Working Papers

Expected Discrimination and Job Search, with Ieda Matavelli and Fernando Secco. Revised and Resubmitted, AER.

Abstract | Paper | World Bank Blog, Nexo

We study how expected discrimination affects job applications and interview performance in three field experiments with 2,167 jobseekers living in Brazilian favelas (urban slums). We focus on antifavela discrimination, which 87% of jobseekers overestimate. Randomizing expected address visibility---or providing information about discrimination in callbacks---does not affect average application rates or interview attendance. However, expecting interviewers to know one's favela address reduces interview performance by 0.13SD, even though interviewers are in fact blind to addresses. Expected discrimination can thus affect labor market matching, especially in hiring processes that involve face-to-face interviews.

Do Virtue Signals Signal Virtue?, with Matt Lowe and The Village Team. Submitted.

Abstract | Paper

We study whether tweets about racial justice predict costly related behaviors. Academics that tweet about racial justice are more likely to favor minority students in an audit experiment, receive higher teaching ratings, work with more Black co-authors, and are more likely to subsequently leave Twitter. Non-academics that tweet about racial justice make larger private donations towards racial justice efforts. However, three pieces of evidence suggest that higher returns to tweeting reduce the predictive value of racial justice tweets. First, tweets became almost completely uninformative during the aftermath of the murder of George Floyd, when more people were tweeting about racial justice. Second, the informativeness of tweets is driven by low-visibility tweet types, like retweets. Third, racial justice retweets are somewhat less informative of donation behavior than private statements of support. Finally, we find that roughly half of surveyed graduate students are overly cynical, believing tweets to be close to uninformative.

Where Does Top Talent Go to College? Global Talent Allocation and Its Consequences, with Ruchir Agarwal and Patrick Gaule. Available upon request.

Abstract

The world's highest-potential students are born everywhere; how many end up at universities below their potential, why, and does it matter for their careers? We answer these questions using 11,424 International Mathematical Olympiad (IMO) participants, linking them to undergraduate institutions and career outcomes. We estimate that only 29% of recent medalists attend a top-10 university. Country of origin is the dominant predictor of attending a top university: the need to migrate imposes a 37–42 percentage point penalty. A regression discontinuity design indicates that informational frictions faced by universities could explain at most one-eighth of the placement gap between developing-country and US/UK medalists. Conditional on IMO performance, top-10 attendance predicts a 32pp increase in top-10 PhD attainment and 5–12pp gains in academic prizes, leading technology firms, and founding a company. Ensuring that international students can access leading universities based on merit could yield high social returns.

The "Missing Nobels", with Ruchir Agarwal and Patrick Gaule. Submitted.

Abstract | Paper

Prestigious recognition prizes—like the Nobel Prizes—can shape scientists' career decisions and how science is seen, yet the landscape of such prizes is not well understood. We screen roughly 2,700 international scientific prizes and rank the 99 most prestigious using a novel prestige index. Three patterns stand out. First, half of today's top prizes were first given after 1980 and one-third after 2000, showing that new awards can still rise to prominence. Second, recognition is unevenly distributed across fields: physics, life sciences, and mathematics are heavily recognized relative to field size, while computer science, engineering, psychology, and the social sciences are under-served. Third, incentive design is narrow: only three of the top 99 prizes target early-career scientists, and most lack mechanisms to promote future research. These findings inform the design of recognition systems that better align with contemporary science.

Work in Progress

Frontier AI and Early-Career Scientific Production

Abstract

Has frontier generative AI changed the early-career production of new scientists—what they work on, how varied their work looks, and how soon they produce it—and do those changes differ across fields by their pre-AI workflow exposure to AI? I combine a common-AI-shock design (the 2024 reasoning-models / agentic-chatbots paradigm shift) with differential pre-AI latent exposure across fields, using OpenAlex on a sample of new PhDs from elite institutions. Pre-AI exposure is operationalized using the Massenkoff–McCrory Observed Exposure dataset as a baseline, with a field-level O*NET-to-academic-discipline crosswalk as the primary methodological question to resolve early in the project. The primary outcomes are time-to-first-publication and within-researcher topic concentration; secondary outcomes include collaboration-network breadth and topic reallocation. AI discussion on Twitter, Pangram scores, AI acknowledgements, and GitHub artifacts serve as triangulating signals for realized use; where they disagree, the disagreement characterizes measurement error rather than being averaged away.

Social Media and Scientific Productivity

Abstract

How does social media use affect scientific productivity? While social media may reduce research time, it could also enhance productivity by facilitating remote and interdisciplinary collaborations, especially for researchers with initially small collaboration networks. I use a matched differences-in-differences approach to explore the effect of joining Twitter on scientific productivity across US academia. My sample includes 28,000 research-active academics from top-150 US institutions, in all fields of study, matched to OpenAlex publication histories via custom author-ID disambiguation, gender classification by name, and institution linking. I estimate the causal effect of joining Twitter on the number and quality of publications, general citations, citations to papers published before joining Twitter, the geographic distance and disciplinary breadth of co-authorship networks, and horizontal and vertical job transitions.

The Effects of Testing on Talent Identification, with Kim Kiekens (SPRING-STOF)

Abstract

Field RCT in Flemish primary and secondary schools randomizing population-referenced cognitive screening (the ZOOV+ adaptive assessment) at the classroom level. Unlike most prior screening studies, the design unbundles the ability signal from any downstream gifted-placement or enrichment program: teachers, parents, and students learn where each child falls in the national ability distribution, and we measure whether the signal alone changes outcomes. We collect teacher and parent perceptions of each child's ability before screening, allowing us to measure the perception gap by gender and SES, and we test how revealing the signal affects STEM interest, olympiad participation, student confidence, and track choice.

The Effects of the Rollout of OBMEP (Brazilian Public-School Math Olympiad) on Talent Creation

Women's Cognitive Load and Labor Market Outcomes in Brazil, with Ieda Matavelli, Beatriz Marcoje, and Jamie McCasland

Resting Papers and Projects

Female Pradhan Autonomy

Abstract

Many countries have implemented electoral gender quotas to improve representation in public decision-making. At the same time, verifying whether such policies are successful -- and not just generating figureheads for male family members, for instance -- is hard, especially at the local level. I propose a novel and scalable measure of female leader autonomy for village leaders in India: whether the female leader owns the phone number used to communicate with higher levels of government. I then explore whether past quotas for females lead to more autonomy and whether autonomy predicts public policy outcomes.

Dynamic Coordination with Network Externalities: Procrastination Can Be Efficient

Abstract | Paper

How does present bias affect welfare when agents want to coordinate over time? To answer that, I analyze a dynamic coordination model under quasi-hyperbolic discounting, documenting a novel mechanism through which present bias can be adaptive. The key trade-off for agents in dynamic coordination models is whether to follow a currently-popular standard (receiving substantial network externalities from other current users) or to adopt a new standard with a higher intrinsic quality, hoping that others will follow. Guimarães and Pereira (2016) showed that exponential discounters begin adopting the new standard when its quality is too low to justify the transition costs, since an early adopter does not account for the negative externality caused on those who stay in the old standard. This paper shows that present bias can act as kludge, since it makes agents overweight their individual transition costs, shifting behavior in the same direction that the planner would suggest.