Research

This paper presents a novel multi-treatment field experiment that investigates the spillover effects of targeted supplementary computer-assisted learning (CAL) and traditional paper-pencil workbook education interventions among 130 boarding schools in rural China. We also discuss the possible channels by which programs may have spillover effects on non-targeted peers' academic outcomes. We find that the paper-pencil workbook program has a negative spillover effect on untargeted non-boarding students’ school performance, but no spillover effect is detected in the CAL group. Our network interference results suggest that the negative spillover effects of the workbook program most strongly affect non-boarding students who have close-boarding peers in the same classroom. 

[Full Draft][Poster at ASSA][Google AI Podcast]

Presented at 2025 AAEA & WAEA Joint Annual Meeting; SEEDEC 2025; 2025 Conference of the International Association for Applied Econometrics; 2025 South East Exchange of Development Studies (SEEDS); 2025 ASSA; 2024 Global GLO-JOPE Conference-North American Job Market Session (Live);  2024 APPAM Fall Research Conference; 2024 CES North America Annual Conference (CES); 2024 Stanford Center on China’s Economy and Institutions (SSC) Young Researcher Workshop.

This paper provides some of the first experimental evidence of the applicability of text messaging technology to improve parenting knowledge, promote stimulating parenting practices, and raise child development outcomes at ages 0-3 in low-income, under-resourced communities. We replicated and expanded a text messaging-based parenting intervention (Tips-by-Text) that has been shown to be effective in the U.S. but has not been tested in low- and middle-income countries (LMICs), where misconceptions about parenting are common and awareness of the importance of stimulating parenting practices is low. Overall, our results show substantial, positive impacts of Tips-by-Text on parenting knowledge (ITT = 0.222 SD, p < 0.01) and some important stimulating parenting practices (including maternal time investment in play activities; counting, drawing, and naming things with the child; and reading or looking at picture books with the child). While the average treatment effects on other parenting practices and on early childhood development outcomes are statistically insignificant in the sample overall, we found large heterogeneities in the treatment effects that were consistent with three behavioral economics concepts: lack of information, inattention, and motivated cognition. 

[Full Draft][Google AI Podcast]

Presented at 2022 Agricultural & Applied Economics Association Annual Conference; 2022–2023 China Education Finance Research Forum for Young Scholars.

This study examines the causal relationship between 3G mobile network coverage, fertility decisions, and infant mortality in Nigeria. Using geo-referenced data from Nigerian Demographic and Health Surveys (2013-2018) matched with mobile coverage information, we implement two-way fixed effects for fertility analysis and a sample selection model for infant mortality. Results show that increased 3G coverage significantly reduces birth rates, with effects approximately twice as strong for adolescent women (15-19 years). The spatial gradient of effects—stronger at closer proximity (20km) and diminishing with distance (40km)—supports a causal interpretation. For infant mortality, our selection-corrected models reveal no statistically significant direct relationship between 3G coverage and child survival outcomes after accounting for fertility decisions. These findings indicate that mobile connectivity primarily influences demographic outcomes through fertility decisions rather than through direct effects on child survival, suggesting telecommunications infrastructure investments may yield substantial demographic benefits primarily through reduced fertility rates, particularly among adolescents.

[Google AI Podcast]

Presented at 2025 AAEA & WAEA Joint Annual Meeting.

This study investigates the causal relationship between peer effects and academic performance in classroom microenvironments, addressing the challenge of endogeneity in peer group selection. Using data from 2,956 primary school students in rural China, we employ network theory to model study group structures and an instrumental variable approach to control for selection bias. Our findings reveal that study groups significantly enhance student achievement by 0.11 standard deviations, with lower-ranked students benefiting more from this effect. Intrinsic motivation emerges as the primary channel through which these peer effects operate. Notably, high-achieving students show no substantial changes in learning outcomes or behavior due to study group participation. The study also uncovers that peer effects are more pronounced in highly cohesive study groups and among male students. These results contribute to our understanding of how effectively assigned peer groups can optimize academic performance and facilitate human capital accumulation, offering valuable insights for educational policy and practice in leveraging peer relationships to enhance learning outcomes. 

[Full Draft][Slides at NEUDC][Google AI Podcast]

Presented at 2024 Agricultural & Applied Economics Associationultural & Applied Economics Association; 2024 Pacific Conference for Development Economics (PacDev); 2023 The North East Universities Development Consortium (NEUDC) Conference; 2021 Chinese Economics Society Annual Meeting; 2021 Chinese Economics Society Annual Meeting. 

Early child development is important to human capital accumulation. Current studies find a positive effect of maternal migration on their early children development due to income increase and parenting knowledge improvements after migration. However, most of those studies lack detailed information on parental migration history. This is first ever study recording maternal monthly migration status. We conducted a survey of 781 households with children aged 1-23 months in rural China and constructed a dose-response estimates using a “generalized propensity score.” This paper uses propensity score matching to identify causal effects of parental migration on early childhood development. Results suggest a statistically significant positive effect of maternal migration on children development outcomes and mental health levels who migrates with them. However, there is no statistically significant effect of maternal migration on left-behind child.

[Poster at AAEA]

Presented at 2023 Southern Agricultural Economics Association Annual Meeting; 2022 Agricultural & Applied Economics Association Annual Conference; WEAI 97th Annual Meeting.

This randomized controlled trial evaluated an early childhood development (ECD) intervention in rural Yunnan Province, China, targeting children aged 6-24 months and their caregivers through biweekly home visits and monthly group activities. Among 1,024 participating children, after one year, the treatment group showed significant improvements across multiple developmental domains compared to controls: cognitive (0.26 SD), language (0.16 SD), motor skills (0.25 SD), and social-emotional development (0.25 SD). The intervention also enhanced caregiving practices, with increased engagement in stimulating activities and improved home learning environments. Caregivers in the treatment group demonstrated better mental health outcomes, including reduced depression and anxiety symptoms. These findings support policy recommendations for strengthening ECD services in rural China through government-led, multi-stakeholder approaches and professional development standards for service providers.

This paper details the development of a Convolu- tional Neural Network (CNN), a shift invariant artificial neural network (SIANN) utilizing convolution operations instead of matrix multiplication, with the goal of classifying ten unique traffic signs. A well-balanced data set of photos with equivalent resolution was used to train and validate the neural network to determine appropriate hyperparameters for optimal perfor- mance, accurate classification greater than ninety percent. The CNN was developed using packages found within the Tensorflow library in Python, including convolution, pooling, and dense layers. Additionally, this paper documents specific experiments conducted during the design and training which led to the final architecture of the neural network. The CNN will be shown to have an accuracy of greater than ninety-four (94) percent during training and validation.

[Full Draft][Code in Github]