Can civilian attitudes predict insurgent violence? Ideology and insurgent tactical choice in civil war

A staple of counterinsurgency theory now holds that winning over civilian ‘hearts and minds’ is the key to defeating insurgents. This approach counsels the adoption of a mix of economic assistance, service delivery, and protection to convince fence-sitting populations to support the counterinsurgent. Civilian behavior follows attitudes; win over hearts and minds, and civilians will provide tips about insurgents hiding among them. Individuals in this account are calculating actors, often supporting whichever side promises the most benefit while disregarding prior ethnic or political loyalties. In brief, the greater the aid distribution, the more information gleaned from the population, and the sharper the corresponding decrease in attacks by insurgent organizations.

These claims generate several observable implications. Most centrally, if these claims are correct, we should observe a decrease in insurgent attacks as relative support for the counterinsurgent grows. The US Army’s Field Manual (May 2014) makes this connection explicit: military offensives and aid campaigns are designed to create safe spaces for the population by reducing insurgent attacks. Each phase of the Army’s ‘shape-clear-hold-build-transition’ framework is associated with decreased insurgent attacks, though not necessarily their complete absence. Indeed, the first metric proposed for measuring hearts and minds effectiveness is the reduction of insurgent attacks and casualties among COIN forces and civilians (Department of the Army, 2014: Section 12-29; see also Figure 9-1).

Scholars remain divided, however, over the central mechanism driving this result. Insurgent violence may decrease, for example, as the provision of tips from civilians better enables counterinsurgents to identify and destroy rebel leaders and networks (Kalyvas, 2006; Department of the Army, 2007; Berman, Shapiro & Felter, 2011). Cash-for-work programs and other forms of economic assistance may also raise the opportunity costs for participating in the insurgency, making recruitment difficult and driving would-be insurgents from the ranks (Blattman & Annan, 2016). Increased contact with counterinsurgents may also convince wary civilians to change their beliefs and to support the counterinsurgents’ cause. Finally, an increase in counterinsurgent troop strength in an area may act as a deterrent, forcing rebels to seek targeting opportunities elsewhere (Department of the Army, 2007: 1–13).

Insights drawn from hearts and minds theorizing have made impressive inroads among academics and policymakers. Yet there are reasons for skepticism. Most existing studies consist of a single-shot impact evaluation with simple pre- and post-aid distribution differences in insurgent attacks. Civilian attitudes are typically not measured; increased support is inferred from reduced insurgent violence (for an important exception, see Beath, Christia & Enikolopov, 2011). Mechanisms, too, remain untested; measures for the quantity and quality of tips from locals are usually unavailable or incomplete, leaving scholars unable to stitch together the causal sequence of aid delivery, changed attitudes, increased information, and then reduced insurgent violence.¹

We contend that insurgent targeting in civil war is a reflection of the underlying spatial distribution of relative support for the combatants. In this view, the insurgent group’s ideology shapes not only the end goal of the insurgency but also the nature of the targets (the ‘who’), the location of these attacks (the ‘where’), and the nature of the tactics and weapons used (the ‘how’). Following Gutiérrez-Sanin & Wood (2014: 214), we define ideology as ‘a set of more or less systematic ideas that identify a constituency, the challenges the group confronts, the objectives to pursue on behalf on that group, and a (perhaps vague) program of action’. In particular, ideology informs tactical choice by establishing the set of appropriate means and targets for achieving desired political goals. Tactics are defined here as the specific techniques that insurgent organizations use to destroy, degrade, or displace their opponents to secure operational objectives (see Drake, 1998).

We anticipate that populations suspected of disloyalty to the insurgent cause are likely to absorb the brunt of insurgent attacks. More specifically, we focus on the political allegiance of local populations, where support is defined as the relative leanings of a given populace toward the insurgency and the counterinsurgent. Contrary to the expectations of ‘hearts and minds’ theorizing, we expect to observe a positive association between relatively higher support for the counterinsurgent and the level of subsequent insurgent attacks. Insurgents may of course be opportunistic in their execution of attacks, but they do so within a bounded set of choices and means delineated by their ideological precepts.

Insurgents’ knowledge of the distribution of civilian attitudes will influence their tactical choice, especially the mix between discriminate and indiscriminate weaponry. Tactics that require networks within villages – notably, the use of improvised explosive devices (IEDs) and suicide bombings – become wasting assets if locals begin to side with the counterinsurgent. Insurgents will feel compelled by a ‘use it or lose it’ dynamic to privilege these tactics in an effort to roll back pro-counterinsurgent support. These tactics are typically less discriminate than other forms of violence, raising the possibility of civilian casualties. Roadside IEDs often miss their intended military targets and instead inflict casualties on locals who had the misfortune of trailing a military convoy, for example. Yet insurgents will relax their prohibitions (if any) on killing civilians if the violence is largely restricted to the pro-counterinsurgent village. In this case, indiscriminate violence underscores the risks associated with abandoning the insurgency. Similarly, civilians who actively side with the counterinsurgent (‘traitors’) will be viewed as legitimate targets by insurgents and their own supporters, mitigating the damage to the insurgent organization’s reputation commonly associated with killing innocents.

These tactical choices are driven by the insurgents’ overall end-goal and are designed to have several effects on counterinsurgent forces and local populations. First, insurgents are seeking to punish, not persuade, pro-counterinsurgent villages. In these situations, they are less concerned about negative effects of killing ‘traitors’. Civilians in swing villages may actually support these actions if they share the view that pro-counterinsurgent individuals or villages have betrayed a particular political cause. Insurgents may even gain support via such actions: if the counterinsurgent is provoked into retaliatory actions that harm civilians, the net effect is likely a shift in relative support toward the insurgent organization. Counterinsurgents are disproportionately punished by populations for civilian victimization if they are drawn from groups outside the insurgent group’s own members (Lyall, Blair & Imai, 2013).

Second, insurgents attack pro-counterinsurgent villages to influence fence-sitting villages. Punishing pro-counterinsurgent villages can forestall the loss of other villages by demonstrating the costs of siding openly with the counterinsurgent. Insurgent targeting against these villages has a twin punish–deter logic; punishing these pro-counterinsurgent villages will have a deterrent effect on other, perhaps wavering, ‘swing’ villages. While insurgents are certainly interested in swing villages, they are less likely to use violence here, hoping to avoid the absolute loss of support that can accompany rebel victimization of civilian populations (Lyall, Blair & Imai, 2013). Insurgents are more likely to use redistributive mechanisms (‘carrots’) in these swing villages; pro-counterinsurgent villages, on the other hand, are more likely to be punished violently (‘sticks’).²

Third, targeting pro-counterinsurgent villages inflicts costs on the counterinsurgent publicly, helping to stem the erosion of insurgent support that can mean the loss of information, resources, and recruits needed to generate the insurgency’s combat power. Loss of these villages may also drive a wedge into the insurgency by creating a ‘loyalist’ faction that undercuts insurgent recruitment along ideological or ethnic lines (see Kalyvas, 2008; Lyall, 2010). As a result, striking pro-counterinsurgent locations preserves the insurgency’s momentum while increasing the counterinsurgent’s cost of defending these centers. If the insurgents are able to create the perception that the counterinsurgent cannot credibly defend the populace, then erstwhile counterinsurgent supporters may curtail their collaboration or defect (back) to the rebels.

This argument rests on two assumptions, namely, that civilian preferences toward the combatants are stable enough to be meaningful and that insurgents devote resources to monitoring these attitudes. We believe these are reasonable assumptions. In particular, the assumption that insurgents possess informational advantages relative to counterinsurgents is shared by nearly all theories of civil war violence (Lyall & Wilson, 2009). It is well recognized that counterinsurgents struggle to solve the ‘identification problem’ – namely, correctly identifying insurgents hiding among civilians (Kalyvas, 2006: 89). We flip the identification problem, however, by exploring how insurgent tactical choice is guided by the spatial distribution of combatant support.

Our theory produces several empirical expectations. Our core claim is that insurgents will disproportionately target locations that express relatively higher pro-counterinsurgent sympathies. Both civilians and military forces should be targeted in these locations at rates higher than in similar, but less pro-counterinsurgent, areas. We expect that these attacks are quite localized (i.e. at the village level). That is, insurgent violence should cluster around a pro-counterinsurgent village; the predictive value of civilian attitudes should diminish as we move away from these villages. Insurgents should also emphasize indiscriminate tactics when attacking these villages given their ability to maximize damage among counterinsurgent forces, their shock value among targeted civilians, and their deterrent value among neighboring villages. Pro-counterinsurgent attitudes should therefore be an important predictor of the use of indiscriminate weapons such as improvised explosive devices. Finally, attitudes should remain important for explaining future attacks even when controlling for alternative explanations such as base location, the distribution of control and aid, and prior patterns of violence.

We measure support for ISAF and the Taliban using a survey of 2,754 adult male respondents from 204 villages in 21 districts of five Pashtun-dominated provinces of Afghanistan (Logar, Kunar, Uruzgan, Helmand, and Khost).³ After two large-scale pilot surveys in these provinces, the survey was conducted between 18 January and 3 February 2011. These pilots were conducted from 25 September to 5 October 2010 and 22 November to 5 December 2010. They were used to test sensitivity to question order, potential design effects (e.g. ceiling and floor effects for the experiment modules), and the nature of logistical challenges.

The sample was constructed using a multistage sampling method. Figure 6 in the Online appendix illustrates the location of the 204 surveyed villages (blue dots) and the 14,606 non-surveyed villages (red dots). First, the five provinces, whose borders are highlighted in the map, were randomly sampled from the 13 Pashtun-majority provinces (colored gray on the map).⁴ Within each of these five provinces, districts, and then villages, households, and finally individuals, were randomly sampled. Households were chosen using the random walk method. Owing to security considerations and cultural constraints, only male respondents aged 16 years and older were randomly selected from the sampled households using a Kish grid.

We obtained an 89% participation rate (2,754 respondents out of 3,094 approached individuals). The average respondent was a 32-year old Pashtun male who was likely married (77%), possibly employed full-time (58%), and possessed little or no formal (government) schooling and only an average of 18 months of madrassa education. Nearly all respondents were Pashtun by ethnicity (93%). Our sampling frame encompassed some of the poorest areas in Afghanistan; respondents reported possessing only 90 minutes of electricity per day, and daily income hovered between US$1 and US$6.

Of the original 204 villages, only four proved inaccessible due to a combination of Taliban hostility, the presence of criminal elements and, in two cases, the inability of enumerators to find the village. In all cases, village elders were first approached by our survey firm, Opinion Research Center of Afghanistan (ORCA), to describe the survey and to obtain permission for enumerators to enter the village. All enumerators were locals; nearly all surveys were conducted in Pashto. Informed consent was obtained in all cases orally; special permission was granted by Yale’s IRB (Institutional Review Board) to waive written consent requirements to avoid linking respondents to particular surveys. This helped minimize the risk of reprisal if the written forms were intercepted by the Taliban. Similarly, the names of respondents were not recorded.⁵

We employ a battery of four endorsement experiments to mitigate the sensitivity of measuring wartime support for ISAF and the Taliban. This indirect approach reduces social desirability bias and item non-response when asking questions about sensitive issues (Bullock, Imai & Shapiro, 2011). Direct questions, by contrast, can endanger enumerators and respondents alike and often result in high non-response rates and biased answers.

To take one example, a recent wave of ISAF’s own Afghan National Quarterly Assessment Report (ANQAR) in November–December 2011 recorded nearly 50% non-response rate as potential respondents refused to participate. Our overall refusal rate was about 5% for our endorsement experiments, a difference we ascribe to the indirect questioning method we employed. Table III in the Online appendix shows the pattern of non-response rates by province. The highest non-response rate is obtained in Helmand province, which has the highest level of insurgent violence and support. We address this non-response problem in our statistical analysis by assuming that the pattern of non-response is random conditional on all observed data.

The mechanics of an endorsement experiment are straightforward. We first randomly divided a sample of respondents into groups. In the ‘control’ group, respondents were asked to rate the level of their support for a particular policy. For those in the ‘treatment’ group, the identical question was asked except that the policy was said to be endorsed by an actor of interest. We then take advantage of subtle cues induced by endorsements and interpret the difference in responses between the treatment and control groups as evidence of support (or lack thereof) for this actor of interest. This is based on the psychological literature, which finds that people tend to more positively (negatively) evaluate an item when paired with another item they like (dislike). In our application, we have two actors of interest, ISAF and the Taliban, and thus the sample was randomly divided into three groups of equal size – Taliban treatment, ISAF treatment, and control – across individual respondents within each sampled village.

Typically, multiple policies in the same domain are selected so that the measurement does not rely on a single instrument. Statistical power is also increased by analyzing multiple items together. In our survey experiment, we ask questions concerning four domestic policies: prison reform, direct election of district councils, a reform of the Independent Election Committee, and the strengthening of anti-corruption policies. The exact question wording appears in the Online appendix.

A Bayesian hierarchical factor analytic model is used to (partially) pool the responses to these four questions together, creating an estimate of individual-level support for ISAF and the Taliban (Bullock, Imai & Shapiro, 2011). We then aggregate individual-level support values by computing the sample mean to create an estimate of village-level support for each combatant. When modeling individual-level support, we use as regressors village-level random effects and individual-level covariates such as the respondent’s age, income level, years of education, years of madrassa schooling, direct exposure to violence by the Taliban or ISAF, experience of encountering the Taliban or ISAF, the frequency in which the respondent encounters ISAF, and whether his tribe was pro-Taliban. The model we use is described in the Online appendix and is fitted through the open-source R package endorse (Shiraito & Imai, 2012).

The resulting measures of village-level support are numerical estimates for each combatant. The support level for the Taliban ranges from $-1.14$ to 1.42 while that for ISAF ranges from $-1.75$ to 0.43 with positive (negative) values, indicating support for (opposition against) the combatant. Figure 7 in the Online appendix presents the distribution of each support measure as well as the distribution of the difference between the two. The figure clearly shows that a majority of the Afghan population supports the Taliban over ISAF. These support measures are standardized on a latent variable scale (so-called ‘ideal points’ and ‘ability’ parameters in the political methodology and psychometrics literatures, respectively), and so only their sign and relative magnitude can be interpreted. In addition, we validate these support measures against another measure based on the item count technique and find these two indirect questioning methods provide essentially identical findings (Blair, Imai & Lyall, 2014).

In our main analysis, we operationalize relative support for ISAF as the difference between ISAF and Taliban support levels and use this measure as the key predictor of insurgent violence. Relative support for ISAF ranges from $-3.01$ to 0.91, suggesting that Afghans are far more supportive of the Taliban than ISAF. Figure 1 presents the spatial distribution of relative ISAF support measure for the 204 surveyed villages in the five sampled provinces. In general, Taliban support is strongest in Uruzgan and especially Helmand, two provinces long associated with the post-2001 re-emergence of the Taliban. The eastern provinces of Khost, Kunar, and Logar present a more mottled picture, with substantial pockets of support for the Taliban and ISAF.

Of particular interest is the fact that neighboring villages sometimes have different relative levels of support for these combatants. Even within the same district, some villages are supportive of ISAF while others are opposed. This local variation underscores the need to embrace disaggregated data; ISAF, by contrast, continues to track support at much more aggregate district or even provincial levels.

We first measure insurgent violence using declassified event data from ISAF’s Combined Information Data Network Exchange (CIDNE). These data record the date, location (using the Military Grid Coordinate System), and nature of insurgent attacks against ISAF forces and installations throughout Afghanistan. Distinct from WikiLeaks’ Afghan War Diary, these data represent the most comprehensive account of insurgent attacks to date, though they are not without their limitations. These ‘Significant acts’ (SIGACTs) rarely cover violence against Afghan National Security Forces and exclude violence against civilians. As such, our main analysis focuses on insurgent attacks against ISAF alone. We use data from 10 months before and after our January–February 2011 survey for our prediction models; a total of 69,841 insurgent attacks were recorded over this period.

CIDNE tracks 11 discrete types of insurgent attacks that are relevant for our purposes here. We aggregate these types into three broad categories. First, we constructed an ‘Improvised explosive device (IED) attacks’ category that records 11,577 events of IED explosions, mine strikes, and premature IED detonations. IEDs represent the most lethal form of insurgent attack against ISAF forces, accounting for 54% of all soldier fatalities since 2007 (iCasualties.org, 2013), and ISAF has devoted billions of dollars in a cat-and-mouse effort to mitigate this threat.⁶

Second, we created an ‘IED found’ category that includes 19,093 events where ISAF forces found and cleared IEDs or mines.⁷ We make a clear distinction between ‘found’ IEDs and ‘detonated’ IEDs to determine whether tips from locals in pro-ISAF villages translate into higher than average timely discovery of IEDs, as anticipated by hearts and minds theory.⁸ This is an imperfect measure: found IEDs may reflect improved detection capabilities via technical means by ISAF that occurred independent of civilian support. Absent classified data on IED detection, we cautiously use this variable as a proxy for tips flowing from civilians.

Lastly, we created a ‘Non-IED attacks’ category that includes 39,171 insurgent attacks using small arms fire, indirect fire (e.g. mortars), and rocket fire against ISAF forces and installations.⁹ We distinguish between IED and non-IED attacks because, as explained earlier, the notion of attitudes-as-targeting-cues suggests an especially strong association between civilian attitudes and certain classes of tactics with maximum destructive value such as IEDs.

In our analysis, these three categories are operationalized as count variables recording the number of relevant events in specified temporal windows before and after the survey in each village. We also test across different spatial radii around villages. We therefore aim to predict the aggregate number of attacks of each category within defined spatial and temporal windows around sampled and then non-sampled villages. Table I in the Online appendix illustrates the pre- and post-survey distribution of insurgent violence across five provinces. The locations of ANSF and ISAF bases and aid projects are also noted. Among the randomly sampled five provinces, most insurgent violence ( $57\%\sim 80\%$) is centered on Helmand, while most ISAF bases ( $31\%$) and aid projects ( $48\%$) are clustered in Kunar.

The core test of our argument lies in observing improvements in predictive accuracy while adjusting for alternative explanations for insurgent violence.

First, the conflict modeling literature has demonstrated that a leading predictor of future violence is simply the prior distribution of violence (e.g. Yonamine, 2013; Zammit-Mangion et al., 2012; Montgomery, Hollenbach & Ward, 2012; Bohorquez et al., 2009). To account for this explanation, we draw on the same declassified CIDNE data as used for our dependent variable and simply mirror the spatial and temporal windows on either side of the survey date for a given village. For example, if we examine the effect of village-level support on the number of IED attacks occurring within a x km radius around each village during y months post-survey, we also include the number of past IED attacks that have occurred within the same spatial and temporal windows.

Second, insurgent targeting may reflect the distribution of counterinsurgent bases and forces within a given area; that is, after all, where the majority of counterinsurgent targets are found. We therefore use geo-referenced base location data to adjust for the number of ANSF and ISAF military installations present in the pre-survey period (June 2007–December 2010) within a three-kilometer radius of each village. The choice of three kilometers is somewhat arbitrary. As a result, in our analysis we use varying radii to examine the robustness of our findings.

We lack a direct measure of ISAF and ANSF force size or patrol rate; these data are either classified or do not exist. That said, we believe our data on base location do a credible job in controlling for ISAF presence. Many of these bases are small combat outposts (COPs) that patrolling forces sortie to and from; a five-kilometer radius is likely to capture all but the most distant (and infrequent) patrol. The Taliban have also expressed a preference for ambushing tired counterinsurgent forces as they arrive back at their bases (Meyerle & Malkasian, 2009: 6). Far from viewing these military bases as oases of calm, they have historically been the site of concerted Taliban efforts to harass or breach their defenses, often via suicide bombing.

Figure 8 in the Online appendix presents the bivariate relationship between ISAF support and the number of ANSF/ISAF bases. We observe that regardless of the choice of radii there is a small degree of positive relationship between them; villages which have a larger number of bases tend to be more supportive of ISAF.¹⁰

Third, pro-counterinsurgent attitudes may be associated with the presence of aid programs designed to influence fence-sitting populations. Insurgents may in turn orchestrate attacks to derail these aid projects. This dynamic was observed in the Philippines, for example, in response to the government’s KALAHI-CIDSS Community Driven Development (CDD) anti-poverty program. Insurgents timed their attacks to the announcement and initial phase of aid programming in CDD-villages to prevent the erosion of their popular support (Crost, Felter & Johnston, 2014). To account for this possibility, we adjust for the number of USAID Community Development Programs (CDP) using geo-referenced data for all projects initiated within one kilometer of each village. The large-scale, $250 million CDP program was designed to foster stability using cash-for-work initiatives in areas where insurgents had been forcibly evicted through military operations during 2010–12. In Figure 9 in the Online appendix, we present the bivariate association between the aid projects (with varying radii from villages) and village level support for ISAF. While there appears to be quite weak correlation between the two variables, we nonetheless include this variable in our base model.

Throughout our in-sample analysis, we use a linear regression model to quantify the association between relative support for ISAF and subsequent insurgent violence after the survey. Note that each village has a slightly different start and end date given their survey order. This results in different temporal windows (of the same size) across villages. Specifically, we regress the number of future insurgent attacks within certain temporal and spatial windows on its relative support level for ISAF as well as the other possible confounders discussed in the Online appendix. We vary the length of time and distance windows in order to examine the robustness of our findings. We also explored the possibility of nonlinearity but we find a simple linear model captured most of the systematic variation.

To further examine the predictive power of civilian attitudes, we evaluate their out-of-sample predictive performance using 14,606 non-surveyed villages. Doing so reduces the possibility of false discovery by ensuring that the results of our in-sample results are not due to over-fitting to a particular sample of surveyed villages.¹⁴ We first predict relative ISAF support for these out-of-sample villages using village-level covariates. We then forecast the number of insurgent attacks with these predicted support levels and then compare our forecasts with actual insurgent attacks. As before, our aim is to examine whether these predicted support levels improve forecasting performance of future insurgent violence even while controlling for prior insurgent attacks and the presence of counterinsurgent bases and aid projects.

The out-of-sample prediction proceeds in two steps. First, using the 204 surveyed villages, we estimate the ISAF support model by regressing the relative ISAF support level and a set of available village- and district-level covariates. These covariates include village population size and elevation as well as several district-level factors, including ISAF’s own measure of its relative control in that district, the existence of Taliban-run sharia courts (as a measure of Taliban control), and whether the district bordered Pakistan (to control for differences arising from cross-border spillover). Using the same set of covariates, we can now estimate ISAF support level for each of the out-of-sample villages based on the fitted model. We rescale the relative support estimate for out-of-sample villages so that their standard deviation is identical to that of the original village sample.

Second, we estimate the Insurgent violence model, which is the same linear regression model as used in the Online appendix. Together with the estimated ISAF support from the previous step, we can now forecast future insurgent attacks for out-of-sample villages. Finally, we compare the predicted insurgent attacks with actual attacks for each out-of-sample village to evaluate our model’s predictive performance.

Figure 4 uses a diagram to summarize our procedure for out-of-sample prediction. Below, we also provide the details about the entire prediction procedure.

We assess the accuracy of our out-of-sample prediction by comparing our forecast with the actual level of insurgent attacks. To measure the degree to which political attitudes improve forecasting performance, we compute the mean absolute forecasting error (MAFE) for two models: (1) one with prior attacks, ANSF/ISAF bases, and CDP projects as the predictors of future violence and (2) one with these three covariates and the estimated support level for ISAF as an additional predictor. This quantity is formally defined as ${\displaystyle {\sum }_{i=1}^N}|{\mathit{\text{Y}}}_i-{\hat{Y}}_i|/N$, where ${\mathit{\text{Y}}}_i$ represents the number of observed future insurgent attacks for an out-of-sample village i and ${\hat{Y}}_i$ is its prediction from a forecasting model.

We then compute the percentage improvement obtained by adding the estimated relative ISAF support level to the model with only prior insurgent attacks, ANSF/ISAF bases, and aid programs. Specifically, we compute $(\text{M}A\mathit{\text{F}}{E}_2-\text{M}A\mathit{\text{F}}{E}_1)/\text{M}A\mathit{\text{F}}{E}_1\times 100$%, where $\text{M}A\mathit{\text{F}}{E}_1$ and $\text{M}A\mathit{\text{F}}{E}_2$ are obtained from the model with and without the estimated relative ISAF support level variable, respectively.