Unemployment Risk by Occupation, Place, and Time.

Example Narratives

---

The demo above allows you to explore these data and unemployment risk for yourself. Here are a few specific examples:

• In January 2018, Construction workers from Texas claimed the most unemployment benefits across all US states (see green map), but, after controlling for labor market size and local employment for construction workers, unemployment risk for Texas Construction workers was close to the average across other states (see orange map). Note that neither plot would be possible with typical total unemployment statistics from LAUS.

  Show Me

• As a second example, the unemployment risk (i.e., the orange map) for Sales and Retail workers in each US state in January and May 2020 captures the effects of lockdowns and social distancing measures on these workers. This temporal and spatial granularity is only possible because of the detailed unemployment statistics in the project and is not possible with other unemployment statistics reported by federal sources.

  Show Me

• From 2018 through 2020, we aggregate Michael Webb's estimates for occupations' exposure to artificial intelligence (AI) patents to assess the average exposure of a worker in each US state. We perform similar calculations for several other technology exposure methods (see blue-to-orange map). For technology exposure measures that are specific to AI or machine learning, California has the lowest exposure. However, California also has the highest unemployment risk for Computer and Math workers.

  Show Me

Data Description

---

Unemployment Risk

From advances in artificial intelligence (AI) to a global pandemic, today's workers face many labor disruptions requiring their adaption to the future of work. Accordingly, new research is required to (1) identify which workers or populations have the greatest risk of unemployment, (2) decompose unemployment into its many possible causes (e.g., pandemic related vs. technology related vs. other factors), and (3) to identify how workers can and do respond to unemployment risk. Although pandemics and AI are ostensibly unrelated, their consequences for workers are similarly determined by workers' skills and abilities. For example, in 2020, knowledge workers with experience using digital technology more easily adapted to work-from-home while frontline workers faced job uncertainty. Digital technologies, including machine learning and artificial intelligence, augment the productivity of knowledge workers while robotics replace physical jobs in manufacturing and agriculture. Given these similarities across different disruptions, how do skills identify adaptable workers and can we identify the workers who are most at risk of unemployment in their local labor market?

To identify the workers who are most likely to require unemployment benefits, we have collected monthly data from the US Department of Education and Training Administration (DETA) detailing the age, gender, ethnicity, and most-recent occupation for the universe of unemployment benefit recipients in each US state. Unlike typical sources of unemployment data—including total unemployment reported by the US Bureau of Labor Statistics (i.e., LAUS)—these data describe unemployment stratified by occupation, ethnicity, gender, and age. These data better describe the labor outcomes of individual workers and the heterogeneous impact of labor disruptions across occupations, states, and time. For example, nationally, women were more likely than men to leave the workforce during the COVID recession and have been slower to return to work during the recovery. Prudent policy and entrepreneurial activity might support child care or retraining to aid these displaced workers if they knew which communities to target. This project asks similar questions by occupation.

Most interestingly, these detailed unemployment statistics enable the unprecedented opportunity to study workers' unemployment risk while controlling for other local labor market factors. For example, it may not be surprising if unemployment for tech workers in California is higher than in other states because of the uniquely high concentration of tech companies in CA's Silicon Valley. We will use LAUS data for monthly total unemployment rates by region in combination with annual employment distributions by occupation and region from the BLS Occupation Employment Statistics (OEWS) to quantify unemployment risk in a given region and year according to the Bayes' theorem approximation $$ p(unemployment|soc) = \frac{p(soc|unemployment)\cdot p(unemployment)}{p(soc)} $$ where \(soc\) is a Standard Occupation Classification (SOC) Major occupation, \(p(soc|unemployment)\) is the probability that an unemployed worker's most recent occupation was \(soc\) according to detailed unemployment data, \(p(unemployment)\) is the LAUS total unemployment rate, and \(p(soc)\) is the probability that a worker in a region is either actively employed in \(soc\) according to OEWS or is unemployed and was most recently employed in \(soc\) according to detailed unemployment statistics (i.e., \(p(soc)\) is relative to the state's labor force). This measure describes the probability of receiving unemployment benefits given your occupation, region, and month.

Data for this demo is available at https://github.com/mrfrank8176/unemploymentRisk.

Estimating States' Automation Risk

Estimations of technology exposure have evolved over the last decade. The first wave of theoretical studies adapted a production function to worker productivity in the presence of automating technologies, arguing that college-educated cognitive high-skill workers were complemented by technology, including computers, while manual low-skill workers were substituted by technologies like robotics. However, current AI technologies threaten cognitive workers as well. As examples, consider that AI surpasses human performance at predicting heart attacks or computer vision applications in radiology and neighborhood safety. The modern, skill-biased technological change framework further argues that both routine manual and cognitive work are ripe for automation—although cognitive workers will tend towards greater productivity with technology while manual workers will tend towards labor substitution (e.g., in manufacturing).

The second wave of studies considered each occupation as a bundle of skill requirements and job tasks. Beyond describing occupations as cognitive or routine, occupations' granular skill requirements are considered, for instance, by leveraging the US Bureau of Labor Statistics (BLS) O*NET database. An Oxford University study subjectively identified "fully automatable" and "not automatable" occupations combined with a subset of O*NET variables representing perception, manipulation, creativity, and social intelligence requirements of occupations. They employed a logistic regression to assign a "probability of computerisation" to each remaining US occupation. Alarmingly, they claimed that 47% of US employment had high-risk of computerization. However, the study only compared their estimates to occupations' education requirements and wages, thus leaving whether this exposure creates unemployment or alters skill demands unclear. A competing study directly estimated the automation risk of skills from the Programme for the International Assessment of Adult Competencies (PIAAC) survey enabling a direct assessment of technological exposure of occupations across OECD countries, finding that only 9% of US workers had high automation risk. Subsequent studies used these occupation estimates in a variety of contexts; for example, finding that automation will affect 35% of employment in Finland, 59% of employment in Germany, 45 to 60% of employment across Europe, and that small US cities face greater impact from automation. To meet the demand for occupation-level automation estimates, the BLS added a Degree of Automation score to occupation profiles in the 2016 O*NET database.

The most-recent third wave of studies directly connects specific technological capabilities to occupations' job tasks to assess each occupations' exposure (i.e., a task-based approach). One study surveyed machine learning (ML) experts on the characteristics of tasks that are suitable for ML and produced a Suitability for ML score for US occupations. Another study surveyed gig workers to establish connections between AI application capabilities and occupations. A more recent study used natural language processing to connect technology patents to job tasks. Although motivated by the risk of technological unemployment, these studies argue that AI exposure will mostly result in labor reorganization through wealth inequality or changing skill demands without major changes to unemployment.

Unfortunately, these estimates are not stratified by state and occupation. Given a per-occupation technology exposure score \(exposure(j)\), we calculate the aggregate exposure for state \(s\) in year \(y\) according to $$ exposure(s,y) = \sum_{j\in SOC} exposure(j)\cdot share_{s,y}(j) $$ where \(j\) is a six-digit Standard Occupation Classification code and \(share_{s,y}(j)\) is the share of employment associated with occupation \(j\) in that state and year according to the BLS.

Future Work:

• We will perform similar calculations to estimate unemployment risk across cities/counties and years by gender, ethnicity and age (e.g., \(p(unemployment|female)\) or \(p(unemployment|asian)\)).
• We will employ small area estimation techniques to project state-level data to sub-state levels and, in particular, to individual cities.

These estimates offer unprecedented granularity into the social aspects and spatial heterogeneity of unemployment risk thus enabling new research and policy interventions that target the most susceptible workers are the community level (e.g., real-time identification of regions where women experience high unemployment risk).