Working papers

• "Language, Identity, and Ethnic Cohesion in Schools" [Link] (with Pepi Pandiloski) Job Market Paper
  
  Abstract

  Distinct group identities can create barriers for trade, investment in public goods, and social learning. This paper studies how integration impacts productivity and cohesion in settings with broad linguistic diversity, where language differences may deepen the divides created by group identities. We randomly assign high school students to same- or different-ethnicity partners to play a cooperative video game in Macedonia, where students are segregated by ethnicity. We measure productivity using video game scores and assess social cohesion with two measures: the willingness to pay for a same-ethnicity partner (via an adaptive choice experiment) and ingroup bias (via a dictator game). We find that inter-ethnic contact improves social cohesion without affecting productivity. To examine the role of language, we exploit random variation in the language skills overlap within pairs. We find that shared language skills are associated with greater inter-ethnic cohesion gains and lower non-pecuniary costs. These results reflect both a match effect (i.e., sharing a language is beneficial) and a correlation between language skills and positive potential outcomes. Additionally, cross-ethnic pairs match on one of the partner's home languages or a third neutral language. Cohesion gains are largest for participants who match on their own home language; if anything, matching on a neutral language is worse than matching on the partner's home language.
  
  
• "The Costs and Benefits of Integration: The Case of Social Learning" [Link] [Code] (with Pepi Pandiloski)
  
  Abstract

  We study how the integration of different groups affects social learning. We designed a social learning game and implemented it in ethnically segregated schools in Macedonia. In the game, 6 players (3 students of 2 distinct ethnicities) get noisy private signals and need to learn the truth with their peers. This game allows us to control the learning environment by randomizing three dimensions: private initial signals, the network on which players observe each other, and bias, which we manipulate by randomizing the number of high-bias players in the 6-person group. We find nuanced effects of integration on learning. First, we find that bias increases the costs of integration on social learning. In low-bias groups, students learn more in the integrated network than in the segregated network, but the same is not true in the high-bias groups. Second, we find that complementary information increases the benefits of integration on social learning. Specifically, integration leads to more learning than segregation only in the games in which one ethnicity has more correct initial signals than the other ethnicity. Students learn from co-ethnics twice as much as they learn from cross-ethnics, and their estimated bias is smaller if they perceive the cross-ethnicity as more educated. We close the paper with exercises that simulate social learning under a counterfactual policy of one integrated classroom per cohort.
  
  
• "Costly Withdrawals Reduce Future College Going for Low-Income Students: Evidence from Return of Title IV Funds" [Link] (with Elizabeth Bell, Oded Gurantz, and Dennis Kramer) Revise & Resubmit, Economics of Education Review
  
  Abstract

  Governments must strike a balance between promoting access to financial aid while at the same time remaining good stewards of taxpayer funds by preventing fraudulent access. This paper focuses on one of the largest-scale and most consequential policies determining whether students maintain access to Title IV aid, the “Return of Title IV” funds policy, referred to as R2T4. Students receiving Title IV aid who withdraw from college before completing the academic term are subject to an R2T4 calculation that could require the student or college to pay back any unearned Title IV funds to the federal government. We estimate the causal impacts of the R2T4 policy on student outcomes in a regression discontinuity design, leveraging a cutoff in the formula that determines whether a student or their college is required to return aid. We find that students at our threshold, who earn 60 percent of the federal aid to which they were entitled, must return $1,600 on average. Such debt makes students almost four percentage points less likely to re-enroll in college the following year and 2.6 percentage points within 4 years. These results are driven by students in the bottom half of the R2T4 income distribution who experience persistent enrollment declines of roughly 5.5 percentage points. Our findings add to a growing body of literature revealing the detrimental impacts of complex administrative processes on student outcomes, particularly for students from marginalized communities interacting with federal policies.
  
  
• "How Local is Local Development? Evidence from Casinos." [Link] (with Jordan Rosenthal-Kay)
  Revise & Resubmit, Regional Science and Urban Economics
  
  Abstract

  One rationale for place-based policy is that local development produces positive productivity spillovers. We examine the employment spillovers from a large local development project: opening a casino. Comparing employment in neighborhoods that won a casino license to runner-up neighborhoods that narrowly lost, we find that casinos create jobs in their immediate vicinity. However, we estimate net job losses overall when considering the broader neighborhood. Employment gains concentrate in the leisure and hospitality industry, suggesting spillovers are industry-specific or are driven by demand-side forces like trip-chaining. We develop theory to show that our estimates imply a rapid spatial decay of productivity spillovers.
  
  

Work in progress

• "The Distributional Impact of Official Communications: Evidence from Two Massive FAFSA Renewal Campaigns" (with Salman Khan and Dennis Kramer)
  Presented at APPAM 2024 & AEFP 2025
  
  Abstract

  We test the efficacy of the U.S. Department of Education's (DOE) communication campaign on Free Application for Federal Student Aid (FAFSA) renewal and explore what types of students respond to a reminder. Who is marginal to completing additional paperwork? Who is marginal to re-enrollment? We address these questions using the results from two massive national-scale RCTs which include roughly 10 million people each. In 2020 and 2023, the DOE randomized the recipients of a typical communication campaign. Unlike other large-scale campaigns, the DOE campaigns move a small, but significant proportion of students to re-submit FAFSA and re-enroll in college. Moreover, while we observe limited variation in who resubmits FAFSA, we see heterogeneity in who re-enrolls, prompting questions about the distributional consequences of the campaigns.
  
  
• "Faking the Grade: Teacher-Student Social Ties and the Evaluation of Ability" (with Pepi Pandiloski)
  
  
• "Differential Impacts of Small Reductions in Pell Grants: Evidence from Two Federal Policy Changes" (with Salman Khan and Dennis Kramer)
  
  

Pre-PhD Research

• "Aequitas - A bias and fairness audit toolkit" [Link] [Code] (with Pedro Saleiro, Benedict Kuester, Loren Hinkson, Jesse London, Abby Stevens, Kit T Rodolfa, and Rayid Ghani) arXiv preprint (2018) arXiv:1811.05577
  
  Abstract

  Recent work has raised concerns on the risk of unintended bias in AI systems being used nowadays that can affect individuals unfairly based on race, gender or religion, among other possible characteristics. While a lot of bias metrics and fairness definitions have been proposed in recent years, there is no consensus on which metric/definition should be used and there are very few available resources to operationalize them. Therefore, despite recent awareness, auditing for bias and fairness when developing and deploying AI systems is not yet a standard practice. We present Aequitas, an open source bias and fairness audit toolkit that was released in 2018 and it is an intuitive and easy to use addition to the machine learning workflow, enabling users to seamlessly test models for several bias and fairness metrics in relation to multiple population sub-groups. Aequitas facilitates informed and equitable decisions around developing and deploying algorithmic decision making systems for both data scientists, machine learning researchers and policymakers
  
  
• "Physical ecology of hypolithic communities in the central Namib Desert: The role of fog, rain, rock habitat, and light" [Link] (with Kimberley A. Warren-Rhodes, Christopher P. McKay, Linda Ng Boyle, Michael R. Wing, Elsita M. Kiekebusch, Don A. Cowan, Francesca Stomeo, Stephen B. Pointing, Kudzai F. Kaseke, Frank Eckardt, Joh R. Henschel, Mary Seely, and Kevin L. Rhodes) Journal of Geophysical Research: Biogeosciences, 118 (2013), pp. 1451-1460.