Introduction
Between 2001 and 2020, the United States conducted over 14,000 confirmed airstrikes using unmanned aerial vehicles (drones) in countries like Yemen, Pakistan, Afghanistan, and Somalia (Bureau of Investigative Journalism 2020), establishing drones as an enduring pillar of US counter-terrorism (Tecott Reference Tecott2017).Footnote 1 Drone adoption is rising globally – twenty-two countries have conducted drone strikes and twenty-two more field armed, re-usable drones (DroneWars.net 2025) – largely because proponents argue that drones are cost effective, reduce risk to military personnel, provide intelligence collection advantages, and deliver precise violence with little collateral damage (Brennan Reference Brennan2012; Sales Reference Sales2022).
Existing research shows that drones can disrupt militant operations (Byman Reference Byman2013; Loidolt Reference Loidolt2022, but see Jordan Reference Jordan2014), but may not be cheaper (Calcara et al. Reference Calcara, Gilli, Gilli, Marchetti and Zaccagnini2022), and may cause ‘blowback’ if collateral damage from strikes fuels resentment and increases support for the targeted groups (Cronin Reference Cronin2013; Boyle Reference Boyle2013; Silverman Reference Silverman2024). Quantitative tests yield mixed results: some studies (primarily in Pakistan) find that drone strikes cause insignificant support for militants or ‘blowback’ violence (Johnston and Sarbahi Reference Johnston and Sarbahi2016; Shah Reference Shah2018; Jaeger and Siddique Reference Jaeger and Siddique2018; Mir Reference Mir2018; Mir and D Moore Reference Mir and Moore2019; Schwartz et al. Reference Schwartz, Fuhrmann and Horowitz2022), while others identify short-term violence spikes (Mahmood and Jetter Reference Mahmood and Jetter2023) or violence diffusion due to degraded militant command structure (Rigterink Reference Rigterink2020; Gartzke and Walsh Reference Gartzke and Walsh2022).
Empirical debates about the costs of drone warfare have overlooked non-lethal disruptions to civilian life, which complicate claims about the ‘surgical’ nature of drone strikes.Footnote 2 First, drone strikes cause some nearby civilians to flee their homes, either temporarily or permanently, undermining physical and mental health (Yip and Sharp Reference Yip and Sharp1993; Bürgin et al. Reference Bürgin, Anagnostopoulos, Vitiello, Sukale, Schmid and Fegert2022; Atrooz et al. Reference Atrooz, Khabour, Almomani, Aljararwah, Alfurjani and Salim2024) and economic security (Kondylis Reference Kondylis2010; Ruiz and Vargas-Silva Reference Ruiz and Vargas-Silva2013; Fiala Reference Fiala2015), as well as the economies of destination communities (Cazabat Reference Cazabat2018; Depetris-Chauvin and Santos Reference Depetris-Chauvin and Santos2018). Second, evidence suggests that information about strikes spreads widely, potentially altering political behavior and attitudes of civilians (Pearlman Reference Pearlman2013; Young Reference Young2018; García-Ponce et al. Reference García-Ponce, Young and Zeitzoff2023). Most existing studies have analyzed data that would not have captured either of these effects. The major exception is work by Christia et al. (Reference Christia, Zoumpoulis, Freedman, Yao and Jadbabaie2021), which finds spikes in overall call volume around drone strikes in Yemen, but does not document civilians’ physical displacement, or calling cascades originating in strike regions.
Using over 12 billion cellphone call detail records (CDRs) from Yemen, we show that seventy-four US-conducted drone strikes between 2010 and 2012 significantly disrupted civilians’ ordinary patterns of movement – whether or not the attack caused recorded civilian casualties. The high spatial and temporal resolution of CDRs make it possible to capture civilian reactions without relying on self-reported data (Fair et al. Reference Fair, Kaltenthaler and Miller2014; Shah Reference Shah2018; Ansari Reference Ansari2022).
We find that mobility for all exposed individuals increases by an average of 24 per cent (6.47 km) above baseline on days with drone strikes, with effects lasting over a month. Across seventy-four strikes, 4,500 individuals (5.4 per cent of exposed individuals) flee their homes immediately, with over 1,000 (1.2 per cent) remaining displaced after a month. We also show that drones are more disruptive to civilian life than other types of political violence at the same magnitude.Footnote 3 Where previous research finds that forced migration (migration due to violence) primarily increases with violence intensity (Davenport et al. Reference Davenport, Moore and Poe2003; Fearon and Shaver Reference Fearon and Shaver2020), our analyses show that drone strikes cause more long-term displacement than similar non-drone events in Yemen. Finally, in the Supplementary Information, we document communication cascades started by exposed individuals, signifying the spread of strike-related information, opinions, and emotions through the population, and suggesting negative consequences for civilian well-being beyond displacement.
Accurately characterizing drone strikes’ non-lethal disruption – including displacement – is important for academic and policy conversations around the costs and benefits of drone campaigns. First, the higher-order effects of forced migration on civilian well-being (cited above) and political stability (see, for example, Salehyan and Gleditsch Reference Salehyan and Gleditsch2006, but see also Zhou and Shaver Reference Zhou and Shaver2021) are potentially large, even if displacement is modest. Therefore, showing that drones cause displacement suggests they likely cause more sociopolitical disruption than previously understood. Second, disruption without long-term displacement can still cause negative psychological and social impacts (Coll Reference Coll2014; Gusterson Reference Gusterson2019; Edney-Browne Reference Edney-Browne2019) that are counterproductive for counter-terrorism policy. Finally, the targeting and approval process (in the United States, at least) has historically weighed ‘anticipated military advantage’ (Lushenko Reference Lushenko2022) against ‘expected collateral damage’ (Cartwright Reference Cartwright2013).Footnote 4 Even if the comparison still favors conducting drone strikes, identifying new forms of non-lethal collateral damage should weigh on future targeting decisions at the margin.
Ultimately, this paper contributes to three themes in the broader study of counter-terrorism and civilian behavior. First, we demonstrate the potential of digital communications data for studying behavior – especially displacement – during conflict (Salah et al. Reference Salah, Pentland, Lepri and Letouzé2019; Buckee et al. Reference Buckee, Balsari and Schroeder2022; Tai et al. Reference Tai, Mehra and Blumenstock2022; Milliff and Christia Reference Milliff and Christia2025). Communication and location data from conflict zones are now available on a massive scale and can be fruitfully used to measure (and minimize) the disruptive effects of conflict on civilians. Second, our analysis sheds light on the US drone campaign in Yemen, which is an understudied theater compared to Afghanistan and Pakistan. With US strikes in Yemen on the rise in 2025, understanding the effects of the early stages of the drone campaign has renewed importance for foreign policy. Finally, our analysis complicates the characterization of drone strikes as ’surgical’ violence. We show that strikes have non-lethal but highly disruptive consequences for civilians that have not previously been documented in quantitative analysis.
The US Drone Campaign in Yemen
The second-ever drone strike occurred in Marib Governorate, Yemen, in November 2002, when a CIA-operated MQ-1 Predator killed an alleged plotter of the 2000 USS Cole attack and five others. Regular strikes in Yemen began in 2009; as of 2023, Bergen et al. (Reference Bergen, Salyk-Virk and Sterman2020) reports at least 375 strikes in Yemen, killing an estimated ninety-one high-value targets and 1,300–1,700 other people.Footnote 5
Between 2010 and 2012, al-Qaeda in the Arabian Peninsula (AQAP) launched conventional offensives, capturing territory across Yemen. US strikes increased in response, peaking in 2012 during Yemen government counter-offensives. Data from the Bureau of Investigative Journalism (2020) and Bergen et al. (Reference Bergen, Salyk-Virk and Sterman2020) count 108 strikes between 1 January 2010 and 31 October 2012 (our data’s end-date). We analyze seventy-four, dropping thirty-four strikes in rural areas without cellphone coverage. Most strikes we analyze did not cause reported civilian casualties, but our results suggest they still caused other types of harm.
The Obama administration heralded the campaign as a success, promoting a ‘Yemen model’ of targeted killing via drones with support for local proxies (Hennessey Reference Hennessey2014; Taub Reference Taub2015). Now rebranded as ‘counterterrorism plus’ (Council on Foreign Relations 2020), the approach promises reduced costs to the United States without imposing major costs on civilians (Plaw and Fricker Reference Plaw and Fricker2012).
Early results looked promising. Government reporting claims that four early strikes (2009–10) killed three high-value AQAP targets. Strike frequency then increased, averaging monthly in 2011, and more than weekly in 2012. Our analysis covers the first ’spike’ in a decade-plus campaign which continued from 2013–22 (see Figure 1), averaging one strike every twelve days. The strikes delivered some benefits over alternative strategies: the median strike caused no civilian casualties per Bureau of Investigative Journalism (2020) data, and the campaign was responsible for an estimated seventy-two civilian deaths over a decade. However, as we show, casualties only tell part of the story. Exposure to a drone strike is about as disruptive as other major violence events in Yemen (like suicide bombings), whether or not civilian casualties are reported.
Figure 1. US drone strikes in Yemen 2002–23 (Bergen et al. Reference Bergen, Salyk-Virk and Sterman2020). Time period analyzed in this paper in red.
Measuring Mobility and Drone Strikes
We use cellphone metadata (CDRs) to measure civilian responses to drone strikes. We analyze records of over 12 billion calls and texts from roughly 6 million subscribers to a major Yemeni telecom provider between 1 January 2010 and 31 October 2012. Each record includes a timestamp, duration, anonymized caller and recipient IDs, and the towers that handled the call, which approximate caller/callee locations when merged with tower coordinates.
Our data covers nearly 50 per cent of Yemen’s adult population – including subscribers in all populated districts – during the first ‘peak’ in the US drone campaign.Footnote 6 Cellphones were Yemen’s most important communication technology in the 2010s. Cell penetration rose during the period we cover, from 49 per cent of total population in 2010 to 58 per cent in 2012 (Statista 2019), while internet (12 per cent, World Bank Reference Bank2019) and landline (5 per cent, International Telecommunications Union 2023) penetration remained low. Though CDRs lack detailed subscriber demographics, widespread phone ownership across regions, income ranges, and genders suggests our data capture a broad cross-section of the population, rather than ‘early adopters’ who would be overwhelmingly male, urban, and wealthy (Aker and Mbiti Reference Aker and Mbiti2010).
We derive two measures of mobility from the CDRs of subscribers exposed to drone strikes. We estimate mobility by calculating the distance (summed daily) between towers used in consecutive records. We also measure each record’s distance from strike co-ordinates to estimate cumulative distance traveled. We also measure communication patterns from the same CDR data, tracking both the behavior of individuals exposed to a strike and individuals who are not physically proximate to a strike but are called by a strike-exposed subscriber. In the supplementary information, we show that the post-strike spike in call volumes observed in Christia et al. (Reference Christia, Zoumpoulis, Freedman, Yao and Jadbabaie2021) persists across multiple degrees of social distance, indicating that some of the disruptive potential of strikes expands beyond the targeted areas.
We focus on civilian mobility/displacement because there are reasons to expect that drone strikes change civilians’ patterns of movement: since physical violence/threats often cause displacement (see, among others, Davenport et al. Reference Davenport, Moore and Poe2003), it is plausible that people would flee from drone strikes. The magnitude of the effect (if any) is less theoretically clear. On one hand, drones are lower intensity and/or shorter duration violence relative to alternatives like ground operations or airstrikes with larger munitions. On the other hand, drones have unique technical capabilities (persistent target surveillance, high-precision strikes with little warning) and are clearly attributable to a very powerful foreign military, creating an unusual level of uncertainty and ‘anticipatory fear’ about the risk of future strikes (Mir Reference Mir2018). Since uncertainty or unpredictability can, in some circumstances, encourage migration (Milliff Reference Milliff2024), drone strikes’ displacement effects may be larger than the small explosive payload would suggest. These competing hypotheses have not been empirically adjudicated, perhaps due to the assumption that drone strikes are small, high-precision interventions.
Drone strike data come from the Bureau of Investigative Journalism (2020) and Bergen et al. (Reference Bergen, Salyk-Virk and Sterman2020). While news-based event data are often incomplete compared to data from governments (Shaver et al. Reference Shaver, Arriaga, Blackburn, de Kadt, Felter, Hyneman, Silva, Lorch, McLean and Prada2023), the opposite is true here; researchers use local reporting to identify and confirm strikes and casualty estimates that are not disclosed by the US government.Footnote 7 Data include strike coordinates, timestamps, and civilian and militant casualty estimates (with uncertainty bounds). We merge strike records with CDR data to identify subscribers exposed to each strike and track prior and subsequent mobility.
Design
We estimate the causal effect of drone strikes on civilian mobility using a panel-data event study design, which regresses outcomes on temporal lag and lead indicators relative to drone strikes in order to estimate dynamic treatment effects, controlling for strike-specific heterogeneity via fixed effects. We use the protocol from Schmidheiny and Siegloch (Reference Schmidheiny and Siegloch2023), which bins endpoints to estimate dynamic treatment effects without untreated units. Our parameters of interest are lag and lead coefficients, indexed − j to j, which compare mobility in period j relative to baseline period j = − 2. We describe notation and setup more fully in the Supplementary Information. We define individuals as exposed to a strike event if they make a call within fifteen miles (24.1 km) of the strike locations during the strike period. This threshold is based on existing literature (Gao et al. Reference Gao, Song, Gao, Barabasi, Bagrow and Wang2014; Christia et al. Reference Christia, Zoumpoulis, Freedman, Yao and Jadbabaie2021), and is robust to alternative radii from five miles to thirty miles (8km–48.3 km, see Supplementary Information).
Causal interpretation of the event study relies on two assumptions. First, drone strikes are assumed to be exogenous shocks with respect to civilian movement. This requires: (1) no simultaneity – for example, mobility changes do not drive strike occurrence – and 2) no joint response – that is, no confounding events cause both strikes and mobility (Greene Reference Greene2018). We evaluate simultaneity by examining pre-treatment trends, finding no evidence of anticipatory changes in mobility (see Figure 2). We also note that US drone strike planning does not rely on aggregate mobility data.Footnote 8 We evaluate the joint response assumption with data from the Critical Threats Project (2019) and Bergen et al. (Reference Bergen, Salyk-Virk and Sterman2020). We find that strikes do not overlap with potentially confounding events like AQAP attacks or known US ground operations.
Beyond these formal assumptions, our design also satisfies the conditions enumerated by Dunning (Reference Dunning2012, 237–9) for ‘as-good-as random’ assignment. Individual cellphone subscribers lack sufficient information to predict strike exposure, so they cannot self-select out of treatment. Furthermore, US strike planners lack incentives to manipulate strike timing or treatment assignment for particular non-targeted civilians. Planners’ focus on civilians is limited to efforts to keep them out of the strike’s blast radius (65 feet / 20 m [Anderson Reference Anderson2014]), well within our fifteen mile exposure threshold. Consequently, civilians’ treatment status is not plausibly self-selected or US-selected based on mobility.
Results
We use CDRs to construct a time series of exposed individuals’ locations and compute daily distance traveled. Using the event study design, we estimate the effect of seventy-four strikes on daily mobility over a seven-day pre-strike and twenty-one-day post-strike window (Figure 2(a)). The average subscriber’s mobility increases by 6.47 km (24 per cent over pre-strike mean, s.e. 1.26) on strike days and remains elevated over the baseline by 2.47 km (s.e. 0.83) twenty-one days afterward. Pre-strike trends are negligible, supporting our identifying assumptions.
Figure 2. Effect of drone strikes on mobility. Figures show parameter estimates from an event study (full model in supplementary information) for exposed individuals’ (a) change in daily distance traveled and (b) distance from the strike region. Shading indicates 95 per cent confidence intervals from strike-clustered standard errors (Abadie et al. Reference Abadie, Athey, Imbens and Wooldridge2017).
Average distance from strike locations also increases over time – from 5.52 km (s.e. 0.89) at one week, to 11.31 km (s.e. 1.33) at twenty-one days, peaking at 13.48 km (s.e. 1.86) at forty-three days (Figure 2(b)). This suggests that a large number of people leave the strike region quickly, dispersing around the country within twenty-four hours (Figure 3). Across the seventy-four strikes, 4,519 subscribers (5.3 per cent of exposed subscribers) leave the area and remain away for at least twenty-four hours. Almost all displaced individuals are likely civilians, not militants.Footnote 9 Some displacement is long term: 51 per cent of displaced subscribers return home within five days, but roughly 25 per cent remain displaced for over a month. Average distance away from strike areas among the displaced is nearly double the full-sample average (see SI), suggesting that people displaced by strikes travel long distances before stopping. The long-term displacement we measure eclipses the number of militants targeted (∼390) by 2.7 times. Our estimate of displacement is almost certainly conservative. First, poverty and poor infrastructure in Yemen likely inhibit some displacement that would otherwise occur in response to strikes. Second, some of the ∼50 per cent of Yemeni adults not in our data – subscribers to other telecoms and/or non-phone owners including family members of subscribers – were likely displaced by drone strikes as well. Even a modest number of conflict-related displacements indicates broader social disruption. Exposed subscribers who do not migrate still experience substantially increased mobility after strikes – behavior which computational social scientists have linked to increased stress and anxiety (Müller et al. Reference Müller, Peters, Matz and WangWand Harari2020). Displacement also negatively impacts destination communities, including the relatives that displaced people likely stayed with. In 2011–12, internally displaced person (IDP) flows in Yemen created severe housing shortages in major destination cities like Aden (IRIN News Reference News2013). In one of the world’s poorest countries, crowding plus migration-induced income loss could create long-term hardship.
Figure 3. Displacement after drone strikes. (a) Subscriber locations at strike time, in blue (strikes as red Xs). Gray shading indicates districts with >30 people/km2. (b) Subscriber locations twenty-four hours post-strike. Light blue arrows show displacement trajectories.
Drone Strikes Displace More Than Other Violence
Displacement from drone strikes (5.3 per cent) is similar in magnitude to displacement after qualitatively larger disasters, like the catastrophic 2010 Haiti Earthquake that killed at least 100,000 Haitians (6 per cent, Lu et al. Reference Lu, Bengtsson and Holme2012), or early COVID-19 lockdowns in New York City in March 2020 (5.5 per cent, Bakker et al. Reference Bakker, Berke, Groh, Pentland and Moro2020). Drone strikes also cause more displacement than similar violent events in Yemen, like suicide bombings. We select five non-drone comparison events that are unexpected, localized (that is, not coordinated with attacks elsewhere), and unusually high-casualty compared to previous local events and to average drone strikes (See Supplementary Information for selection details).
The mobility of exposed subscribers (daily distance traveled) spikes on the event day for most comparison events (Figure 4). Event day average mobility increased 7.5 km (84 per cent, s.e. 0.40), 0.4 km (8 per cent, s.e. 0.06), 3.9 km (52 per cent, s.e. 0.37), 1.4 km (23 per cent, s.e. 0.06), and 3.0 km (38 per cent, s.e. 0.14) respectively, but events produced smaller magnitude and less durable displacement. After the events, an average of 1.7 per cent of exposed subscribers were displaced, but no individuals remained away for more than thirty days across all five events.
Figure 4. Mobility after five comparison events. The figures display average daily distance traveled by exposed subscribers.
Civilian Casualties Matter at the Margin
Drone strikes have different consequences depending on the reported numbers of militants and civilians killed. All strikes are disruptive, but strikes causing more civilian casualties are, on balance more disruptive, causing more displacement. We regress the number of individuals who remain displaced thirty days post-strike on predictors including the number of civilian and militant casualties, the time of day, the district population, and the number of strikes in the district in the preceding month (all data from Bureau of Investigative Journalism (2020) and Bergen et al. (Reference Bergen, Salyk-Virk and Sterman2020)).
Table 1 shows that reported civilian casualties are strongly associated with higher long-term displacement: a strike killing ten civilians results in 40.8 more long-term displacements, compared to a strike causing no civilian casualties.Footnote 10
Table 1. Effects of civilian casualties on displacement
Note: *p < 0.1; **p < 0.05; ***p < 0.01.
This suggests that precision targeting matters at the margins: strikes causing more collateral damage are associated with more civilian displacement. At the same time, the event study results are not solely driven by civilian casualties – 90 per cent of strikes we analyze report no civilian casualties but still cause displacement and increased mobility for non-displaced subscribers. These results have clear implications for operational planning and broader counter-terrorism policy: while drone campaigns negatively affect civilian well-being in any circumstance, minimizing civilian casualties reduces their disruptive impacts.
Why Are Drones So Disruptive?
Despite the US drone campaign’s ‘precision’ reputation, disruptive effects on civilians’ mobility are substantial and surprisingly durable even for civilians who are not physically harmed by the strikes. Why are drone strikes, hardly Yemen’s highest-intensity violence between 2010 and 2012, so disruptive? Our results are consistent with various mechanisms, which may operate in parallel.Footnote 11
First, evidence from Pakistan suggests that drone strikes cause anticipatory fear, which reduces targeted militants’ movement and communication (Mir Reference Mir2018). We posit that anticipatory fear causes different behaviors for civilians v. the militants that Mir (Reference Mir2018) studies: fearing possible future strikes and aware of ongoing drone activity, civilians might migrate away from areas they expect to be targeted again (WH Moore and Shellman Reference Moore and Shellman2004). High uncertainty about future strikes (compared to uncertainty around conventional front lines elsewhere in Yemen, for example) might further drive displacement (Milliff Reference Milliff2024). Higher civilian casualties could amplify fear and uncertainty. Anticipatory fear might blunt the effects of increasing familiarity or habituation to drone strikes over time, which ceteris paribus might decrease response to later strikes (though not in Yemen; see Supplementary Information).
Second, drone strikes may have a platform-specific reputation in Yemen.Footnote 12 Reporting suggests that civilians in Yemen were aware of drones as a separate technology of conflict – and one particular to the United States – in the early 2010s (Terrill Reference Terrill2013), likely thanks to militant groups’ use of drone strikes in recruiting propaganda. Civilians’ particular associations with drone strikes may explain different responses compared to other types of violence. Analyses of calling cascades (see Supplementary Information) are consistent with this ‘reputation’ effect: information about strikes spreads rapidly across Yemen, reaching many people who aren’t directly exposed and potentially conditioning their future behavior.
Conclusion
In this article, we study the effects of drone strikes on civilians, and show that a purportedly high-precision counter-terrorism strategy causes previously unmeasured disruption. Using an event study framework and a dataset on the location and communication of nearly 6 million Yemeni cellphone users over thirty-four months, we provide causally identified evidence that drone strikes trigger rapid, substantial, and durable changes in civilian mobility.
Across seventy-four strikes between 2010 and 2012 in Yemen, we find that thousands of people quickly flee the area around the strike, with over 1,000 remaining displaced for at least a month. Our estimates might represent a lower-bound estimate of civilian responses to drones in other contexts, given Yemen-specific impediments like poor transport infrastructure (Yemen has <10,000 km of paved roads) and widespread poverty.
Our findings raise new questions about the costs of drone campaigns. Even when drone strikes do not cause civilian casualties, they create substantial collateral damage because displacement during conflict has severe consequences for physical and mental health (Toole and Waldman Reference Toole and Waldman1993; Owoaje et al. Reference Owoaje, Uchendu, Ajayi and Cadmus2016), household finances (Fiala Reference Fiala2015), and long-term economic well-being and human capital (Kondylis Reference Kondylis2010; Ruiz and Vargas-Silva Reference Ruiz and Vargas-Silva2013). Prior research finds that US drone strikes can acquire a negative ‘reputation’ that spreads beyond the battlefield (Silverman Reference Silverman2019), which is especially concerning given a large body of research emphasizing the importance of civilian attitudes in asymmetric conflicts (Berman et al. Reference Berman, Felter and Shapiro2018, among many). In 2019 (Arab Barometer has since been unable to survey Yemen), 84 per cent of Yemenis said attacks against the United States were a ‘logical consequence’ of American intervention in the region, and 93 per cent said America’s foreign policy was bad for Yemen (Arab Barometer Reference Barometer2019). More research is needed to understand how drone strikes shaped these opinions, but our results suggest some ways they may have contributed to negative attitudes.
Finally, new evidence of the disruption caused by the US drone campaign in Yemen ought to change the ethical calculus for using the platform in the future. Despite touted benefits like reducing civilian casualties compared to non-drone interventions, lowering costs, and minimizing US casualties, we show that drones create less of a ‘win–win’ than advertised. Decision makers should account for these costs when evaluating the strategic utility of drones.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0007123425100975.
Data availability statement
Replication data for this paper can be found in Harvard Dataverse at: https://doi.org/10.7910/DVN/HH1YUN.
Acknowledgments
We thank Alberto Abadie, Dean Eckles, Adeline Lo, Asfandyar Mir, Teppei Yamamoto, and audiences at the 2019 Conference of the Society for Political Methodology, 2019 MIT Statistics and Data Science Conference, 2019 Households in Conflict Network Workshop, 2020 NBER Summer Institute, Emory University, and Florida State University for helpful comments and discussion. We also thank the MIT Supercloud and Lincoln Laboratory Supercomputing Center for providing high performance computing and database resources.
Financial support
This research was supported by the Army Research Office’s Multidisciplinary University Research Initiative (MURI) Grant No. W911NF-12-1-0509 and the Department of Defense’s Vannevar Bush Fellowship.
Competing interests
None declared.
Ethical standards
The research meets ethical guidelines, including adherence to the legal requirements of the study country. Respondent privacy, specifically, is discussed in greater detail in the supplementary information.
Open access
