Exploring interactions between socioeconomic context and natural hazards on human population displacement

Climate change is intensifying extreme weather events, leading to increased human displacement. While the relationship between weather and population movement is complex and multi-causal, involving a tangled web of interacting factors, understanding this relationship is crucial for effective mitigation and adaptation. Existing research often overlooks the interplay between hazard, exposure, and vulnerability, often employing simplistic linear models. This study addresses these limitations by using data-driven machine learning (ML) to analyze global internal displacement data (2016-2021) from the Internal Displacement Monitoring Centre (IDMC), focusing on low and middle-income countries particularly vulnerable to climate change. The study utilizes ensemble models (random forests and gradient boosting machines) alongside explainable AI (XAI) and causality-based methods to identify key drivers and interactions, providing a more comprehensive understanding of disaster-induced displacement.

Existing literature acknowledges the complex and multi-causal nature of human mobility, highlighting the interaction of environmental, social, and economic factors. However, past research has often been limited by: (1) the use of simplified linear models that fail to capture non-linear relationships; (2) a lack of comprehensive datasets encompassing all three dimensions of disaster risk (hazard, exposure, vulnerability); and (3) data limitations in terms of availability, completeness, and reliability. While some studies focus on international migration, this research centers on internal displacement, recognizing its prevalence and the challenges faced by low- and middle-income countries in coping with climate stressors. The review highlights the need for more sophisticated modeling approaches and high-quality data to accurately assess the complex dynamics of disaster-induced displacement.

This study employs a data-driven approach using machine learning (ML) algorithms, specifically random forests (RFs) and gradient boosting machines (GBMs), to predict the number of new displacements (NDP) associated with floods, storms, and landslides during 2016–2021. The dataset, compiled from IDMC's Global Internal Displacement Database, includes various socioeconomic and environmental covariates at national and disaster-specific levels. These covariates include measures of hazard (precipitation and wind speed), exposure (population, affected area, agricultural land, vegetation index), and vulnerability (Absolute Wealth Index (AWI), education expenditure, human modification index, conflict fatalities). The models were trained and evaluated using bootstrapping techniques to ensure robustness. Explainable AI (XAI) methods, including Shapley values, were used to interpret model predictions and identify the most influential factors. Causal forest algorithms were employed to estimate causal treatment effects and account for potential confounding factors. The analysis focuses on low- and middle-income countries due to data availability constraints associated with the AWI. Data preprocessing involved standardization and log transformation of the target variable (NDP) to handle skewness and outliers.

The ML models (RFs and GBMs) outperformed a linear regression baseline in predicting NDP, demonstrating the importance of non-linear relationships. XAI techniques revealed that precipitation and AWI are consistently among the most influential predictors. Poorer areas (lower AWI) experience significantly higher NDP per disaster, while higher AWI is associated with lower displacement. Areas with high precipitation levels experience more displacement. The affected area is another important factor, likely capturing the impact on infrastructure and livelihoods. Conflict fatalities also show a significant positive association with NDP. While other factors like elevation, agricultural land, and vegetation index play a role, their influence is less pronounced. The interaction effects between AWI, precipitation, and area are highlighted: high precipitation leads to greater NDP in poorer regions, and among regions with similar affected areas, those with lower AWI show greater displacement. Causal forest analysis supports the findings from Shapley values, although statistical significance for individual causal effects was not reached with the current dataset. The analysis reveals the non-linear relationship between NDP and hazard/exposure factors, with potential saturation effects at high precipitation levels and large affected areas.

The findings support the concept of differential vulnerability, showing that the impact of environmental stressors on displacement is strongly mediated by socioeconomic conditions, hazard exposure, and additional factors like conflict. The study's data-driven approach, using ML and XAI, avoids strong assumptions and reveals complex interactions overlooked by traditional models. The significant role of precipitation and AWI highlights the need for targeted interventions in vulnerable regions to reduce displacement risk. The non-linear relationships identified underscore the limitations of simplistic models and the necessity of using advanced techniques to accurately assess the complex dynamics of climate-induced displacement. Future research could explore alternative ways of characterizing the hydro-climatic dimension, incorporate lagged effects, use more accurate geolocated displacement data, and incorporate factors like coping capacity. The limitations of the current study, including data gaps, should be addressed with improved data quality and resolution for more robust results.

This study provides empirical evidence for differential vulnerability to disaster-induced displacement, quantified through data-driven machine learning. The findings confirm the importance of socioeconomic conditions and their interaction with hazard and exposure. This research contributes to a better understanding of the complex interplay among the dimensions of disaster risk, informing more effective mitigation and adaptation strategies. Future research should focus on improving data quality and resolution, exploring alternative modeling approaches and incorporating a broader range of socioeconomic and environmental factors to refine our understanding and promote more effective intervention strategies.

The study's analysis is limited by data availability. The focus on low- and middle-income countries due to AWI data constraints may introduce income-based bias in the global assessment of hazard impacts. Some variables serve as proxies for other underlying factors (e.g., governance quality, disaster response effectiveness) that are difficult to capture fully with existing data. The definition of affected areas might also introduce some bias, as the method used to extract polygons may not perfectly match the actual impacted region. While the study uses advanced techniques, its findings are still based on observed patterns and may not capture the full complexity of individual decision-making processes involved in displacement.