 days.