Working Papers

Abstract: A central concern regarding artificial intelligence (AI) is its potential to replace jobs and exacerbate economic inequality. However, recent research argue that AI may provide a path to decrease inequality through a Turing Transformation process: AI simplifies work, reduces barriers to job entry, and consequently widens job opportunities for more workers. In this paper, we empirically test the Turing Transformation theory by examining AI's deskilling and job opportunity effects. We develop a novel occupational AI exposure index using a sentence transformer model to compare the semantic similarity between the occupation descriptions (what people do) and AI patents (what AI technologies do). We find that, on average, occupations with higher AI exposure experience a decrease in the importance for a wide range of work activities, along with an increase in job postings and employment. This provides the first empirical evidence for the existence of the Turing Transformation process. However, the beneficial job opportunity expansion effects are absent for low skill occupations. For high skill occupations, as their AI exposure increases, we observe large increases in job postings but little changes in actual employment, suggesting a talent gap for high-skilled workers due to AI.

Abstract: We study the changes in job requirements for workers when companies adopt Machine Learning (ML) related technologies. Using over $51$ million job postings of S\&P 500 companies from $2011$ to $2023$, we find that firms utilizing ML also increased job requirements for their workers in prior work experiences and decision-making related skills. %Such effects are evident for non-ML jobs. We provide a causal link with a Shift-Share IV estimation strategy based on the text of occupation description and AI patents. Further, the ML's job requirements increasing effects are strong not only for knowledge workers (e.g., Managers), but also for lower skill occupations that do not require a college degree (e.g., Retail Salespersons and Customer Service Representatives). Our results indicate that, with the ability to handle straightforward cases and processes, ML technologies may have shifted the ``mass'' of work for people to handle difficult and novel tasks that necessitate greater domain expertise, decision-making, and subjective judgment. Consequently, the proliferation of ML could lead to a reduction in entry-level job opportunities across a broad spectrum of occupations. With the ever-growing adoption of ML and related technologies, these consequences require proactive career development strategies for individuals and institutions.

Abstract: Despite the well-recognized benefits of AI for businesses, its actual adoption rate remains low. This study explores human capital availability as a barrier to AI integration. We utilize the newly launched Analytics and Data Science (ADS) programs over the past decade as a supply shock and examine their impact on firms' ADS postings. The key idea behind our empirical strategy is that, if human capital availability constrains firms' AI adoption, job postings should increase following such supply shocks. Our findings reveal a significant human capital constraint, particularly among local firms hindered by the limited AI talent in the local job markets. These results underscore potential AI inequality across firms and regions and highlight the critical role of education in mitigating these disparities. Further, we observe larger effects from business than non-business ADS programs. We discuss the implications of our findings for employees, employers, educators, and policymakers in workforce development in the age of AI.

Abstract: What effect do skilled immigrants have on the retention of native U.S. workers? With granular career information of over 20,000 employees, we examine this question by teasing out the effect of a firm's proportion of skilled immigrants on its native employees' turnover probability. We apply unsupervised machine learning to categorize employees' self-reported skills and find that skilled immigrants disproportionately specialize in IT. In contrast, native workers predominantly focus on management and analyst skills. As suggested by the labor supply-demand framework and the human capital externality theory, the skilled immigrants at a firm could increase the productivity of the native workers with complementary skills at the firm. As a result, the firm may be more likely to retain such native employees and the same employees are also more likely to stay at the firm in which they experience the improved productivity, leading to a decrease in their turnover probability. Utilizing an IV based on the randomness in the H-1B visa lottery and a 2SLS design, we find that a one percentage point increase in a firm's proportion of skilled immigrant employees leads to a decrease of 0.69 percentage points in a native employee's turnover risk. However, this beneficial crowding-in effect varies for native workers. As an example, native workers with management skills derive beneficial crowding-in effect from the increased proportion of IT skilled immigrants but incur a steep crowding-out effect from the increased proportion of immigrants with management skills. We discuss the implications of such heterogeneity on immigration and HR policies. 

Abstract: Given its economic impact on firms, employee turnover has attracted interest from management practitioners and researchers for over a century.  However, the existing literature has paid little attention to the prediction and identification of employees who are most likely to turnover. This gap likely emanates from the lack of large-scale individual-level career data that can be used to leverage more recent developments in data science. In this paper, we use publicly available online employee profiles and study the prediction and identification of the most quit-prone employees. In particular, by tracking employee job histories across different firms over time, we can observe an employee's social ties with other employees in external firms resulting in a coworker network structure. We construct network embeddedness metrics and evaluate their utility in predicting employee turnover above and beyond control predictors derived from the literature. While using various predictive performance criteria, we find that the proposed network embeddedness metrics are essential in predicting most quit-prone employees. To the best of our knowledge, we are the first to investigate the role of external firm social ties through network embeddedness in predicting employee job turnover. We also discuss the practical implication of our turnover prediction framework for both internal and external stakeholders of firms.

Abstract: This study examines whether digitalization increases firm resilience in the presence of a negative exogenous event. We propose a measure of firm digitalization that is based on the intensity of its investment in IT-related human capital using an extensive job postings database. The measure enables us to quantify digitalization for a large cross-section of firms through time. We test the efficacy of firm digitalization as a prophylaxis against the disruption created by the COVID-19 pandemic using two different identification strategies. The first strategy uses a difference-in-differences styled model with two-way fixed effects, and the second uses the synthetic control method. Through both empirical strategies, we consistently find that higher levels of firm digitalization result in greater firm resilience during COVID-19 disruptions. We also document a complementary prophylactic effect of the scale and the scope of firm digitalization. Furthermore, there is considerable heterogeneity in these effects across different types of IT-related human capital. Our study highlights a new dimension of IT business value — digital prophylaxis — and estimates it empirically for firm resilience.  The results have implications for managers, shareholders, and regulators while positioning IT to the center of preparedness plans for future environmental disruptions.

Abstract: In estimating causal effects with observational data, many methods rely on the unconfoundedness assumption. A popular tool for detecting bias due to unobserved confounding is the use of Negative Control Outcomes (NCO). In this paper, we present a procedure for bias correction when data on many candidate NCOs are available. We first propose an algorithm to construct a good NCO, which we call ``CoNeCo''. Then we apply the transformed outcome (Y - CoNeCo) to the Synthetic Control Method. Using simulated data, we show that our method can recover the treatment effects under challenging scenarios such as when pre-treatment time periods are small and good pre-treatment outcome matches are unavailable. Our method extends the Negative Control Outcomes from bias detection to bias correction with a specific application to the Synthetic Control Method.