Climate Change and Migration in Central America: Evidence from New Environmental Event Data

Introduction

Climate change, climate variability, and resulting resource scarcity have created deep social challenges (IPCC). In many regions, these challenges undermine agricultural livelihoods and drive migration, which can be categorized as a form of adaptation (Gemenne and Blocher 2027).

To date, subnational and high frequency data on these social and political adaptations has been missing. Indeed, one of the largest challenges to understanding climate impacts on agricultural and social systems is the lack of high-resolution, high-frequency data on how people and governments are responding to climate change (NASEM 2024).

To address this knowledge gap, we have developed the MLEED (Machine Learning for Environmental Event Detection) dataset. We deploy an AI-driven approach that generates original data on climate adaptation at the individual, community, and governmental levels, alongside enhanced data on environmental disasters.

Using fine-tuned large language models (LLMs) applied to 25 million Central American media articles since 2012, we produce a monthly, sub-national dataset that captures13 distinct types of environmental adaptations at an unprecedented scale. 

Combining the MLEED adaptation indicators with climate, disaster, agricultural, and migration data, we are able to better understand climate impacts in the Central American Dry Corridor (CADC), comprised of El Salvador, Guatemala, Honduras, and Nicaragua. Key findings include:

• Our original MLEED data on sudden-onset climate disasters (e.g., floods, cyclones) has important advantages over the EM-DAT data that many climate researchers use.
• Climate variability is impacting agriculture across Central America most directly via the cyclical El Niño–La Niña pattern. El Niño is associated with reduced agricultural productivity and higher prices for key agricultural commodities, particularly in the CADC.
• Slow-onset disasters (e.g., drought) are associated with lower levels of international migration in subsequent months. Alternatively, sudden-onset disasters are associated with higher international migration in later months.
• Higher prices for key agricultural commodities are associated with higher international migration in subsequent months. We interpret this as an income effect whereby higher agricultural incomes are needed to fund expensive migratory trips.
• Climate disasters are sometimes associated with subsequent violence reflecting resource competition. Environmentally motivated violence is concentrated in areas where agriculture is the dominant mode of production and is associated with increased international migration in later months.

Generating the Novel MLEED Dataset

Kleinman support has enabled us to create the original MLEED dataset, which provides fine-grained data on human adaptations to climate change at unprecedented temporal and geographic scales. We define adaptation as any effort by individuals, social groups, or governments to respond to the negative effects of climate change—for example, prioritizing vulnerable populations in social safety nets or promoting the use of more drought-resilient seeds. Table 1 lists the 13 adaptations captured in the MLEED dataset, along with three types of environmental disasters: sudden-onset, slow-onset, and human-induced.  

To generate MLEED data, we scrape approximately 25 million online news articles published in several languages between 2012 and 2024 across Central America, collected from dozens of international, regional, and local news outlets as well as 19 specialized environmental sources. Using a fine-tuned LLM (ModernBERT), we classify individual articles into different disaster and adaptation categories.  The model performs comparably to human coders, with most discrepancies occurring in areas where human coders themselves often disagree.¹ To add subnational granularity, we develop a new geoparsing protocol using LLMs to identify all geographic locations mentioned in each article, which are then mapped to their respective country, region/state, and district/city levels.

While media-based data have well-known limitations, such as reporting bias and uneven coverage (Daphi et al. 2025; Earl et al. 2004), they remain the most consistent and comprehensive source for monitoring events across a wide range of temporal and spatial contexts.

While platforms like radio and social media are important, they often rely on content originally produced by traditional news outlets, which are generally more trusted (Fotopoulos 2023) and provide more comprehensive coverage of political events (Lee 2002; Schäfer and Schemer 2024). Our background technical paper explains how we overcome many common shortcomings in LLM-generated event data, including duplicate stories, date parsing, as well as limitations of the approach.

Advantages of MLEED Disaster Data

We compare MLEED’s monthly counts of sudden-onset environmental disasters with events from EM-DAT, a global disaster database compiled from sources including UN agencies, NGOs, insurance companies, and the media (see Figure 1, where MLEED disasters appear as the black time series and EM-DAT events as point overlays). The close alignment between peaks in MLEED reporting and EM-DAT events provides strong external validation for MLEED’s detection of sudden-onset environmental disasters.

At the same time, MLEED offers important advantages. It captures a larger number of disasters (i.e. spikes not reflected in EM-DAT) and the volume of MLEED reports can provide a proxy for disaster severity, a feature largely missing from EM-DAT.

MLEED also improves coverage in data-scarce regions by overcoming EM-DAT’s reliance on insurance data and national disaster monitoring systems, which are often limited in the Global South, and by including subnational geographic detail. Together, MLEED’s original measures of disaster events represent a major advance for research on climate and agricultural systems.

Climate, Agricultural Productivity, and Food Prices

The CADC is among the region’s most climate-vulnerable areas, marked by high temperatures, variable rainfall, and distinct wet and dry seasons (Imbach et al. 2017).  Rainfall peaks in spring and late summer/early fall, separated by a mid-summer dry season known as the canícula (Anderson et al. 2019). This pattern is influenced by both natural El Niño–Southern Oscillation (ENSO) variability and climate change (Hidalgo et al. 2019).

El Niño events typically bring hotter, drier conditions and a more intense canícula, reducing precipitation and disrupting crop development. These effects appear in temperature, rainfall, and drought indicators such as the Standardized Precipitation-Evapotranspiration Index (SPEI), which is strongly negatively correlated with ENSO. In contrast, La Niña events bring cooler temperatures, above-average rainfall, and more frequent floods and storms.

Vegetation data mirror these trends, with agricultural and ecosystem productivity tending to decline during El Niño years and recover during La Niña periods (see Figure 2 below). Long-term climate change exacerbates these risks by raising baseline temperatures, intensifying drought and extreme rainfall, and altering seasonal cycles, which may deepen the canícula and delay the second rainy season (Maurer et al. 2017).

Agriculture in the CADC is vulnerable to climate variability. Farming is dominated by rainfed subsistence crops and coffee for export. Productivity is constrained by limited access to drought-resistant seeds and irrigation, insecure land tenure, and widespread poverty.

During El Niño events, reduced rainfall is typically associated with lower crop yields, driving up food prices and threatening farmer livelihoods. ENSO cycles are closely linked to acute malnutrition, as prolonged drought during El Niño years coincides with volatile market prices for staple crops such as beans (Beveridge et al. 2019).

As shown in Figure 3 below, El Niño periods consistently correspond to widespread drought and higher agricultural prices. In short, as climate shocks reduce productivity, market prices rise, which highlights the importance of farmer adaptation to these pressures

Climate and Migration in Central America

Climate-related disruptions, particularly drought, are a major driver of displacement and migration, reducing income and food security for rural households reliant on subsistence farming. When agricultural conditions deteriorate, migration often becomes a coping strategy (Tuholske et al. 2024). This is especially evident in the CADC, where households face growing climate stresses at the same time as structural barriers limit adaptation (Daoust and Selby 2024). 

Repeated shocks such as the 2014–2016 droughts and Hurricanes Eta and Iota coincided with severe agricultural losses, deepening food insecurity and forcing families to adopt extreme coping strategies, including selling land or livestock. For some, international migration, mainly towards the U.S, becomes the only viable option (Pons 2021).

These pressures are worsened by limited access to timely, accurate weather data, and mismatches between farmer perceptions and actual rainfall, which lead to poorly informed planting and harvesting decisions (Anderson et al. 2025). Migration is therefore driven not only by physical shocks but also by households’ ability to respond to perceived risks.

Internal migration, particularly from rural to urban areas, is the most common form in the CADC, but international migration remains significant.² Wealthier and better-educated households are more likely to be represented among international migrants due to the high costs involved (Klaiber 2014), while the most vulnerable are often unable to leave despite mounting pressures (NASEM 2024).

Beyond climate, high levels of violence in the Northern Triangle countries (Honduras, El Salvador, and Guatemala) further contribute to migration, making these some of the most dangerous areas outside active warzones (Huber et al. 2023). As a result, people from this region now represent the fastest-growing segment of unauthorized immigration to the United States (Reichman 2022).

Migration and Slow vs. Sudden-Onset Disasters

Given this context, we further examine the relationships between climate variability, agricultural prices, and migration. We merge original data from the U.S. Department of Homeland Security (DHS) on monthly encounters between border agents and migrants at the southwest U.S. border with the SPEI drought indicator, highlighting the complex spatial and temporal patterns linking drought and migration (see Figure 4 below).

To investigate these patterns systematically, we use Elastic Net regression models to identify climate, agricultural, and MLEED variables most strongly associated with U.S. Border Patrol encounters.³ Figures 5 and 6 below show which factors are linked to monthly spikes in southwest border encounters, based on one model that includes slow-onset disasters like drought (measured by unusually dry conditions over the preceding six months), and another model that includes sudden-onset disasters like storms or floods.

In the first model, we find that a six-month lag best captures migration responses to slow-onset climate shocks, indicating that drought impacts in CADC countries tend to appear at the U.S. border about six months later.  We find that drought is negatively associated with migration surges, possibly because prolonged drought limits people’s ability to move in the short- and medium-term.⁴

For agricultural outcomes, higher prices for staple crops (beans and maize) are positively associated with migration surges at the U.S. border. One possible interpretation for this relationship is an income effect, where higher agricultural earnings may enable more households to finance costly international migration.

We also find that environmentally motivated violence correlates with higher migratory surges, which aligns with the fact that such violence often occurs in agricultural areas most affected by climate change and poorly regulated industrial agriculture. Additional notable results include associations with U.S. administrations (e.g., more encounters observed under the Trump and Biden administrations, and fewer under the Obama administration).

Regarding the second model, we find a positive relationship between the MLEED measure of sudden-onset disasters and border encounters. Similar to the previous model, a six-month lag provides the strongest relationship, suggesting that people might require several months to mobilize the necessary resources to attempt an international move. Other explanatory variables show similar relationships to the previous model.

Conclusion

Overall, our results highlight how different types of climate shocks are associated with varying patterns of mobility over time. Slow-onset disasters like drought appear linked to reduced movement, whereas sudden-onset disasters are associated with increased migration toward the U.S. in subsequent months.

While our analysis focuses on identifying patterns rather than establishing causality, the findings still offer several policy insights:

1. Strengthening drought monitoring and early warning systems could help anticipate migration pressures several months in advance.
2. Expanding access to climate information, drought-resistant seeds, and social protection in rural areas of the CADC would improve households’ capacity to adapt in place.
3. Integrating climate, agricultural, and migration data could support coordinated regional planning to manage displacement more effectively.

Future analysis will explore the causal mechanisms underlying these relationships between climate shocks, agricultural markets, violence, and subsequent migration. We also plan to disaggregate the analysis to the subnational level to identify more localized patterns.

Better understanding these relationships is essential, especially as climate-induced displacement is often overlooked in global immigration debates. Recognizing how climate change and variability drive migration in the CADC is key to creating long-term environmental and migration policy solutions.

Reference

Anderson, Talia G., Kevin J. Anchukaitis, Diego Pons, and Matthew Taylor. 2019. “Multiscale Trends and Precipitation Extremes in the Central American Midsummer Drought.” Environmental Research Letters 14, no. 12: 124016. https://doi.org/10.1088/1748-9326/ab5023.

Anderson, Talia G., Diego Pons, Matthew Taylor, Antonia Xuruc, Hugo H. Rodríguez Salvatierra, Zack Guido, Jonathan A. Sullivan, Diana Liverman, and Kevin J. Anchukaitis. 2025. “Complexity and Mediating Factors in Farmers’ Climate Perceptions and Agricultural Adaptation Strategies in the Guatemalan Dry Corridor.” Climatic Change 178, no. 7: 139. https://doi.org/10.21203/rs.3.rs-4824595/v1.

Beveridge, Louise, Stephen Whitfield, Simon Fraval, Mark van Wijk, Jacob van Etten, Leida Mercado, James Hammond, Luz Davila Cortez, Jose Gabriel Suchini, and Andrew Challinor. 2019. “Experiences and Drivers of Food Insecurity in Guatemala’s Dry Corridor: Insights from the Integration of Ethnographic and Household Survey Data.” Frontiers in Sustainable Food Systems 3: 65. https://doi.org/10.3389/fsufs.2019.00065.

Daoust, Gabrielle, and Jan Selby. 2024. “Climate Change and Migration: A Review and New Framework for Analysis.” Wiley Interdisciplinary Reviews: Climate Change 15, no. 4: e886. https://doi.org/10.1002/wcc.886

Daphi, Priska, Jan Matti Dollbaum, Sebastian Haunss, and Larissa Meier. 2025. “Local Protest Event Analysis: Providing a More Comprehensive Picture?” West European Politics 48, no. 2: 449-463. https://doi.org/10.1080/01402382.2024.2363709.

Denny, Elaine K., David Dow, Gabriella Levy, and Mateo Villamizar-Chaparro. 2022. Extortion and Civic Engagement among Guatemalan Deportees. World Bank. https://documents1.worldbank.org/curated/en/099730304262235605/pdf/IDU02026f74207fd304e610a1160b8141ae76d28.pdf.

Earl, Jennifer, Andrew Martin, John D. McCarthy, and Sarah A. Soule. 2004. “The Use of Newspaper Data in the Study of Collective Action.” Annu. Rev. Sociol. 30, no. 1: 65-80. https://doi.org/10.1146/annurev.soc.30.012703.110603.

Food and Agriculture Organization (FAO). 2019. “Adverse Climate Events in the Central American Dry Corridor Leave 1.4 million People in Need of Urgent Food Assistance.” FAO News Release. https://www.fao.org/americas/news/news-detail/Adverse-climate-events-in-the-Central-American-Dry-Corridor-leave-1-4-million-people-in-need-of-urgent-food-assistance/.

Fotopoulos, Stergios. 2023. “Traditional Media Versus New Media: Between Trust and Use.” European View 22, no. 2: 277-286. https://doi.org/10.1177/17816858231204738.

Gemenne, François, and Julia Blocher. 2017. “How Can Migration Serve Adaptation to Climate Change? Challenges to Fleshing Out a Policy Ideal.” The Geographical Journal 183, no. 4: 336-347. https://doi.org/10.1111/geoj.12205.

Hidalgo, Hugo G., Eric J. Alfaro, Jorge A. Amador, and Álvaro Bastidas. 2019. “Precursors of Quasi-Decadal Dry-Spells in the Central America Dry Corridor.” Climate Dynamics 53, no. 3: 1307-1322. https://doi.org/10.1007/s00382-019-04638-y.

Huber, Jona, Ignacio Madurga-Lopez, Una Murray, Peter C. McKeown, Grazia Pacillo, Peter Laderach, and Charles Spillane. 2023. “Climate-Related Migration and the Climate-Security-M Nexus in the Central American Dry Corridor.” Climatic Change 176, no. 6: 79. https://doi.org/10.1007/s10584-023-03549-6.

Imbach, Pablo, Megan Beardsley, Claudia Bouroncle, Claudia Medellin, Peter Läderach, Hugo Hidalgo, Eric Alfaro et al. 2017. “Climate Change, Ecosystems and Smallholder Agriculture in Central America: An Introduction to the Special Issue.” Climatic Change 141, no. 1: 1-12.https://doi.org/10.1007/s10584-017-1920-5.

Intergovernmental Panel on Climate Change (IPCC). n.d. https://www.ipcc.ch/.

Klaiber, H. Allen. 2014. “Migration and Household Adaptation to Climate: A Review of Empirical Research.” Energy Economics 46: 539-547. https://doi.org/10.1016/j.eneco.2014.04.001.

Lee, Hye-Kyung. 2022. “Rethinking Creativity: Creative Industries, AI and Everyday Creativity.” Media, Culture & Society 44, no. 3: 601-612. https://doi.org/10.1177/01634437221077009.

Maurer, Edwin P., Nicholas Roby, Iris T. Stewart-Frey, and Christopher M. Bacon. 2017. “Projected Twenty-First-Century Changes in the Central American Mid-Summer Drought using Statistically Downscaled Climate Projections.” Regional Environmental Change 17, no. 8: 2421-2432. https://doi.org/10.1007/s10113-017-1177-6.

National Academies of Sciences, Engineering, and Medicine (NASEM). 2024. “Climate Change and Human Migration: An Earth Systems Science Perspective: Proceedings of a Workshop.” https://doi.org/10.17226/27930.

Pons, Diego. 2021. “Climate Extremes, Food Insecurity, and Migration in Central America: A Complicated Nexus.” Migration Policy Institute 18. https://www.migrationpolicy.org/article/climate-food-insecurity-migration-central-america-guatemala. 

Reichman, Daniel R. 2022. “Putting Climate-Induced Migration in Context: The Case of Honduran Migration to the USA.” Regional Environmental Change 22, no. 3: 91. https://doi.org/10.1007/s10113-022-01946-8.

Schäfer, Svenja, and Christian Schemer. 2024. “Informed Participation? An Investigation of the Relationship Between Exposure to Different News Channels and Participation Mediated through Actual and Perceived Knowledge.” Frontiers in Psychology 14: 1251379. https://doi.org/10.3389/fpsyg.2023.1251379.

Tuholske, Cascade, Maria Agustina Di Landro, Weston Anderson, Robbin Jan van Duijne, and Alex de Sherbinin. 2024. “A Framework to Link Climate Change, Food Security, and Migration: Unpacking the Agricultural Pathway.” Population and Environment 46, no. 1: 8. https://doi.org/10.1007/s11111-024-00446-7.

1. The model’s F1-score, a common measure of accuracy that combines precision and recall, is similar to the human coders. [↩]
2. This is clear in everything from systematic survey of deportees from the U.S. (Denny et al. 2022) to the long-running EMIF-Sur surveys carried out at Mexico’s southern border. [↩]
3. Elastic net regression looks for patterns in data to make predictions. It identifies which factors or variables are most strongly associated with an outcome, but it does not prove that one factor causes another. In other words, the model can tell us which variables are useful for predicting what will happen, but it cannot tell us why those relationships exist or whether changing one factor would change the outcome. [↩]
4. Drought’s impact on migration likely works through agricultural productivity and commodity prices. In subsequent work we will more carefully examine these pathways. [↩]