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

The study investigates how violent events affect internal displacement in Afghanistan, where decades of conflict have produced large populations of internally displaced people who are difficult to observe in conventional data. The central research question is to quantify the causal effect of violence on population movements within the country, over short and long horizons, and to understand how effects vary by type and location of violence. The paper’s purpose is twofold: methodologically, to demonstrate how high-frequency digital trace data from mobile phones can support causal inference on displacement dynamics using a panel event study design; substantively, to provide rich, quantitative evidence on displacement patterns in Afghanistan, complementing traditional surveys and administrative data. The authors emphasize that the high temporal resolution and scale of mobile phone data enable identification strategies that are infeasible with standard survey approaches, allowing estimates of discontinuities in mobility that coincide with precisely timed violent events. They argue this approach can improve research and policy for conflict-affected populations that are often underrepresented in official statistics.

The paper situates its contribution within several strands of literature: (1) conflict and displacement research that has used qualitative, survey-based, and observational methods to link violence to internal displacement in fragile settings; (2) emerging work using non-traditional digital trace data (e.g., mobile phone records, social media) to study human mobility at population scale, including disaster response and disease spread; and (3) event-study and panel regression methodologies for causal inference with longitudinal data. The authors note that prior work documents links between violence, perceived threat, and mobility decisions, but lacks the temporal granularity and coverage afforded by mobile phone data. The study also engages with literature on differences across armed actors (e.g., Islamic State vs Taliban), the role of casualties and conflict intensity, and urban-rural differences in responses to violence. Finally, it references ongoing debates about biases in conflict event datasets, representativity and privacy issues in mobile data, and best practices for conflict prediction and data collection.

Data sources: (1) Violent events from UCDP Georeferenced Event Dataset Global v19.1, filtered to events with district-level location and day-level timing. Of 5,984 events in 2013–2017, 4,740 fall within April 2013–March 2017; 3,354 occur in districts on days with mobile activity and are used. (2) Mobile phone call detail records (CDR) from one of Afghanistan’s largest operators, covering ~20 billion events from April 2013 to March 2017, with pseudonymized subscriber IDs, timestamps, and tower IDs geolocated to approximate positions (~500 m urban, ~10 km rural). Towers (13,315) are grouped into 1,439 tower groups. Only districts with towers are analyzed. Privacy safeguards include pseudonymization and aggregation at district-day level. Measuring migration: The authors infer daily modal locations by (a) computing each subscriber’s most used tower per hour; (b) for each 06:00–06:00 day, taking the mode of hourly modal towers; (c) mapping towers to districts. They then apply an unsupervised scanning algorithm (migration_detector) to segment each subscriber’s timeline into contiguous residence spells and detect migration events as discontinuities between districts, tuned to capture movements with at least five full days in origin and destination (roughly week-long or longer stays). This yields interdistrict migration events. Validation: District-level subscriber counts correlate strongly with official population estimates (Pearson r = 0.94; p < 0.001). Aggregated 2016 province shares of outgoing and incoming movement correlate with IOM displacement metrics (Spearman ρ = 0.49 for outgoing; ρ = 0.56 for incoming). Causal design: A district-day panel event study estimates the effect of violence on excess out-migration. For each k in 1–120, the outcome Y_dkt is the proportion of those in district d at time t−k who are in a different district at time t. The model is a beta regression with logit link: g(E(Y_dkt|X_dt,T_dt)) = α + Σ_{r=-30}^{180} β_r T_{d,t+r} + γ_d + δ_t, where T indicates violence in district d at lags r days (−30 to +180), with district and day fixed effects, and standard errors clustered by district. Coefficients are interpreted as multiplicative changes in odds. Identification assumes positivity, consistency, conditional exchangeability (no unobserved time-varying confounders beyond fixed effects and treatment history), correct model specification, no spatial spillovers, and carryover limited to 180 days. Robustness checks test predictability of residual violence using outcome trends and subscriber counts (R^2 ≤ 0.00028), add region×month fixed effects, and exclude district-days with outcomes of 0, 1, or missing due to potential data sparsity/outages. Heterogeneity: Separate treatment indicators are created to estimate differential effects by (a) actor (Islamic State vs Taliban), (b) severity (≥11 casualties vs ≤10), (c) recency (fewer than 60 days of peace vs ≥60), and (d) location (provincial capital vs non-capital). To jointly assess multiple characteristics, the authors estimate event-specific 30-day displacement profiles (means over lag windows 1–15, 16–30, 31–45, 46–60, 61–75, 76–90 days) and regress these on event covariates: provincial capital, log district population, IS involvement, high casualties (≥11), and ≥60 days of prior peace. Destinations: For k = 7, 30, 90, the outcome is the fraction of movers whose destination falls into categories defined by province (same vs different) and district type (top five cities: Kabul, Kandahar, Hirat, Mazari Sharif, Jalalabad; other capitals; non-capitals). Separate regressions are fit for origins in capitals vs non-capitals, comparing violent vs non-violent days. Missing violence data: To assess bias from spatial imprecision, events are split by location precision (exact point, within 25 km, district centroid), and displacement responses are compared; patterns are similar with overlapping CIs, though unobserved events cannot be tested.

- Violence causes measurable internal displacement. Among subscribers present on the day of a violent event, the odds of being in a different district increase immediately and peak around 10 days post-event at +4.0% (95% CI: 2.6%, 5.5%; p < 0.001). Effects persist: at 120 days post-event, odds remain ~+2.0% (95% CI: 0.5%, 2.9%; p = 0.007). - Anticipatory movement: Approximately five days before reported violent events, subscribers begin leaving impacted regions; 30-day displacement relative to 30 days prior shows elevated odds in the days leading up to violence. - Actor heterogeneity: IS-linked events trigger substantially more displacement than Taliban-linked events. One day after the event, IS increases odds of displacement by 12.7% (95% CI: 5.7%, 20.2%; p < 0.001) vs Taliban 1.9% (95% CI: −0.4%, 4.3%; p = 0.111). The IS–Taliban difference in coefficient estimates is 0.10 (95% CI: 0.04, 0.17; p = 0.002), most pronounced immediately after events. - Severity: High-casualty events (≥11 casualties; ~top 10% by severity) produce larger displacement at all post-event horizons than lower-casualty events (paired t-test over 120 horizon-by-horizon comparisons: p < 0.001; mean difference 0.034; 95% CI: 0.031, 0.036). - Conflict recency: Violence following fewer than 60 days of peace induces more displacement than violence after ≥60 days of peace (p < 0.001; mean difference 0.030; 95% CI: 0.029, 0.031). - Location: Provincial capitals show muted and often insignificant displacement responses, whereas non-capital districts exhibit large and persistent effects (p < 0.001; mean difference 0.026; 95% CI: 0.024, 0.028). Capitals appear relatively resilient and may serve as safer destinations. - Joint heterogeneity: Holding other factors constant, IS involvement has the largest and most persistent association with 30-day displacement across lag windows; for events 1–15 days past, the IS vs recent-violence difference in coefficients is 0.194 (95% CI: 0.032, 0.356; p = 0.019). - Baseline flows on non-violent days: From capitals, 73.5% of movers go to a different province; 47.0% move to other capitals/major cities. From non-capitals, 50.6% move to another province; 30.0% move to the provincial capital in the same province. - Destination shifts on violent days: When violence strikes non-capitals, movers are more likely to relocate to their provincial capital and less likely to go to major cities outside their origin province. When violence strikes capitals, movers are more likely to leave the home province for major cities or non-capital districts elsewhere.

The findings address the core question by quantifying how precisely timed violent events alter internal mobility patterns. The causal event-study approach shows that violence leads to immediate and sustained excess out-migration relative to normal mobility, with effects varying by actor, severity, recency, and location. The results underscore the role of provincial capitals as protective hubs that attract displaced persons from rural districts, likely due to perceived security from concentrated government forces, better access to humanitarian assistance, urban economic opportunities, and social connections. The heightened displacement following IS attacks, high-casualty incidents, and in chronically violent contexts is consistent with elevated perceived risk and fear prompting flight. The anticipatory movements observed prior to recorded events suggest populations respond to signals of impending violence, such as warnings, unrest, or unreported skirmishes, with information spreading rapidly through social networks. These insights are relevant to humanitarian operations and policy, indicating where and when displacement pressures will materialize and highlighting the importance of urban centers in providing refuge.

The study introduces a scalable, high-frequency method to measure conflict-induced internal displacement using anonymized mobile phone data and a panel event study framework. Applied to Afghanistan, it quantifies both short- and long-term displacement effects of violence, reveals strong heterogeneity by perpetrator, severity, conflict recency, and location, and documents the critical role of provincial capitals as displacement destinations. Despite limitations, the approach complements traditional surveys and administrative data, offering timely, granular evidence to inform policy and humanitarian response. Future work could expand to other contexts, integrate richer measures of territorial control and non-fatal conflict dynamics, improve representativity adjustments, and combine multiple data sources to better distinguish voluntary from forced movement and to account for within-district and cross-border displacement.

- Coverage of movement: Only interdistrict movements within Afghanistan are captured; within-district and international displacement are missed, likely underestimating total displacement. - Representativity: Data reflect mobility of active subscribers on one commercial network, not the full population. Vulnerable groups (women, children, lower-income individuals) are underrepresented in mobile ownership and usage, potentially biasing results. Phone sharing, power-off periods, and intermittent use introduce measurement error; lower usage among vulnerable populations may exacerbate errors. - Data sparsity/outages: Intermittent call activity and possible tower outages can affect inferred locations; the analysis excludes district-days with outcomes of 0, 1, or missing to mitigate sparsity artifacts. - Access and privacy: Proprietary CDR require data-sharing partnerships; strict privacy safeguards are necessary. The study uses pseudonymized, aggregated data but broader ethical considerations remain. - Conflict data bias: UCDP relies on media and other sources; reporting is incomplete and non-random across space and event types. Spatial and temporal precision filters exclude ~47% of events, potentially biasing estimates (e.g., higher reporting thresholds in rural areas). Robustness shows similar displacement responses across varying spatial precision, but missing/unrecorded events cannot be tested. - Contextual data limitations: Limited panel data on territorial control restrict robust analysis of control/contest dynamics over time; cross-sectional proxies may be inadequate. - Mechanism ambiguity: The approach cannot distinguish voluntary vs forced movement or specify which components of violence trigger mobility. Aggregation may mask instances of immobility or trapping. - Generalizability: Effects may be context-specific to Afghanistan’s protracted conflict, demographics, and social structures; magnitudes and heterogeneity may differ elsewhere.