Logo: to the web site of Uppsala University

In 2023, the United Nations High Commissioner for Refugees reported over 110 million displaced individuals globally, many in regions facing extreme weather and violence. Here we examine how these crises interact to shape household mobility in Bangladesh. Using data linking local conflict events, natural hazards, and household characteristics from 2011 to 2018, we apply machine learning models to capture complex, non-linear relationships between these risks. We find that combining conflict and hazard information improves predictions of household mobility. While exposure to violence or disasters increases mobility, households with remittances are more likely to move, whereas those with loans often remain. Interactions, such as between one-sided violence and landslides, further amplify movement, highlighting the importance of understanding how multiple stressors jointly influence household decisions.

Electoral violence remains a significant challenge worldwide. It not only threatens to undermine the legitimacy and fairness of electoral outcomes, but often has serious repercussions on political stability broadly. The ability to prevent electoral violence is critical for safeguarding democracy and ensuring peaceful transitions of political power. Predicting which elections are at risk of violence is a crucial step in effective prevention. In this study, we develop a set of machine-learning models to forecast the likelihood of electoral violence worldwide. Using diverse data sources, which include economic indicators, the history of electoral violence, political instability, and digital vulnerability, we predict the risk of electoral violence on a scale ranging from no violence to severe violence. Our final forecasts are produced by combining constituent models into an ensemble using a genetic algorithm. Out-of-sample evaluation of the system shows that the final model accurately distinguishes between different levels of risk. After validating our system on historical data, we generate out-of-sample probabilistic forecasts for national-level elections in 2025 and 2026. This research contributes to the field of political violence prediction by providing a medium-term data-driven forecasting tool for electoral violence.

Protest is a low-intensity form of political conflict that can precipitate intrastate armed conflict. Data on protests should therefore be informative in systems that provide early warnings of armed conflict. However, since most protests do not escalate to armed conflict, we first need theory to inform our prediction models. We identify three theoretical explanations relating to protest-repression dynamics, political institutions and economic development as the basis for our models. Based on theory, we operationalize nine models and leverage the political Violence Early Warning System (ViEWS) to generate subnational forecasts for intrastate armed conflict in Africa. Results show that protest data substantially improves conflict incidence and onset predictions compared to baseline models that account for conflict history. Moreover, the results underline the centrality of theory for conflict forecasting: our theoretically informed protest models outperform naive models that treat all protests equally.

This dissertation contributes to the literature on conflict-related mobility by examining how and under what conditions political violence shapes mobility. Each of the four essays approaches this question from a different angle by analysing how the temporal, spatial, and actor-related characteristics of political violence shape mobility decisions and outcomes.  Essay I shows that disaggregated temporal dynamics (persistence, escalation, volatility), spatial patterns (scope, location), and actor-related characteristics (relative strength, fragmentation, government targeting of civilians) substantially improve one-year-ahead predictions of refugee and asylum seeker outflows, especially for large outflows. Essay II develops a theoretical and empirical framework distinguishing observed violence from violence risk, modelled via spatio-temporal and network-weighted proximity, and provides evidence of anticipatory mobility in Somalia, suggesting that civilians move in response to violence occurring in other districts. Essay III quantifies the longer-term demographic consequences of organised violence by using a machine learning-based hurdle model with counterfactual simulations to estimate how state-based and non-state conflict reshape five-year net migration across Africa and the Middle East; the analysis indicates that state-based conflict is associated with an estimated 2.2 million net migration differences between 2015 and 2020 relative to a counterfactual peace scenario, corresponding to about 3.9% of total absolute predicted net migration. Essay IV examines compound pressures at the household level in Bangladesh and shows, using interpretable machine learning, that models incorporating both political violence and natural hazards outperform simpler specifications, with effects conditioned by household resources in ways that generate both mobility and immobility. The dissertation advances research on conflict-related mobility theoretically by demonstrating that mobility is shaped by how different forms of violence unfold and are perceived, rather than by their mere presence, and methodologically by integrating predictive and explanatory approaches through theory-informed, machine learning–based forecasting frameworks. By examining refugee and asylum seeker outflows, internal displacement, net migration patterns, and household-level responses within a unified conceptual lens, it bridges refugee studies and migration research and reveals how conflict reshapes population distributions through multiple, often overlapping, pathways.

The detrimental impacts of wars on human development are well documented across research domains, from public health to micro-economics. However, these impacts are studied in compartmentalized silos, which limits a comprehensive understanding of the consequences of conflicts, hampering our ability to effectively sustain human development. This article takes a first step in addressing this gap by reviewing the literature on conflict impacts through the lens of an inter-disciplinary theoretical framework. We review the literature on the consequences of conflicts across 9 dimensions of human development: health, schooling, livelihood and income, growth and investments, political institutions, migration and displacement, socio-psychological wellbeing and capital, water access, and food security. The study focuses on both direct and indirect impacts of violence, reviews the existing evidence on how impacts on different dimensions of societal wellbeing and development may intertwine, and suggests plausible mechanisms to explain how these connections materialize. This exercise leads to the identification of critical research gaps and reveals that systematic empirical testing of how the impacts of war spread across sectors is severely lacking. By streamlining the literature on the impacts of war across multiple domains, this review represents a first step to build a common language that can overcome disciplinary silos and achieve a deeper understanding of how the effects of war reverberate across society. This multidisciplinary understanding of conflict impacts may eventually help to reconcile divergent estimates and enable forward-looking policies that minimize the costs of war.

One way of improving forecasts is through better data. We explore how much we can improve predictions of conflict violence by introducing data reflecting third-party efforts to manage violence. By leveraging new sub-national data on all UN peacekeeping deployments in Africa, 1994–2020, from the Geocoded Peacekeeping (Geo-PKO) dataset, we predict changes in violence at the local level. The advantage of data on peacekeeping deployments is that these vary over time and space, as opposed to many structural variables commonly used. We present two peacekeeping models that contain several local peacekeeping features, each with a separate set of additional variables that form the respective benchmark. The mean errors of our predictions only improve marginally. However, comparing observed and predicted changes in violence, the peacekeeping features improve our ability to identify the correct sign of the change. These results are particularly strong when we limit the sample to countries that have seen peacekeeping deployments. For an ambitious forecasting project, like ViEWS, it may thus be highly relevant to incorporate fine-grained and frequently updated data on peacekeeping troops.

This article presents an update to the ViEWS political Violence Early-Warning System. This update introduces (1) a new infrastructure for training, evaluating, and weighting models that allows us to more optimally combine constituent models into ensembles, and (2) a number of new forecasting models that contribute to improve overall performance, in particular with respect to effectively classifying high- and low-risk cases. Our improved evaluation procedures allow us to develop models that specialize in either the immediate or the more distant future. We also present a formal, 'retrospective' evaluation of how well ViEWS has done since we started publishing our forecasts from July 2018 up to December 2019. Our metrics show that ViEWS is performing well when compared to previous out-of-sample forecasts for the 2015-17 period. Finally, we present our new forecasts for the January 2020-December 2022 period. We continue to predict a near-constant situation of conflict in Nigeria, Somalia, and DRC, but see some signs of decreased risk in Cameroon and Mozambique.