explaining support for combatants during wartime- a survey exp in afghanistan

8/12/2019 Explaining Support for Combatants During Wartime- A Survey Exp in Afghanistan

1/27

American Political Science Review Vol. 107, No. 4 November 2013

doi:10.1017/S0003055413000403 c American Political Science Association 2013

Explaining Support for Combatants during Wartime:A Survey Experiment in AfghanistanJASON LYALL Yale UniversityGRAEME BLAIR Princeton UniversityKOSUKE IMAI Princeton University

H ow are civilian attitudes toward combatants affected by wartime victimization? Are these effectsconditional on which combatant inicted the harm? We investigate the determinants of wartimecivilian attitudes towards combatants using a survey experiment across 204 villages in vePashtun-dominated provinces of Afghanistanthe heart of the Taliban insurgency. We use endorsement experiments to indirectly elicit truthful answers to sensitive questions about support for different combat-ants. We demonstrate that civilian attitudes are asymmetric in nature. Harm inicted by the International Security Assistance Force (ISAF) is met with reduced support for ISAF and increased support for the Taliban, but Taliban-inicted harm does not translate into greater ISAF support. We combine amultistage sampling design with hierarchical modeling to estimate ISAF and Taliban support at theindividual, village, and district levels, permitting a more ne-grained analysis of wartime attitudes than previously possible.

H ow does the victimization of civilians affecttheir support for combatants during wartime?Arethe effects of violence onattitudesuniformacross warring parties, or are they conditional on whoinicted the harm? Civilians, existing accounts tacitlyassume, are simply guided by a logic of survival that

Jason Lyall is Associate Professor of Political Science, Departmentof Political Science, Yale University, New Haven, CT 06520 ([email protected]), http://www.jasonlyall.com.

Graeme Blair is a Ph.D. candidate, Department of Politics,Princeton University, Princeton NJ 08544 ([email protected]),http://graemeblair.com.

Kosuke Imai is Professor, Department of Politics, PrincetonUniversity, Princeton NJ 08544 ([email protected]), http://imai.princeton.edu.

An unabbreviated version of this article won the Pi Sigma Alphaaward for the best paper presented at the 2012 Midwest PoliticalScience Association annual meeting and is available at the authorswebsites. We thank the Opinion Research Center of Afghanistan(ORCA), and especially Raq Kakar, Abdul Nabi Barekzai, Mr.Soor Gul, Zabihullah Usmani, and Mr. Asadi, along with districtmanagers and the 149 enumerators who conducted the survey, forhelpful feedback and excellent work under trying conditions. Ourprogram manager, Prakhar Sharma, deserves special thanks.We alsothank Will Bullock, Sarah Chayes, Jeff Checkel, Christina Davis,Dan Gingerich, Don Green, Betsy Levy Paluck, and Abbey Steele,along with two APSR editors and three anonymous reviewers, forhelpful comments on an earlier version. The survey instruments,

pretest data, or earlier versionsof thearticle beneted from feedbackfrom seminar participants at George Washington University, DukeUniversity, the University of Pennsylvania, the Ohio State Univer-sity, University of California San Diego, the Juan March Institute,the University of British Columbia, Cornell University, HarvardUniversity, Princeton University, and the University of Washington.Financial support for the survey from Yales Institute for Social andPolicy Studiess Field Experiment Initiative and the Macmillan Cen-ter for International and Area Studies is gratefully acknowledged.Additional support from the Air Force Ofce of Scientic Research(Lyall; GrantNo. FA9550-09-1-0314) and the National Science Foun-dation (Imai; Grant No. SES0849715) is also acknowledged. Thisresearch was approved by Yales Human Subjects Committee underIRB protocol no. 1006006952. In the Appendix we present furtherdetails about the survey and analyses, and replication materials areavailable at http://hdl.handle.net/1902.1/20368.

denies them the luxury of taking into considerationprior ethnic or ideological attachments. This seeminglyeliminates the need to study wartime attitudes.

We challenge this view by arguing that the effects of violence on civilian attitudes are conditional on com-batant identity (Lyall 2010). We contend that inter-group biasthe systematic tendency to interpret theactions of ones own in-group in a more favorable lightthan those of the out-groupshould produce robustand observable asymmetries in how combatant actionsaffectcivilian attitudes. Simplyput, it is likelythat harmby ones own group carries a different set of implica-tions and effects than victimization by members of an

out-group. We anticipate that in-group harm does notlead to either increased support for the out-group orto signicant loss of support for in-group combatants.Harm by the out-group, however, is likely to increaseout-group antipathy while heightening support for in-group combatants.

To test this argument, we conducted a survey in thevery heart of the current Taliban insurgency (Bareld2010; Crews and Tarzi 2008; Giustozzi 2008; Jones2009) that utilizes multiple endorsement experimentsto measure civilian attitudes toward the Taliban andthe International Security Assistance Force (ISAF) inAfghanistan.We combinea multistagesamplingdesignwith multilevel statistical modeling (Bullock, Imai, andShapiro 2011) to examine attitudes among 2,754 malerespondents in 204 villages within 21 districts of vePashtun-dominated provinces.

Three main ndings emerge. First, there is clearevidence that victimization by ISAF and the Talibanhas asymmetrical effects on individual attitudes. Harminicted by ISAF is met with reduced support forISAF and increased support for the Taliban. In con-trast, Taliban-inicted harm does not translate intogreater support for ISAF and has only a marginallynegative effecton Taliban support.Second, subsequentefforts by each combatant to mitigate the effects of the harm they caused among aggrieved individuals

679


2/27

A Survey Experiment in Afghanistan November 2013

actually appear to be successful, though in ISAFs casethis nding rests on a small subset of selected individ-uals given ISAFs haphazard approach to respondingto civilian casualties. 1 Third, we nd little support foralternative explanations that privilege prior patterns of violence, the current distribution of territorial controlby the combatants, or the role of economic assistancein determining civilian attitudes. Taken together, these

ndings illustrate the need to broaden our theoriesto consider the psychological mechanisms that drivecivilian behavior in wartime.

THEORY AND HYPOTHESES

Motivation: Rational Peasants, CivilianAgency, and Wartime Attitudes

A near consensus now exists among practitionersaround the notion that counterinsurgency wars are de-cided by the relative success each combatant enjoys inwinning popular support from the civilian population(Kilcullen 2009; Mao 1961; Thompson 1966; Trinquier2006; U.S. Army 2007). From the counterinsurgentsperspective, victory is obtained through a combina-tion of service provision, material assistance, informa-tion campaigns, and, above all, restraint on the useof ones force with the goal of minimizing civiliancasualties. From the Vietnam-era Hamlet EvaluationSystem (HES) to contemporary efforts in Iraq andAfghanistan, winning the hearts and minds of thecivilian population, or at least an important subset of it, is a central element of counterinsurgency campaigns.

Surprisingly, however, there are few rigorous empir-ical studies of civilian attitudes towardcombatants dur-ingwartime (see Beath,Christia, and Enikolopov 2011,

for an exception). To be sure, this gap is due in part tothe logistical and ethical issues that accompany surveyresearch in a wartime setting. Even if surveys could beconducted safely, methodological issues abound: socialdesirabilitybias, highrefusal rates, andpreference falsi-cation mayfrustrateefforts to obtain reliable answers,especially if sensitive issues such as support for com-batants are investigated using direct questions (e.g.,Berinsky 2004; DeMaio 1984). As a result, one leadingscholar hasargued weshouldbracket [the] questionof individual motives and attitudes (Kalyvas 2006, 101)in favor of theories that emphasize the more easilyobservable structure of opportunities and constraintsfacing civilians in wartime (see also Leites and Wolf 1970, 45).

Indeed, theprevailingviewofcivilians in thecivilwarliteraturesuggeststheyarerationalpeasants(Popkin1979), that is, utility-maximizing agents driven by theshort-term dictates of survival. Without strong ethnicor ideological attachments, civilians seek to minimizetheir exposure to harm from the combatants whilemaximizing the benets offered by combatants, includ-ing material assistance and basic security. Based on

1 In fact, only 164 of our respondents were approached by ISAFafter victimization, while the Taliban approached 535 respondents.See Figure 10 in the Appendix for the full distribution of responses.

thisconsensus, our theoriesconcentratedon explainingwhich factors drive civilian behavior : the relative bal-ance of control exercisedbycombatants (Kalyvas2006;Kalyvas and Kocher 2007; Leites and Wolf 1970), com-petitive service provision (Akerlof and Yellen 1994;Berman, Shapiro, and Felter 2011; Crost and Johnston2010), andrelative levelsof civilian victimization (Con-dra and Shapiro 2012; Kocher, Pepinsky, and Kalyvas

2011; Stoll 1993), to name only three. Prior allegiances,if present at all, are thought to be easily recast in theface of a given mixture of (threatened) punishmentsand(promised) rewards levied by combatants whoaimto exercise control over the civilian population.

Despite privileging different independent variables,these theories converge on the same empirical predic-tion: civiliansshould respondsymmetrically to combat-ant actions. Put differently, civilians should be indiffer-ent to which combatant provides material assistance(Berman, Shapiro, and Felter 2011, 776; Collier andHoefer 2004, 569; Mampilly 2011, 545) or victim-izes them (Condra and Shapiro 2012, 1834; Kalyvas2006, 11119; Stoll 1993, 20). What matters is the typeand amount of aid or violence, not which combatantis responsible for these actions, since civilians cannotafford to discriminate when their survival is at stake.We would therefore expect civilians to respond uni-formly to similar assistance provided by foreign anddomesticnongovernmental organizations, for example.Similarly, these theories predict that civilians will pro-vide information to each side equally and routinely,with the choice of which combatant to inform dictatedonly by prevailing levels of control or violence.

Yet the empirical ndings in these studies are of-ten inconsistent with this baseline expectation of con-stant causal effects across combatants. The effects of

aid programs on violence reduction in Afghanistan,for example, appear to hinge on the ethnic compo-sition of a district, not the level of control or priorviolence (Beath, Christia, and Enikolopov 2011). Sim-ilarly, small-scale aid programs in Iraq have heteroge-neous effects across Sunni and non-Sunni dominateddistricts (Berman, Shapiro, and Felter 2011). In ad-dition, Condra and Shapiro (2012)s ndings suggestthat Sunni districts exhibit markedly different pre- andpostcivilian casualty patterns of violence than Shia ormixed districts. In each case, it appears that civiliansareresponding asymmetrically , rather than uniformly, tocombatants. If responses are conditional on combatantidentities, then we need to theorize how heterogeneityin the effects of aid and violencevariation in who isrewarding or punishing civiliansaffects attitudes andsubsequent behavior asymmetrically.

Second, these theories rely on indicators of civil-ian behavior that are observationally equivalent withmultiple causal mechanisms. As a result, we are leftunable to adjudicate between competing accounts or,more simply, to explain why a particular relationshipexists. Kalyvas (2006) acknowledges, for example, hiscentral independent variable, the relative distributionof combatant control, generates its effects via at leastsix different mechanisms, three of them attitudinal innature (12432).Observational datacannot teaseapart

680


3/27

American Political Science Review Vol. 107, No. 4

thesedifferent mechanisms; onlyby directly measuringattitudes can the relative inuence of each mechanismbe determined. 2 Getting the mechanisms right also hasimportant policy implications: it matters whether civil-ians are providing tips to insurgents because they fearpunishment for noncooperation or if they genuinelysupport the rebel cause since the policy response islikely to be different even if rebel control is constant in

these scenarios.

Why Identity Matters:Intergroup Bias and Support for Combatants

We argue that wartime attitudes are shaped by inter-group biases thatcreate durableexpectations about theresponsibility and blame for combatant actions towardthe civilian population. These biases lead civilians tocondition their interpretation of events on the perpe-trators identity (Lyall 2010). As a result, we expectthat support for combatants will depend on the com-batants identity, as will the effects of actions taken by

the combatants.We dene intergroup bias as the systematic ten-dency by individuals to evaluate ones own member-ship group (the in-group) more favorably than agroup one does not belong to (the out-group) (Hew-stone, Rubin, and Willis 2002, 576; Tajfel 1970; Tajfeland Turner 1979). 3 Positive actions by ones own in-group are therefore viewed as arising from the innatedisposition of in-group members while similar actionsby the out-group are interpreted as situational in na-ture: the out-group was compelled by the situation toundertake a positive act, whereas the in-group did sobecause of its inherent nature. Negative actions by anin-group, by contrast, are understood as situational innature (forced to be bad), while negative actions bythe out-group conrm biases about the out-group andits members as bad actors.

Laboratory experiments have shown that intergroupbias is especially prevalent in situations where the out-group possesses a signicant power advantage andmembers of the in-group feel threatened by the out-group (Brewer 1999; Esses et al. 2001; Hewstone, Ru-bin and Willis 2002, 5856). Such conditions charac-terize counterinsurgency wars, in which members of agroup are challenging a much stronger out-group suchas a national government or external occupier in a bidfor independence, autonomy, or state capture. 4

2 The same criticism holds true for work on civilian victimization,where mechanisms purporting to explain civilian behavior have pro-liferated but research designs capable of testing them have not.3 The literature on social categorization theory, and related eldssuch as prejudice, discrimination, and biased assimilation theory, isenormous. For overviews, see Hewstone, Rubin, and Willis (2002),Lord, Ross, and Lepper (1979), Paluck and Green (2009), and Tajfel(2010).4 This discussion presumes a single in/out-group cleavage, but thiscouldbe complicatedby introducing multiple out-groupsor in-groupssures (on this point, see below).

TABLE 1. Home Team Discount: TheAsymmetry of Combatant Support in Civil War

Expected Effect onIndividuals Support for

Action Actor Out-Group In-Group

Negative Out-Group Negative PositiveNegative In-Group Neutral Neutral/WeakPositive

One potential basis for in-group identication isethnicity. 5 Ethnicity can perform two functions underwartime conditions. First, it acts as a shorthand or so-cial radar (Hale 2008, 34) that allows forward-lookingindividuals to anticipate the nature of in- and out-group actions based on prior interactions or informa-tion about ethnic types. In particular, ethnicity cancon-vey information about often unobservable characteris-

tics about other relevant actors like soldiersandrebels,such as the credibility of their threats and assurancesfor (non)compliance. Biases about (non)coethnics canthus bedrawnontofacilitate wartime risk managementby allowing individuals to quickly attach probabilities,if only crudely, to the expected (re)actions of combat-ants (Lyall 2010, 15).

Second, and central to our argument, group mem-bership also moderates beliefs about responsibility andblame for past actions toward the individual and hergroup. Viewed through the lens of intergroup bias,individuals are more likely to punish out-groups fortransgressions, simply conrming biases about the out-groups disposition. Harm inicted by the in-group,however, carries a different meaning: victimized in-dividuals, and the community at large, may be moreforgiving, since such acts are justied by appeal toextenuating circumstances that forced the in-groupshand. This bias helps explain why external occupiersare often blamed by victimized individuals for harmthat was actually inicted by the rebels themselves.

This discussion generates several conditional hy-potheses, whicharecharacterizedby twotypesof asym-metry. Table 1 summarizes these hypotheses. First, weanticipate that harm inicted by the out-group will beassociated with a sharply negative effect on civilianattitudes toward the out-group (Hypothesis 1) and a

correspondingly large rally effect for insurgents drawnfrom the in-group (Hypothesis 2). Taking Hypotheses1 and 2 together, we should observe a clear asymmetryof effect on attitudes that is conditional on the jointidentities of the perpetrator and victim(s).

Second, the in-groups victimization of civilians willnot result in a transfer of support to the out-group(Hypothesis 3) but will instead generate only a neu-tral effect on out-group support. We may observe a

5 Ethnicity is dened as an identity category in which descent-basedattributes are necessary for membership (Chandra and Wilkinson2008, 517).

681


4/27


modest negative effect on in-group support (Hypothe-sis 4), though this should be substantially smaller thanthe negative effect on out-group support after civilianvictimization. Here we observe a second, more sub-tle, asymmetry: Unlike out-group victimization, whichleads to support for the in-group, in-group victimiza-tion does not lead to a corresponding increase for theout-group. We term this the home team discount:

insurgents may be less constrained in their abuse of their fellow in-group members since such actions donot lead to a corresponding positive effect on supportfor the out-group. 6

We also hypothesize that as we move from an in-dividuals direct experience with the combatants toindirect experiences shared by other in-group mem-bers, the effects of victimization on combatant supportshould attenuate. Put differently, the magnitude of theactions effect, but not its direction, should diminish aswe move from direct experiences to indirect ones pro-vided by secondhandaccounts, combatant propaganda,or rumors (Hypothesis 5).

Simply put, we anticipate that the effects of combat-ant violence on civilian attitudes are asymmetrical, notuniform. Violence by the out-group toward civiliansshould have negative effects on support for that group(in many contexts the counterinsurgent or nationalgovernment) while having a positive effect on supportfor the in-group. In-group violence, by contrast, shouldresult in only a modest negative effect on in-group sup-port (if any effect is recorded) and should not yield atransfer of support to the out-group.

THE SURVEY

We adopt a survey methodology that employs a bat-

tery of indirect endorsement experiments to measureattitudes. Endorsement experiments are especially aptfor environments such as Afghanistan, where concernsabout safety, social desirability bias, and nonrandomrefusals to participate are present in spades (Blairet al. 2013; Bullock, Imai and Shapiro 2011). In ad-dition, there are Afghanistan-specic reasons for con-cern. First, ISAF, along with other organizations, havespent billions of dollars over the past decade seekingto win hearts and minds. This situation creates andreinforces incentives to provide answers perceived tobe acceptable to ISAF, especially if individuals believefuture receipt of aid is conditional oncurrent responses.Moreover, for cultural reasons individuals are typicallysurveyed in public settings. We recorded an averageof two additional individuals present at each surveysession, excluding the respondent and enumerator. 7Finally, our survey rm negotiated access to villageswith localelders, arbaki (militia)commanders, and Tal-iban leaders. In such situations, indirect questions are

6 We anticipate that similar asymmetry should result when individ-uals judge positive actions, i.e., giving aid, but a proper test of thisconjecture is beyond the scope of this article.7 Despite the group setting, only one individuals answers wererecorded, and enumerators were trained not to interview thesepassersby to avoid creating a convenience sample.

preferable because they yield a higher rate of accep-tance among these gatekeepers, thus avoiding biases inthe choice of interview locations.

The Endorsement Experiment

The mechanics of a survey endorsement experimentare straightforward. Randomly selected respondentsare assigned to a treatment group and asked to expresstheir opinion toward a policy endorsed by a specicactor whose support level we wish to measure (here,ISAF and the Taliban). These responses are then con-trasted with those from a control group of respondentsthat answered an identical question without the en-dorsement. Higher levels of enthusiasm for a policywith an endorsement relative to those without it areviewed as evidence of support for the endorsing actor.Since each respondent is assigned only one conditionfor any endorsement experiment, it is impossible forenumerators or others to comparesupportlevelsacrossdifferent conditions for any individual respondent.

Drawing on electronic and print media, as well asfocus groups and two pretest surveys conducted insampled districts, we identied four policies with theproperties desired for an endorsement experiment.Successful endorsement experiments share four prop-erties (Bullock, Imai, andShapiro 2011). First, selectedinitiatives should be in the same policy space so thatthey can be combined for statistical analysis. We selectpolicies that are all related to domestic policy reforms:prison reform, direct election of district councils, areform of the Independent Election Committee, andthe strengthening of anticorruption policies. Second,these initiatives should be well knownby individuals tominimize Dont Know responses and to differentiate

support for an endorser from learning about a policyfrom theendorsement itself. In thesurvey, only 4.8% of respondents replied Dont Know, while refusal rateswere lowin allprovinces. Third, these initiatives shouldactually be endorsed by the particular actors in ques-tion so that the questions are realistic and respondentstake them seriously (Barabas and Jerit 2010). 8 Finally,a wide range of views was held by the general publicabout these initiatives, enabling us to detect supportfor endorsers without suffering from ceiling and ooreffects.

Here, we reproduce one endorsement question, call-ing for reform of Afghanistans notoriously corruptprison system, below. The text of the remaining ques-tions is provided in the Appendix.

Control Condition: A recent proposal calls for thesweeping reform of the Afghan prison system, in-cluding the construction of new prisons in everydistrict to help alleviate overcrowding in existingfacilities. Though expensive, new programs for in-mates would also be offered, and new judges andprosecutors would be trained. How do you feelabout this proposal?

8 Deception is also avoided with this approach.

682


5/27


Treatment Condition: A recent proposal by theTaliban [or foreign forces] calls for the sweepingreform of the Afghan prison system, includingthe construction of new prisons in every districtto help alleviate overcrowding in existing facili-ties. Though expensive, new programs for inmateswould also be offered, and new judges and pros-ecutors would be trained. How do you feel about

this proposal?Plagued by allegations of widespread overcrowd-

ing and systematic torture, among other human rightsabuses (Human Rights First 2011), the Afghan prisonsystem has been publicly criticized by ISAF, the Tal-iban, and outside observers alike. When polled, only48% of Afghans claim to place any faith in the Afghanprison system (The Asia Foundation 2010). As a result,both ISAF and the Taliban have at various occasionsoffered remarkably similar reform programs (Nixon2011; International Crisis Group 2008). 9

We also asked whether individuals would support aproposal to allow Afghans to vote in direct electionswhen selecting leaders for district councils. While tech-nically permitted under Afghanistans 2004 ElectoralLaw, theseelectionshavenot been held to date.In early2010, the Independent Directorate for Local Gover-nance (IDLG)a government agency led by a presi-dential appointee that works with ISAF to coordinatesubnational governance policiespublicly oated theidea of direct elections, to be held in March 2011 (theywere not). Surprisingly, the Taliban seized upon thisidea in their own propaganda, suggesting that Karzaisclaims of being democratically elected were hollowsince the Electoral Law was not being followed (Nixon2011).

Third, we gauged support for reform of Afghanistans Independent Election Committee(IEC), a much-maligned institution that failed toprevent widespread fraud in the 2009 Presidentialand 2010 Parliamentary elections. The Afghan publichas also soured on the IEC; only 54% approved of its performance in 2010 (The Asia Foundation 2010).ISAF and various international organizations spentmuch of the post-September 2010 election periodpublicly discussing various reform proposals. Even theTaliban, which had sought to derail these electionsthrough violence, raised the IEC as additionalevidence of the Karzai administrations democracydecit (Nixon 2011).

Our nal question asked whether individuals wouldsupport strengthening the new Ofce of Oversight forAnti-Corruption. Alongside security concerns, there isperhaps no more salient issue in the minds of Afghanstoday than corruption; a full 76% rated it a majorproblem in 2010, ranking it among the top two do-mestic concerns of individuals (The Asia Foundation2010). Both ISAF and the Taliban are aware of the

9 After consultation with our eld teams as well as our surveypretests, we used the term foreign forces rather than ISAF sincethis is the most common term used by Afghans. The term ISAF,by contrast, was unfamiliar to nearly all respondents.

issues salience; each has issued repeated public state-mentsconcerningcorruptionin general and theneed tostrengthen existing institutions, including the Ofce of Oversight, in particular (Konrad-Adenauer-Stiftung,Royal United Services Institute and Transparency In-ternational UK 2011; Ofce of the Special InspectorGeneral for Afghanistan Reconstruction 2009).

For all four endorsement questions, respondents

were asked to assess their level of support for eachproposal on a ve-point scale: I strongly agree withthis proposal; I somewhat agree with this proposal;I am indifferent to this proposal; I disagree with thisproposal; andI strongly disagree with this proposal.Respondents were also permitted to answer DontKnow or Refuse to Answer.

Figure 1 displays the overall (top row) and provin-cial distributions of responses to the four endorsementquestions. Three patterns are worth noting. First, thereis substantial heterogeneity of interprovince supportfor these policies, even independent of particular en-dorsements. This is encouraging since it suggests thatthe questions, taken separately or together, possessstrong discriminatory power. Compare higher supportfor prison reform in Logar, for example, with Khost, oranticorruption efforts in Logar with Helmand. Second,support differs substantially across Taliban and ISAFendorsements. In some provinces including the Tal-iban stronghold of Helmand or the warlord-controlledUruzgan, support for Taliban endorsed policies is farhigher than ISAF endorsed policies. In fewer cases,notably in the key battleground province of Kunar,ISAF endorsement translates into highersupport foratleast twoproposals. To further validate whether theen-dorsement experiments are measuring the support fortheTaliban andISAF, the Appendix presentsa detailed

descriptive analysis, which argues that the patterns weobserve across provinces and districts are consistentwith conditions in each province at the time of thesurvey.

Sampling Design and Target Population

Our survey experiment was conducted between 18January and 3 February 2011 in ve provinces of Afghanistan. 10 We chose to concentrate on Pashtun-majority areas rather than a nationwide sample forthree reasons. First, estimation at the village and dis-trict level requires a sufcient number of individualrespondents and selected villages, respectively, mak-ing a nationwide sample prohibitively expensive if wescaled up with the same level of coverage. Second,these areas are not only substantively importantthey

10 The survey was implemented by the Opinion Research Center of Afghanistan (ORCA), an Afghan-owned rm that recruits all of itsenumerators locally. Two pretests were also run, from 25 Septemberto 5 October and22 November to 5 December 2010, to pilot differentendorsement policies, to assess sensitivity to question order (nonewas found), and to obtain feedback from enumerators about respon-dent reactions and security issues. Taken together, the pretests had600 respondents in 40 villages from 10 districts within our samplingframe. These villages were later removed from our sampling frameto avoid contaminating results.

683


6/27


FIGURE 1. Overall and Within-Province Distribution of Responses from the EndorsementExperiments

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Control

ISAF

Taliban

Direct Elections Prison ReformIndependent Election

CommissionAntiCorr uption

Reform

O v e r a

l l

( N =

2 7 5 4 )

H e

l m a n

d

( N =

8 5 5 )

K h o s

t

( N =

6 3 0 )

K u n a r

( N =

3 9 6 )

L o g a r

( N =

4 8 6 )

U r o z g a n

( N =

3 8 7 )

Strongly agree Agree Indifferent Disgree

Strongly disagree Don't Know Ref used

Notes : Plots depict the distri bution of responses to four policy questions (columns) across three groups (Tali ban/ISAF endorsementgroups and control group) for the overall sample (top ro w) and each province ( bottom ve ro ws). Sample sizes are also sho wn.

encompass theTaliban heartland,andareconsideredthe battleground provinces of the Afghan warbut arealso areas where intergroup bias between Pashtuns andISAF should be present. Third, these areas have expe-rience with both ISAF- and Taliban-initiated violence.Indeed, they represent the high end of the distributionof violence.

We devised a multistage sampling design to generateestimates of support for combatants at the individ-ual, village, and district levels. First, ve provincesHelmand, Khost, Kunar, Logar, and Uruzganwererandomly selected from the 13 Pashtun-majorityprovinces. 11 Next, we randomly selected one-third of each provinces districts for a total of 21 selected dis-tricts. Third, villages were randomly sampled from

11 The remaining eight provinces are Ghazni, Kandahar, Laghman,Nangarhar, Paktia, Paktika, Wardak, and Zabul.

these districts with the restriction that at least 10%of each districts villages had to be selected, yielding204villages. Householdswere then chosenwithinthesevillages using the random walk method. Finally, malerespondents aged 16 years and older were randomlyselected from these households using a Kish grid. Wesurveyed nine respondents from villages with popula-tions below the mean of sampled villages (about 680individuals) and 18 in villages with above-mean pop-ulations. Table 2 summarizes our sampling design andthe realized sample.

We obtained an 89% participation rate (2,754 outof 3,097 attempts). Figure 2 presents the distribu-tions of basic demographic variables for our sampleof 2,754 respondents. Our average respondent wasa 32-year-old Pashtun male who was likely married(77% answered yes), possibly employed (only 58% re-portedholdinga job, withthemostfrequentoccupation

684


7/27


TABLE 2. Overview of Multistage Sampling Design

Violent eventsDistricts Villages Individuals initiated by

Provinces total sample total sample total sample Tali ban ISAF

Helmand 13 5 1,57 8 61 1,411,506 855 11, 806 2,074

Khost 13 5 88 0 45 754,262 630 779 257Kunar 15 5 818 30 54 8 ,199 396 1,015 166Logar 7 3 641 40 3 84,417 4 86 6 81 137Urozgan 5 3 514 2 8 324,100 3 87 849 314

Total 53 21 4,431 204 3,422,4 84 2,754 15,130 2,94 8

8 nonsampled Pashtun provinces 112 0 10,3 83 0 6,156,571 0 10,007 2,135

Other 21 provinces 233 0 20,7 86 0 14,903,729 0 3, 829 1,225

Notes : The sampling design was conducted as follo ws: First, ve provinces sho wn in this ta ble were randomly sampledfrom a total of 13 provinces with a Pashtun majority. Second, districts were randomly chosen within these districts. Third,villages were then randomly sampled from w ithin selected districts. Fourth, households were randomly selected w ithineach of the selected villages. Finally, one male respondent 16 years or older was randomly sampled within each of theselected households. The ta b le also displays the num ber of Tali ban- and ISAF-initiated violent events during one year

prior to the survey (from 1 8 January 1 8 2010 to 17 January 2011) inicted at the provincial level.

FIGURE 2. Distribution of Demographic Characteristics

60

0

200

400

600

800

1000

1200

1400

Median

0 2 4 6 8 10 12+

0

200

400

600

800

1000

1200

1400

0 1 2 3 4 5 6 7+

0

200

400

600

800

1000

1200

1400

30k

0

200

400

600

800

1000

1200

1400

Age Years of Ed ucationYears of Madrassa

Schooling Income (Afghanis)

Notes : Bar plots indicate the num ber of respondents in each category for demographic varia bles. Dashed lines represent median values.

being farmer), and earning between US $1.40 and$6.97 a day. Respondents typically had little or nogovernment-provided schooling, and on average had18 months of madrassa religious schooling. In keep-ing with the high poverty rates of the sample area,respondents on average had less than 90 minutes of electricity a day. Less than half (48%) possessed cell-phones, while nearly all (90%) owned a radio, the prin-cipal means of obtaining information. Ethnic minori-ties, mostly Tajiks, made up 6% of the respondentsand had comparatively higher rates of education andemployment.

Since the vast majority of Afghans live in small ru-ral settlements, we avoided sampling from large urbancenters such as district capitals. Our villages thus rangefrom 29 to 6509 inhabitants (mean: 681 individuals).We also made thedifcultdecisionto include only malerespondents giventhe cultural and logistical challengesof interviewingwomenin these violent anddeeplycon-servative areas. Our protocol included a special con-sent form for those aged 1618 years. We elected to

include these individuals since Afghanistans medianage is only 18, a fact typically overlooked by existingsurveys of public opinion.

Of the original 204 villages, only four proved in-accessible due to a combination of Taliban hostility,the presence of criminal elements and, in two cases,the inability of the enumerators to locate the selectedvillage. 12 In all cases, village elders, who may havebeen members of the Taliban, were rst approachedbyORCA district supervisorswiththerelevantconnec-tions to describe the survey and to receive assurancesof enumerator safety. Despite our extraordinary levelof access to Taliban-controlled areas, a testament toORCAs connections as well as its decision to hire lo-cals enumerators, we nonetheless experienced myriaddelays and logistical challenges. These included move-ment restrictions due to open war-ghting (especially

12 Each village wasmatched with a similar replacement that the enu-merators could select if conditions warranted. These replacementswere used in the four cases mentioned.

685


8/27


FIGURE 3. Sampled Villages

Notes : Black dots represent randomly sampled villages within ve Pashtun-dominated provinces. Red dots represent Tali ban- or ISAF-initiated violence in 2010. Three colors are used to distinguish Tali ban-controlled (red), government-controlled ( blue), or contested(green) districts.

in Helmand); Taliban and militia roadblocks (typicallyappearing after 2 p.m. in most of our districts); and at-tempted highway robbery. In one case, a clean-shavenenumerator was pulled from his car in Kunar where,along with other motorists, his face was blackened withmotor oil by local Taliban to signify his sinfulness. 13Our most serious incident involved a district managerin Uruzgan, who was wounded by a roadside impro-vised explosive device (IED) the day after placing thelast bundle of completed surveys on a fruit truck (ourmethod of returning surveys to Kabul).

Measuring Exposure to ViolenceTable 2 outlines just how violent these provinceshave been using two different datasets. The numberof attacks initiated by the Taliban and ISAF from18 January 2010 to 17 January 2011 was extractedfrom ISAFs Combined Information Data NetworkExchange, which records the date, location, and type

13 In a sign of how routine this practice has become, roadside shop-keepers had set up a thriving business charging these unlucky indi-viduals a 100 Afghani per face charge for soap and water.

of event from 17 specic categories. 14 These data showthat the sampled provinces have experienced variablelevels of violence but, on average, are considerablymore violent than other non-Pashtun provinces.

Figure 3 illustrates this pattern by overlaying thelocation of sampled villages (as black dots) on the dis-tribution of Taliban and ISAF violence (as red dots)during 2010. Two spatial patterns emerge immediately.First, while some of our surveyedvillages arewithinex-tremely violent areas, other sitesexperienced relativelyfew attacks nearby, giving rise to considerable hetero-geneity in the level and type of harm experienced by

respondents. Second, we overlay the spatial distribu-tion of relative control by combatants, as measuredby ISAF, at the district level. Nine of the 21 districtsare dened as under Taliban control (red areas), whileanother nine were coded as contested (green areas).Only three districts in our sample are thus consideredunder central government or pro-ISAF local control

14 The specic event categories: ISAF (Cache Found, Direct Fire,Escalation of Force, and Search and Attack) and Taliban (As-sassination, Attack, Direct Fire, IED Explosion, IED False, IEDFounded/Cleared, IED Hoax, Indirect Fire, Mine Found, MineStrike, SAFIRE, Security Breach, and Unexploded Ordinance).

686


9/27


(blue areas). These data underscore both the difcul-ties of research in these areas as well as the fact that,for nearly all respondents, experience with harm at thehands of ISAF or the Taliban (or both) is not an ab-stract notion but a realistic possibilityand, for many,a reality. In the Appendix, we demonstrate that thoughthese provinces are more violent than non-Pashtunprovinces, the sampled villages exhibit violence levels

similar tononsampledvillages in thesedistricts. We alsoshow that sampled villages are similar to nonsampledvillages on other substantively important dimensions.As expected with multistage sampling, in-sample bal-ance at the district and province level is slightly worse,given the small samples of districts (21) and provinces(5).

We measure our central explanatory variable, expo-sure to harm at the hands of ISAF or the Taliban, withseveral questions. We distinguish between two modelsof harm: direct exposure, in which an individual or hisfamilyaredirectlyaffected in thepastyear; andindirectexposure, in which individuals are asked about theirawareness of ISAF- or Taliban-inicted harm withintheir manteqa (area) over the same time period. Theseself-reports, elicited prior to the endorsement experi-ments, were preceded by a script that dened harmas both physical injury and property damage. Thesequestions and detailed descriptive analyses appear inthe Appendix.

Our respondents reported a staggering degree of exposure to ISAF and Taliban violence. Some 37%of respondents indicated that they had personally ex-perienced victimization by ISAF, while a similar 33%acknowledgedTaliban victimization (left corner plot inFigure 10(a) in the Appendix). About 19% of respon-dents reported that they had been victimized by both

combatants. Nearly 70% of respondents had heard of ISAF harming civilians in their manteqa , while about40% had similarly heard of Taliban-inicted harm intheir area over the past year (left corner plot in Figure10(b)).

These self-reported measures were followed byquestions about whether the individual had directlyexperienced or was indirectly aware of the efforts bythe responsible combatant to make amends for harminicted. These questions are designed to test whetherintergroup biases are irretrievably xed or, instead,there remains the possibility of softening, if only par-tially, these biases in a wartime setting. We are there-fore able to classify individual experiences by exposureto harm (yes/no), by combatant, and whether the in-dividuals were directly approached by the combatantsfollowing victimization or had heard of such effortswithin their manteqa .

To be approached by ISAF signies that an indi-vidual or family received a one-time solatia payment,typicallyaround $2,500, that absolved ISAF of criminalliability for civilian casualtiesor property damage. 15 Bycontrast, to be approached by the Taliban signies that

15 In 2010, theU.S. spent $4.44 million on 1114condolence andbattledamage payments. For our sample, Helmand had the greatest shareof payments ( N = 518, $2.8 million), followed by Kunar ( N = 48,

the aggrieved party received a funeral oration by theTaliban extolling the virtues of the fallen individual(s)as well as modest monthly payments or basic staplessuch as kerosene or foodstuffs.

Yet canwe rely on these self-reports, especially thoseobtained via direct questions? At least two objectionsmight be raised. First, we may be concerned that ques-tions of wartime harm are simply too sensitive to

elicit truthful answers. This, too, was our initial con-cern, but one subsequently mitigated when enumera-tors reported that respondents, far from being reticentabout sharing their experiences, were only too eagerto describe them. In fact, the combined rate of DontKnow and Refuse to Answer for these questionswas only 2% for Taliban and 1% for ISAF-inictedharm questions. 16

Second, there may be endogeneity between the cer-tain covariatesfor example, the level of combatantcontrol, or an individuals socioeconomic and culturalattributesand survey responses. Individuals did re-port a much greater (second-hand) awareness of ISAFvictimization than Taliban misdeeds, for example. Thismay suggesta greater willingness tospeakopenlyaboutISAF, especially in the Taliban-controlled districts thatrepresent 43% of our sample. Coethnic bias itself mayalso skew these self-reports: Pashtuns could conceiv-ably be more likely to report ISAF victimization thanharm by the almost-exclusively Pashtun Taliban.

Without dismissing thedangersof endogeneity, thereare reasons to believe such concerns may be muted.First, minimal nonresponse rates were recorded; re-spondents clearly were not hiding their answers behindappeals to Dont Know and Refuse to Answer.Second, almost identical levels of victimization wererecorded at the hands of ISAF and the Taliban. Third,

substantial variation exists within Taliban-controlledand contested districts (and villages), suggesting thata heavy Taliban presence does not necessarily dictatepro-Taliban answers. Fourth, the large discrepancy be-tween indirect exposure to ISAF and Taliban victim-ization is likely due to the Talibans own impressiveand far-reaching information campaign. Unlike ISAF,the Taliban is adept at disseminating news of ISAF-inicted civilian casualties (real or imagined) via cellphone videos and SMS as well as routine visits to vil-lages. The efforts of ISAF lack the reach of those of theTaliban and remain largely conned to urban centersand traditional media.

Finally, to complement respondent self-reports, wecalculated the number of attacks initiated by the Tal-iban and ISAF occurring in or near each village. Tobe consistent with our survey questions, we count onlythe attacks that occurred within a ve kilometer radius

$129,000), Khost ( N = 40, $66,197), Urozgan ( N = 15, $76,000), andLogar ( N = 12, $44,000). The mean disbursement was$3,982,thoughpayment per individual was far lower.16 The dictates of ethical research ruled out collecting data on thespecic nature of harm inicted since these details could, in theory,be used to identify individuals in a given village if these data werecompromised. Note that the emphasis here on physical and prop-erty damage may underestimate other forms of intimidation such asTaliban night letters.

687


10/27


of each village during one year prior to the survey.This event data offers an extraordinarily ne-grainedmeasure of violence that can act as an important, if limited, alternative measure to our self-reported mea-sures of exposure to harm. Intriguingly, we nd littlecorrelation between these individual self-reports andISAFs own dataset of violent events. We constructedmeasures of violence using the number of Taliban and

ISAF-initiated events during the year prior to the sur-vey at different radii around these villages and foundlittle correlation. 17 Moreover, there is almost no cor-relation between these self-reports and ISAFs owncivilian casualty data. ISAFs data suggest that theTaliban are responsible for inicting 89% of incidentsinvolving the death or wounding of civilians in 2010(or, by UNAMAs estimate, 83% of civilian deaths).By contrast, our self-report data reveal a much moreeven distribution of harm inicted, with a slightly largerproportion of respondents reporting victimization atthe hands of ISAF compared to the Taliban.

Several factors may account for this discrepancy.First, thetwomeasuresoperationalize harmdifferently.ISAFs data focuses solely on deaths and wounded,while our self-reports measure both physical harm andpropertydamage.The lattercategory, which ISAF doesnot collect systematically, is critical because it repre-sents a far more frequent occurrence than the raremass-casualty events that command media attention.Second, ISAFs coding rules for assigning responsibil-ity for a particular event rely on objective indicators(e.g., if a Taliban-placed IED kills civilians, then theTaliban is assigned responsibility for the event) whilethe population may not be making the same calcula-tion (e.g., ISAF is blamedfor IED-conicted casualtiessince ISAFs presence made the IED likely). Finally,

these self-reports measure individual-level exposure,while ISAFs event data in present form can only bepegged to the village level. For these reasons, we be-lieve that our self-reports offer a more comprehensivemeasure of harm inicted than ISAFs own data.

STATISTICAL MODELING

Since all four endorsement experiment questions oc-cupy the same policy space, we employ the methodol-ogy developed by Bullock, Imai, and Shapiro (2011)to combine responses from the four endorsement ex-periment questions to uncover the underlying latentlevels of support for each combatant. We exploit themultistage sampling design of our survey experimentand construct a multilevel model where each level di-rectly corresponds to each stage of sampling. Underthis model, for example, a village-level parameter is

17 The correlations between mean self-reported harm to the individ-ual or family and to others in the respondents manteqa area and thecount of the numberof violent events at 1-km, 5-km, and10-km radiifrom villages are less than 0.1 in absolute value with few exceptions.The only strong relationship with violent events is with respondentsperceptions of harm inicted to other individuals in their province.Those responses are strongly correlated (between 0.2 and 0.25) withviolent events within 5 and 10 km of villages.

assumed to be randomly drawn from a distribution de-ned at the district level in which the village is situated.This corresponds to the sampling process, in whichvillages are randomly selected from the populationof villages within each of the sampled districts. Ourmultilevel modeling approach is thus justied by thesampling design and allows us to estimate (via partialpooling) support levels of the Taliban and ISAF and

their difference at all four levelsindividuals, villages,districts, and provincesproviding a wealth of infor-mation about the distribution of support. Finally, ourmodel can also accommodate predictors at each of thefour levels including individual-level survey measuresand village-level violence data. 18 The model also al-lows for spatial correlation at the village, district, andprovince levels through the introduction of randomeffects at each level.

We now formally dene our statistical model. Let iindex survey respondents in our sample ( N = 2,754).In the endorsement experiment, each respondent israndomly assigned to one of the three conditions withequal probabilities: Taliban endorsement, ISAF en-dorsement, and no endorsement (control condition).We use T i to denote this randomized treatment as-signment, which takes one of the three values, i.e., 2 =Taliban, 1 = ISAF, and 0 = Control. Now, respondentisanswertothe j thpolicy questionis representedby Y ij for j = 1, . . . , 4 and is recorded as the ve-category or-dinal response variable, 5 = Strongly agree, 4 = Agree,3 = Indifferent, 2 = Disagree, and 1 = Strongly Dis-agree. Following Bullock, Imai, andShapiro (2011), wepartially pool all policy questions in order to identifythe common underlying pattern across all questionsrather than idiosyncratic noise associated with a par-ticular question.

First, the individual level is based on the orderedprobit item response model:

Pr( Y ij l | T i = k) = ( j l j ( xi + sijk ))

where j 1 = 0 and jl < j ,l + 1 for any j and l . In thismodel, the latent variable xi represents the degree towhich respondent i is pro or antireform while sijk mea-sures his support level for group k = 1, 2 where sij 0 =0 and a greater value of sijk indicating a higher level of support. In addition, the item difculty parameter jl controls how likely respondents are to agree ordisagree with policy j without endorsement, and the

discrimination parameter j operationalizes the de-gree to which policy j differentiatesbetween proreformand antireformrespondents. 19 We assume j > 0 for any j because we have written policy questions such that arespondentsagreementwould imply he/she is support-ive of the reform. Thus, in our model, the probabilityof agreeing with a reform proposal is a function of the

18 The standard regression approach with cluster robust standarderror cannot incorporate the sampling design into the estimationand also is not suited for data with a complex correlation structure.19 We nd that the estimates of j are similar across policy questions,indicating that each question made a roughly equal contribution tothe nal estimates.

688


11/27


two factors: how proreform a respondent is in generaland how supportive he/she is of the assigned group. Inaddition, the degree to which responses are affected bythese factors depends on how popular the reform pro-posal is and how theproposal candiscriminate betweenproreform and antireform respondents.

We model the two key latent parameters, xi and sijk ,using the multilevel modeling strategy. Formally, the

model is specied as follows:

xiindep . N village[i] + Z i

Z , 1

sijkindep . N k, village[i] + Z i

Z k ,

2k, village

village[i]indep . N district[ i] + V village[i]

V , 2district

k, village[i]indep . N k, district[ i] + V village[i]

V k ,

2k, district

district[ i]indep . N province[ i] + W district[ i]

W k ,

2province

k, district[ i]indep . N k, province[ i] + W district[ i]

W k ,

2k, province

where Z , V , and W are optional covariates at the in-dividual, village, and district levels, respectively. Themodel is completed by assigning noninformative priordistribution on each parameter (see Table 3 in the ap-pendix).

Specically, we assume that each respondent is ran-domly drawn from a population distribution in his ownvillage, each village is randomly drawn from a dis-tribution of its district, and each district is randomlydrawn from a distribution of its province. At each

level, covariates can be incorporated and the normaldistributionis used. We use Markovchain Monte Carloto obtain random draws from the joint posterior dis-tribution of relevant parameters. For each model t,which is accomplished via JAGS (Plummer 2009), wemonitor convergence by running three parallel chainswith overdispersed starting values.

Finally, though the number of nonresponses is small,the pattern of missingness is not completely random.Thus, we impute these missing values as part of ourmodel by assuming that the data are missing at randomconditional onvillage andother covariates at individualand more aggregate levels. This mitigates the bias andinefciency that typically result from list-wise deletion.

Model Specication. We t two models: a direct ex-posure model that incorporates a respondents (or hisfamilys) direct exposure to harm and our key ex-planatory variables, and an indirect exposure modelthat draws on the same covariates, but substitutes di-rect exposure to harm for an individuals awarenessof civilians being harmed within his larger manteqa .Both models include a common set of theoreticallyrelevant variables (see the Appendix). At the individ-ual level, we include a battery of socioeconomic anddemographic covariates that have been cited as impor-tant determinants of support for terrorist or insurgent

groups. These covariates include the respondents age,marital status (Berrebi 2007), income level, years of state, and madrassa education (Bueno de Mesquita2005; Krueger and Maleckova 2003; United StatesAgency for InternationalDevelopment 2011 b), ethnic-ity, whether the respondent is a Pashtun, and whetherthe tribe he afliates himself with was deemed alreadypro-Taliban (Giustozzi 2008, 5269). 20

The models also include village- and district-levelcovariates, which include the settlements altitude toaccount for difculty of state control (Fearon andLaitin 2003), population size, and the number of ISAF-and Taliban-initiated violent events within ve kilo-meters of the villages center one year prior to thesurveys launch. 21 District-level covariates include to-tal expenditures on ISAF Commanders EmergencyResponse Program (CERP) short-term aid projectsin 2010 (ISAF CERP Data, 2005-10); the number of villages in a given district that had received NSP-runcommunity development projects by the conclusion of 2010; a dummy variable that indicated whether the dis-trict was home to a Taliban-run sharia court system;the area in a given district under opium cultivation(in hectares) in 2010 (United Nations Ofce of DrugControl 2010, 112); the length of paved roads, a proxyfor general economic development; a dummy variableindicating whether the district neighbored Pakistan,to control for cross-border ows of arms, insurgents,and funds; and ISAFs own measure of the control itexercised over a given district. We draw on ISAFsfourfold index of control, which which rates districtsfrom government control or dominant inuence tolocalcontrol or dominant inuence to contested toTaliban control or dominant inuence. These ratingswere assigned in September 2010, roughly four months

before our survey was elded.22

ISAF- and Taliban-directed aidandservice provision variablesaccount forthe possibly offsetting positive actions of combatantson the battleeld. 23

EMPIRICAL FINDINGS

We rst examine the relationship between individual-level traits and attitudes before providing evidence of the asymmetric effects of violence on combatant sup-port. We then add nuance to our discussion by examin-ing how intra-Pashtun tribal heterogeneity, along withvariation within the Taliban itself, also condition atti-tudes. Finally, we explore whether postharmmitigationefforts by ISAF and the Taliban are capable of over-coming intergroup bias in a wartime setting. Given ourquantities of interest, we rely on graphs rather than

20 For our purposes, the pro-Taliban tribes are the Noorzai, Zadran,Gheljay, and Durrani.21 These results were robust to the substitution of violent eventswithin one- or ten-kilometer radii, of violent events within ve kilo-meters for a longer period of time from 2006 to 2010, and of civiliancasualties for the same JanuaryDecember 2010 period.22 ISAF Insurgent Focus Brieng Slide, dated September 2010.23 As a robustness check, we reran our models without the villageand district-level covariates. The results are substantively similar andtherefore we present results from the hierarchical models only.

689


12/27


FIGURE 4. Bivariate Relationships between Demographic Variables and Individual-level Supportfor the Taliban and ISAF

10 20 30 40 50 60 70 80

3

2

1

0

1

2

10 20 30 40 50 60 70 80

3

2

1

0

1

2

0 2 4 6 8 10 12+

3

2

1

0

1

2

0 2 4 6 8 10 12+

3

2

1

0

1

2

0 1 2 3 4 5 6 7+

3

2

1

0

1

2

0 1 2 3 4 5 6 7+

3

2

1

0

1

2

30k

3

2

1

0

1

2

30k

3

2

1

0

1

2

Age Years of Ed ucationYears of Madrassa

Schooling Income (Afghanis)

S u p p o r t

f o r T a

l i b a n

S u p p o r t

f o r

I S A F

Notes : Support estimates are given in terms of the standard deviation of ideal points. For age (the left column), dots representrespondents and the lo wess curve is tted as a solid line. For remaining plots, box plots are used: the width of each box is proportionalto (the square root of) the num ber of respondents for each value of the demographic varia bles.

coefcient tables to demonstrate our main empiricalndings. 24

Main Results

Our multilevel models generate estimates of a respon-dents level of support for the particular combatant;estimates are measured in terms of the (posterior)standard deviation of respondents preferences withina singledimensional policy space (i.e., ideal points). If arespondent is assignedto thecontrolgroup,noestimateis generated. Imagine, for example that a respondentis neutral toward the proposed reform initiative. If thisindividuals support level for the Taliban was a 1,then it implies that a Taliban endorsement can shift

this individuals ideal point by one standard deviationin the proreform direction.

Attitudes Toward Combatants. Whois more likely tosupport the Taliban or ISAF? As an initial test of thevalidity of our survey instrument, we plot individual-level support estimates from thedirectexposuremodelagainst key demographic covariates in Figure 4. Wenote that estimated support for both combatants de-creases with age, though the Taliban enjoy more sup-port among the youth than ISAF. This may be some-

24 The posterior mean and standard deviation for each coefcient inour multilevel models are reported Table 3 in the Appendix.

what surprising given the Talibans public image as aconservative gerontocracy, but the war has created op-portunities for advancement among the youth, and theTaliban appears to be given credit for this mobility. Pre-dictably, supportfor ISAFor,more bluntly, decreaseddislike for ISAFis associated with increased educa-tion and with fewer years spent in madrassa education.The reverse is true for the Taliban: its estimated sup-port levels rise with years of madrassa education andfall with years of government education. Finally, percapita income is positively associated with estimatedISAF support (or, again, less dislike of ISAF) whilethere appears to be no clear pattern between incomeand Taliban support.

Attitudes by Levels of Victimization. What are theaverage support levelsforISAF andthe Talibanamongthose harmed by these combatants? As an initial in-vestigation of the effects of victimization, we estimatethe level of support for each combatant according towhether respondents reported harm by ISAF only,the Taliban only, by both combatants, or by neither(Figure 5). 25

25 In our empirical analysis, we assume that the effects of harm byeach combatant are additive in nature when estimating the level of support for those harmed by both combatants. For the sake of clarityof analysis, we did not include an interaction term in our models.Future research could relax this assumption, however, and generate

690


13/27


FIGURE 5. Estimated Support Levels for Each Combatant by Victimization Type

1.0

0.5

0.0

0.5

1.0

S u p p o r t

f o r

I S A F

None TalibanOnly

ISAFOnly

Both

Victimization

1.0

0.5

0.0

0.5

1.0

S u p p o r t

f o r

t h e

T a l i b a n

None TalibanOnly

ISAFOnly

Both

Victimization

Notes : We derive and then plot the posterior predicted mean support level from our multilevel models with 95% condence intervals.

FIGURE 6. Estimated Effects of ISAF and Taliban Victimization on Support Levels for EachCombatant

1.0

0.5

0.0

0.5

1.0

C h a n g e

i n S u p p o r t

f o r

I S A F

Self orfamily

Mante qa(area)

Victimization

by ISAF by Taliban

1.0

0.5

0.0

0.5

1.0

C h a n g e

i n S u p p o r t

f o r

T a l i b a n

Victimization

by ISAF by Taliban

1.0

0.5

0.0

0.5

1.0

D i f f e r e n c e

i n C h a n g e

i n S u p p o r t

( T a

l i b a n I

S A F

)

Victimization

by ISAF by Taliban

Notes : The right panel presents the differences bet ween the results in the middle and left panels. Posterior means of coefcients derivedfrom multilevel models are plotted with 95% condence intervals.

We nd that most respondents are not supporters of either combatant: among those not harmed by eithercombatant, respondents exhibit a strongly negative ef-fect towardISAF (a 0.62 effect in standarddeviationsof the ideal points) and are only slightly less nega-tive about the Taliban (at 0.29 standard deviations of the ideal points). Interestingly, those respondents whowere harmed by ISAF only or by both ISAF and theTaliban exhibit a positive effect toward the Taliban (at0.27 standard deviations and 0.15 standard deviations,respectively). Respondents who had beenharmed onlyby ISAF or the Taliban were the least supportive of these combatants (at 0.89 standard deviations forISAF and 0.40 standard deviations for the Taliban),though the magnitude of the effect is clearly asymmet-rical across combatants.

Asymmetrical Effects of Victimization. We now pro-vide the key evidence to support our central claim thatthe effects of violence on civilian attitudes are condi-tional on combatant identity. Figure 6 illustrates theestimated average effects of ISAF (left panel) and Tal-

both theoretical predictions and empirical tests that examine howexposure to harm by multiple combatants affects the direction andmagnitude of the asymmetry we observe here.

iban victimization (middle panel) on ISAF andTalibansupport at two levels: the individual/family (solid cir-cles) and the manteqa (open circles). In the right panel,the net differences between the effects of victimizationon support for the Taliban and ISAF are also provided.

A number of important observations standout. First,as expected by Hypothesis 1, we note that ISAF vic-timization is associated with a signicant negative ef-fect on ISAF support. ISAF victimization is also as-sociated with a substantively large positive effect onTaliban support (Hypothesis 2). These data conrmthat harmed individuals are both punishing the out-group for inicting harm and are transferring supportto the out-group as expected if intergroup biases are inplay.

Second, Taliban victimization has no clear effecton support for ISAF. As expected by Hypothesis3, we do not observe a blossoming of support forISAF after Taliban-inicted harm, suggesting that theout-group is not rewarded for its (relative) restraint.Moreover, Taliban victimization itself only slightly de-creases Taliban support and certainly does not ap-proach the penalty on ISAF support witnessed afterISAF victimization. While it would be incorrect toconclude that the Taliban can victimize civilians withimpunity, they appear to enjoya much largercushionof

691


14/27


public support than does ISAF, for whom civilian ca-sualties are associated with a steep negative effect onsupport.

Third, we observe the hypothesized asymmetry of effects between ISAF victimization on Taliban support(middle panel, left column) and Taliban victimizationon ISAF support (left panel, right column). This dif-ference is statistically signicant and is estimated to

be 0.37 standard deviation of the ideal points (withthe 95% condence interval of [0.02, 0.73]) at theindividual/family level. The right panel in Figure 6 un-derscores the asymmetry between the effects of ISAFvictimization on relative support for the combatants(Taliban support minus ISAF support) and those asso-ciated with Taliban violence. Once again, it is appar-ent that while ISAF victimization creates a substantialpositive effect on net Taliban support, Taliban victim-ization generates at best only a very modest positiveeffect on relative support for the out-group.

Are these results driven by personal exposure toharm, as suggested by Hypothesis 5, or are indirectinuencesequally as important? Returning to Figure 6,we observe that manteqa -level effects (as representedby open circles) of ISAF victimization are attenuatedwhen compared with individual self-reports. We ndthat the net difference in the effect of ISAF violenceon ISAF support between the individual and manteqalevel is nearlystatistically signicant, with a 0.203 (95%condence interval at [0, 0.65]). We do not observe asimilar pattern of attenuation for Taliban violence, inpart because the individual-level effects are alreadysmall. While the general pattern of effects is thereforesimilar across individual and manteqa levels, the mag-nitude of estimated effects on support levels for eachcombatant is typically diminished at the manteqa level,

and signicantly so for ISAF.One interesting wrinkle also emerges from thesendings: both Taliban and ISAF endorsements are, insome cases and regions, associated with reductions insupport for the proposed reform relative to the con-trol endorsement. 26 While caution is warranted, oneinterpretation of this result is simply that the controlmost closely resembles an ofcial (government) en-dorsement. What we may be measuring, then, is therelative shift in support for Taliban and ISAF com-batants associated with victimization against a moregeneral backdrop of war weariness in which neithercombatant is especially favored. That said, however,since support in war is relative, not absolute, it is alsoclear that the Taliban do enjoy a greater measure of support, and more tolerance of inicting harm on civil-ians, than ISAF.

Tribal Analysis. Support for combatants is best castas a continuum (Petersen 2001, 8), rather than a sim-ple (and misleading) choice of pro-Taliban or anti-ISAF. We therefore add nuance to our discussion

26 Our breakdown of responses by question and province are pro-vided in the Appendix. Since the control provides the baseline forassessing ISAF and Taliban endorsements, support for the control ismeasured as the reverse of the Taliban or ISAF endorsement.

of intergroup bias by exploring intragroup variation.Several Pashtun tribes, for example, have largely (andpublicly) declared in favor of the Taliban, while oth-ers, often in conict with pro-Taliban tribes, have re-mained neutral or in a few instancessided against thesetribes. While we cannot assume homogeneity of viewsamong all members of a particular tribe, we can ex-plore whether the attitudes of individuals from these

pro-Taliban Pashtun tribes diverge from those of theirnon-Taliban aligned counterparts.In particular, we expect that individuals from pro-

Taliban tribesare more likely to be forgiving of Talibanvictimization given their prior identication with theTaliban. The negative effects of ISAF victimization arealso likely to be greater in magnitude for this subset,though if ISAF support is already low we may observea oor effect given that support for ISAF may not sinkany further. In each case, victimization is likely to con-rm the intergroup bias that negative in-group actionsare situational in nature while those of the out-groupand its members are dispositional.

We begin with the effects of ISAF victimization onsupport for ISAF and the Taliban among pro-Talibantribe (solid circles) and non-pro-Taliban tribe (solidsquares) members, then turn to how Taliban violenceaffects support for ISAF and the Taliban itself. Weprovide a graphic interpretation of these results inFigure 11 in the Appendix.

We nd, for example, that ISAF victimization hasa more pronounced negative effect on ISAF supportamong pro-Taliban tribe members than their non-pro-Taliban tribal counterparts. Notably, however, the dif-ference itself is fairly modest at 0.328 standard devia-tions (95% condence interval at [ 0.904, 0.168]) andstatistically insignicant. The effects of ISAF violence

on support for the Taliban are also different acrosssubgroups. Here, it is the non-pro-Pashtun tribes thatrecord the largest positive effect on support for theTal-iban after experiencing ISAF victimization. This resultis likely due to two factors. First, it is probable thatsince pro-Taliban tribes have already decided for theTaliban, their support is subject to a ceiling effect thatlimits the magnitude of the positive shock to Talibansupport. Second, thisresult suggests that ISAFviolencehas an outsized effect on individuals who believe thatthey are eitherneutral or (more rarely) whohave sidedwith ISAF. The difference between these two groupsis striking at 0.604 standard deviations (with a 95%condence interval of [ 1.09, 0.23]), suggesting thatISAF-inicted harm has a large effect on members of non-Taliban-aligned tribes and their support for theTaliban.

By contrast, Taliban victimization has broadly simi-lar effects on ISAF support for both pro-Taliban andnonaligned Pashtuns. This result is consistent with ourexpectation that intergroup bias prevents a transfer of support from the in-group to the out-group. Wherewe observe divergence between these two subgroupsis in the effects of Taliban violence on Taliban sup-port. Members of pro-Taliban tribes appear not onlyto tolerate but embrace harm, as witnessed by the positive effect of Taliban harm on Taliban support. By

692


15/27


FIGURE 7. Estimated Effects of Victimization and Subsequent Approach by Combatants onSupport

1.0

0.5

0.0

0.5

1.0

C h a n g e

i n S

u p p o r t

f o r

I S A F

Experience with Violenceby ISAF by Taliban

1.0

0.5

0.0

0.5

1.0

C h a n g e

i n S

u p p o r t

f o r

T a l i b a n


1.0

0.5

0.0

0.5

1.0

D i f f e r e n c e

i n C

h a n g e

i n

S u p p o r t

( T a

l i b a n I

S A F )


Victimization with No Approach Effect of Being Approached Victimization and Approach

Notes : The right panel presents the differences bet ween the results in the middle and left panels. Posterior means of the correspondingcoefcients are plotted along with 95% condence intervals.

contrast, Taliban violence has a slightly negative effecton non-pro-Taliban tribal members support for theinsurgent organization. 27

These ndings deepen our previous discussion bysuggesting that combatant violence has asymmetric ef-fects across different Pashtun tribal afliations. ISAFvictimization clearly has a disproportionate effect onsupport for the Taliban among non-pro-Taliban tribalmembers. Taliban victimization, by contrast, has a pos-itive (and large) effect on pro-Taliban tribal members,suggesting that the Taliban possess a far higher free-dom of actionto include causing civilian casualtiesamong their supporters than ISAF.

Effects of Post-Harm Mitigation Efforts. Can com-batants offset the negative effects of their vio-lence by rendering postharm assistance? We offera preliminary exploration of this issue in Figure 7,which presents the estimated combined effect of be-ing harmed but not subsequently approached (cir-cle), the effect of being approached (square), andthe combined effect of victimization and receiv-ing post-harm assistance on attitudes toward thecombatants.

Two broad conclusions stand out. First, postharmmitigation efforts by both combatants are associatedwith a marked negative effect on support for the other combatant. Taliban victimization alone has a small

positive effect on attitudes toward ISAF (as reportedabove) but, once aid has been proffered, the effectis a substantially negative effect on attitudes towardsISAF. Similarly, experience with ISAF victimizationalone is associated with a positive effect on supportfor the Taliban. Yet if aid has been provided by ISAF,the result is a large negative effect on Taliban sup-port (second column). Indeed, the net difference ineffect between individuals who were harmed but not

27 The difference between the two groups is statistically signicantat 0.57 standard deviations (with a 95% condence interval of [0.05,1.14]).

approached by ISAF and those who were harmed andthen approached is a highly signicant 1.00 standarddeviation movement away from a pro-Taliban position(95% condence interval of [ 1.43, 0.53])

An optimistic take on these results would suggestthat hearts and minds can in fact be swayed, if notbought outright, via postharm mitigation efforts. Yet asecond nding is also important: these efforts appearto have little effect on support for the actual com-batant that rendered the assistance. While ISAF doesappear more successful at generating modest positiveeffects after itsefforts, there is much less positivemove-ment in attitudes toward the combatant responsiblefor the victimization than there is a shift away from

the other combatant. In this situation, postharm aidappears less about persuading civilians to move towardones side as it is about shifting away from the rivalcombatant.

We must also be careful not to overstate the ndingabout the effects of ISAF assistance. There is a severeselection problem at work here, because ISAF man-aged to approach only 16% of those who claimedthat they had been harmed by ISAF. By contrast, theTaliban approached over 60% of those who self-identied as suffering Taliban victimization. ISAFsdecision-making on the extension of condolence andbattle damage payments is often haphazard, but is atleast partly conditioned on anticipated reception in agiven village (Campaign for Innocent Victims in Con-ict 2010b). The more likely a unit is to receive armedresistance from aggrievedvillagers, the less likely ISAFwill return to disburse payments. Given closer mediascrutiny, mass casualty events are more likely to befollowed by postharm mitigation efforts. Aggrieved in-dividuals in small-claims cases, by contrast, are oftenforced to go to military bases to receive their funds, dis-couraging all but the most determined and risk accep-tant claimants(Campaign for Innocent Victimsin Con-ict 2010a). As a result, the nding regarding ISAFsability to use condolence payments to overcomeintergroup bias relies on a small subset of individuals

693


16/27


whowerespecically selected to receive thisassistance.Indeed, these individuals may be the ones most likelyto be swayed by ISAFs efforts. 28

Summary. Taking these ndings as a whole, we un-cover substantial evidence to support claims that theeffects of violence on attitudes are conditional on per-petrators identity. ISAFs violence is strongly associ-ated with a negative effect on attitudes toward ISAF,while Taliban violence only has a marginally negativeeffect onTaliban support. Similarly, ISAFvictimizationleads to a marked positive effect on Taliban supportbut, consistent with our expectations, Taliban victim-ization results in only a small positive effect on ISAFsupport. These asymmetries are often powerful amonga subset of individuals, namely, those who claim mem-bership in publicly pro-Taliban Pashtun tribes. Finally,these results appear to be largely driven by personalexperiences with harm, rather than indirect knowledgeof harm inicted within an individuals manteqa , es-pecially for ISAF violence. This nding reinforces theneed to devise and test our arguments about civilian

victimization at the most microlevel unit possible, thatof the individual, rather than assuming that macroleveleffects scale down to the individuals in a given geo-graphic space.

Alternative Explanations

Our ndings stronglysuggest that intergroup biasesareresponsible for the observed asymmetries of effects of combatant actions on civilian attitudes. Yet what of alternative explanations? We consider four alternativeidentity-based theories and three additional explana-tions that emphasize battleeld factors such as the dis-

tribution of aid, relative control by thecombatants, andTaliban service provision.

Other Identity-Based Accounts. We might imagine,for example, that the observed asymmetry of effectsarises from a sense of betrayal among civilians. Per-haps these individuals expect that ISAF, but not theinsurgents, will protect them, a belief that leads them topunish ISAF for violating its perceived commitment tosafeguardthe population.While intuitive, it is apparentthat few believe ISAF will protect them from harm; infact, the opposite is true. We posed the question Inyour view, how often do foreign forces [Taliban] takeprecautions to avoid killing or injuring innocent civil-

ians during their operations? Possible answers wereas follows: always (3), sometimes (2), rarely (1), andnever (0). The mean response for ISAF was 0.56; forthe Taliban, a 1.92, a substantial difference. 29 These

28 Intriguingly, a study of 155 recipients of USAIDs Afghan CivilianAssistance Program I (ACAP I)which provides cash grants andsmall bundles of household items to victims of ISAF and Talibanviolencefound that satisfaction with this aid hinged on who hadinicted the violence. Only 45% of recipients were satised withACAP assistance after experiencing ISAF victimization, comparedwith nearly 85% of recipients after experiencing insurgent violence.See United States Agency for International Development 2011 a, 10.29 This difference is signicant at p = 0.000, t 5506 = 56.55.

ndings arestunningwhenwe consider that theTalibankilledan estimated2,080 Afghancivilians in 2010 alonewhile ISAF was deemed responsible for 440. Yet theperception remains that ISAF, rather than the Taliban,is wielding violence indiscriminately, suggesting thatindividuals in our samplearea donot believe that ISAFis protecting them. 30

These ndings are also consistent with arguments

that the post-2001 status reversal experienced by Pash-tuns in Afghanistans ethnic hierarchy has created highlevels of support for the Taliban (for status reversalarguments, see especially Petersen 2002, 4061). Yetwhile it is true that foreign-imposed regime changesoften upend existing institutions, this is less the casein Afghanistan than elsewhere. The Karzai-led gov-ernment created by the Bonn Agreement (Decem-ber 2001) actually represented a return to prior pat-terns of ethnic hierarchy, not a decisive break. Thecurrent Pashtun-dominated executive has echoes inearlier Durrani Pashtun regimes, including the ruleof the Iron Amir (18801901), under MohammedNadir Shah (193079), and the Taliban era itself. De-spite (modest) local score settling, Pashtuns were notcollectively punished for their earlier support of theTaliban, andthe newAfghanconstitution itselfwasrat-ied unanimously (Bareld 2010, 27293). Given thatdissatisfaction with the Karzai government transcendsethnic lines, it is unlikely that current Taliban supportamong Pashtunsitself variable, not constant, in ourdatais due to status reversal.

A third alternative explanation privileges revengemotives arising fromthe Pashtun-specic code of ethicsknown as Pashtunwali . Consisting of nine principles,Pashtunwali norms call for aggrieved individuals totake revenge ( badal ) upon a wrongdoer while acting

bravely ( tureh ) to defend their property, family (espe-cially women), and honor (Johnson and Mason 2008;Tomsen 2011, 4754). This argument about culturedictating attitudes stumbles, however, over two issues.First, it is not clear why attitudes would be asymmetri-cal among harmedindividuals without invoking a priorclaim about intergroup bias. In principle, revenge seek-ing should lead to symmetrical effects of violence oncivilian attitudes since the obligation to seek redressis not directed solely against non-Pashtuns. Second,there is considerable intra-Pashtun tribal differencesin how their attitudes are inuenced by combatant vio-lence andsubsequentrestitution efforts. It is neither thecase that Pashtun attitudes are monolithic toward thecombatants, nor are these attitudes uniformly affectedby combatant actions.

Finally, these asymmetrical effects remain regardlessof an individuals level of prior exposure to ISAF. Con-tact theory (Allport 1954; Cook 1971) suggests that,

30 According to UNAMAs gures, the Taliban has been responsiblefor the bulk of civilian deaths since at least 2008. In 2008, ISAF wasresponsible for 828 deaths to 1,160 by the Taliban; in 2009, ISAFinicted 596 deaths, compared with 1,631 by the Taliban (see UnitedNations Assistance Mission in Afghanistan 2011). This raises a keymethodological problem for observational studies if the objectivecoding of responsibility and the perceived blame by the victimizedindividuals are not highly correlated.

694


17/27


under certain conditions, intergroup biasand, specif-ically, the lessening of out-group derogationcan bereduced via frequent, positive interaction with mem-bersof theout-group. Current ISAFcounterinsurgencydoctrinealso stresses the importance of face-to-face in-teractionwithlocalpopulationsto build trust andshapeattitudes. There is little evidence to support this claimin our context, however. To begin with, the conditions

cited as necessaryforcontact theoryto be operative aredaunting. These include equal statusbetween the inter-acting parties, shared goals, sustained intimate contact,and the absence of competition (Paluck and Green2009, 346), a set of conditions absent in Afghanistan.We also directly tested this claim with a question thatmeasured the respondents frequency of prior inter-action with ISAF forces ( Frequency ). Frequency wasstatistically signicant and negatively associated withISAF support, exactly the opposite of contact theoryspredictions.

Battleeld Dynamics: Aid, Control, and Taliban Ser-vice Provision. How do other theoretical explana-

tions fare in explaining civilian attitudes? Surprisingly,we nd little evidence that district-level variables holdweight in explaining attitudestoward either combatant.Caution is warrantedwhen interpreting these results, tobe sure, because there are only 21 districts in our sam-ple. Moreover, while our sample is randomly drawn,combatant d

explaining support for combatants during wartime- a survey exp in afghanistan

Documents