Visa Policy, Global Education and Labor Markets

with Yangfan Wang

We study how post-study work visa policy jointly shapes international students' decisions of where to study and where to work. We quantify the effects of reduced job access due to a visa shrinkage reform in 2012 UK on international students using a novel longitudinal panel of LinkedIn profile data. Focusing on Indian students, we leverage variation across undergraduate institutions in students' likelihood of working in the UK, and find that worsening job prospects deter students from enrolling in UK master's programs. Importantly, these enrollment responses are similar whether past graduates stayed to work in the UK for high- or low-skill jobs. We then develop a two-stage discrete choice model of education and post-graduation work. We find that most of the labor supply decline in bachelor-degree occupations is due to the direct search friction raised by the reform, which also drives a 53% drop in student enrollment and thereby indirectly reduces the labor supply of advanced-degree occupations by 58%. Our results highlight how a reform targeting low-skill employment can generate negative spillovers to high-skill sectors through both the employment and the education enrollment margins.

Perceived Discrimination and Job Search among First- and Second-Generation Americans

with Arnaud Maurel, Angelino Viceisza, and Daniel Y. Xu

This paper investigates the relationship between perceived discrimination and job search among first- and second-generation immigrants. We elicit beliefs about discrimination in two domains, the current workplace and the job search process. We also conduct an information provision experiment to assess whether population-level earnings from the U.S. Census shift beliefs and job search. We find that both first- and second-generation immigrants perceive labor market disadvantages relative to otherwise identical native-born workers, with larger perceived gaps among first-generation immigrants, younger immigrants, recent arrivals, and less educated workers. Both groups also overestimate the native–immigrant earnings gap in their occupation. Providing respondents with objective occupation-specific earnings information reduces perceived discrimination among those with more pessimistic prior beliefs and shifts beliefs beyond earnings to promotion prospects and job-search outcomes. Finally, perceived discrimination is positively associated with job-search effort, and instrumental-variables estimates suggest that belief changes translate into more job applications and interviews in the months following the intervention. The findings suggest that immigrants respond to perceived discrimination by intensifying job search rather than becoming discouraged.

Generative AI and Demand for Foreign High-Skill Labor in the U.S.: Evidence from the H-1B Program

with Boyang Bai

We study how generative AI is changing firms' demand for high-skill foreign labor, combining H-1B Labor Condition Applications for fiscal years 2020–2025 with a usage-based measure of AI exposure from the Anthropic Economic Index. We first document that foreign skilled workers are highly exposed to generative AI: 85.7 percent of H-1B positions are in the most AI-exposed quartile of occupations, and AI usage per worker in the average H-1B occupation is 6.5 times that of the average U.S. job. Since the release of ChatGPT, the total demand for foreign skilled workers has not declined, but its composition has shifted. By FY2025, filings in occupations where AI use is mostly automation are 15 percent below their FY2022 level, while filings for AI-builder positions are 37 percent above it. The shift appears in both hires for fresh graduates and experienced workers. Regression-based analysis shows that within a firm and year, filings in AI-substitute occupations fall by 12 percent relative to the firm's other filings and the decline is entirely driven by software developers. We find increases in filings and offered wages for external hires in AI-builder positions.

College Accessibility, Quota Relocation and Job Location Choice

with Yihong Liu

Access to higher education significantly influences social equality and intergenerational mobility, which is directly governed by the regional college admission quota system in China. The Chinese government initiated the "Cooperative Plan of College Enrollment in the Mid-western Province" (CPCE) in 2008, mandating some colleges in provinces with rich education resources to allocate admission quotas to Mid-western provinces. Using the Difference-in-Difference strategy, we find that while the policy led to lower entrance test scores for students from recipient provinces, these individuals did not exhibit inferior academic performance compared to their counterparts enrolled before the policy. Moreover, the policy induced students with stronger desire to work in metropolitan labor markets post-graduation to enroll in colleges located in eastern provinces, resulting in more Mid-western students working in China's top-four cities.

Disaster under control: droughts, locust outbreaks, and state capacity in premodern China, 1644-1911

with Yu Hao

Undergraduate Thesis, 2019

Locust outbreaks, which are normally caused by drought, pose a major threat to agrarian societies. Nonetheless, state intervention could effectively mitigate the impacts of droughts on locusts by coordinating and enforcing antilocust actions. Based on a provincial panel dataset for Qing China (1644-1911), we find locust outbreaks to be positively related to droughts in current and previous years. A within-province centralization reform imposed in the 1720s mitigated the impacts of drought on locust outbreaks in current and previous years. A further investigation shows that the mitigation effect is nonmonotonic with the degree of centralization, suggesting that both central coordination and local information were important. In addition, the reform enhanced bureaucratic control over local officials in times of crisis, leading to more effective locust control.

Bail Decision and Racial Discrimination

with Bocar A. Ba, Xavier D'Haultfœuille, and Arnaud Maurel

This paper investigates the determinants of bail decisions. We focus on the determinants of bail amounts set by the judges, a dimension that has not received much attention in the literature. To do so, we specify a structural model of the optimal bail amount, in which judges set the bail amount that minimizes the welfare cost of releasing the defendant. Our model allows for multiple dimensions of defendant-level unobserved heterogeneity, some being private information to the defendant and some being observed by the judge but not by econometricians. We establish constructive identification of the model parameters, including the welfare weights assigned by the judges to different groups of defendants, and derive a simple estimation procedure. We estimate our model using data from Chicago, focusing on whether the observed bail amounts are consistent with racial bias.

Model Misspecification in Dynamic Discrete Choice Models with Choice Friction

In static discrete choice models, a friction that probabilistically removes alternatives from agents' choice sets is observationally equivalent to a degradation of those alternatives' deterministic utilities. We analyze a two-stage discrete choice model in which a downstream friction restricts access to a preferred alternative with regime-specific probability $\lambda$, and contrast it with a naive specification that assumes a full choice set and attributes the same observed restrictions to a change in underlying utility. We show that under i.i.d. log-concave error distribution, the naive specification always understates the level of the inclusive value in any single regime. Furthermore, its bias on the change in the inclusive value for different $\lambda$, and hence on the predicted upstream response, can go in either direction, over- or under-predicting depending on the parameters. Under Type-I Extreme Value errors, we derive closed-form expressions for the implied upstream (Stage-1) inclusive value response. We then show that the misspecification is testable under Type-I Extreme Value errors: when the econometrician observes the upstream choice response, all structural parameters—latent utility, pre- and post-reform arrival rates, and the upstream taste parameter—are point-identified in closed form from four observed choice probabilities under a calibrated discount factor.