Working papers
"Student demand and the supply of college courses" [Draft]
Media coverage: Education Next, Marginal Revolution
In an era of rapid technological and social change, how do universities adapt? To examine this, I extracted the information contained in the course catalogs of over 700 US colleges and universities, observing courses dating back to 1998. When there are changes in student demand, universities adjust course quantity substantially less than one-for-one, particularly in fields experiencing declining demand. Research universities (R1) are somewhat more responsive in expanding high-demand fields and less research-intensive universities in reducing declining fields. Using Natural Language Processing, I further show that while the content of existing courses remains largely unchanged, newly created courses incorporate topics related to social justice and job skills. R1 and Liberal Arts universities exhibit the most pronounced content changes, with a notable increase in emphasis on social justice.
"Anatomy of labor market distress" (with Eric Hanushek, Simon Jansenn, and Lisa Simon) [Draft] [NBER working paper]
Earnings losses after job displacement are highly skewed: a small number of workers experience catastrophic losses, while most workers recover quickly. Thus, average earnings losses as commonly estimated by event studies significantly overstate the losses for a majority of displaced workers. This paper documents the heterogeneity in earnings losses after job displacement and the behavioral differences underlying these adjustment differences. We study workers from firms in West Germany that closed between 2000-2005. By creating a synthetic control for each laid-off worker from similar workers who were not laid off, we can estimate the full distribution of economic losses. As found in other analytic approaches, older, less educated, and female workers suffer larger average losses, but a key result in our analysis is the remarkable overlap of the demographic loss distributions. Fixed characteristics do not predict which workers will experience the greatest losses; instead, these losses are associated with post-layoff adaptability, such as switching professions or geographic relocation.
"Value-Added in Postsecondary Education" (with Merrill Warnick and Anthony Yim) [Draft]
Media coverage: Marginal Revolution
Estimating post-secondary instructors’ value-added is challenging because college students select their courses and instructors. In the absence of sound measures of value-added, universities use subjective student evaluations to make personnel decisions. In this paper, we develop a method to estimate instructor value-added at any university. The method groups together students who have previously taken similar courses and estimates value-added based on differences in outcomes for students in the same group and same course who have different instructors. Using a unique policy at a large public university in Indiana, we show that our non-experimental method controls for selection just as well as methods that exploit conditional random assignment of students to courses. We next show that our method reduces forecast bias in a wider variety of institutions using data from nearly all public universities in Texas. We find that individual instructors matter for students’ future grades and post-college earnings in many subjects and courses. On average, moving to a 1 standard deviation better instructor would increase a student’s next semester GPA by 0.13 points, and earnings six years after college entry by 17%. Strikingly, value-added is only weakly correlated with student evaluations. An instructor retention policy based on value-added would result in 2.7% higher earnings for students attending Texas universities.
"How exposed is higher education to artificial intelligence?" [Draft]
Media coverage: Business Insider, Education Next, Show-Me Institute
Generative AI changes not only the labor market value of skills, but also the process through which those skills are developed. I construct a new measure of curricular exposure to large language models by combining task-level estimates of LLM capabilities with course descriptions from more than 1,000 U.S. colleges and universities. The measure reveals broad overlap across the curriculum: exposure is highest in fields intensive in writing and quantitative analysis --- including Statistics/Data Science, English, Computer Science --- and lowest in fields that emphasize physical, clinical, or interpersonal tasks. Average exposure has changed little over the past fifteen years, suggesting that LLM capabilities overlap with a longstanding set of tasks taught in college rather than a recently emerging curricular niche. I then use 1.1 million syllabi from 27 institutions to study how instructors responded to ChatGPT. AI policies spread rapidly, appearing in a majority of syllabi by Fall 2025, but observable changes in exam modality and grading structure were modest. The findings show that colleges have recognized the instructional challenge posed by generative AI but have made limited observable changes to how student learning is assessed.
Work in progress
"Estimating public returns to higher education" (with Natalie Millar and Merrill Warnick)
"Political polarization in college courses" (with Gideon Moore and Sam Thau)
Published and forthcoming work
"Long-run Trends in the U.S. SES—Achievement Gap" (with Eric Hanushek, Paul Peterson, Laura Talpey, and Ludger Woessmann), Fall 2022 Education Finance and Policy
Other
"The Research Challenges of the AI Labor Market Challenges" (with Eric Hanushek and Lisa Simon)