Research

I study how the party affiliation of elected sheriffs affects traffic stop enforcement in North Carolina. Using a difference-in-differences design, I find that Democratic-to-Republican sheriff turnovers, compared to Democratic-to-Democratic transitions, increase the share of Black drivers in traffic stops by 3.8 percentage points (15.7%). The increase is concentrated in moving-violation stops and reflects both broad changes among incumbent officers and personnel reshuffling. Racial disparities in searches widen within moving-violation stops, while unconditional contraband finding rates and vehicle crash counts remain unchanged. Overall, the results underscore the role of leadership in shaping frontline policing.

We study the impact of partisan leaders on the political composition of law enforcement agencies in the United States using elected sheriffs in North Carolina as a case study. Using a difference-in-differences design, we find that offices shifting from a Democrat to a Republican sheriff experience a 6.8 percentage point (18%) decrease in the Democratic share of sheriff deputies relative to counties experiencing no party turnovers in the sheriff’s party affiliation. This change is driven both by existing Democratic deputies disproportionately leaving the agency and changing their party registration. We do not find significant changes in the race and gender composition of deputies. Decomposing the political party composition along race and gender dimensions, we find that the changes are driven entirely within White officers (84% of the personnel). Changes within male (83% of the personnel) and female officers contribute proportionally to the overall changes.

This paper examines the causal effects of textbook content on individuals' national identity, by exploiting a curriculum reform that introduced a new perspective on Taiwan's history for students entering junior high school after September 1997. Using a repeated nationally representative survey and a regression discontinuity design, we show that students exposed to the new textbooks were more likely to hold exclusive Taiwanese identity rather than dual identity (i.e. Taiwanese and Chinese). The effect was greater for academic track students and those living in neighborhoods where fewer people identify as Taiwanese. In addition, our results suggest that the new curriculum had little impact on people's political preferences related to Taiwan independence. Finally, we find that the probability of reporting as Taiwanese among old textbook readers converges with that of people reading new textbooks in the long run since the perspectives of old textbooks are in conflict with the recent social trends.

Discussion of targeting the poor mostly focuses on the performance of specific methods on the general population, while the performance across groups is rarely examined. This project examines the performance of the proxy-means-testing (PMT) algorithms across internal migrants and locals in Indonesia. Using the same household expenditure sample and household characteristic variables, I reproduce the PMT algorithm adopted by the Indonesian central government. I find that, conditional on the same true household expenditure level, the PMT algorithm over-predicts internal migrants' expenditure compared to locals. Zooming in on the poor households the central government aims to target, whose expenditure falls within the bottom 25% of the population, the PMT algorithm's over-prediction results in an 8 percentage-point increase in exclusion error for migrants compared to locals. Preliminary analysis indicates that among the PMT variables, household head education and occupation, as well as residential neighborhood characteristics, contribute the most to the over-prediction bias. Future work will explore whether the second stage of targeting — discretionary targeting by local village heads — mitigates or worsens the disparities introduced by the algorithm.

Composite scores are widely used to allocate resources and target services, for example, proxy-means tests in social protection programs or risk-assessment scales in health and social services. Existing empirical work focuses on the one-dimensional composite score and estimates average treatment effects at the cutoff. We propose a new approach that exploits the underlying multi-dimensional score construction and estimates conditional treatment effects along the entire classification boundary in the input space. Heterogeneity along this boundary signals that the weights of the input variables are not efficient for targeting. In such cases, reweighting can increase welfare by reallocating treatment toward individuals with higher treatment effects. We illustrate the method using Taiwan's Intimate Partner Violence Danger Assessment (TIPVDA), an 18-item risk scale used by social workers to classify high-risk domestic violence reports. We have secured linked administrative data that include (i) the item-level TIPVDA responses, (ii) social-worker treatment decisions, (iii) health-care utilization from the National Health Insurance system, and (iv) earnings from tax records.

Governments increasingly deploy algorithmic risk scores to help social workers make frontline decisions in health and social services. We study the introduction of an ML-based risk classification system in Taiwan's elder abuse service workflow in 2020. Social workers receive a case-level risk category (high, medium, or low) derived from visit information and administrative data. The risk category is intended to inform follow-up service decisions. We study how this tool shapes frontline decision-making and service allocation. First, we test whether social workers' service provision decisions respond to the algorithm's risk categories. Leveraging the algorithm's score thresholds, we implement a regression discontinuity design to estimate the causal effects of follow-up services on elders' long-term health outcomes. Second, we examine how the introduction of algorithmic risk classification altered patterns of resource allocation across cases and whether the tool's influence differs by worker experience. We have secured linked administrative data that includes (i) the inputs to the ML algorithm, (ii) the algorithm itself and the risk category thresholds, (iii) social worker follow-up decisions, and (iv) health-care utilization from the National Health Insurance system, and we are currently processing the data for analysis.

This paper studies the impact of electoral incentives on frontline public finance arrangements. I examine the electoral cycles of informal tax collection and rice subsidy provision in Indonesian villages. In the election year, lower-expenditure households pay less in informal taxes, and average households receive more subsidized rice. As a result, village heads make the informal tax system more progressive and reduce overall leakage in the subsidy program right before the village head elections. The results suggest that electoral incentives faced by local agents are important in shaping the local public finance system.

Local infrastructure is a critical input to economic development, yet in many low- and middle-income countries it remains scarce due to limited fiscal capacity. This project examines whether expanding local fiscal resources fosters growth, and through which mechanisms, by studying Indonesia's Village Fund program (Dana Desa). The program's allocation formula placed heavy weight on a fixed per-village component, thereby generating substantial quasi-random variation in per-capita transfers. Currently, we analyze at the district level. Exploiting variation in the number of villages per capita, we use a continuous difference-in-differences design to estimate the causal impacts of the transfer on local government spending and the fiscal multipliers on household consumption. We are extending the analysis to examine village-level infrastructure, land use, and deforestation outcomes to trace how fiscal resources translate into production and environmental changes. This project contributes to understanding how expanding fiscal capacity shapes local economic activity and provides one of the first estimates of fiscal multipliers using within-country variation in middle- and low-income countries.

This study examines the impact of residential care subsidies on mortality among aged residential care (ARC) facility residents in New Zealand. All applicants for residential care subsidies must undergo an interRAI Home Care (HC) assessment before entering an ARC facility, providing a unique snapshot of individuals' health status before facility entry. Linking interRAI data with mortality records in the Integrated Data Infrastructure (IDI), I estimate a linear regression model controlling for demographic, socioeconomic, and pre-entry health characteristics captured in the HC assessment. Preliminary findings suggest that subsidy recipients experience significantly lower mortality rates than non-recipients: respectively 5.3, 3.7, 3, and 2.3 percentage points lower (17.1%, 7.5%, 4.8%, and 3.1% compared to the mortality rates among the non-recipients) in the first through fourth years following HC assessment. Future analysis will explore residential care facility choices and differential healthcare utilization between subsidy recipients and non-recipients to shed light on potential mechanisms.

Funded through New Zealand's 2019 Wellbeing Budget, the Integrated Primary Mental Health and Addiction (IPMHA) program is expanding integrated mental health services in GP clinics, yet as of March 2025, 30% of the enrolled population still lacks access. We conduct surveys at the GP clinic level to assess service readiness and integration quality, to document service take-up obstacles, and to estimate costs of service provision and benefits to the enrolled population's health and labor market participation. Findings will shed light on how practice-level integration quality shapes the effectiveness of IPMHA.