The Impact of US Drone Strikes on Terrorism in Pakistan and Afghanistan

Patrick B. Johnston RAND Corporation Anoop K. Sarbahi Stanford University

July 14, 2013

Abstract This study analyzes the eects of US drone strikes on terrorism in Pakistan and Afghanistan. Some theories suggest that drone strikes anger Muslim populations, and that consequent blowback incites Islamist terrorism. Others argue that drone strikes disrupt and degrade terrorist organizations, reducing their ability to conduct attacks. We use detailed data on U.S. drone strikes and terrorism in Pakistan and Afghanistan from 2004-2011 to test each theorys implications. We nd that drone strikes are associated with decreases in the incidence and lethality of terrorist attacks, as well as decreases in particularly intimidating and deadly terrorist tactics, including suicide and improvised explosive devices (IED) attacks. These results lend credence to the argument that drone strikes, while unpopular, have bolstered U.S. counterterrorism eorts in Pakistan and cast doubt on claims that drone strikes are militarily ineective.

Earlier versions of this article were presented at the 2011 Annual Meetings of the American Political Science Association, the Belfer Center for Science and International Aairs at Harvard Universitys Kennedy School of Government, and the New America Foundation. For helpful feedback on earlier versions, we thank Peter Bergen, James Dobbins, C. Christine Fair, Melissa Willard-Foster, Seth G. Jones, Jennifer Keister, Akbar Khan, Peter Krause, Sean Lynn-Jones, Steven E. Miller, Jacob N. Shapiro, Arthur Stein, Katherine Tiedemann and Jeremy Weinstein. Johnston acknowledge nancial support from AFOSR Award #FA9550-09-1-0314.

Introduction

Do drone strikes against terrorists reduce the threat posed by terrorist organizations, or do they unintentionally increase support for anti-U.S. militants and thus fuel terrorism?
1

. Empirical studies of targeted killings and civilian casualties in counterinsurgency

and counterterrorism show that both outcomes are possible.2 Strikes conducted by remotely piloted aircraft may undermine counterterrorism eorts or enhance them depending on the nature of the violence, the intentionality attributed to it, or the precision with which it is applied.3 Existing research has studied the eects of coercive airpower,4 targeted killings,5 and civilian victimization,6 but social scientists have
Examples of arguments that drone strikes are ineective or counterproductive include Living Under Drones: Death, Injury, and Trauma to Civilians From US Drone Practices in Pakistan, Stanford Law School and NYU School of Law, September 2012; and Audrey Kurth Cronin, Why Drones Fail, Foreign Aairs, July 1, 2013. Examples of arguments that drone strikes are eective include C. Christine Fair, Drone Wars, Foreign Policy, May 28, 2010; Fair, For Now, Drones Are the Best Option, New York Times, September 26, 2012; and Daniel Byman, Why Drones Work, Foreign Aairs, July 1, 2013. 2 Benjamin Valentino, Paul Huth, and Dylan Balch-Lindsay, Draining the Sea: Mass Killing and Guerrilla Warfare, International Organization, Vol. 58, No. 2 (Spring 2004): 375407; Alexander B. Downes, Draining the Sea by Filling the Graves: Investigating the Eectiveness of Indiscriminate Violence as a Counterinsurgency Strategy, Civil Wars, Vol. 9, No. 4 (December 2007): 420444; Jessica Stanton, Strategies of Restraint in Civil War (New York: Columbia University, 2009); Jenna Jordan, When Heads Roll: Assessing the Eectiveness of Leadership Decapitation, Security Studies, Vol. 18, No. 4 (December 2009), pp. 719-755. 3 Stathis N. Kalyvas, The Logic of Violence in Civil War (New York: Cambridge University Press, 2006); Downes, Draining the Sea by Filling the Graves: Investigating the Eectiveness of Indiscriminate Violence as a Counterinsurgency Strategy; Matthew Adam Kocher, Thomas B. Pepinsky, and Stathis N. Kalyvas, Aerial Bombing and Counterinsurgency in the Vietnam War, American Journal of Political Science, Vol. 55, No. 2 (April 2011), pp. 201-218. 4 Robert A. Pape, Bombing to Win: Air Power and Coercion in War (Ithaca, NY: Cornell University Press, 1996); Michael Horowitz and Dan Reiter, When Does Aerial Bombing Work? Quantitative Empirical Tests, 1917-1999, Journal of Conict Resolution, Vol. 45, No. 2 (April 2001), pp. 147173. 5 David A. Jaeger, The Shape of Things to Come? On the Dynamics of Suicide Attacks and Targeted Killings, Quarterly Journal of Political Science, Vol. 4, No. 4 (December 2009), pp. 315342; Jordan, When Heads Roll; Patrick B. Johnston, Does Decapitation Work? Assessing the Eectiveness of Leadership Targeting in Counterinsurgency Campaigns, International Security, Vol. 36, No. 4 (Spring 2012), pp. 47-79; and Bryan Price, Targeting Top Terrorists: How Leadership Decapitation Contributes To Counterterrorism, International Security, Vol. 36, No. 4 (Spring 2012), pp. 9-46. 6 Kalyvas, The Logic of Violence in Civil War; Jason Lyall, Does Indiscriminate Violence Incite Insurgent Attacks?: Evidence from Chechnya, Journal of Conict Resolution, Vol. 53, No. 3 (February 2009), pp. 331362; and Luke Condra and Jacob N. Shapiro, Who Takes the Blame? The Strategic Eects of Collateral Damage, American Journal of Political Science, Vol. 56, No. 1 (January 2012), 167187.
1

conducted little empirical analysis of the eects of drone strikes.7 This lack of attention is unfortunate: unmanned aerial vehicles, and their lethal targeting capabilities, are likely to represent a critical aspect of current and future counterterrorism eorts. The consequences of drone strikes are a critical policy concern. The United States has frequently been called upon to cease drone strikes in Pakistan in order to protect noncombatants, but instead it has expanded its use of drones to other countries in which al Qaida-aliated militants are believed to operate, such as Somalia and Yemen.8 The laws governing international armed conict codify and strengthen norms against targeted killings, yet other interpretations of the laws of war leave civilian ocials and military commanders with substantial latitude to target enemy combatants believed to be aliated with terrorist organizations against which the U.S. has declared war.9 Liberal democratic states face substantial pressures to protect civilians in war, but at the same time are often confronted with substantial uncertainty as to what abiding by legal principles such as discrimination the obligation of military forces to select means of attack that minimize the prospect of civilian casualties actually entails.10 Drone strikes are not the only instrument the U.S. can use to ght al Qaida terrorists; states have used other methods to ght terrorism for centuries. The eectiveness of drone strikes at countering terrorism lies at the core of U.S. policymakers arguments for their continued use. Yet because of the drone programs secretive nature and wide
7 Exceptions include David A. Jaeger and Zahra Siddique, Are Drone Strikes Eective in Afghanistan and Pakistan? On the Dynamics of Violence between the United States and the Taliban, IZA Discussion Paper No. 6262, November 2011; and Megan Smith and James Igoe Walsh, Do Drone Strikes Degrade Al Qaeda? Evidence From Propaganda Output, Terrorism and Political Violence, Vol. 25, No. 2 (February 2013), pp. 311-327. 8 For excellent descriptions of the drone wars expansion, see Mark Mazzetti, The Way of the Knife: The CIA, a Secret Army, and a War at the Ends of the Earth (New York: Penguin Press, 2013); and Jeremy Scahill, Dirty Wars: The World Is a Battleeld (New York: Nation Books, 2013). 9 Christine D. Gray, International Law and the Use of Force (New York: Oxford University Press, 2000). 10 Neta C. Crawford, Just War Theory and the U.S. Counterterror War, Perspectives on Politics, Vol. 1, No. 1 (March 2003), pp. 5-25; and Michael Walzer, Just and Unjust Wars: A Moral Argument with Historical Illustrations (New York: Basic Books, 2006).

disagreement about the eects of drone strikes on terrorist organizations and civilian populations, U.S. government ocials and human rights advocates have both failed to present compelling, systematic evidence in support of their positions. What is needed is a rigorous, evidence-based assessment of drone strikes impact on terrorism. Such an assessment should sharpen the debate on drone strikes and help counterterrorism ocials and critics alike to evaluate the tradeos associated with drone warfare. The present study provides such an assessment by using a data-driven approach to analyze the consequences of drone strikes. Based on detailed data on both drone strikes and terrorism in Pakistan throughout the course of the U.S. drone campaign there, the study examines how drone strikes have aected terrorist violence in northwest Pakistan and bordering areas of Afghanistan. In order to provide the most comprehensive analysis possible, this study investigates the relationship between drone strikes and a wide range of militant activities and tactics, including terrorist attack patterns, terrorist attack lethality, and especially deadly and intimidating tactics such as suicide and improvised explosive device (IED) attacks. A systematic analysis of the data reveals that drone strikes have succeeded in curbing deadly terrorist attacks in Pakistan. Specically, the key ndings of our study show that drone strikes are associated with substantial reductions in terrorist violence along four key dimensions. First, drone strikes are generally associated with a reduction in the rate of terrorist attacks. Second, drone strikes are also associated with a reduction in the number of people killed as a result of terrorist attacks. Third, drone strikes tend to be linked to decreases in the use of particularly lethal and intimidating tactics, including suicide and IED attacks. Fourth, the study nds that this reduction in terrorism is not the result of militants leaving unsafe areas and conducting attacks elsewhere in the region; on the contrary, there is some evidence to suggest that drone strikes have a small violence-reducing eect in areas near those struck by drones. Taken together, these ndings strongly suggest that despite drone

strikes unpopularity, ocial claims that drones have aided U.S. counterterrorism eorts in Pakistan appear to be credible and should not be dismissed out of hand. The remainder of the article proceeds as follows. In Section 2, we outline the range of relevant hypotheses on the eects of drone strikes, and briey discuss the theoretical logics that undergird them. In Section 3, we describe our dataset and the methodology used to assess the eects of drone strikes on terrorism. In Section 4, we discuss the results of our empirical analysis and our interpretation of key ndings. Finally, Section 5 concludes with a discussion of our ndings implications for policy and the future of counterterrorism.

Hypotheses on Drone Strikes and Terrorism

Two contradictory arguments characterize the debate concerning the eectiveness of drone strikes. The rst focuses on how drone strikes aect the attitudes of the civilian population, while the second focuses on the impact of drone strikes on insurgent and terrorist organizations. Below, each argument is discussed in turn.

2.1

Drone Strikes, the Civilian Population, and Militant Mobilization

The rst argument is that drone strikes do little to curb terrorism and might increase it. Critics have suggested, for example, that drone attacks are ineective or counterproductive to the U.S. strategy of disrupting and dismantling al Qaida and other terrorist networks because they are unpopular among the Pakistani population, largely because they occasionally inict civilian casualties. Consistent with this argument, Smith and Walsh nd no evidence that drone strikes degrade al Qaida propaganda eorts.11 Taking this argument a step further, others
11

Smith and Walsh, Do Drone Strikes Degrade Al Qaeda? Evidence From Propaganda Output,

argue drone strikes are the wrong tool to curb militancy in fact, they may worsen it because the tactic itself breeds a counterproductive desire for revenge among Pakistanis who might otherwise harbor no hostilities toward the United States. In the words of David Kilcullen and Andrew Exum, Every one of these dead noncombatants represents an alienated family, a new desire for revenge, and more recruits for a militant movement that has grown exponentially even as drone strikes have increased.12 Given the expected anticipated anti-U.S. mobilization and desire for revenge among the civilian population suggested by this logic, we elaborate the following hypothesis: H1: All else equal, drone strikes increase terrorist violence.

2.2

Disruption, Degradation, and Militant Capabilities

The second argument, which is common among U.S. counterterrorism ocials but rarely buttressed with specic empirical evidence due to the drone programs secrecy, contends that drone strikes are eective at reducing the terrorist threat posed by targeted groups. Two mechanisms are frequently cited: disruption and degradation.

2.2.1

Disruption

The rst mechanism counterterrorism ocials cite involves disruption of militant operations. The disruption mechanism suggests drone strikes reduce militants ability to operate in a cohesive, eective manner and erode their ability to exercise sovereign control over local areas. Even if an insurgent or terrorist organization is the only armed actor on the ground, as they often are in FATA, where state authority is extremely weak, the greater the threat from above, the more costly it is for the militants to exercise de facto control in that area. The standard logic of violence would predict that this innovation should lead
pp. 318-324 12 David Kilcullen and Andrew McDonald Exum, Death From Above, Outrage Down Below, New York Times, May 17, 2009, p. WK13.

us to anticipate an increase in terrorist violence as a result of their eorts to deter defection.13 In contrast, our argument predicts that, in this scenario, militant violence should decrease, both in terms of its frequency and its lethality. The reason is that drone strikes in an area represent a meaningful indication of an increased security risk to militants operating in that area. The increased risk associated with continuing to operate in the targeted areas should apply to any type of militant activity that is vulnerable to drone capabilities, including conducting terror attacks, regardless of whether militants would otherwise conduct operations at their average rate and level of lethality (i.e., the null hypothesis), or if they would otherwise escalate the frequency and lethality of their operations to deter potential defectors (i.e., the alternative logic of violence hypothesis). We thus advance the following hypothesis: H2: All else equal, drone strikes decrease terrorist violence.

2.2.2

Degradation

The second mechanism by which drones can reduce terrorism is through a degradation eect. According to this argument, drone strikes reduce terrorism by taking terrorist group leaders and other high-value individuals (HVIs) o the battleeld, consequently hindering the terrorists ability to produce violence at a sustained rate. Killing operational leaders of al Qaida and its aliated movements is the primary objective of drone strikes.14 Indeed, drone strikes have resulted in the deaths of many top terrorists. According to an Obama administration ocial, the U.S. eliminated at least 20 of al Qaidas 30 top leaders from 2009 to 2012 in Pakistan and Afghanistan.15 In Pakistan alone,
Kalyvas, The Logic of Violence Remarks of President Barack Obama, speech delivered at National Defense University, May 23, 2013. Accessed online at http://www.whitehouse.gov/the-press-office/2013/05/23/ remarks-president-barack-obama. Last accessed on July 5, 2013. 15 Two-Thirds of Top Qaeda Leaders Removed Since 2009: Obama Aide, Reuters, December 18, 2012. Quoted in International Crisis Group, Drones: Myths and Reality in Pakistan, Crisis Group Asia Report N 247, May 21, 2013, p. 22.
14 13

according to the New America Foundation, drone strikes killed 51 militant leaders, including 28 senior al Qaida operatives, between 2004 and early 2013.16 They have also killed several high-level Pakistani and Afghan Taliban and al Qaida-aliated leaders.17 An emerging political science literature investigates the eects of leadership decapitation the killing or capture of militant leaders or other HVIswith a focus on evaluating the group-level eects of killing or capturing top insurgent or terrorist leaders, usually on outcomes such as rates of group collapse or group success.18 The ndings of this literature are mixed. On the one hand, using a large-N approach, Johnston and Price both nd evidence that removing the top leaders of insurgent and terrorist groups helps degrade these organizations, rendering them less lethal, more vulnerable to defeat, and more likely to end quickly than groups that did not suer leadership decapitation.19 On the other hand, Jordan argues that decapitations of terrorist organizations rarely collapse a group quickly or degrade terrorist group capabilities to conduct attacks. Jordan suggests decapitation can have counterproductive eects when performed against larger and older organizations, as well as against religious and separatist organizations.20 We expect drone strikes that kill terrorist leaders will be associated with reductions in terrorist attacks. Previous research convincingly demonstrates that conducting eective terrorist attacks requires skilled individuals, many of whom are well-educated and come from upper middle-class backgrounds.21 Indeed, captured documents conThe Year of the Drone: Leaders Killed, New America Foundation. The data reect gures until January 6, 2013. 17 International Crisis Group, Drones: Myths and Realities, p. 22. 18 Scholars disagree about the conceptualization and measurement of these variables. On leadership decapitation and terrorist group collapse, see Jordan, When Heads Roll, pp. 731-733. On decapitation and group mortality, see Price, Targeting Top Terrorists, pp. 26-33. For a critique of empirical strategies of leadership decapitation scholarship, see Johnston, Does Decapitation Work?, pp. 47-50. 19 Johnston, Does Decapitation Work?; Price, Targeting Top Terrorists. 20 Jordan, When Heads Roll. 21 Alan B. Krueger, What Makes a Terrorist: Economics and the Roots of Terrorism (Princeton, NJ: Princeton University Press, 2007); Ethan Bueno de Mesquita, The Quality of Terror, American
16

taining detailed biographical data on foreign al Qaida militants in Iraq illustrate that among the foreign terrorists who are conventionally known to be more sophisticated than local ghters their most commonly listed occupation prior to arriving in Iraq was that of student. For militants for whom information on experience was available, computers was the most commonly listed experience type, just ahead of weapons.22 In the context of northwest Pakistan, where militant freedom of movement is limited by the threat of drone strikes, we expect that militant groups will be unable to replace senior leaders killed in drone strikes because recruiting and deploying them, perhaps from a foreign country with a Sala jihadist base, will be costly and dicult. This is not to say that leaders killed in drone strikes are irreplaceable. On the contrary, other militants are likely to be elevated within their organization to replace them. But we also anticipate that those elevated to replace killed leaders will be, on average, of lower quality to the organization than their predecessors. Thus, we predict that the loss of leaders will be associated with the degradation of terrorists ability to produce violence. This logic implies Hypothesis 3: H3: All else equal, drone strikes that kill one or more terrorist leader(s) will lead to a decrease in terrorist violence. Based on the contradictory theories and ndings in the literature, however, we cannot dismiss the possibility that killing terrorist leadership might have a counterproductive eect. We thus elaborate Hypothesis 4: H4: All else equal, drone strikes that kill one or more terrorist leader(s) will lead to an increase in terrorist violence.
Journal of Political Science, Vol. 49, No. 3 (July 2005), pp. 515530; Efraim Benmelech, Claude Berrebi, and Esteban F. Klor, Economic Conditions and the Quality of Suicide Terrorism, The Journal of Politics, Vol. 74, No. 1 (January 2012), pp. 113128. 22 Joseph Felter and Brian Fishman, The Demographics of Recruitment, Finances, and Suicide, in Brian Fishman, ed., Bombers, Bank Accounts, and Bleedout: Al-Qadas Road In and Out of Iraq (West Point, NY: Combating Terrorism Center, 2008), pp. 4244.

2.3

Diversion

Another possibility is that drone strikes disrupt terrorist activities in their FATA strongholds by diverting militants to other areas where these activities can be continued. The terrorists themselves have documented the threat of drones and devised countermeasures to mitigate the threat. Captured al Qaida documents show diversion as a recommended strategic response to drones. Interestingly, as a counterintelligence strategy, diversion could push terrorists into rural or urban areas. Each can oer militants a dierent type of protection. Rural areas especially ones with rugged, mountainous terrain oer favorable geography for insurgency and, perhaps, a measure of protection from drones. Urban areas might oer terrorists human camouage, enabling them to blend into the population and limiting the U.S. ability to conduct lethal targeting due to concerns about civilian casualties. This theory implies that drone strikes in FATA might increase militant violence in rural or urban areas. In documents captured from Osama bin Ladens compound in Abbottabad, Pakistan itself an urban area outside of Islamabad, where the al Qaida leader had been hiding since 2005 bin Laden advised al Qaida members there to move to Afghanistans Kunar province for protection from U.S. drones: Kunar is more fortied due to its rougher terrain and many mountains, rivers and trees, and it can accommodate hundreds of the brothers without being spotted by the enemy, wrote bin Laden. This will defend the brothers from the aircraft.23 Other militants have taken refuge in urban areas to elude drone targeting.24 Dozens of al Qaida and Afghan Taliban have been arrested in Balochistan since 2009, when the drone war in FATA escalated.25
Osama bin Laden, Letter dated 7 August 2010 from Zamarai (Usama bin Ladin) to Mukhtar Abu al-Zubayr, SOCOM-2012-0000015-HT, May 2012. Accessed online at http://www.ctc.usma. edu/posts/socom-2012-0000015-english. Last accessed July 2, 2013 24 See, for instance, a report in The Times, dated August 8, 2009, which was accessed at http:// www.thetimes.co.uk/tto/news/world/asia/article2611093.ece. Last accessed June 11, 2013. 25 These statistics came from an assessment by the Institute for Conict Management, a South Asian think tank, based primarily on reporting from Pakistani newspapers. It was accessed online
23

If drone strikes systematically divert militants to other locations, spatial patterns of observed violence in areas around FATA should increase. This argument implies the following testable hypothesis: H5: All else equal, drone strikes increase militant violence in neighboring areas not targeted by drones.

2.4

Duration

Finally, there is also considerable debate about drone strikes short-term versus long- term utility. Some suggest any eect of drone strikes is tactical and short term. In this view, a drone strike might aect a militant groups operations for several days, but generally speaking these strikes do not signicantly curtail militant activities. Others suggest, however, that drone strikes have longer-lasting operational or strategic eects. In this view, drone strikes serve to weaken or strengthen militants over time. The former argue that because of drones persistent surveillance and targeting capabilities, drones are a game changer that have signicantly enhanced counterterrorism capabilities and eectiveness. The latter argue that drone strikes result in boons in militant mobilization that enhance militant groups overall ability to conduct violent attacks. These contrasting arguments generate two additional hypotheses: H6: Drone strikes have an extended violence-reducing eect. H7: Drone strikes have an extended violence-increasing eect.26
at http://www.satp.org/satporgtp/countries/pakistan/Balochistan/index.html. Last accessed on June 10, 2013. 26 For both hypotheses, extended is dened as longer than one week.

10

Empirical Strategy

In this section, we describe our methodology for evaluating the eects of drones. Our study spans from January 2007 through September 2011. We analyze how drone strikes in the FATA region of Pakistan aect militant violence both in FATA and in other parts of Pakistan and neighboring areas of Afghanistan We use the agency-week as our unit of analysis. Agencies in FATA are akin to districts in many other countries. In the present study, they include Bajaur, Khyber, Kurram, Mohmand, North Waziristan, Orakzai and South Waziristan. Agencies correspond with the geographic distribution of militant groups in FATA more closely than does any other administrative unit, making agency-level analysis useful for tracking secular dierences in violence that might arise because of heterogeneity in the militant groups operating in the region.27 Indeed, as Figure 3 shows, FATAs seven agencies suered varying levels of violence over time.28 Our empirical approach also includes spatial analysis, specically, tests for a spillover eect of drone strikes. We examine whether drone strikes eect militant violence in neighboring areas in both Pakistan and Afghanistan using varying radii from the center of each agency. We increase the radius of the neighborhood for spatial
27 On variation in militant organizations across FATA agencies, see, for example, Shuja Nawaz, FATA A Most Dangerous Place: Meeting the Challenge of Militancy and Terror in the Federally Administered Tribal Areas of Pakistan, Center for Strategic and International Studies, January 2009; Imtiaz Gul, The Most Dangerous Place: Pakistans Lawless Frontier (New York: Viking, 2010); and Brian Fishman, The Battle for Pakistan: Militancy and Conict across the FATA and NWFP, New America Foundation, April 2010. 28 Although the rst documented drone strike in FATA occurred in June 2004, our analysis focuses primarily on events between early 2007 through late 2011. Through the end of 2006, only six drone strikes were reported. The number of strikes in 2007 ve nearly equaled the number that had been conducted in the entire previous history of the war. This number would increase dramatically in the following years, peaking in 2010 at 122 and declining to 73 and 48 in 2011 and 2012, respectively. Temporal variation in drone targeting at the local level during the period under study is an important part of our identication strategy. Likewise, 2007 is also an ideal starting point because, unlike in previous years when levels of violence in the region were fairly at, there was signicant variation in militant violence starting in 2007 both across agencies and in FATA overall due to conict escalation largely unrelated to drone strikes. Our data allow us to trace this violence to particular locations and times, giving us some ability to assess possible endogeneity in the statistical results.

11

analysis from 25 km to 150 km in increments of 25 km.29 This approach enables us to examine how far any spillover eects of drone strikes appear to extend and track changes, if any, in the eect of drone strikes on militant activities in response to increasing distance from the targeted area.

3.1

Identifying Assumptions

Our empirical strategy is motivated by the fact that the week-to-week timing of drone strikes in FATAs agencies is subject to a range of quasi-random factors. Many factors unrelated to militant violence are likely to inuence whether a drone is used in a given week. Drone strikes clearly are not conducted at random, but there is reason to believe the week-to-week incidence of drone strikes our temporal unit of analysis is only weakly related to levels of terrorist violence. This is because in practice, the ability to conduct drone strikes depends on a complex range of factors meteorological, bureaucratic, and technological, among them whose unpredictability from week-to-week means that a drone strike on a terrorist target identied this week might be conducted this, next week, the following week, or not at all. We describe seven such complicating factors below. First, weather patterns play a signicant role in drone operators ability to identify and strike targets, for example, introducing a random component into the timing of drone strikes when they are examined in relatively modest intervals. This random element in the timing of drone strikes is not only observed by journalists, but also by al Qaidas leadership in multiple theaters of operation.30 Recently declassied al Qaida documents show, for example, that Osama bin Laden once advised operatives
The average radius of a FATA agency is 32 kilometers. Nelly Lahoud, Stuart Caudill, Liam Collins, Gabriel Koehler-Derrick, Don Rassler, and Muhammad al Ubaydi, Letters from Abbottabad: Bin Ladin Sidelined? (West Point, NY: Combating Terrorism Center, 2012); Associated Press, The Al-Qaida Papers Drones, February 21, 2013. Accessed online at http://hosted.ap.org/specials/interactives/_international/ _pdfs/al-qaida-papers-drones.pdf. Last accessed July 5, 2013.
30 29

12

not to move from their safe houses on clear days.31 This is consistent with information from the U.S. sources that cloudy days obscure satellites and make it more dicult to view objects on the ground. Moisture and electrical interference from storms may also hinder operations.32 Second, drones are a scarce commodity and are in high demand across the theaters in which the U.S. conducts counterterrorism missions. Thus, the availability of drones in FATA whether for intelligence, surveillance, and reconnaissance (ISR) missions to identify terrorist targets, or for lethal targeting itself varies with changing ISR requirements and priorities assigned to other theaters.33 Third, key technological aspects required to conduct drone strikes can, and reportedly have, varied, at times semi-randomly. A UAV operator relies on imagery from onboard sensors for target detection, but the quality and timeliness of this imagery depends on data link and bandwidth limits. Low update rates and long communication delays, for example, will produce slow and discontinuous imagery to a UAV pilot, encouraging operators to adopt a go-and-wait strategy.34 Fourth, human factors specic to drone operations introduce another possible element of randomness into the incidence and timing of strikes. A recent study found that humans ability to operate drones as planned has varied due to the dierent types of drones that may be deployed because not all have the same software for mission planning and some are equipped with diering user interfaces with which a pilot might be more or less familiar. As with any kind of software, familiarity matters; a pilots
Letter dated 7 August 2010 from Zamarai (Usama bin Ladin) to Mukhtar Abu al-Zubayr, SOCOM-2012-0000015-HT, May 2012, pp. 2-3. 32 Robert Tilford, Al-Qaedas Anti Drone Tactics Discussed In Bin Laden Letter, The Examiner, March 3, 2012. Accessed online at http://www.examiner.com/article/ al-qaeda-s-anti-drone-tactics-discussed-bin-laden-letter; for a detailed analysis of the bin Laden documents, see Lahoud et al., Letters from Abbottabad, pp. 32, 4647. 33 Greg Miller, Military Drones Aid CIAs Mission, Washington Post, October 3, 2010, p. A1; and Adam Entous, Julian E. Barnes, and Siobhan Gorman, CIA Escalates in Pakistan: Pentagon Diverts Drones From Afghanistan to Bolster U.S. Campaign Next Door, Wall Street Journal, October 2, 2010. 34 Jason S. McCarley and Christopher D. Wickens, Human Factors: Implications of UAVs in the National Airspace, (Urbana-Champaign, IL: University of Illinois, 2005), p. 7.
31

13

ability to conduct a given strike may depend on which software suite he or she was trained.35 Fifth, bureaucratic and logistical factors as mundane as the work schedules of key lawyers and decision-makers in the United States, who are required to provide legal counsel and authorization before a strike can occur, might lead a strike to happen or not for reasons that have little to do with the availability of a strike target. Key principals are many time zones away from Pakistan in Washington, D.C., and authorizations apparently can take hours or days to receiveif they are received at all.36 Sixth, the timing of when a known terrorist presents a clean shot is likely to be largely random on a week-to-week basis, meaning the treatment could plausibly have occurred in the preceding or following agency-week as in the current one, making weekly comparisons of dierences in violence across agencies a credible causal estimate. Seventh, and nally, a key to identication based on any of these factors is to make the unit-of- analysis relatively small temporally. As the temporal unit of aggregation increases, the validity of the identifying assumption goes down. The longer the window, the less factors like the ones described above will matter, consequently reducing condence that the relationship identied is causal. As a result, we analyze the eects of drone strikes at the weekly level instead of a higher level of aggregation, such as the month or quarter.
35 Kevin W. Williams, A Summary of Unmanned Aircraft Accident/Incident Data: Human Factors Implications, (Washington, DC: United States Department of Transportation Federal Aviation Administration, 2004), p. 12. 36 Afsheen John Radsan and Richard W. Murphy, Due Process and Targeted Killing of Terrorists, Cardozo Law Review, Vol. 31, No. 2 (May 2009), pp. 412-413.

14

3.2

Estimation

In the analysis presented below, we estimate two-level xed-eect models with both agency and temporal (week) xed eects and a spatial lag of drone strikes (2FESL).37 Fixed-eects regression is a standard econometric approach to panel data analysis.38 Letting i denote the cross sectional index (i.e., the agency) and t the time index (i.e., the week), a two-level xed eect equation is given by:

yit = i + xit + ht +

it

(1)

where y measures the incidence of terrorism, x is the number of drone strikes, i are unobserved agency xed eects, and ht are time (week) xed eects. Agency xed eects account for all the time-invariant dierences between agencies, such as terrain and elevation, which could otherwise confound cross-sectional analysis. In practice, the xed eects are included to control for unobserved factors that might vary by agency, as well as secular quarterly trends in levels of conict violence. Week xed eects allow us to control for time-specic dierences such as heavy snow, ooded terrain, natural disasters, and religious festivals, which could potentially determine combatant activity. In addition to the xed-eects regressions described above, we also estimate models that include a spatial lag. Phillips and Sul (2003, 2007) have shown that cross-sectional dependence may cause panel OLS estimates to be biased and inconsistent. Including a spatial lag enables us to directly model cross-sectional dependence in the regression.39 A spatial lag model with two-level xed eects assumes
The spatial lag in spatial econometrics is equivalent of the temporal lag in time-series analysis. It is the value of the dependent variable for the unit(s) that constitute(s) the space of the observation under consideration, which in this article is formed by all agencies or districts in Afghanistan and Pakistan falling within a certain distance from the centroid of the agency under consideration. 38 See especially Jerey M. Wooldridge, Econometric Analysis of Cross Section and Panel Data (Cambridge, Mass.: MIT Press, 2002); and Joshua D. Angrist and Jorn-Steen Pischke, Mostly Harmless Econometrics: An Empiricists Companion (Princeton, NJ: Princeton University Press, 2009). 39 See, for instance, R.J. Franzese, Jr. and J. C. Hays, Spatial Econometric Models of CrossSectional Interdependence in Political Science Panel and Time-Series Cross-Section Data, Political
37

15

the following form:

yit = i +
j =i

wij yjt + xit + ht +

it .

(2)

where is the spatial autoregressive coecient, which measures the general strength of spatial dependence, wij is an element of the spatial weight matrix reecting the degree of connection between two units i and j, yjt is the measure of militant violence for unit j during time period t, xit is the number of drone strikes in unit i at time t, i are unobserved agency-specic eects, and ht are quarterly time eects.

3.3

Data and Variables

To examine the eect of drone strikes, we combined detailed data on US drone strikes in FATA originally collected by researchers at the New America Foundation (NAF)40 with incident-level data on terrorist activities in FATA during the same time period compiled in the National Counterterrorism Centers (NCTC) Worldwide Incidents Tracking System (WITS)41 and incidents of militant violence against tribal elders compiled by the South Asia Terrorism Portal.42 Incidents from each data source were georeferenced according to the reported locations of the incidents in the media accounts used to track and cross-reference each drone strike and militant attack. The NAF data on drone strikes include information on the incidence, date, and location of each strike, the high and low estimates of fatalities that have occurred in each strike, deaths of militant leaders in drone strikes, and the sources of information
Analysis, Vol. 15, No. 2 (2007), pp. 140164. We also performed the Pesaran cross-sectional dependence (CD) test on the residuals of the estimated models. See M. Hashem Pesaran, A Simple Panel Unit Root Test in the Presence of Cross-Section Dependence, Journal of Applied Econometrics, Vol. 22, No. 2 (March 2007), pp. 265312. The results of the CD test are available upon request. 40 Peter Bergen and Katherine Tiedemann, New America Foundation Drones Database, New America Foundation, 2011. 41 Worldwide Incidents Tracking System, National Counterterrorism Center, 2012. 42 The SATP data were accessed online at http://www.satp.org/satporgtp/countries/ pakistan/database/Tribalelders.htm. Last accessed June 15, 2013.

16

that were used to compile each summary. The data were compiled from reports in reputed international and Pakistani news media sources. The WITS database uses fairly standard criteria in coding incidents as terrorist attacks. To be included as a terrorist attack in the WITS database, activities were required to be incidents in which sub-national or clandestine groups or individuals deliberately or recklessly attacked civilians or non-combatants, including military personnel and assets outside war zones.43 Moreover, attacks have to be initiated and executed by non-state militants. Spontaneous violence, hate crimes and genocides are excluded from the database. Using data that focuses on terrorist incidentsviolence against civilian rather than military targetsis justiable for both theoretical and empirical reasons. Theoretically, Kalyvas (2006) argues that the combatants are likely to target civilians selectively in their zones of control as a result of real or perceived spying by civilians. A similar narrative is often used to describe militant responses to drone strikes in FATA: militants believe drone strikes are the result of informant betrayal, and thus target suspected informants44 Along these lines, tribal elders typically associated with a local incumbency have been cited as particularly common targets.45 We use data on militant attacks on tribal elders in Pakistan from 2005 through 2011 compiled by SATP.
46

The inclusion

of this variable is warranted by the suggestion that drone strikes increase attacks on tribal elders whom militants suspect of collaborating with U.S. or Pakistani military or intelligence services. Table 1 summarizes the variables and data sources used in our analysis. We focus
Worldwide Incidents Tracking System, National Counterterrorism Center, 2012. Dashiell Bennett, Pakistani Death Squads Target Informants Who Help Drone Attacks, The Atlantic Wires, December 29 2011. 45 Brian Fishman, The Battle for Pakistan: Militancy and Conict across the FATA and NWFP, Counterterrorism Strategy Initiative Policy Paper. The New America Foundation. Last accessed April 2010. 46 The SATP data were compiled from open-source media reports, primarily from south Asian sources, by the Institute of Conict Management, New Delhi.
44 43

17

on drone strikes and four key measures of terrorist activity. Our data set contains information on the following variables at the agency-week level: UAV: The number of drone strikes in a given agency and week. HVT: The number of senior leaders killed by drone strikes in a given agency and week. (Source: New America Foundation) Incidents: The number of militant incidents or attacks in a given agency and week. Lethality: The number of dead and wounded in terrorist incidents or attacks in a given agency and week. IED Attacks: The number of IED attacks conducted in a given agency and week. Suicide Attacks: The number of suicide attacks conducted in a agency and week. Attack on Tribal Elder(s): The number of militant attacks against tribal elders in a given agency and week.

3.4

Descriptive Statistics

For this study, we constructed an agency-week dataset. The time-series spans from January 1, 2007 through September 30, 2011. Descriptive statistics of key variables over this time period are shown in Table 1. Figures 13 illustrate the variation in terrorist attacks and drone strikes over space (Figure 1) and time for all of FATA (Figure 2 ) and for its constituent agencies (Figure 3). Figure 2 shows the monthly time trend of drone strikes and terrorist attacks for all of FATA from 2007 through September 2011. Militant attacks began trending upward 18

in mid-2007, peaking in early 2009 before declining back to roughly mid-2007 levels by Fall 2011. Drone strikes (left axis) were relatively rare until Fall 2008 before August 2008, when four strikes were conducted, there had never been more than one strike in a month. At the agency level, gure 3 shows that North Waziristan closely mirrors the macro trend, with trends uctuating more in South Waziristan and Khyber while being relatively rare elsewhere in FATA.

Empirical Results

A cursory look might suggest the former: as gure 2 shows, violence rose from 2007 until 2009 and was as high in September 2011, when our time-series ends, as in any year since 2007. Yet gure 2 also shows that the rise of drone strikes appears to have been a response to a deteriorating environment in which terrorist violence was increasing dramatically. It is thus plausible that the drone wars escalation occurred as a result of real and anticipated increases in terrorist violence. Given the upward trend in terrorist violence prior to the escalation of the drone campaign, and the observed variation in terrorist attacks across agencies, we use both week- and agency-xed eects to mitigate confounding impacts of secular time trends in terrorist violence and of agency-specic dierences, using these within regressions to estimate the average eect of drone strikes within agencies over time.47
As a robustness test, we also ran regressions using a series of model specications including ordinary least square (OLS) and involving temporal lags, spatial lags, and rst-dierences, both with and without xed eects. We also conducted two panel unit-root tests, the Breitung and Pesaran tests, which both allow for cross-sectional dependence. Results of these tests are available on request. Jorg Breitung, The Local Power Of Some Unit Root Tests For Panel Data, in Advances in Econometrics, Vol. 15: Nonstationary Panels, Panel Cointegration, and Dynamic Panels (New York: JAI Press, 2000), pp. 161-178; and Pesaran, A Simple Panel Unit Root Test.
47

19

4.1

Disruption

Table 2 presents the 2FESL estimates of drone strikes on four measures of militant violence. The spatial lag included in the 2FESL models measures the value of our dependent variables in the districts falling within 75 km of the centroid of the agency in which strikes occurred. To test Hypotheses 1 and 2, we examine ve dierent measures of militant violence: the frequency of attacks, the lethality of attacks, the number of IED attacks, the number of suicide attacks, and the number of attacks on tribal elders. The results do not support Hypothesis 1 that drone strikes are associated with increased terrorism. On the contrary, they support our hypothesis, (Hypothesis 2), that that drone strikes are associated with decreases in militant violence. We nd no evidence in support of the competing hypothesis (Hypothesis 1) that drone strikes increase violence. We discuss these results in more detail below. The 2FESL estimates in column 2 of table 2 show that drone strikes are associated with a decrease in militant attacks of approximately 24 percentage pointsa result that is statistically signicant at the one percent level. From 2007 through 2011, the average agency suered roughly 0.88 militant attacks per week. During weeks in which a drone strike occurred, agencies suered an average of about 0.68 attacks. Given that drone strikes are associated with reductions in insurgent attacks in the areas where they occur, it makes sense that drone strikes might also be negatively associated with the lethality, or quality, of attacks in those same areas. Consistent with Hypothesis 2, the estimates presented in column 2 of table 2 suggest that the lethality of militant attacks declines by more than 36.5 percent as a result of a drone strike in a given week. On average, 2.77 people were killed or injured in militant attacks in FATA between 2007 and the end of the third quarter of 2011. This gure would decline substantially to 1.76 per week as a result of a single drone strike if the

20

number of drone strikes would increase by one per agency-week.48 The disruption hypothesis also implies that drone strikes should reduce militants ability to conduct complex and coordinated attacks like IED and suicide attacks. We nd support for these propositions in our econometric tests. Drone strikes are negatively associated with the number of IED attacks in FATA during the period studied. Based on the estimates in column 3 in table 2, a drone strike is associated with a 21-percentage point reduction in IED attacks. The marginal eect translates into an estimated decrease in IED attacks from an average of 0.32 per agency-week to 0.25 per agency-week when there is one drone strike. Regarding suicide attacks, the coecient in column 4 of table 2 suggests that drone strikes are also associated with reductions in these tactics. This result is signicant at the one percent level. Suicide attacks are relatively rare but extremely high-prole events: the mean number of suicide attacks per agency per week is 0.02, or about one per agency every year. The point estimate appears small, but the marginal eect translates into an almost 67 percent decline in the number of suicide attacks in a week with one drone strike. Thus, the average number of weekly suicide attacks in FATA, which is 0.14 per week during the period under consideration, would decline to 0.05 per week as a result of one drone strike per agency-week. On balance, the evidence is clearly consistent with Hypothesis 2 the disruption hypothesis and not with the argument that drone strikes trigger increased violence (Hypothesis 1).

4.2

Degradation

Given that killing terrorist leaders or HVIs in terrorist organizations is the purpose of drone strikes, we evaluate whether patterns of militant attacks dier following strikes
It is important to note that the estimate of decline in lethality of militant attacks is based on an assumption of a constant linear relationshipan assumption that may or may not be correct. The predicted decline is probably an overstatement of the impact drones could realistically have, simply because even at the peak of the drone campaign in 2010, when the number of drone strikes was two and a half times larger than the previous year (119 in 2010, versus 53 in 2009), the number of drones per campaign-week in 2010 was 0.33, while it was 0.14 in 2009.
48

21

in which a militant leader was killed. Table 3 provides tests of Hypotheses 3 and 4 against the four metrics of militant violence examined here using the same 2FESL specications as in table 2. The results are largely consistent with Hypothesis 3 that killing militant leaders is associated with decreased violence. There is little support for Hypothesis 4, that killing HVIs has counterproductive eects on violence. Controlling for the number of drone strikes per agency-week, the rst column of table 3 shows that drone strikes that kill a HVI are associated with reductions in the number of militant incidents that occur. This result is statistically signicant at the one-percent level. There is, however, weaker evidence that HVI removals reduce militant lethality and IED attacks.49 Overall, the evidence is somewhat consistent with the argument that individuals matter for a terrorist organizations ability to produce violence at sustained rates. Along with other evidence from macro-level studies of leadership decapitation, the present results suggest that critics who argue against the ecacy of removing key gures may be overemphasizing the extent to which such individuals can be readily replaced.50

4.3

Diversion

A potential concern with the previous ndings is that it is possible that drone strikes do not actually reduce terrorist violence, but rather displace it. While drone strikes might cause militant activities to decline in the targeted agencies, they may cause an escalation in militant violence in proximate areas that are not subject to drone strikes if militants move their operations in response to UAV targeting in FATA. The concern
These estimates may be more imprecise than the statistical results suggest, as a result of heterogeneity in the measurement of the HVI variable. Although U.S. government ocials consider terrorists targeted by drone strikes target as senior leaders or high-value individuals (HVI), the U.S. government has not publicly stated the criteria it uses to identify individual terrorists as senior leaders or HVIs. Available information on individuals identied as leaders killed in drone attacks suggests a degree of heterogeneity 50 Johnston, Does Decapitation Work?; and Price, Targeting Top Terrorists.
49

22

with spillover eects is not just academic; media reporting points to it as a possible policy concern.51 To address this issue, we extend the above analysis by estimating the eect of drone strikes beyond the seven FATA agencies in neighboring areas within various distances of agencies where strikes have occurred. To do this, we vary the radius of struck agencys neighborhood from 25 km to 150 km by increments of 25 km. By testing the eect of drone strikes on militant violence in geographic units that expand outward to varying distances, we assess how drone strikes aect militancy beyond FATA. The results of potential spillover eects are presented in table 4. Each column in these tables presents estimates of the eect of drone strikes on militant violence in a neighborhood of a particular radius, beginning with a radius of 25 km in column 1 and ending with a radius of 150 km in column 6. In the rst two rows of table 3, we present the estimates of the eect of drone strikes on the number of militant attacks in the neighborhood. The sign of the drone strike estimate is negative up to 125 km and is statistically signicant at 25 km and 100 km. The coecient becomes positive at a radius of 150 km, but the positive coecients are statistically insignicant. The estimates of the eect of drone strikes on the lethality of militant attacks and IED attacks in the neighborhood display a pattern similar to the estimates of militant attacks. For suicide attacks, the results deviate slightly from the trend observed for the other dependent variables. Unlike the other dependent variables, the coecient associated with suicide attacks does not change signs from negative to positivethe results remain negative for each of the radii tested. The evidence in support of a favorable spillover eect on suicide attacks is somewhat inconclusive, however: only the coecient associated with the 25 km radius variable is statistically signicant at conventional levels.
Alex Rodriguez, U.S. Concerns Grow as Militants Move Bases Along Pakistan Border, LA Times, November 7, 2010.
51

23

Duration

If the evidence indicates that drone strikes help disrupt and degrade terrorist group operations in Pakistan, a nal question of considerable importance is whether the eect is long-lasting. There is some evidence that it extends over a few weeks and is thus relatively long in duration. Using a model that includes ve one-week lags of drone strikes, there is a signicant negative relationship between strikes that occurred ve weeks earlier and both attack lethality and suicide attacks. Moreover, the sign of the coecients of UAV at t-5 are negative for both number of incidents and IED attacks but are not statistically signicant at conventional levels. There is also limited evidence that drone strikes might have a deterrent eect that lasts between two and ve weeks. Indeed, in agencies contiguous to those that were struck, the lethality of militant attacks has decreased, on average, in the weeks following a drone strike in a neighboring agency. Several of these results are statistically signicant at the ve-percent level. The evidence certainly suggests that drones disruptive eect is strongest in the week when strikes occur, but the negative results on the lagged variables both in agencies where drone strikes occurred as well as neighboring agencies suggests a possible deterrent mechanism at work as well. Overall, these ndings are broadly consistent with Hypothesis 6, and at odds with Hypothesis 7.

Implications

This paper oers a systematic analysis of the relationship between U.S. drone strikes and militant violence in northwestern Pakistan and eastern Afghanistan. Our analysis suggests that drone strikes are negatively associated with various measures of militant violence, both within individual FATA agencies and their immediate neighborhoods. There is also evidence to suggest that the negative association between drone strikes 24

and three measures of militant violence incidents, lethality and IED attacks changes sign as we increase the neighborhood radius to exceed 125 km. This may or may not be indicative of drone strikes causing militant activities to move farther away from FATA. With the current research design and data, we are unable to make any denitive conclusion regarding the spillover eects. We are also not in a position to make strong causal claims about the impact of drone strikes on militant violence. There is evidence of a strong negative contemporaneous correlation between drone strikes and various measures of militant violence. This may indicate that that drone strikes have important counterterrorism dividends, but caution should be exercised in inferring causality due to the selection bias inherent in the data despite the econometric techniques used to mitigate selection bias in our regression estimates. Still, our ndings appear consistent with the hypothesis that new technologies specically, remote means of surveillance, reconnaissance and targetingare able, at least in certain key areas of northwest Pakistan and eastern Afghanistan, to disrupt and degrade militants in ways that compensate for an incumbent governments lack of physical presence in and control over these areas, and can consequently limit both the frequency and the lethality of militant attacks. This suggests that new technologies that provide information previously available only to actors with a strong physical presence in a geographic area can alter conventionally accepted logics of violence in civil war.52 The implication of these ndings, of course, is that as technology continues to become increasingly sophisticated, warfare is likely to become increasingly virtual but not bloodless. Adversaries not only governments, but also non-state actors such as insurgents, terrorists and criminal organizations will adapt their organization and behavior to reduce their vulnerability to adversaries countermeasures, and some are likely to try leveraging these technologies for their own use against their state
52

Kalyvas, The Logic of Violence in Civil War.

25

and non-state enemies. In the near term, for example, insurgents may increasingly abandon rural areas like FATA in favor of urban areas of the sort that insurgents have traditionally eschewed, but that may now oer greater protection from drones and other sophisticated countermeasures. However, the operational constraints on urban operations might restrain militants use of violence, just as they have for state actors.53

On urban insurgency, see Brian Michael Jenkins, The Five Stages of Urban Guerrilla Warfare: Challenge Of The 1970s (Santa Monica: Rand Corporation, 1971); Fair, Urban Battle Fields of South Asia: Lessons Learned from Sri Lanka, India and Pakistan (Santa Monica, CA: RAND Corporation, 2005); and Paul Staniland, Cities on Fire: Social Mobilization, State Policy, and Urban Insurgency, Comparative Political Studies, Vol. 43, No. 12 (December 2010), pp. 1623-1649

53

26

Figure 1: Drone Strikes and Militant Attacks in FATA & its Neighborhood

27

Figure 2: Time Trends in Drone Strikes and Terrorist Attacks

28

Figure 3: Time Trends in Drone Strikes and Militant Attacks by Agency

29

Table 1: Summary Statistics: FATA & Neighborhood

FATA Mean 0.153 .0231 0.880 2.777 0.316 0.020 0.013 1729 50822 0.605 0.181 1.333 14.019 0.734 0.149 0.112 0 0 0 0 0 0 0 8 3 13 285 7 2 1 0.183 0.689 0.063 0.009 0.732 6.759 0.360 0.102 0 0 0 0 17 361 11 4 0.681 2.148 0.229 0.038 3.044 21.982 1.592 0.559 37791 S.D.* Min. Max. Mean S.D.* Min. Max. Mean S.D.* Neighborhood Afghanistan Min. Max. 0 0 0 0 77 1305 70 28 Mean 1.824 7.696 0.658 0.085 Pakistan S.D.* 5.500 61.135 2.674 0.744 13091 Min. 0 0 0 0 Max. 91 2219 49 21

Variable

30

UAV HVI Incidents Lethality IED Attacks Suicide Attacks Attacks on Tribal Elders

Number of Observations

* Standard Deviation

Table 2: Drone Strikes and Terrorist Violence: 2FESL Estimates

Incidents UAV Constant Observations AIC BIC -0.048*** (0.010) 0.023 (0.020) 1729 473.224 620.517 Lethality -0.247*** (0.090) 0.136 (0.200) 1729 8998.330 9145.623 IED Suicide Attacks on Elders -0.001** (0.001) 0.005** (0.002) 1729 -7594.435 -7447.142

-0.016*** -0.003*** (0.005) (0.001) 0.004 (0.010) 0.000 (0.002)

1729 1729 -1448.116 -6737.893 -1300.823 -6590.600

Standard errors in parentheses * p<0.10, ** p<0.05, *** p<0.01

Table 3: Leaders Killed and Militant Violence: 2FESL Estimates Incidents UAV HVI Constant Observations AIC BIC -0.012 (0.010) -0.092*** (0.040) 0.205*** (0.008) 1729 417.207 433.572 Lethality -0.136 (0.100) -0.057 (0.200) 0.649*** (0.08) 1729 8751.883 8768.249 IED -0.022*** (0.007) -0.002 (0.01) 0.075*** (0.004) Suicide -0.001 (0.001) -0.001 (0.002) 0.005*** (0.001)

1729 1729 -1606.601 -6977.664 -1590.235 -6961.298

Standard errors in parentheses * p<0.10, ** p<0.05, *** p<0.01

31

Table 4: Drone Strikes and Neighborhood Militant Violence

Neighborhood Radius Dependent Variable 25 km Incidents Lethality IED Attacks Suicide Attacks Observations -0.042*** (0.010) -0.252*** (0.090) -0.014*** (0.005) -0.003*** (0.001) 50 km -0.022 (0.010) -0.152* (0.080) -0.002 (0.009) -0.003 (0.003) 75 km -0.009 (0.006) -0.037 (0.040) -0.002 (0.003) -0.001 (0.001) 100 km -0.007* (0.004) 0.081 (0.050) -0.001 (0.002) -0.000 (0.0008) 1722 125 km -0.004 (0.004) 0.055 (0.040) -0.001 (0.002) -0.000 (0.0006) 150 km 0.002 (0.003) 0.038 (0.030) 0.002 (0.002) -0.000 (0.0004)

Standard errors in parentheses. Coecient estimates for drone strike (UAV) variable. Intercept estimates not presented. * p<0.10, ** p<0.05, *** p<0.01

32

Table 5: The Duration of the Eect of Drone Strikes

(1) Incidents -0.030*** (0.0049) 0.0056 (0.010) -0.0011 (0.0088) 0.017* (0.0084) -0.0034 (0.012) -0.0087 (0.0098) -0.0045 (0.028) -0.0047 (0.019) 0.0043 (0.059) -0.015 (0.059) -0.0047 (0.044) 0.077 (0.041) 0.16*** (0.034) 0.13*** (0.034) 0.077 (0.041) (2) Lethality -0.11*** (0.0086) -0.033 (0.044) -0.045 (0.032) -0.061 (0.035) 0.088 (0.14) -0.16*** (0.027) 0.46 (0.56) -0.34*** (0.061) -0.20 (0.32) -0.24 (0.34) -0.20** (0.061) (3) IED (4) Suicide

VARIABLES UAV UAVt-1 UAVt-2 UAVt-3 UAVt-4 UAVt-5 Neigborhood UAVt-1 Neigborhood UAVt-1 Neigborhood UAVt-2 Neigborhood UAVt-3 Neigborhood UAVt-4 Neigborhood UAVt-5 Incidentt-1 Incidentt-2 Neighbohood Incidentt-1 Lethalityt-1 Lethalityt-2 Neighbohood Lethalityt-1 IEDt-1 IEDt-2 Neighborhood IEDt-1 Suicidet-1 Suicidet-2 Neighborhood Suicidet-1 Constant

-0.011** -0.0023*** (0.0038) (0.000076) 0.0054 -0.0010** (0.0053) (0.00040) 0.0080 0.00047 (0.0066) (0.0010) 0.016** -0.0014*** (0.0064) (0.00019) -0.0090** 0.0019 (0.0029) (0.0015) -0.0013 -0.0033*** (0.0096) (0.00028) 0.0047 -0.00019 (0.013) (0.0043) 0.011 0.012* (0.017) (0.0050) -0.016 -0.0058 (0.016) (0.0040) -0.0062 -0.0011 (0.025) (0.0038) -0.000024 0.0077 (0.025) (0.0059)

0.0053 (0.024) 0.00087 (0.0098) -0.043 (0.079) 0.23*** (0.043) 0.049*** (0.013) 0.038 (0.039) 0.050 (0.055) -0.017 (0.020) 0.014 (0.044) 0.0050*** (0.00053) 1,694

0.14*** 0.72*** (0.010)33 (0.057) 1,694 1,694

0.049*** (0.0057) 1,694

Observations

Appendix A: Robustness Tests

Here we evaluate whether the results are sensitive to certain time periods. The drone war escalated signicantly in 2008 relative to previous years; drone strikes increased again in both 2009 and 2010, and remained higher in 2011 than in 2008. Given that we cannot rule out that unobserved changes in FATA, starting approximately in 2008, drive this change, we restrict the sample to 2008 and later to test whether the patterns that we observed in the previously discussed results hold during this later period. Table A-1 shows that the main ndings do hold when we estimate the 2FESL specication for each of the measures of violence with the sample restricted to observations after 2007. In Table A-2, we extend our analysis to an additional three years by starting from the beginning of 2004, the year of the rst-known drone strike in FATA. The results are remarkably similar to the main ndings. Table A-1: Drone Strikes and Terrorist Militant Violence: 2008-2011
Incidents UAV Constant Observations AIC BIC -0.034*** (0.142) 0.079*** (0.025) 1456 480.277 607.080 Lethality -0.194*** (0.089) 1.137*** (0.534) 1456 7792.078 7918.881 IED Attacks -0.012** (0.005) 0.040** (0.015) 1456 -1051.775 -924.9727 Suicide Attacks -0.001*** (0.001) 0.004*** (0.001) 1456 -5902.082 -5891.515 Attacks on Elders -0.001* (0.001) 0.005*** (0.002) 1456 -6176.432 -6049.629

Standard errors in parentheses * p<0.10, ** p<0.05, *** p<0.01

Table A-2: Drone Strikes and Militant Violence: 2004-2011

Incidents UAV Constant -0.051*** (0.010) 0.120 (0.012) Lethality -0.227*** (0.076) 0.035 (0.086) 2912 13654.120 13893.180 IED Attacks -0.021*** (0.005) 0.005 (0.006) 2912 -3737.697 -3498.633 Suicide Attacks Attacks on Elders -0.003*** (0.001) 0.0003 (0.001) 2912 -12867.330 -12628.270 -0.002*** (0.001) 0.002** (0.001) 2912 -13228.340 -12989.270

Observations 2912 AIC -273.484 BIC -34.42016

Standard errors in parentheses * p<0.10, ** p<0.05, *** p<0.01

34