COVID-19-Induced Automation: An Exploratory Study of Critical Occupations

Abstract

The COVID-19 pandemic may have spurred automation, especially in critical occupations. This article explores the potential of each detailed Standard Occupational Classification System (SOC) occupation being automated due to COVID-19. The authors explore two key elements of each occupation: its exposure to diseases such as COVID-19 and the probability of that occupation being automated. The results reveal that food preparation, service, and cleaning-related occupations have a higher chance of pandemic-induced automation. Using monthly U.S. job postings from 2016 to 2021, the estimates show that the potential pandemic-induced automation is associated with a statistically significant decrease in job postings. A higher Automation Index is associated with fewer job postings since the pandemic. Such trends remain robust after accounting for posting duration and excluding health-related occupations. These findings contribute to the early assessment of the impact of COVID-19 on the potential integration of automation in the labor force and offer insights into building a resilient and labor-centric post-pandemic labor market.

The ongoing technological advancement of automation procedures and their integration has been accelerated due to the COVID-19 pandemic shocks and could influence occupations, especially those that require individuals to perform job-related tasks in close proximity. In this context, automation refers to computer-assisted machines, robotics, and artificial Intelligence (AI; Acemoglu & Restrepo, 2018). AI and robotic technologies are garnering increased attention from industry and business leaders who are witnessing a declining pool of workers who are available, or willing, to work during the extended period of uncertainty and risks associated with the COVID-19 pandemic. Especially hard hit are workers employed in critical occupations that entail higher risks of exposure due to their job-related functions and proximity to co-workers and customers. For example, Walmart utilized more cleaning and stocking robots (CNN, 2020) and many robots were used to recycle trash (Corkery & Gelles, 2020). Fast-food restaurants expanded their installation of kiosks (Zietz, 2020) while hospitals are deployed robots to measure patients’ body temperature and distribute hand sanitizers (The Guardian, 2020). In addition, more chatbots performed customer service work (Howard & Borenstein, 2020) while semiautonomous robots delivered food (OECD, 2020). For example, Purdue University, the largest university campus to offer automated delivery of mobile robots (Purdue Exponent, 2019), has experienced a surge in demand for such automated food deliveries since the COVID-19 pandemic. Although the adoption of these technologies was taking place prior to the pandemic, what is worthy of careful examination is whether COVID-19 accelerated the growth of contactless working strategies through the adoption of automation.

While the coronavirus is likely to taper off at some point and the coping strategies adopted as a result of the pandemic may decline, some suggest that the pandemic is likely to alter consumer and labor force preferences over the long run, potentially redefining work and expanding the demand for robot workers in the near future (Thomas, 2020). At the beginning of the pandemic, a Brookings Institution analysis (Muro & Maxim, 2020), for example, predicted that the coronavirus may lead to a recession that could create a spike in labor-replacing technologies. More recent observations show that the economy has largely recovered from the short-term pandemic-induced recession yet the disruption to the labor market remains due to the supply shock and unwillingness of some laborers to rejoin the workforce. Note that unemployment rates in different communities have reached pre-pandemic levels, but not the labor participation rates.

Without question, there was an underlying movement of increased automation, robotics, and AI in various industries prior to the pandemic. Previous studies have agreed that while new opportunities may emerge because of automation, the downside is that individuals employed in certain occupations are likely to be displaced. Automation may affect the structure of an occupation's tasks, income inequality, and job polarization. For example, a National Bureau of Economic Research (NBER) study found that several routine jobs vanished during the Great Recession of 2008 to 2009 because of automation (Jaimovich & Siu, 2012). A study by the International Labour Organization (ILO) postulated that the eventual impacts of technological changes on labor markets will depend on the interactions and interrelationships between “economic, institutional, political, and business-related factors and design and application of these technologies” (Ernst et al., 2019). The ILO study found that neither costs, such as unprecedented job losses due to the replacement of laborers by technology, nor benefits, such as a reduction in inequality and poverty, have yet to materialize from these technological advances. Given the situation of COVID-19, it is an opportune time to consider AI's potential influence on different occupations and evaluate which occupational groups are most likely to be impacted. Lessons can be learned that will help establish new norms and offer a long-term vision that prepares the labor market for possible shocks while realizing the potential of AI.

Our article examines the following key question: What occupations are more likely to be automated due to COVID-19? We hypothesize that the occupations at greater risk of exposure to COVID-19, and that also place higher on the Automation Index (an index that measures the chances of an occupation being automated based on task content derived from the Occupational Information Network [O*NET]), are the most likely to be automated. We explore our research question by developing a proximity and contagious disease risk-based assessment of occupations. We examine the propensity of occupations being automated along two strands: (1) the disease risk associated with a given occupation and (2) the capacity for the job to be automated. We create a quadrant plot to visualize and distinguish occupations’ tendency for pandemic-induced automation. Finally, we estimate the change in occupation demand using extensive monthly job posting panel data for the United States.

We find that food preparation, service, and cleaning-related occupations have a high chance of pandemic-induced automation, while medical and health care professions are far less likely to be automated, even though workers face a high risk of exposure to disease. The occupations that face a high risk of pandemic-induced automation also have many workers with relatively low wages. In other words, pandemic-induced automation might impact low-skilled and low-wage workers disproportionately. Furthermore, fixed-effect regression shows significant (albeit weak) evidence that a higher pandemic automation risk is associated with fewer job postings during the pandemic, suggesting evidence of the labor market's response.

The primary contributions of this article are the development of pandemic-induced automation indexed at a detailed Standard Occupational Classification System (SOC) level, the visualization of the results for each occupation, and an empirical examination of whether pandemic-induced automation, approximated by the interaction of automation risk and exposure risk, is associated with changes in the number of job postings since the pandemic. It allows one to observe significant heterogeneity among occupations. Our work is similar to Chernoff and Warman (2020) who developed the Automation Index to determine occupations that have both automation potential and a high risk of viral infection, and then linked these occupations with American Community Survey (ACS) data. But, unlike our study, the researchers did not provide a detailed analysis of each occupation nor report on changes in job postings. Our article also contributes to a small emerging literature on COVID-19's early impacts on the labor market by giving specific attention to the influence of automation. Given the explosion of Omicron cases and other new variants, we know that the risk of the pandemic remains. We expect that our findings will hold and that the empirical association between pandemic-induced automation and labor market responses might be magnified.

Literature Review

This article is related to several different, but closely related, strands of literature. The first are studies that explore the impacts of automation on the labor market. Previous investigations have examined this issue in depth, and the general findings were that automation (e.g., robotics, AI, and computerization) can boost productivity while not leading to net job losses. Autor (2015) pointed out that although some forms of automation might be a substitution for human labor, many are complementary, suggesting that the pandemic-induced automation may increase productivity and wages in some occupations. Counterviews argued that many tasks still favor human skills over automation (Acemoglu & Autor, 2011). But at the same time, studies have shown that automation may cause polarization of wages and replace workers, especially those in more routine low- and middle-income jobs (Acemoglu & Restrepo, 2018; Autor & Salomons, 2018; Frey & Osborne, 2017; Kumar et al., 2020). In terms of the scale of potential impact, Arntz et al. (2016) estimated the job propensity to be automated for 21 OECD countries using a task-based approach that accounts for the heterogeneity of workers’ responsibilities within occupations. The results show that, on average, across the 21 OECD countries, 9% of all jobs could be automated. Yet, these studies did not address the impact of sudden, external shocks like COVID-19, and their short- and long-term impacts on labor market demand.

The second line of literature demonstrated the heterogeneous impact of a demand shock on the workforce. These studies segregated workers by skill, industry–occupation pairs, spatial locations, and/or skill–location pairs. Autor and Reynolds (2020) discussed the possible long-term impact of the pandemic on low-skill workers by reducing demand for building cleaners, security, and maintenance services; hotel workers and restaurant staff; urban transportation service; and those who participate in food preparation, transportation, clothing, entertainment, and other occupations that are relevant to people who spend a considerable amount of their time outside the home for business reasons. A Brookings report showed that different occupations, industries, and locations affect labor mobility, suggesting that certain workers are more vulnerable to changes in labor market demand (Berube, 2019). Shah and Shearer (2018) also noted that labor mobility dynamics vary greatly across local labor markets. Weise et al. (2019) studied the skills of local workers in various locations and derived practical strategies that help better align changes in workforce demand with talent supply needs. A recent strand of work in professional labor market databases has proposed that “geography matters” in exploring occupations and skills. The past efforts have integrated industries and occupations, noting how occupations are clustered within the industries and their geographical variations (Kumar et al., 2020; Nolan et al., 2011). Exploring occupations and talents as a bundle of skills and exploring the distribution of, and variation in, skills over geographic space, represents relatively new areas of research (Weise et al., 2019).

Finally, there are an increasing number of studies focused on the impact of COVID-19 on the labor market. Recent research has identified occupations with the highest risk of exposure (Zipper, 2020) and its regional variation (Chernoff & Warman, 2020). There are a handful of other studies that have conducted initial assessments of the labor market's response to COVID-19, such as unemployment (Cajner, 2020), vacant office buildings (Bernstein, et al., 2020), labor force participation (Coibion et al., 2020), and job postings (Kahn et al., 2020). Yet, none provided a detailed analysis of individual occupations or the change in job postings. The pandemic experience is very likely to accelerate the automation and digitalization wave across different sectors of the economy (Giordani & Rullani, 2020).

Building on existing literature, this article uses job posting data to explore the potential impacts of the pandemic on the labor market through the accelerated adoption of job-related automation processes.

Method for Classifying High-Risk Occupations

Data

Our main data source is Lightcast (formerly known as Emsi Burning Glass: Economic Modeling Specialists International and Burning Glass). Lightcast gathers and integrates economic, labor market, demographic, education, profile, and job posting data from dozens of government and private sector sources, creating a comprehensive and current data set that includes both published data and detailed estimates with full coverage of the United States (Lightcast, 2020c).

Our pandemic-induced automation measure begins with the SOC six-digit occupation codes. The SOC system is used to sort workers into one of 775 occupational categories. Our list of SOC occupations is obtained from Lightcast. Lightcast follows the Bureau of Labor Statistics’ (BLS) SOC structure with slight modifications (Lightcast, 2020b). We evaluate the position of each occupation along two dimensions: (1) The Exposure Index, which measures how susceptible an occupation is to close contact with other humans and proximity to infection and diseases, and hence, the possible exposure to contagious and virulent diseases such as the coronavirus-based SARS, MERS, and COVID-19; and (2) the Automation Index, which measures the likelihood of each occupation being computerized, irrespective of its pandemic-related risk.

The Exposure Index

The Exposure Index measures the risk of a worker in each occupation being exposed to contagious diseases. The data being utilized to help create the index were obtained from O*NET. The O*NET database includes ratings of work characteristics for individual occupations. We use the O*NET ratings of two important work characteristics that could make an occupation vulnerable to the negative impacts of the pandemic: (1) the risk of workers being exposed to diseases or infections due to the job's tasks; and (2) the extent to which a job requires workers to be in close physical proximity to other workers.

To calculate the Exposure Index, we first match each O*NET occupation to the relevant SOC occupation. We use O*NET data because they provide more detailed occupational information than SOC. There are usually multiple O*NET occupations associated with each SOC. For example, SOC 29-1071 is Physician Assistants, which is further classified in O*NET as 29-1071.00, Physician Assistants, and 29-1071.01, Anesthesiologist Assistants. BLS gathers employment and wage data at the SOC level and O*Net incorporates occupational characteristics, such as knowledge, skills, abilities, and work conditions. For purposes of our study, we apply the average ratings of O*NET for each SOC. In total, we have information on 775 SOC five-digit occupations. Next, we calculate the average of two scores: (1) the risk of exposure to diseases and infections and (2) physical proximity to other people. The Exposure Index is the resulting average. Our employment of the two scores is consistent with a recent study that sought to assess the health impact of the pandemic on different occupations (Zipper, 2020).

Risk of Exposure to Diseases and Infections

This scale has a range of 0 to 100, where 0 means workers have no exposure to disease and infections 25 indicates the potential for workers to be exposed once a year or more but not every month; 50 signifies that exposure could occur once a month or more but not every week; 75 implies that exposure could be once a week or more but not every day; and 100 indicates that the risk of exposure could be a daily concern. Acute care nurses, dental hygienists, family and general practitioners, and general internists score 100 points on the exposure to disease or infection scale.

Physical Proximity to Other People

This metric provides a score ranging from 0 to 100 for each SOC code. A higher score indicates that the job requires greater contact with others. A score of 0 means that workers operate at least 100 feet apart from others. A score of 25 indicates workers are situated in private offices, 50 means that workers’ activities are conducted in a shared office, a 75 implies that coworkers operate at arm's length from each other, and 100 means workers are touching, or nearly touching other humans. Choreographers, dental hygienists, physical therapists, and sports medicine physicians score 100 on this scale. Slaughterers and meat packers, an occupational group that experienced high rates of COVID-19 infection (especially in rural communities), tend to score around 73 on this scale.

The Exposure Index is the average of the two ratings described above. Table 1 illustrates one example of how the Exposure Index is calculated for each SOC occupation.

Table 1.

Example of how the Exposure Index is Calculated.

SOC	O*NET	Disease exposure	Average	Physical proximity	Average	Exposure Index
29-1071 Physician Assistant
	29-1071.00 Physician Assistants	94	93	88	88.5	$\frac{(93+88.5)}{2}=90.8$
	29-1071.01 Anesthesiologist Assistants	92		89

Note. O*NET = Occupational Information Network; SOC = Standard Occupational Classification System.

Source: Authors’ elaboration.

The Automation Index

The Automation Index associated with each occupation is obtained from Lightcast.

The construction of the Automation Index starts with task time shares derived from O*NET work activities. The tasks are regressed for each occupation on Frey and Osborne's published “computerization probabilities” (2017), which helps to identify which tasks are positively and negatively correlated with automation risk. This classification is linked with the task time shares to identify the share of each occupation's time spent in high- and low-risk work, from an automation perspective. Then the method uses occupation compatibility scores, by looking at all similar roles (defined as having an O*NET compatibility score over 75) and finds the percentage of jobs in similar roles that are at risk of automation. Finally, using staffing pattern data, the share of an occupation's jobs in three-Industry Classification System (NAICS) industries is multiplied by that industry's share of at-risk jobs to calculate the overall industry's automation risk. It is then standardized and scaled so that 100 represents the “average worker” (Lightcast, 2020a).

The Automation Index ranges from 72 (corresponding to SOC 27-1014 Multimedia Artists and Animators) to 139.1 (corresponding to SOC 47-2042 Floor Layers, except Carpet, Wood, and Hard Tiles,) with a mean of 101 and a SD of 15.8. Figure 1 shows the box plot for automation score by SOC two-digit occupations. The black dots represent outliers–occupations that have an Automation Index score outside 1.5 times the interquartile range above the upper quartile and below the lower quartile for each SOC two-digit cluster. From the distribution of the Automation Index, we notice that there is substantial heterogeneity both within and between occupation clusters, with the two highest Automation Index scores being for SOC 35 Food Preparation and Serving Related, and SOC 47 Construction and Extraction. The lowest Automation Index scores are for SOC 11 Management, and SOC 21 Community and Social Services.

Figure 1.

Automation Index by Standard Occupation Classification System (SOC) two-digit occupations.

Source: Lightcast.

Visualization of the Classification

Based on these two indices, we classify an occupation as high risk if it scores above average on both indices. Then, we locate the position of each occupation along the two dimensions—Exposure Index and Automation Index—using a quadrant plot.

In the quadrant plot, each bubble represents an occupation and its position along the two dimensions. The horizontal axis shows the Automation Index while the vertical axis represents the Exposure Index.

The dashed gray reference lines display the average along each dimension. The average score is 41.4 for the Exposure Index and 100 for the Automation Index. Note that the U.S. average Automation Index (developed by Lightcast) score is 100, which means an occupation scoring higher than 100 has a greater automation propensity than one scoring less than 100. These two reference lines separate the plot into four quadrants. The top right quadrant highlights occupations that have higher than average scores along both dimensions. We define those occupations as ones with high pandemic-induced automation probability. Our argument is that if an occupation is highly likely to expose workers to a contagious disease like COVID-19, and was at risk of being automated prior to the pandemic, then the chances are strong that it will be automated as a response to the pandemic. There are a total of 124 six-digit SOC occupations that fall into this category, representing 37 million jobs (24% of all U.S. jobs).

The size of the bubble refers to the number of jobs associated with a given occupation in the United States as of 2019. Occupation job counts are generated by taking industry job counts from the Quarterly Census of Employment and Wages (QCEW) and combining them with staffing patterns from the BLS's Occupational Employment Statistics (OES) data set. The results show a larger number of jobs associated with Laborers (3,216,984 jobs), Registered Nurses (3,104,530 jobs), Heavy and Tractor-Trailer Truck |Drivers (2,768,560 jobs), and so on. The color of the bubble represents the two-digit SOC codes associated with each occupation. A total of 22 occupation groups is highlighted in different shapes and gray scales in Figure 2.

Figure 2.

Pandemic-induced automation quadrant: Number of jobs in the United States.

Source: Authors’ elaboration.

The following are some key observations related to the quadrant plot presented in Figure 2. First, Janitors and Cleaners, Waiters and Waitresses, Dishwashers, Cashiers, Food Servers, and Maintenance and Repair workers have high scores on both the Exposure Index and the Automation Index, suggesting a relatively high propensity for pandemic-induced automation. These occupations also tend to pay lower wages, something that could limit workers’ ability to purchase the personal protection equipment necessary to better protect themselves from exposure and infections (unless these protective devices are provided by their employers). They are also more likely to pay hourly wages and normally do not offer paid sick days and other benefits.

Fixing the level of the Automation Index around the mean (100), we are able to focus on the Exposure Index of the occupations along the vertical axis. This allows us to examine the level of pandemic-induced automation at a given level of automation. Occupations with high risks include Dental Hygienists and Nursing Assistants. Hospital robots have been used for both these occupations to help in helping combat the ongoing wave of burnout among health care workers (The New York Times, 2020).

Worthy of note is the fact that these occupations with high risks encompass a sizable number of workers (as shown by the large bubble size). It is important to be aware of the potential shock that automation could portend for this sizable workforce. A majority of Health Care Practitioner and Health Care Support occupations (marked in circles and squares in medium gray colors) are at high risk of exposure to disease and infection (located on the top left corner in the quadrant) but score low on the Automation Index. However, there are exceptions noted for Orderlies and Pharmacy Technicians, both with high exposure risk and high automation indices, and thus more likely to face pandemic-induced automation. Given the critical role of most health care occupations in providing much-needed services to patients and the low likelihood of automation, employers would be prudent to pursue strategies that offer their workers the highest level of protection possible from exposure to disease and infection.

The lion's share of Food Preparation and Building and Ground Cleaning occupations place high on both the pandemic-induced Automation Index and the Exposure Index. A sizable proportion of transportation occupations, which employ many workers in the United States, have high automation scores but face low risks of exposure. As a result, they have a lower propensity for being automated as an outcome of the COVID-19 pandemic. A reason for high automation scores in the transportation industry is the ongoing technology-induced automation, such as connected and automated vehicles or self-driving cars and trucks. A list of the top 20 occupations in each quadrant, sorted by the total number of jobs in the United States, can be found in online Appendix 1.

A Look at Median Hourly Earnings by Occupation

In examining the median hourly earnings associated with various occupations, we find that jobs that are at the highest risk of automation and exposure provide the lowest hourly compensation. The hourly income data are obtained from Lightcast (2020c). Figure 3 shows the median hourly income by the size of the bubble. The minimum median hourly wage in our data set is $5.90 per hour (SOC 27-1012 Craft Artists). The average median hourly wage for occupations with high automation and high exposure scores (top right quadrant) is $18.10. For those with low automation and low exposure scores (bottom left quadrant), the average median hourly wage is $33.10. It is $18.70 for occupations with high automation, low exposure scores (bottom left quadrant), and $32.30 for those having low automation, high exposure scores.

Figure 3.

Exposure–Automation Index quadrant: Median hourly income.

Source: Authors’ elaboration.

In summary, our initial analysis suggests that occupations in the food preparation and service occupations, as well as cleaning-related occupations, have a higher chance of pandemic-induced automation, while medical and health care occupations are less likely to be automated, despite having high exposure scores. Of course, this study is exploratory and does not account for constraints, such as the financial capacity of the firms to undertake an expensive automation project. Nor does it account for the unintended distortions of public policies such as the Paycheck Protection Program (PPP), extended unemployment benefits, or other short-term support activities of federal and state governments. Table 2 shows the descriptive statistics for the top five occupations that are categorized as having high automation-high exposure scores (top right quadrant in Figure 2).

Table 2.

Top Five Occupations by Number of Jobs in High Pandemic-Induced Automation Quadrant.

SOC	Occupation	U.S. jobs 2019 (million)	Percentage of U.S. jobs	U.S. median wage ($)
43-9061	Office Clerks, General	3.526	2.4%	15.74
41-2011	Cashiers	3.670	2.5%	10.77
37-2011	Janitors and Cleaners	3.211	2.2%	12.66
35-3031	Waiters and Waitresses	2.649	1.8%	10.47
49-9071	Maintenance and Repair Workers, General	1.646	1.1%	18.31

Note. SOC = Standard Occupational Classification System.

Source: Authors’ elaboration.

Demand for High-Risk Occupations: The job Posting Analysis

Data and Specifications

Having constructed indices to measure potential pandemic-induced automation, we now use panel regressions to estimate if shifts have been occurring in the demand of jobs for occupations that have different levels of risk for automation and exposure because of the pandemic. We use the monthly U.S. job posting data between September 2016 and August 2021 (60 months in total). We define April 2020 through August 2021 as the pandemic period. ¹ The job posting data are obtained from Lightcast's Unique Job postings, which is the number of deduplicated job vacancy advertisements scraped from over 45,000 websites (Lightcast, 2021). Deduplication is the process of identifying duplicate job postings and counting only one of the duplicates. The unique posting count is the count of postings after the deduplication process has been completed. The total posting count is the number of postings before deduplication. For example, if a user runs a report that returns 12 total job postings and two unique job postings, this means that the 12 postings contained 10 duplicates and only two unique job advertisements.

It is important to note that the raw data of job postings from Lightcast are gathered by scraping over 45,000 websites, including company career sites, national and local job boards, and job posting aggregators. Its job posting data are a measure of employers purportedly looking to fill job vacancies. This makes it a valid proxy for labor market demand. But it is not necessarily the same as the actual total job vacancies or job openings (which might be more precise in characterizing labor market demand) because not all job postings are available online.

We use the monthly job postings as the dependent variable and estimate if the Automation Index and Exposure Index are associated with a change in labor demand. We start with the fixed-effect regression. The fixed-effect model controls for the within-occupation time-constant unobservables that might correlate with job postings (e.g., occupation-specific demand). It estimates the change in the relationship between the pandemic-induced automation propensity and the changes in the number of job postings before and during the pandemic. The period before the pandemic is the pre-period and the period during the pandemic is post-period. The estimation is based on the following specification:

$$Jo{b}_{ot}={\beta }_{1}Automatio{n}_o\times Pandemi{c}_t+{\beta }_{2}Exposur{e}_o\times Pandemi{c}_t+{\beta }_{3}Automatio{n}_o\times Exposur{e}_o\times Pandemi{c}_t+{\alpha }_o+M_t+{\epsilon }_{ot}$$ (1)

Here $Jo{b}_{ot}$ is the number of unique job postings for occupation o in time t, $Automatio{n}_o$ is the Automation Index for occupation o; $Exposur{e}_o$ is the disease Exposure Index for occupation o; $Pandemi{c}_t$ is the time indicator for the pandemic period and equals 1 for periods after April 2020. ${\alpha }_o$ is the indicator for occupation o (i.e., the fixed effect) that captures the individual occupation heterogeneity, and ${\epsilon }_{ot}$ is the error term. To account for fluctuation, we include indicator $M_t$ for each month. The key variable of interest is the three-way interaction term of $Automation$ and $Exposure$ and $Pandemi{c}_t$ . This interaction term characterizes the possibility that the impact of the pandemic depends on both automation and exposure level.

To account for occupation-specific growth in job demand, we also include occupation-month trend. The specification is shown in Equation (2).

$$Jo{b}_{ot}={\beta }_{1}Automatio{n}_o\times Pandemi{c}_t+{\beta }_{1}Exposur{e}_o\times Pandemi{c}_t+{\beta }_{3}Automatio{n}_o\times Exposur{e}_o\times Pandemi{c}_t+{\alpha }_o+Tren{d}_{ot}+{\epsilon }_{ot}$$ (2)

Job Posting Analysis: Results

We start by summarizing the unconditional correlation. The results show that before COVID-19, the correlation coefficient between the Automation Index and the job posting is negative at a smaller magnitude (−0.063) and remains negative but at a greater magnitude (−0.078) post-pandemic. These negative coefficients indicate that occupations with higher automation risks tend to be associated with fewer job postings, and since the pandemic, the negative relationship became stronger, as we anticipated. In terms of the Exposure Index, we find, interestingly, that the index is positive, with the correlation coefficient being 0.0017 before the pandemic and 0.022 since the pandemic, indicating a stronger positive correlation that is consistent with the panel regression results.

Table 3 summarizes the results from the fixed-effect panel regressions. We find that there is a significant decrease in the number of unique job postings post-pandemic for occupations with a higher Automation Index score. On the other hand, we note the significant increase in postings for occupations with a higher Exposure Index score. The key variable of interest is the interaction term of the Automation Index, Exposure Index, and pandemic indicators. It shows significant effect in model 3, and an insignificant (but still negative) coefficient in model 4 where we control for the two-way fixed effect. This indicates that since the pandemic, occupations with higher automation and/or higher Exposure Index scores are experiencing a decrease in the number of job postings, but the size of the influence is relatively small or statistically not different from zero. Overall, the results indicate that there is little evidence, to date, to suggest that a substantial demand shock to the labor market has occurred because of pandemic-induced automation.

Table 3.

Panel Regression for the Monthly job Posting.

	Model (1)	Model (2)	Model (3)	Model (4)
$Automation\times Pandemic$	−18.20***	−52.19***	−11.29*	−87.52***
	(3.72)	(11.20)	(4.50)	(31.98)
$Exposure\times Pandemic$	78.46	64.83***	150.29***	−22.19
	(8.51)	(9.29)	(27.78)	(74.68)
$Automation\times Exposure\times Pandemic$	/	/	−0.86**	−0.92
	/	/	(0.31)	(0.71)
$Occupation\times Timetrend$	Yes	Yes	Yes	Yes
$OccupationFE$	Yes	Yes	Yes	Yes
$Year-monthFE$	No	Yes	No	Yes
$Fstatistics$	2.74***	3.31***	3.90***	7.31***
$NumebrofObservations$	34,800	34,800	34,800	34,800

Note: Dependent variable: Unique job posting by month. The standard error are shown in parentheses, all clustered at the SOC two-digit level and allow AR (1) serial correlation. All models include both occupation effects and year-month effects. Occupation× Time trend represents the occupation-specific growth in job posting. Significance level: *** p<0.01, ** p<0.05, * p<0.1. Via Hausman specification tests, we reject the null hypothesis of random effect models.

We find that on average, since the pandemic, a score that is one point higher in the Automation Index is associated with around 30 fewer unique job postings per month in the United States relative to the pre-pandemic period. This might suggest a substitution effect. Moreover, holding other variables constant, on average, since the pandemic, a score that is one point higher in the Exposure Index is associated with around 80 more unique job postings each month. This might be driven by higher demand for medical-related occupations during the pandemic. However, if an occupation scores one point higher in the pandemic-induced automation score, there is one less unique job posting each month post-pandemic, although it is not significant in the preferred model specification. ² In other words, the relationship between the Automation Index and postings is dependent on the Exposure Index.

At the first glance, the effect of pandemic-induced automation seems economically small but notice that the marginal effect of the Automation Index (Exposure Index) is a function of the Exposure Index (Automation Index). Therefore, we visualize the marginal effects of the Automation Index (Exposure Index) on monthly job postings, evaluated at different levels of the Exposure Index (Automation Index) using the preferred model specification. The results are presented in Figure 4. The dashed vertical lines indicate the average level of each index. In the left panel, the marginal effect of the Automation Index is negative and significant. With a higher score on the Exposure Index, the effect of a higher Automation Index score has a stronger association with fewer monthly job postings, suggesting that the pandemic-induced automation might have a correlation with decreasing labor demand. Evaluated at the mean of the observed Exposure Index, this effect translates into approximately 28 fewer monthly postings. Similarly, in the right panel, the marginal effect of the Exposure Index is positive and significant along most of the Automation Index spectrum. With a higher Automation Index score, the effect of a higher Exposure Index score has a positive but smaller correlation with fewer monthly job postings.

Figure 4.

Marginal effects of the Automation Index and Exposure Index on job posting.

Source: Authors’ elaboration.

It is also possible that the effect of pandemic-induced automation on job demand may display a time-lagged effect, and that the effect of pandemic-induced automation on job postings may change over time. We have extended the analysis to different time lags after April 2020. First, we started with a placebo month—January 2020—when the pandemic began in other countries but not in the United States. At that time, the coefficient of the interaction term (Automation Index and Exposure Index) is not significant. Moreover, with longer time lags starting from April 2020, the estimated effect remains negative, and the overall trend is toward a greater impact, suggesting some extent of time lag in the effect of pandemic-induced automation, although the effect is not significant after February 2021. Figure 5 shows the estimated impact of automation–exposure interaction on monthly job postings. The effect, in general, is getting larger and remains negative, indicating that the higher the multiplication of the Automation and Exposure indices, the lower the monthly job postings, consistent with our expectation, although the effect is insignificant after February 2021.

Figure 5.

Time trend of the estimated effect.

Source: Authors’ elaboration.

Since the duration of active job postings varies substantially across occupations, we account for different durations in time as a robustness check. In our data, the median job posting duration ranges from 6 days to 52 days, with the mean duration at 30.5 days. A short posting duration can represent a different market demand than a long duration for the same number of postings. Therefore, we divide the number of unique job postings by the median posting duration, such that the dependent variable is interpreted as a unique posting per day during which the posting remains active. For example, from September 2017 to October 2017, both Mathematician and Sound Engineering Technicians have 121 unique job postings. The median durations, however, are 33 days and 19 days, respectively. Therefore, the re-weighted postings are 3.6 and 6.4, respectively, and are interpreted as the number of unique postings per active posting per day. The average of the median posting duration in our data is 29 days, with a SD of 13 days.

With the new dependent variable of unique job postings per day, we estimate the fixed-effect regression following the same specifications. The results are reported in Table 4. The estimated coefficient remains largely robust. After the pandemic, occupations with a higher Automation Index score tend to have fewer postings per duration day, and occupations with both a higher Automation and Exposure Index also have fewer postings.

Table 4.

Panel Regression for the Monthly job Posting Weighted by Duration.

	Model (1)	Model (2)	Model (3)	Model (4)
$Automation\times Pandemic$	−1.65***	−1.38***	−1.77*	−2.81***
	(0.16)	(0.43)	(0.19)	(1.25)
$Exposure\times Pandemic$	2.36	2.46***	1.21	−1.07
	(0.33)	(0.36)	(1.09)	(2.91)
$Automation\times Exposure\times Pandemic$	/	/	−0.01	−0.04*
	/	/	(0.01)	(0.02)
$Occupation\times Timetrend$	Yes	Yes	Yes	Yes
$OccupationFE$	Yes	Yes	Yes	Yes
$Year-monthFE$	No	Yes	No	Yes
$Fstatistics$	123.85***	30.99***	93.18***	21.16***
$NumebrofObservations$	34,800	34,800	34,800	34,800

Note: Dependent variable: Unique job posting by month divided by median duration of posting. Hausman test rejects the null hypothesis of random effect at 1%. The standard error are shown in parentheses, all clustered at the SOC two-digit level and allow AR (1) serial correlation. All models include both occupation effects and year-month effects. Occupation× Time trend represents the occupation-specific growth in job posting. Significance level: *** p<0.01, ** p<0.05, * p<0.1. Via Hausman specification tests, we reject the null hypothesis of random effect models.

Job Posting Analysis for Non-Health-Related Occupations

It is possible that pandemic-induced automation affects non-health and health-related occupations differently. On the one hand, the health-related occupations might experience greater digitalization and more applications of AI technologies due to COVID-19. Since the pandemic's onset, innovations have been seen in the health care sector, including detection of outbreaks, vaccine discovery, facilitation of diagnostic and prognostic decision support, identification of people with fevers, and epidemiological monitoring and prediction, and telehealth, demonstrating that AI holds much promise for helping health care workers respond to the health crisis and future medical functions. On the other hand, some studies have pointed out that even with the pandemic, overall automation technologies have not yet been scalable in medical occupations, due to technical, regulatory, and ethical issues (Mckinsey & Company, 2022).

As an extension of the main model, which uses a full sample, we estimate the preferred model specification, but exclude SOC 29 Healthcare Practitioners and Technical Occupations and 31 Healthcare Support Occupations. Table 5 summarizes the results. Notice that the direction of influence of the main coefficients of interest—the interaction of the Automation Index with the pandemic indicator, as well as the pandemic indicator interacting with the Automation and Exposure Indices, remains robust. Moreover, the estimated coefficients show a larger magnitude than the baseline results reported in Table 1. On the other hand, using the subsample of only health-related occupations, we find an insignificant effect on job postings (results not reported here). One possible explanation is that most automation technologies in health occupations demand more human skills rather than human replacements. In fact, many new medical treatments related to COVID-19 diagnosis and treatment is likely to require skilled labor (Mariano & Wu, 2021).

Table 5.

Panel Regression for the Monthly job Posting (Excluding Health-Related Occupations).

	Model (1)	Model (2)	Model (3)	Model (4)
$Automation\times Pandemic$	−38.62***	−41.12***	−40.82***	−145.37***
	(4.84)	(10.74)	(5.48)	(34.69)
$Exposure\times Pandemic$	48.75***	47.61***	25.19	−235.59***
	(12.29)	(13.01)	(30.06)	(90.87)
$Automation\times Exposure\times Pandemic$	/	/	0.27	−2.88*
	/	/	(0.32)	(0.91)
$Occupation\times Timetrend$	Yes	Yes	Yes	Yes
$OccupationFE$	Yes	Yes	Yes	Yes
$Year-monthFE$	No	Yes	No	Yes
$Fstatistics$	133.39***	13.21***	100.23***	12.14***
$NumebrofObservations$	31,620	31,620	31,620	31,620

Note: Dependent variable: Unique job posting by month for non-health-related occupations. The standard errors are shown in parentheses, all clustered at the SOC two-digit level and allow AR (1) serial correlation. All models include both occupation effects and year-month effects. Occupation× Time trend represents the occupation-specific growth in job posting. Significance level: *** p<0.01, ** p<0.05, * p<0.1. Via Hausman specification tests, we reject the null hypothesis of random effect models.

Caveats

The Automation Index and Exposure Index used in this article are selected and constructed based on careful economic and contextual considerations. While it is appropriate for our research objective, they may have the following limitations, which suggest directions for future research. To begin, data constraints make it difficult to draw conclusions about the temporal change in the Automation Index itself since COVID-19. Additionally, both the Automation Index and Exposure Index do not vary across geographical units, implicitly assuming spatially homogenous risks. Yet, it is also reasonable to expect that automation risk or exposure risk may vary across rural and urban areas, due to different policies and social responses. If the task share of occupations was available at a more disaggregated level, a spatial-variant Automation Index could be estimated and visualized more granularly from location to location. Similarly, we are unable to disaggregate the index by different demographic variables, such as education, age, and other socioeconomic variables. Additional case studies in conjunction with qualitative evidence will help elucidate if some groups are more prone than others to pandemic-induced automation risks.

Note that job postings data may exclude job openings that use analog methods for advertisement such as flyers, billboards, and word of mouth. There may be a rural-urban divide between the coverage of online job postings in rural areas that lack Internet access and urban areas that have good access. However, these limitations show opportunities and directions for future research.

Conclusion

The COVID-19 pandemic has necessitated interventions that reduce physical contact among people. Occupations that require close contact impose a shadow cost on labor that may accelerate the development and adoption of new technologies to automate human work. Our study sought to explore the potential impacts of the pandemic on the labor market through the accelerated adoption of job-related automation processes. Recent studies have pointed out that COVID-19 may serve as a tipping point for automation technology (Coombs, 2020). Even after COVID-19 passes, companies may opt to pandemic-proof their operations (Coombs, 2020). As such, greater demand for automation in critical occupations could increase.

This article serves as an early assessment of the propensity of different occupations to be automated because of COVID-19. Consistent with anecdotal evidence, we found that food preparation, service, and cleaning-related occupations have the highest chance of pandemic-induced automation, while medical and health care professions, although facing high contagious-disease exposure risk, are less likely to be automated. The list of occupations that are deemed to be at high risk of disease exposure and automation may provide local and state economic development leaders and policy makers with an early signal of the possible need to put in place a mix of programs and policies to minimize the potential for displacement of local workers. This could include the upskilling of displaced workers so that they can fill existing job openings in their locality, and/or providing them with the training needed to qualify for the newer, higher-paying jobs that are likely to be created because of automation.

Using monthly job posting data, we find that the potential for pandemic-induced automation has been somewhat translated into a statistically significant, although economically weak, change in job postings, suggesting that the labor market demand might have been somewhat influenced by the application of automation technology since the onset of the pandemic. Institutions and policy makers should support efforts and investment in AI that complements human workers. Therefore, it is important to understand the occupations that would be more prone to pandemic-driven automation, and help workers be more prepared and more productive in tasks that are not easy to automate.

On the other hand, the adoption of new technologies and robotic systems can help firms to overcome the negative effects of the pandemic shock more quickly while keeping their workforce safe. However, some businesses may lack the financial resources to adopt automation technology. This may partially explain the quantitatively weak empirical relationship we find between the pandemic-induced automation and job postings during the pandemic. Attention needs to be paid to this as businesses that do not adopt these technologies might have a competitive disadvantage during the pandemic, resulting in a loss of jobs that may be inaccurately blamed on automation directly. Proper stimulus packages, including broadband development (PCRD, 2022), may help businesses adopt digital solutions as well as provide the necessary fiscal incentives for investments. A World Bank report emphasized that the COVID-19 pandemic has revealed the critical importance of digital stimuli, including digital infrastructure, technologies, and services during times of crisis that enable governments, businesses, and communities to continue to function (World Bank, 2020).

This study has other limitations, too. Due to the data constraint, we could only look at the short-term impacts on labor market demand. In the long run, as markets adjust, changes in labor demand are mainly reflected in the adjustment of wages rather than a change in the number of job postings or jobs. In addition, the level of analysis of our study is occupation. But in practice, each occupation incorporates a variety of work tasks. Tasks are automated when an automated system can perform them more efficiently and at a lower cost than human laborers. Some tasks can be easily automated while some others are comparatively difficult to automate. One direction for future researchers is to analyze the pandemic-induced automation at the task level. Finally, the study period reflects conditions before the start of the Omicron wave of the virus. Given that the pandemic situation is still rapidly changing in the United States, it is important to continuously monitor its impact on the labor market.

Evidence from more recent studies has suggested that workers are exiting the job market in hordes. There are more job openings than available labor. Fatigue from the pandemic, risk, and fear, are some of the reasons that might impact the number and duration of job postings in the coming months. Recent evidence has shown that big businesses are doing better than small ones in this great reshuffling of workers (Beaulieu, 2022; Cirera et al., 2021). The biggest occupation gains and losses point to a blue collar–white collar divide, with finance and insurance leading in new job openings and accommodation and food service having the greatest number of job losses.

Author Biographies

Chun Song was a research assistant at the Purdue Center for Regional Development at the time of this work. She holds a PhD in agricultural economics from Purdue University. She is currently a spatial econometrician at CGIAR.

Lionel J. Beaulieu is a professor emeritus of rural and regional development with the Purdue Center for Regional Development and the Department of Agricultural Economics. His research interests include education/human capital development and asset-based community development.

Indraneel Kumar is the principal regional planner at the Purdue Center for Regional Development, and currently has an additional role as the interim assistant director of research. His research interests intersect regional economics, transportation planning, and renewable energy.

Roberto Gallardo is the director of the Purdue Center for Regional Development and an associate professor in the Agricultural Economics Department. His research interests include regional, economic, and community development and digital inclusion and transformation.