Research

with Laurent Bossavie, Joseph-Simon Georlach, and Caglar Ozden

Abstract: New survey data from Bangladesh provide direct evidence that temporary migration is an effective strategy for workers to accumulate capital and finance self-employment activities back home. We estimate a dynamic model of temporary migration and entrepreneurial investment under constraints, which we use for policy analysis. Lowering of migration costs increases emigration, reduces the age at which workers depart and the duration of their time abroad, which together lead to higher savings and domestic self-employment. Cutting the cost of migration by one half boosts business creation by 8%. Reducing the interest rate for entrepreneurial loans lowers migration and savings repatriation, undercutting the positive effects on business creation at home. This highlights the need to investigate migration and investment jointly, since policies targeting either choice may be enhanced or undercut by endogenous responses in the other.

with Laurent Bossavie, Joseph-Simon Georlach, and Caglar Ozden

Abstract: This paper examines international temporary migration as an intermediary step among aspiring entrepreneurs to accumulate the needed capital when they face credit constraints at home. The analysis is based on a representative dataset of lifetime employment histories of return migrants from Bangladesh. After establishing the credit constraints that potential entrepreneurs face, the paper shows that non-agricultural self-employment rates are significantly higher among returning migrants — over half versus around 20% of non-migrants. Most migrants transition into self-employment by using their savings from abroad as the main source of financing. The paper then offers, for the first time, a detailed account of the financial costs and benefits of international migration. Our findings suggest that temporary migration can contribute to structural transformation of lower-income countries by enabling credit-constrained workers to enter into non-agricultural entrepreneurship.

with Yang Yao and Yue Zhou

Abstract: We study China’s municipal bond markets to find whether mayors’ political human capital —measured by their abilities to grow their cities’ economy—is recognized and priced by market investors. We find an increase of one standard deviation in a mayor’s ability reduces municipal bond yields (investors’ bid yields) by 11.4 basis points, equivalent to saving 38 million yuan during an average mayor’s tenure. Ability has stronger impacts in the first year of a mayor’s tenure, on bonds issued by issuers with lower ratings, and by economically or financially less developed cities.

with Christine Zhenwei Qiang, and L. Colin Xu

Abstract: We investigate how the effects of the business environment depend on whether the measure is de jure or de facto, and how the business environment effects differ by ownership. Four aspects of the business environment are found to be relatively robust by multiple data sources: access to finance, electricity, internet, and human capital. The effects of de jure business environment indicators on firm performances depend on measures of contract enforcement. Foreign-owned firms benefit more from the maintenance of physical safety and ease in obtaining construction permits, and gain competitive advantage in productivity when domestic infrastructure or access to finance is worse.

with Yang Yao, Tianyang Xi, Lixing Li, Feng Wan, Qian Zhang, Songrui Liu, and Shundong Zhang

Abstract: The promotion of officials is one of the key features of China's political system. Using the personal career data provided by the CCER Officials Dataset for officials who served as citiy officials in the period 1995-2017, we describe a full picture of China's promotion system and discovers important trends regarding three key components of promotion, namely, selection, training and incentivizing. While selection on ability and education has always been important, the role of training has been enhanced and incentivizing has become less important. These findings are consistent with the changes of China's development stage and the shift of the party's priorities.

Key words: political selection, sponsored promotion, local government incentives

with Songrui Liu, and Tianyang Xi

Abstract: Abstract:Local governments often have to grapple with multiple policy tasks, ranging from primary objectives such as economic growth and revenue collection to secondary objectives such as public health and environmental protection. This paper argues that the performance and resource allocation on secondary objectives by local leaders are critically shaped by political incentives. We propose a theoretical framework of administrative bidding, which allows local governments to voluntarily respond to the initiatives of the central government and subscribe themselves to a performance competition around secondary policy agendas that would otherwise be underplayed following local governments' routines of policy implementation. Using the evaluation of National Health City by the state council from the 1990s as a case in point, our paper investigates the interplay among local leaders' political incentives and their performance improvement on public health and environment through the campaign for National Health City. The empirical analysis using city-year level data comes down to three main findings. First, cities being rewarded the title of National Health City register a significant enhancement in the performance indicators on public health and environment. Second, cities are most likely to obtain the title of National Health City when their rankings of growth in gross domestic product (GDP)are located in the intermediate range within the province. Third, local leaders presiding over the cities through the campaign for winning National Health City are 8.8 percentage points more likely to be promoted in the long run. Altogether, these results attest to a model of administrative bidding, which helps align the incentives of the central and local governments on secondary policy objectives through a non-conventional, optional, performance competition. The mechanism of administrative bidding is conducive to a balanced approach for development and enhances the efficacy of political selection.

Key words: Administrative bidding, Secondary Tasks, Governance Performance, Political Incentives, National Health City

with Caglar Ozden

Abstract: How do large-scale infrastructure projects shape internal migration and labor mobility in developing countries? India’s Golden Quadrilateral (GQ) highway, with construction starting in 2001 and largely completed by 2006 to connect major economic hubs, provides a unique case to answer this question. Using 2001 and 2011 census data, this paper explores the causal link between reduced travel times and internal migration patterns. We begin by documenting how proximity to the GQ influenced migration rates using a difference-in-differences approach with instrumental variables, estimating a 17% increase in migration for corridors with destination and origin within 50 km to the GQ. Using a modified gravity model, we, then, establish the causal effects of travel time savings on migration flows. Notably, the effect varies by travel distance, as improved infrastructure disproportionately lowers barriers for longer-distance migration. Additional analysis explores demographic heterogeneity and finds that older and lower-educated workers benefit more from the reduced travel time. Overall, we estimate that excluding population growth, 32% of the total increase in migration flows between 2001 and 2011 is directly attributable to the GQ.

Conference presentations:

5th Annual Infra4Dev Conference, Guangzhou, China, May 2025

Asian Economic Development Conference, Beijing, China, July 2025

with Yan Liu

Abstract: This paper offers the first comprehensive, global analysis of generative AI adoption and usage, using novel data sources including website traffic and Google Trends. The paper also examines country-level factors driving the uptake and early impacts of generative artificial intelligence on online activities. As of March 2024, the top 40 generative artificial intelligence tools attract nearly 3 billion visits per month from hundreds of millions of users. ChatGPT alone commanded over 80 percent of the traffic, yet its reach remains less than two percent of Google’s. Generative artificial intelligence users skew young, highly educated, and male, particularly for video generation tools, with usage patterns strongly indicating productivity-related activities. Generative artificial intelligence has achieved unprecedentedly rapid global diffusion, reaching almost all economies worldwide within 16 months of ChatGPT’s release. Strikingly, middle-income economies account for over half of global generative AI traffic, a disproportionately high share relative to their economic size, while low-income economies contribute less than 1 percent. Country level adoption intensity is strongly correlated with the share of youth population, digital infrastructure, English fluency, foreign direct investment inflows, services’ share of GDP, and human capital. Finally, the paper also documents disruptions in online traffic patterns and emphasizes the need for targeted investments in digital infrastructure and skills development to fully realize the potential of artificial intelligence.

Conference presentation:

HKMA-Aof/HKIMR-CEPR GenAI conference, Hongkong, April 2025

with Chen Cheng, Christopher Li, Zitong Zhang, and Weifeng Zhong

Abstract: Economic volatility varies substantially across democracies. We study how political decentralization can be a source of this variation. First, we document a negative correlation between decentralization and the volatility of both economic growth and fiscal policy. We then present a dynamic growth model with a public sector to explain these stylized facts. Contributions by the central government to local public goods is uncertain due to uncertainty about the identity of the winning coalition in a legislature of district representatives. On the other hand, local government spending is more stable. In equilibrium, the decentralization of policy-making powers can mitigate overall policy uncertainty. This implies less volatility in fiscal policy and consequently less volatility in economic growth.

with Laurent Bossavie

Abstract: Despite the importance of temporary migration from South Asia to overseas destinations, the labor market outcomes of return migrants in this region of the world have been understudied. This paper fills this gap by examining systematic differences between the labor market outcomes of return migrants and nonmigrants in Bangladesh, Nepal and Pakistan with nationally-representative survyes which include information on past migration overseas. We use regression analysis and study the following four dimensions: (i) labor force status (ii) sectoral choice (iii) employment type (iv) earnings. The paper finds that return migrants are somewhat less likely to be employed than nonmigrants, which is mainly driven by returnees close to retirement who returned at an older age. As evidenced in other contexts, return migrants are more likely to become self-employed compared to nonmigrants in Bangladesh and Pakistan, and self-employed returnees are also more likely to hire paid employees and to be engaged in non-farm activities, compared to nonmigrant entrepreneurs. Among the minority of returnees who become paid employees, the wage premium earned relative to nonmigrants is positive but small compared to estimates reported in other contexts. The returnee wage premium, however, is larger in the construction sector where the majority of return migrants were employed overseas.

with Maggie Liu and Caglar Ozden

Abstract: How are local labor markets linked and how are wage spillovers transmitted across markets? More importantly, what is the role of the inter-market labor flows in this process? In order to answer these questions, we construct a spatial equilibrium model of endogenous labor mobility where the migration decisions are based on the standard random utility model. We show that the resulting equilibrium distribution of wages across markets can be estimated by an endogenous spatial autoregressive (SAR) specification where the elements of the weight matrix are functions of the bilateral migration flows. Next, we use internal migration and local wage data from India to estimate the predictions of the analytical model. More specifically, we explore the extent of the spatial interdependence of wages while accounting for the endogeneity of migration patterns. Using the estimated parameters, we simulate several counterfactual scenarios. With an exogenous productivity shock at the origin market, we find greater productivity shocks spillover to other markets that have stronger migration linkages with the affected market.

Conference Presentation:

10th South Asia Economic Policy Network Conference on Migration, Nepal, November 2022

with Yan Liu and Jingyun Huang

Abstract: Nearly three years after ChatGPT's launch, the generative AI landscape remains in rapid flux. Using high-frequency website traffic data from Semrush, we track global adoption patterns for the 60 most-visited GenAI tools through mid-2025. Five key findings emerge. First, fierce competition drives continuous innovation: two of 2025's top five tools—DeepSeek and Grok—are new entrants, while development rapidly diversifies into multi-modal capabilities, reasoning, and specialized applications. Second, ChatGPT maintains dominance despite competition, with adoption increasingly youth-driven. Third, GenAI usage has exploded since mid-2024: ChatGPT traffic grew 123% year-over-year, driven by 42% user growth and 50% increases in visits per user, with session duration doubling. Fourth, high-income countries are pulling decisively ahead, creating stark global divides. While 25% of internet users in HICs use ChatGPT, penetration drops to 5.6% in upper-middle-income countries, 3.6% in lower-middle-income countries and just 0.3% in low-income countries. Regression analysis confirms GDP per capita strongly predicts adoption growth. Fifth, localization shapes competitive advantage: non-U.S. tools concentrate heavily in home markets, with Le Chat drawing 69% of traffic from Europe and several Chinese tools exceeding 90% domestic usage. These patterns reveal an AI landscape characterized by intense innovation, persistent market leadership, accelerating growth, and deepening global inequality, underscoring the needs for inclusive policies as GenAI becomes central to economic participation.

with Yan Liu and Shu Yu

Abstract: This paper examines the impact of Generative Artificial Intelligence on U.S. labor demand. Leveraging the near universe of online job posting data from 2018 to early 2025, and the public release of ChatGPT in November 2022 as an exogenous shock, this paper reveals that job postings in GenAI-susceptible occupations plummeted by 18% by early 2025 based on a difference-in-differences approach, with the decline intensifying over time. The impact is especially stark for entry-level and less-skilled roles, as well as in professional services and a few other industries. 

with Nan Sandi

Abstract: This study examines the impact of internet infrastructure expansion on female employment and digital adoption in Thailand. Using nationally representative labor force and economic census data, we find that improved internet access leads to greater labor market participation as well as digital adoption among women, particularly in the service sector. Our results highlight the role of digital connectivity in shaping inclusive labor market outcomes and underscore the importance of infrastructure investments for narrowing gender gaps in employment in emerging economies.

Key Messages

·       Data powers AI: The quantity, quality, and diversity of data are essential to making AI systems effective and adaptable, enabling them to evolve and perform reliably in diverse real-world scenarios.

·        The data divide grows as the AI training data industry booms globally, yet private sector investment in LICs and MICs remains negligible: In 2023, cumulative VC investments in the AI training data industry reached US$32.4 billion, predominantly concentrated in advanced economies—56% in the US, 15% in the EU, and 2% in the UK. China and India contributed 17% and 3%, respectively, with the rest of the world making up the remaining 6%.

·        The dominance of English in text data excludes many countries while opportunities in non-text data are expanding:

o   The language divide affects the usefulness of GenAI models in non-English speaking countries. English dominates high-quality text data due to historical, economic, and technological factors.

o   English is spoken by 19% of the global population, but makes up 44% of global URLs, 56% of open-source datasets on leading AI development platform Hugging Face, and nearly 98% of scientific papers.

o   For non-text data, the language distribution tends to be more diverse, offering opportunities for non-English speaking countries. Only 21% of YouTube videos are estimated to be in English, with Hindi and Spanish accounting for 7.6% and 6.7%, respectively.

·        Market failures highlight the need for government intervention to bridge the data divide:

o   While the private sector is adopting innovative approaches such as synthetic data to fill data gaps, market failures such as reusability, vague property rights, and challenges in valuing data often lead to underinvestment and under-sharing.

o   Governments play a critical role in addressing these gaps by curating and publishing government data, facilitating secure and ethical data exchange, developing and promoting interoperable data formats and taxonomies. Civil societies can contribute to enhance AI data diversity through open-source datasets and community involvement.

·        Open-source AI is proliferating and allows developing countries to adapt global solutions to their local needs:

o   Developing countries could face challenges in adopting AI due to model misalignment with local needs. Open-source AI models can lower the barrier to entry for developing countries, enabling them to adapt and customize AI solutions to local needs without incurring high licensing costs.

o   Open-source models are growing rapidly and narrowing the performance gaps with proprietary models: among the 308 notable AI models published during 2022-June 2025, around half have open weights, and a quarter have both open weights and open-source training code. Open-source models are also increasingly challenging proprietary models in terms of performance.  

with Jieun Choi, Jonathan Marskell, and Anchidtha Roonguthai

Thailand stands at a digital crossroads with bold ambitions to transition from a traditionalmanufacturing-based economy to a dynamic, high-income digital economy. This transformationis not only about adopting new technologies, cutting costs or streamlining operations; it is also aboutfundamentally reimagining how businesses operate, how citizens access essential services, and howthe entire nation innovates, enters new markets and competes on the global stage. By strategicallyinvesting in its digital future through a robust enabling environment, a strong digital infrastructure,regulatory frameworks and digital skills, Thailand can unlock productivity gains, create higher-valuejobs, and build a more resilient and equitable society for all.

with Jieun Choi

The Digital Data Infrastructure Roadmap (DDIR) for 2025–2029 aims to provide strategic recommendations to enhance Thailand’s digital transformation through improved data infrastructure and systems, incorporating international best practices for data governance. The topics for each chapter on data infrastructure, regulation, institutional arrangement, and their usage by government, the private sector and individuals (in social protection) were jointly decided by the Electronic Transactions Development Agency (ETDA) and the World Bank, based on practical needs and reflecting discussions among related agencies.