Mobile phone data reveal the effects of violence on internal displacement in Afghanistan

The global scale of internal displacement due to conflict, persecution, and human rights violations—affecting nearly 50 million people—presents significant challenges for research and policy. The difficulty in observing and accurately representing displaced populations in surveys and official statistics, particularly in developing countries and insecure environments, hinders effective policy design. This paper addresses these challenges by developing a novel approach using anonymized, high-frequency mobile phone data to study the relationship between violence and internal displacement in Afghanistan, a country with a long history of conflict and over two million internally displaced people. The study's methodological contribution lies in demonstrating how high-frequency digital trace data can enable new approaches to identifying and estimating the causal effects of violent events on displacement. The high-frequency nature of the data allows for causal inferences based on discontinuities in spatio-temporal trajectories coinciding with specific violent events, a feat not easily accomplished with traditional survey data. The study complements existing qualitative, survey-based, and observational studies on conflict and displacement and builds upon recent work leveraging digital data to analyze human population movements. The paper's second contribution lies in providing rich quantitative evidence on the nature of violence-induced displacement in Afghanistan, offering a new perspective to complement traditional approaches based on surveys and administrative reporting.

The paper reviews existing literature on the challenges of studying internal displacement, highlighting the limitations of traditional survey methods, especially in conflict zones and developing countries. It then cites previous research using non-traditional data sources, such as mobile phone data, to study human mobility patterns. The authors note that their work complements existing qualitative and survey-based research on conflict and displacement by providing a quantitative perspective.

The analysis uses call detail records (CDR) from Afghanistan's largest mobile phone operator covering approximately ten million subscribers over four years (April 2013-March 2017). The data are anonymized and contain metadata for each call and data packet transfer, including subscriber ID, date, time, and the cell tower used. The location of each tower allows for approximate geolocation of subscribers. Geo-coded information on violent events in Afghanistan from the Uppsala Conflict Data Program (UCDP), derived from public media reports, is used to pinpoint violent events. The methodology involves several key steps. First, the researchers determine each subscriber's home district using their cell tower usage patterns. Next, they identify migration events as changes in a subscriber's home district lasting at least a week, adopting the International Organization for Migration (IOM)'s definition of migration. Individual migration events are then aggregated to calculate daily population flows between districts. Finally, a high-frequency panel event study design is employed to estimate the causal effect of violence on migration. This involves regressing total out-migration from each district on each day on binary indicators for whether violence occurred in that district on that day, in the preceding 180 days, and in the following 30 days. District and day fixed effects are controlled for, and standard errors are clustered at the district level. This approach allows estimation of the 'average displacement effect' of violence. The authors acknowledge that their statistical notion of displacement differs from that used by international organizations and address potential data and modeling issues, such as population representativeness and data biases, through various robustness checks and data processing techniques. Data validation is performed by comparing the derived migration metrics from CDR with displacement measures published by the IOM, finding a positive correlation between the two, despite inherent differences in how the measures are collected and defined. Additional analyses explore the heterogeneity of displacement responses across different types of violence, locations, and levels of past violence.

The study finds a statistically significant and immediate increase in the likelihood of leaving a district following a violent event. This increase peaks around ten days after the event, with a persistent effect even 120 days later. Interestingly, the analysis reveals an anticipatory effect, with subscribers leaving impacted regions roughly five days before violent events are reported. Heterogeneity in displacement responses is observed, with violence involving the Islamic State (IS) causing significantly more displacement than violence involving the Taliban. High-casualty events and violence in regions with recently experienced violence also lead to greater displacement. The response to violence varies by location, with provincial capitals exhibiting relative resilience while non-capital areas show larger and more persistent displacement effects. Provincial capitals act as magnets for those fleeing violence in non-capital regions, while violence in capitals drives people to other major cities or rural non-capital districts. The analysis of mobile phone metadata provides granular insights into displacement destinations, showing how the equilibrium pattern of movement shifts in response to violence.

The findings corroborate prior research showing that violence causes displacement but offer a finer-grained, quantitative analysis. The study highlights the crucial role of provincial capitals, serving as relatively safe havens due to higher concentrations of security forces, better access to humanitarian aid, potential economic opportunities, and existing family connections. The greater displacement associated with IS violence, high-casualty events, and chronic violence is explained by the perceived risk level influencing the decision to flee. The anticipatory displacement effect is discussed in terms of preemptive warnings from both NATO forces and the Taliban, responses to general unrest, or the spread of rumors about potential violence. The limitations of the UCDP data on violence are acknowledged, given the reliance on media reports, which may exhibit bias towards populous areas and salient events.

The paper concludes that mobile phone data offer significant potential for studying internal displacement, providing insights into the nature of violence-induced displacement that are difficult to obtain through traditional methods. Although limitations regarding data representativeness, access, and privacy concerns exist, this approach can complement traditional perspectives on displacement and contribute to the development of more effective policies for prevention and mitigation.

The study acknowledges several limitations. The data reflect the displacement patterns of only mobile phone owners on one specific network, potentially underrepresenting vulnerable populations. The data provide approximate locations, introducing measurement error. The UCDP data on violence may be biased towards more populous regions and newsworthy events, and data on insurgent control are limited, restricting the scope of some analyses. The analysis focuses on interdistrict movement, potentially underestimating total displacement. Finally, the study cannot determine whether displacement is voluntary or involuntary or the specific aspects of violence driving displacement decisions.