Does ethnicity matter for explaining patterns of violence during civil wars? To date, the empirical record remains mixed at best. Ethnicity has, for example, been invoked to account for numerous wartime dynamics, including insurgent recruitment and distinctive patterns of civilian victimization (Humphreys and Weinstein Reference Humphreys and Weinstein2006; Weinstein Reference Weinstein2007); the durability and alliance patterns of rebel organizations (Sinno Reference Sinno2008); and the location and intensity of fighting along ethnic faultlines (Kaufmann Reference Kaufmann1996; Petersen Reference Petersen2001; Slack and Doyon Reference Slack and Doyon2001; Toft Reference Toft2003; Weidmann Reference Weidmann2009).Footnote ¹ Yet ethnicity's purported causal effects have also been challenged in light of a historical record littered with examples of individuals “defecting” to fight against their own coethnics (Kalyvas Reference Kalyvas2008), “ethnic” alliances that were little more than temporary marriages of convenience (Christia Reference Christia2008), and identities that shifted endogenously under wartime pressures (Kalyvas Reference Kalyvas2006; Wood Reference Wood2008).

Answering the question of how and why ethnicity shapes patterns of wartime violence is important because a substantial portion of civil wars are believed to be “ethnic” in nature. By one count, more than two thirds of the 127 civil wars fought since 1945 have been fought entirely or partially between, and sometimes within, warring ethnic groups (Fearon and Laitin Reference Fearon and Laitin2003; see also Sambanis Reference Sambanis2001, 269). Similarly, about 40% of the 286 insurgencies fought since 1800 have featured external intervention by a state that differed ethnically from the local population (Lyall and Wilson Reference Lyall2009).

An enduring feature of most, if not all, of these wars has been the creation of state-sponsored auxiliary forces drawn from the same ethnic population that is fighting the state. Here, too, the debate over the impact of these militia is divided. Existing research has staked out two contradictory claims. Counterinsurgency theorists and practitioners, for example, often laud the value of these forces because they possess local knowledge and language skills, rendering them more effective than non-coethnic forces (Byman Reference Byman2006, 87–8; Felter Reference Felter2005; Galula Reference Galula2006; Gompert Reference Gompert2008; Gwynn Reference Gwynn1934). These local allies are also valuable, it is argued, for legitimating the actions of the external occupier or government by creating a “loyalist” faction that undercuts insurgent claims to represent the sole voice of the targeted minority. Others, however, suggest that such militia are typically plagued by chronic defection and rebel subversion. These militia may also use their coercive abilities to pursue their own agendas, including score settling and ethnic cleansing, that may set in motion an escalatory logic of retaliation within the ethnic group or against other ethnic rivals (Kalyvas Reference Kalyvas2006, 107–9; Mason and krane Reference Mason and Krane1989, 185; Stoll Reference Stoll1993, 98–103; Stubbs Reference Stubbs1989, 70–2).

Although seemingly a relatively narrow issue, the widespread creation of coethnic militia in civil wars offers one empirical strategy for addressing the broader question of ethnicity's impact on patterns of violence. In this article, I exploit variation in the ethnicity of soldiers conducting so-called sweep operations (zachistka, plural zachistki) during part of the Second Chechen War (2000–5) to test whether insurgent responses are conditional on soldier identity. More specifically, the large-scale defection of Chechen rebels to the Russian side enables us to compare changes in patterns of insurgent violence after Russian-only, pro-Russian Chechen only, and joint operations. While ethnicity cannot be directly manipulated, these sweeps are matched in a bid to isolate ethnicity's causal effects by controlling for observable pre sweep differences.

I find substantial evidence to support the claim that insurgent violence is in fact conditional on the ethnicity of the sweeping soldiers. Three findings stand out. First, there is nearly a 40% average decrease in the number of insurgent attacks following Chechen-only sweeps compared with similar Russian-only operations. Second, Chechen insurgents display markedly different timing in their attacks conditional on identity of sweepers, with Russian sweeps being met by much swifter retaliation. Finally, the frequency and timing of insurgent attacks after joint Russian-Chechen operations resembles those observed after Russian-only, not pro-Russian Chechen-only, operations, suggesting that coethnics' informational advantages are not readily transferred across ethnic divisions.

I proceed as follows. I first detail the Second Chechen War, the empirical strategy, and the data and matching procedures used. Next, I present the main findings from comparison of matched Russian-only, Chechen-only, and joint operations for the 2000–5 time period. I then explore the possible channels through which ethnicity may, and may not, account for observed changes in insurgent behavior. A final section concludes with suggestions for future research on the link between ethnicity and action during civil wars.

ETHNICITY, VIOLENCE, AND THE SECOND CHECHEN WAR

As with many civil wars, the seeds of the Second Chechen War (August 1999–ongoing) were sown by its predecessor's ambiguous conclusion. The first Chechen War witnessed a clumsy Russian attempt to squash Chechnya's secessionist bid in December 1994 that quickly degenerated into a grinding counterinsurgency campaign that stretched until August 1996. Highly unpopular on the homefront, the war was brought to some semblance of a conclusion when vastly outnumbered Chechen insurgents staged a surprise attack that not only wrestled Chechnya's capital, Groznyy, from Russian hands but also forced a stunned and embattled Kremlin to the negotiation table (Lyall Reference Lyall2006). The resulting Khasavyurt Accords (signed August 30, 1996) mandated the withdrawal of all federal military units but deferred a final decision on Chechnya's constitutional status until 2001. In the meantime, however, the war bequeathed to the quasi-independent Chechen republic a shattered economy, still-armed rebel leaders with few job prospects or skills, endemic corruption, and a ravaged population that had suffered an estimated 35,000 to 50,000 war-related deaths—nearly 5% of Chechnya's prewar population. Persecution of remaining Russians led to their outmigration by 1998, further lending an air of crisis to the now monoethnic republic (Dunlop Reference Dunlop1998; Evangelista Reference Evangelista2002; Lieven Reference Lieven1998; Tishkov Reference Tishkov2004).

The second war thus broke as an unwelcome, but not wholly unexpected, storm. In August 1999, Shamil Basayev, a prominent rebel commander, led a 2,000-strong formation into neighboring Dagestan in a bid to carve out a regional “Islamic Khanate.” Set against the backdrop of exploding apartment buildings in three Russian cities in September—blamed on Chechen rebels, although never proven conclusively—Basayev's raid led the still unknown Prime Minister Vladimir Putin to counsel an overwhelming military response to the “Chechen problem.” After withering aerial bombardment once again sent Chechens fleeing for neighboring Ingushetia, Russian armed columns moved into Chechnya from the north in October (Evangelista Reference Evangelista2002; Kramer Reference Kramer2005/2006; Souleimanov Reference Souleimanov2007).

The ensuing war would be marked by two distinct military phases. First, Russian forces battled with an estimated 5,000 to 7,500 Chechen insurgents in a series of fixed battles in key cities (October–March) in an ultimately futile attempt to stem the steady southward march of federal forces. With the fall of their last redoubt at Komsomolskoye (Saadi-Kotar) in March 2000, Chechen insurgents, now badly bloodied, shifted to a guerilla war campaign characterized by near-daily hit-and-run attacks on Russian forces. Split into nationalist and Islamic factions, the insurgency was bitterly torn over tactics, with Islamic units under Basayev favoring the use of suicide bombing and mass hostage taking (both inside and beyond Chechnya) that nationalists largely abjured (Hughes Reference Hughes2007; Souleimanov Reference Souleimanov2007; Wilhelmsen Reference Wilhelmsen2005).

Despite inflicting substantial losses on Russian and pro-Russian Chechen forces—at least 5,500 soldiers were killed between 1999 and 2008—insurgent ranks slowly dwindled (Zürcher Reference Zürcher2007, 100). By 2008, remaining rebels, probably fewer than 500, were pinned largely in Chechnya's mountainous south, and a state of uneasy normalcy had returned to most of Chechnya as reconstruction funds poured into the war-ravaged republic. In April 2009, Vladimir Putin declared the war officially over, though the months after his statement were punctuated with continued insurgent attacks and, even more alarmingly, a return to once abandoned suicide bombing tactics.

Scholars have often cited the second Chechen War as an archetypal “ethnic” civil war (Fearon and Laitin Reference Fearon and Laitin2003; Sambanis Reference Sambanis2000, 499; Toft Reference Toft2003, 127–48). These characteristics include a marginalized ethnic minority with a distinct sense of, and desire for, its own national homeland; a lengthy history—some 200 years—of struggle for this homeland (Gammer Reference Gammer2006; Gapurov, Izraiilov, and Tovsultanov Reference Gapurov, Izraiilov and Tovsultanov2007); territorial concentration in a peripheral section of a federal state dominated by a different ethnic group (Buhaug Cederman, and Rod Reference Buhaug, Cederman and Rod2008); and clear ascriptive, linguistic, and religious differences between Orthodox Russians and Sufi Chechens.

Yet the fact that so many Chechens crossed “ethnic lines” to collaborate with Russian forces should give us pause about blanket statements concerning the war's “ethnic” nature. In fact, the centerpiece of the Russian counterinsurgency campaign was “Chechenization,” the calculated devolution of political power to approved Chechen officials who supported the Kremlin's efforts to keep Chechnya within Russia's legal fold (Russell Reference Russell2008; Sokirianskaia Reference Sokirianskaia and Gammer2007). Steps included a referendum on Chechnya's constitutional status (March 2003) and partly free presidential and parliamentary elections in October 2003 and November 2005, respectively. Substantial political power was first vested in the figure of Akhmed Kadyrov, an ambitious mufti who had fought against Russian forces in the First (but not Second) Chechen War, and subsequently in his son Ramzan's hands after the elder Kadyrov's assassination by rebels while watching a Victory Day parade in a Groznyy stadium in May 2004.

EMPIRICAL STRATEGY

The second pillar of Chechenization was the creation of Chechen-only combat formations to conduct sweep operations. Until early 2003, Russian military units had borne the brunt of the counterinsurgency campaign. Seeking to stamp a Chechen face on these counterinsurgency efforts, Russians first experimented by conducting joint patrols with Chechen police formations in 2002 before training and equipping dedicated Chechen-only Ministry of Defense units—Special Battalions Vostok and Zapad—in early 2003.Footnote ² The Kadyrovs also took the opportunity to construct their own irregular formations, collectively known as the kadyrovtsy, who were also tasked with conducting sweep operations against suspected rebels and their supporters.Footnote ³

While ethnicity cannot be directly manipulated, these sweep operations offer one empirical strategy for exploiting variation in soldier identity to test how ethnicity shapes subsequent insurgent violence. In addition, the concurrent nature of Russian-only, Chechen-only, and joint operations permit comparison across these sweep operations while holding a variety of background factors, including the war's own temporal dynamics, constant.

But what exactly is a “sweep operation?” Long a staple of counterinsurgency, sweep operations consist of armed house-by-house identification checks to sift insurgents from noncombatants once a targeted neighborhood, village, or city has been blockaded by soldiers. To “attack an enemy who is invisible, fluid, uncatchable,” a noted French counterinsurgency theorist argued, “we have no alternative but to throw a net of fine mesh over the entire area in which the bands move” (Trinquier Reference Trinquier2006, 74). By controlling all entry and egress points, soldiers restrict insurgent movement, facilitating their discovery while potentially uncovering new sources of intelligence or weapon caches.

In Chechnya, sweeps typically involved several hundred soldiers and lasted between three and five days on average, although on rare occasions a few unfortunate villages were “swept” for as long as four weeks.Footnote ⁴ Despite their putative military utility, it should be noted these operations have also been driven by motives other than strict utilitarian logic. These operations have become synonymous with excesses by both Russian forces and, later, their pro-Russian Chechen allies. Human Rights Watch, the European Court of Human Rights, and local nongovernmental organizations (NGOs) have issued a stream of reports decrying the use of indiscriminate violence, forced disappearances (about 5,000 since 1999), and extrajudicial killings (Human Rights Watch 2001, 2002a, 2002b, 2006) in these operations. Theft, torture, kidnapping, and sheer wanton destruction, especially of items such as automobiles that soldiers could not readily steal, have been reported as well.

On the Russian side, sweeps were conducted by soldiers drawn from either the Ministry of Defense or the Interior Ministry, with representatives of at least six additional security agencies also present. Sweep operations were often staffed by a mixture of 18- and 19-year-old conscripts and older short-term contract-based volunteers (the so-called kontraktniki) who, despite their much better (and more regular) pay, were often cited as the least disciplined by local inhabitants.

Who joined the pro-Russian Chechen units is a more difficult question to answer given existing data limitations. We do know, however, that Zapad was staffed principally by Chechens who had not taken up arms against Russian forces; Vostok, on the other hand, was comprised mostly of former insurgents who were recruited for a variety of motives, including monetary payments, fear of retaliation against relatives, and disillusionment at prospects of victory. The various kadyrovtsy also drew heavily from insurgent ranks and, by 2008, would encompass at least 10,000 men.

In total, about 20,000 Chechens had joined Vostok, Zapad, or a Kadyrov-affiliated organization by the end of 2005, when large-scale sweep operations were phased out as Ramzan Kadyrov consolidated power.Footnote ⁵ Since his ascent to power in 2006, Ramzan Kadyrov has carefully cultivated a climate of fear through selective disappearances, targeted assassination of regime critics, and the nighttime burning of suspected insurgents' homes (Human Rights Watch 2009a; Russell Reference Russell2008).

DATA

The data set consists of known Russian, Chechen, and joint sweep operations that occurred in Chechnya between January 2000 and December 2005. In total, 680 sweep operations were identified (Table 1). A sweep operation was defined as an operation that seeks to separate insurgents from the noncombatant population by (1) isolating a specific location, typically by surrounding it with armed forces that control points of entry and egress; and then (2) by detaining, killing, or forcing the withdrawal of suspected insurgents through armed patrols, identification checks, and the searching of local residences. Such operations were staffed by >10 soldiers (drawn from Russian, Chechen, or joint units) and may or may not involve the physical destruction of specific individuals or sites within the targeted population.Footnote ⁶

TABLE 1.

Sweep Operations in Chechnya, 2000–5, by Soldier Identity

Data were drawn from nearly 35 Russian, English, and French language sources, including human rights organizations (Memorial, Human Rights Watch, Demos, Amnesty International, and Prima News Agency); the European Court of Human Rights; official Russian press releases (including the Ministry of Defense and Interior Ministry); rebel websites; and local, national, and international newspapers.Footnote ⁷ Using these same sources, a second data set was constructed that recorded nearly 3,500 insurgent attacks at the village level in Chechnya (1999–2005). The two data sets were then merged into a new geographic information systems (GIS) map of 409 village, town, and city locations. Of these known settlements, 119 experienced at least one sweep operation during 2000–5 (Figure 1).

FIGURE 1.

Frequency of Sweep Operations in Chechnya by Village, 2000–5

Note: 680 sweep operations, 2000–5. Groznyy was the site of at least 80 sweeps.

In total, the data set recorded 22,253 abuses inflicted during these sweep operations. This figure should be viewed as the floor, not the ceiling, of abuse estimates.Footnote ⁸ All reported abuses required at least two independent sources to be included in the data set. In cases of discrepant reporting, we adopted the lower of the two estimates. Eleven categories of abuse were coded: extrajudicial killing or wounding of an individual; forced disappearance; kidnapping for ransom; forced detention at the site of the sweep operation; sexual assault; torture; forced displacement; property damage; farm damage; and theft.

One key facet of the identification strategy requires some elaboration. Sweep operations are only a valid instrument insofar as the identity of the soldiers involved can be clearly established for both the researcher and the swept population. Can we assume that the operations are correctly coded and perceived, especially in the face of soldiers' efforts to mask, often literally if haphazardly, their own identities?

In point of fact, Chechen human rights activists and the general populace are readily able to distinguish between Russian and Chechen soldiers. Two important clues are commonly cited in reports and testimonials as clearly demarcating Russian-staffed sweeps from comparable Chechen ones.

First, since few Russian soldiers speak Chechen, and so many Chechens speak Russian with an accent, observers were consistently able to sort identity types based on the demonstrated proficiency of the language spoken. Second, local populations proved remarkably adept at identifying contextual clues that specified whether a unit was Russian or Chechen staffed. Indeed, although sweeps often used comparable numbers of soldiers, there was clear variation in the type of equipment employed. Better-equipped Russian units relied on armored vehicles (BTRs and BMPs), for example, whereas Chechen units, especially the kadyrovtsy, made due with heavy-duty Niva jeeps and trucks. Human rights spotters were especially attuned to these differences: in some cases, they managed to record the serial numbers painted on vehicle sides with such accuracy that individual military units can be identified. It is a reasonable assumption, then, that the identity of sweeping soldiers was common knowledge among targeted populations.Footnote ⁹

VARIABLES

The dependent variable, Insurgent Violence, is defined as insurgent-initiated attacks against Russian and pro-Russian Chechen military units, government officials, civilians, and infrastructure in or near populated settlements in Chechnya. Insurgent violence is operationalized in two ways. First, the difference between the number of insurgent attacks 90 days after and before a sweep operation in a given populated settlement is used to capture the change in the amount of insurgent violence. Second, the timing of insurgent attacks is operationalized as the lag in days between the sweep operation and the first insurgent attack from that populated settlement in the initial 90 days following the military operation.

I adopted 90-day treatment windows for two reasons. First, prevailing theories assume a tight temporal link between action and reaction, suggesting these windows are sufficient to capture treatment effects (Berrebi and Lakdawalla Reference Berrebi and Lakdawalla2007; Lyall Reference Lyall and Wilson2009). Second, difference-in-difference estimates of treatment effects are most reliable in the short-to-medium term (Duflo, Glennerster, and Kremer Reference Duflo, Glennerster and Kremer2007, 17). As the length between treatment and observed response increases, confidence in our estimates is diminished since (unobserved) events are increasingly likely to intervene. A 90-day pretreatment window also smoothes “spikes” in presweep insurgent violence that may be the immediate reason for the operation in the first place. These windows represent a pragmatic compromise: long enough to establish treatment effects, but not so long that causal claims become tenuous.

The independent variable, Treatment, is a binary variable that denotes whether a sweep operation was conducted by Russian or Chechen military units. As a supplemental analysis, I also use a binary Russian-only/joint patrol treatment variable as a finer-grained test of whether sweeper ethnicity shapes insurgent violence. Russian-only zachistki are coded as the “control” observations for each version of Treatment.

Sixteen demographic, spatial, and conflict-related covariates are also considered here. In most cases, covariates were measured well before the creation of Chechen-only battalions, ensuring that these covariates are not confounded with the treatment itself.

Population is the log of a settlement's estimated population in 2000. Population estimates were provided by the Danish Refugee Council (DRC), which conducted a Household Economic Survey in March–April 2000 (DRC 2002). Missing values were imputed from World Health Organization (WHO) estimates of prewar ambulatory facilities and expected caseloads of regional clinics and, in rare cases, from the size of wheat shipments by humanitarian organizations (WHO 2003). Wartime population estimates, although often unreliable, are important to include because insurgent violence is often positively correlated with population size: the larger the population, the greater the potential pool of insurgent recruits, and the greater the ability of insurgents to blend into the broader (noncombatant) population.Footnote ¹⁰

Second, while Chechnya's population is overwhelmingly Sufi Muslim, it is nonetheless divided internally into two brotherhoods (or Tariqa), the Naqshbandiyya and the Qadiriyya, that have their own historical experiences with Russia. Naqshbandiyya teachings were first introduced into Dagestan and Chechnya in the early 1800s and quickly became the basis for anti-Russian resistance. The 1864 defeat of its most famous adherent, Shamil of Gimry, created an opening for the spread of pacifist Qadiriyya teachings that had been circulating since the 1840s. Its pro-coexistence stance was quickly abandoned in the 1870s, however, in the face of Russian repression of leading Qadiriyya leaders. Now Qadiriyya adherents assumed the anti-Russian mantle, with Naqshbandi populations favoring collaboration. By the 1890s, the Qadiri had replaced the Naqshbandi in most, but not all, regions of Chechnya. These roles have remained mostly, although not entirely, stable across the ensuing decades and wars (Gammer Reference Gammer1994, 39–46; Reference Gammer2006, 45–52, 68–81; Zelkina Reference Zelkina2000, 121–35, 169–85).

At present, it is estimated that the Naqshbandi represent 10% of Chechnya's population. They are geographically concentrated in Chechnya's northern districts and have sizable populations in two large towns, Urus-Martan and Tolstoy-Yurt. Using these historical settlement patterns, as well as known Naqshbandi shrine locations, I created Tariqa, a binary variable that records whether a populated place is dominated by Naqshbandi.

Third, it is likely that the willingness of inhabitants to consider taking up arms is positively correlated with poverty levels. To capture relative poverty levels, I draw on the DRC's Household Economic Survey, which recorded the daily caloric intake of a large portion of the population (N = 228,021) (Bonke Reference Bonke2003, 3, 67–9). Following the council's codings, I created Poverty, a threefold variable that classifies a settlement's need of humanitarian assistance from none (“1”) to moderate (“2”) to severe (“3”).

Spatial characteristics might also influence insurgent attack patterns. Most studies of insurgency, for example, conclude that rough terrain facilitates guerrilla warfare (Collier and Hoeffler Reference Collier and Hoeffler2004; Fearon and Laitin Reference Fearon and Laitin2003; Galula Reference Galula2006, 23–5). As a result, Elevation records the settlement's elevation in meters (logged). Isolation measures the number of settlements that are found within five kilometers of the swept or control village. This captures the belief held among practitioners that isolated villages are easier to suppress because insurgents have few or no options when seeking to escape (U.S. Army Field Manual No. 3-24 2007, 185). The size and symbolic importance of the capital city, Groznyy, may also generate different dynamics than in Chechnya's plains or mountains. I therefore include a dummy variable for sweeps that occur in one of its four districts (Groznyy).

War-induced dynamics may also condition the nature of insurgent violence. Refugee centers, for example, arise endogenously from wartime violence but can shape subsequent attack patterns if they serve as incubators for a continued insurgent presence (Lischer Reference Lischer2005; Salehyan and Gleditsch Reference Salehyan and Gleditsch2006). A binary variable, TAC, therefore records whether a populated settlement hosted a UNHCR-sponsored temporary accommodation center sponsored by the United Nations High Commissioner for Refugees (UNHCR), by April 2000. Similarly, the location of military garrisons can either deter subsequent attacks or incite them by providing insurgents with a fixed target of symbolic and military importance. As a result, I draw on published reports and satellite imagery to record the location of military bases and/or interrogation facilities (Garrison) as of March–April 2000 (Human Rights Watch 2006; Memorial and Demos Center 2007). Finally, the insurgents themselves were divided in the 2000–5 time period between Shamil Basayev's radical Islam-inspired groups and the more nationalist forces of Doku Umarov. Because it is plausible that rebel organizations respond differently to state violence, I code whether the swept settlement was located in a district (raion) controlled by Basayev (Rebel).

ADDRESSING SELECTION EFFECTS

Given the observational nature of these data, we must take care to guard against mistaken inferences that might arise from selection bias given the nonrandom nature of these sweep operations. Indeed, the issue of treatment assignment looms large here because we are clearly not privy to the private information that guided how these soldiers chose which villages to cordon and search. Perhaps, for example, Russian and Chechen units identified villages according to different criteria—Russian units may concentrate on only the most violent locations, for exampleFootnote ¹¹—thus creating selection bias since these units are not sweeping comparable settlements.

Concern over selection effects is mitigated, however, by the nature of Russian and pro-Russian Chechen military deployments in Chechnya. Vostok (East) and Zapad (West), as their names imply, did have their own areas of responsibility in Eastern and Western Chechnya, areas that together encompassed the bulk of the Chechen population. The kadyrovtsy also possessed bases and detention centers that spanned much of Chechnya, whereas Russian forces were garrisoned throughout Chechnya. As a result, populated settlements in Chechnya were at risk of being swept by Russian, pro-Russian Chechen, and joint patrols because no villages fell within an exclusive “Russian” or “Chechen” zone. Unsurprisingly, many villages in the sample were swept by all three types of military units, indicating that they were not necessarily chosen according to unit-specific selection criteria (or, alternatively, that these units had similar selection criteria).

We cannot, of course, exclude entirely the possibility that villages are being targeted for sweep operations due to private information yielded by an unobserved covariate. We can, however, reduce the possibility of selection bias arising from three different observable sources. First, it is possible that presweep levels of insurgent violence condition the type of unit that is tasked with “sweeping” particular villages. Second, sweeping units may have heterogenous motives for choosing sweep locations that are conditional on soldier ethnicity, meaning that targeted populations experience different types of sweep operations. Third, sweep operations may be conditional on the past responses of targeted villages, suggesting that Russian and pro-Russian Chechen units may be assigned differently given expectations about anticipated insurgent behavior.

To control for selection bias arising from differential presweep levels of insurgent violence, I recorded the presweep number of insurgent attacks against Russian and Chechen forces, civilians, and infrastructure in a given settlement during the 90 days preceding the sweep operation (Prior attacks). This creates identical 90-day treatment windows before and after the sweep operation, permitting difference-in-difference estimation of treatment effects when similar swept and nonswept villages are paired.

The second, perhaps more subtle, form of potential selection bias stemming from heterogenous motives is more challenging to address. Indeed, Russian soldiers, often bereft of proper provisions (especially in the war's early years) and military discipline, are believed to use sweeps to engage periodically in large-scale looting and to inflict vengeance on populations for lost comrades (Human Rights Watch 2001, 7–20; 2002a, 13–18; Politkovskaya Reference Politkovskaya2003, 116–19). There is ample evidence for these claims. Organized theft by Russian soldiers—ranging from small household appliances to automobiles—was reported in nearly 18% of all Russian sweeps (82/458) but occurred in only 3.4% of Chechen-only operations (5/145). Similarly, at least one citizen was killed during 25% of total Russian sweeps (113/458), a level of extrajudicial killing not matched by the 11.7% recorded during Chechen-only operations (17/145).

The question of sweeper motivation poses a difficult methodological challenge since we do not possess an ex ante measure of motives (or, more broadly, the information available to the soldiers), and can therefore only impute them from revealed behavior. Moreover, it is apparent that in at least some cases, the sweep operations had only a tangential relationship to events in the targeted villages. “Zachistkas often take place at the whim of tactical [local] group commanders,” Emil Souleimanov has argued, “who can justify their actions with intelligence reports, whether real or falsified” (Souleimanov Reference Souleimanov2007, 176).

I address this potential selection problem by creating two dummy variables—Killing and Large-Scale Theft—that record whether at least one individual was extrajudicially killed and whether organized theft encompassing at least 10 residences occurred during the sweep operation. The addition of these covariates helps ensure that insurgents are responding to the same class of events. At the same time, subsequent matching on these proxies for motives poses an additional hurdle for demonstrating ethnicity's independent effects if Chechen soldiers kill and steal at lower rates because of their ethnic identity.Footnote ¹²

In addition, the level of abuse inflicted by Russian and Chechen forces may also condition the subsequent nature of insurgent violence. I therefore created Abuse, which is the logged estimated number of individuals abused during a sweep operation. “Abuse” refers to an act directed against a specific individual rather than a more diffuse sense of collective punishment. A village of 500 individuals may be collectively swept, for example, but it is only the acts attached to individuals—say, an extrajudicial killing or three disappearances—that are recorded as abuses. This restrictive coding rule likely underestimates the true extent of the abuses suffered by a population, but has the advantage of being verifiable while avoiding the assumption that all inhabitants of a settlement are abused uniformly during a sweep.Footnote ¹³

Finally, because it is likely that past patterns of state and insurgent violence may condition both the selection of sweep locations and the subsequent insurgent response (Wood Reference Wood2003, 237–9), I created History, which records the number of sweeps a village has endured prior to the current sweep operation. Nearly all civil war studies have concluded that an incumbent's use of indiscriminate repression fosters grievances among the victimized population that translate into greater insurgent violence (i.e., Azam and Hoeffler Reference Azam and Hoeffler2002; Kalyvas Reference Kalyvas2006, 146–72; Tse-tung Reference Tse-tung2000, 93; Valentino Huth, and Balch-Lindsay Reference Valentino, Huth and Balch-Lindsay2004; but see Lyall Reference Lyall and Wilson2009). As a result, we might expect that insurgent violence increases with the number of sweeps a population has already experienced.

MATCHING

Matching is a method of data preprocessing in which treated cases are paired with similar, if not identical, control cases to separate treatment effects from shared background covariates (Ho et al. Reference Ho, Imai, King and Stuart2007; Morgan and Winship Reference Morgan and Winship2007; Rubin Reference Rubin2006). By squeezing out variance among possibly confounding covariates across groups, matching simultaneously reduces bias in our estimates of treatment effects and model dependency because it removes “extreme counterfactuals,” that is, control observations with no analogue among the treated cases (King and Zeng Reference King and Zeng2006). Pairs were first identified using MatchIt (Ho et al. Reference Ho, Imai, King and Stuart2006) and 1:1 nearest neighbor matching with replacement, creating 145 pairs of control and treated observations from the original sample of 603 Russian- and Chechen-staffed zachistki. I then matched manually within these results to reduce reliance on any one control observation—none was used more than four times—and to privilege certain variables, including Prior Attacks, Killing, and Large-Scale Theft, where concern over selection effects makes it imperative to remove imbalance across groups.

Table 2 reports the closeness of the matched groups using three different balance tests.Footnote ¹⁴ Standardized bias is the difference in means of the treated and control groups, divided by the standard deviation of the treated group. A value of ≤0.25—signifying that the remaining difference between groups is less than one fourth of a standard deviation apart is considered a “good match” (Ho et al. Reference Ho, Imai, King and Stuart2007, 23fn15). Wilcoxon rank-sum test values are also provided to determine if we can reject the null hypothesis of equal population medians. Finally, Kolmogorov-Smirnov equality of distribution tests are also generated for continuous variables; values ≤.1 suggest that the distribution of means is highly dissimilar, whereas values approaching 1 signify increasing similar distributions (Sekhon Reference Sekhon2006).

TABLE 2.

Balance Summary Statistics and Tests: Russian and Chechen Sweeps

Notes: 145 matched pairs. Matching with replacement.

As Table 2 confirms, these pairs are closely matched, meeting or exceeding every standard for variance across all three balance tests. Closeness of fit between groups is especially important for two clusters of covariates, namely, those dealing with why the sweep was conducted (treatment assignment) and the level of violence inflicted by the sweeping soldiers. Prior attacks, for example, are almost identical across groups, removing the concern that Russian and Chechen units are selecting into different threat environments.Footnote ¹⁵ Russian and Chechen operations are also characterized by similar reported levels of large-scale theft and deaths among the targeted populations, thereby controlling, if only partially, for the possibility that these operations were driven by different motives.

We must also ensure that these pairs are closely matched on the level of abuse inflicted by sweeping soldiers if we are to separate the effects of ethnicity from the magnitude of violence visited on the targeted populations. Here, too, the control and treated groups are highly similar. For example, the average number of individuals abused per sweep is comparable across units, with Russians abusing 10 individuals, and Chechens 11, during each operation. Moreover, Chechen-swept villages had been the site of 7.72 prior operations on average, whereas Russian-swept villages had similarly been “swept” an average of 8.7 times in the past. Remaining differences in sociodemographic and spatial variables are negligible: Russian-swept villages are .44 meters higher on average than Chechen-swept counterparts, for example, and possess an average of 136 more individuals.Footnote ¹⁶

Finally, these observations were also matched on identical 90 day pre- and posttreatment windows within the same year to control for maturation effects or any potential bias created by a common trend not produced by the treatment itself. In actuality, the paired sweeps are occurring about two weeks apart in time.Footnote ¹⁷ This close temporal matching has the added advantage of controlling for variation in climatic conditions. Although typically overlooked, climatic and seasonal variance can affect the military operations of both insurgents and incumbents, creating nonrandom changes in their behavior. In Chechnya, for example, the combination of unpaved roads and rainfall can blunt sweep operations that rely on heavy mechanized formations for movement. Similarly, heavy snowfall can force rebels to winter in towns, whereas verdant summers can enhance insurgent mobility by providing cover. By matching on time, however, we control for these nonrandom biases.Footnote ¹⁸

FINDINGS

The empirical analysis unfolds over two stages. First, difference-in-difference estimation is used to assess the impact of sweeper ethnicity on insurgent attacks in paired settlements.Footnote ¹⁹ Second, hazard models are adopted to estimate how sweeper ethnicity affects the timing and conditional probability of observing the initial postsweep insurgent attack.

Patterns of Insurgent Violence

Do we observe a difference in postsweep levels of insurgent violence after Russian- and Chechen-only operations? As Table 3 illustrates, the answer is clearly yes. While beginning from nearly identical presweep levels of insurgent attacks, these settlements exhibit sharply different post-sweep levels of violence, conditional on the identity of the sweeping soldiers. Insurgent attacks increase by about 7% in the 90 days following a Russian sweep; such attacks decrease by nearly 33% after Chechen-staffed zachistki. Using difference-in-difference estimation, the treatment effect of shifting sweeper ethnicity from Russian to Chechen is a 1.22 decrease in the mean number of attacks in the 90 days postsweep treatment window. Put differently, there is nearly a 40% decrease in the mean number of insurgent attacks when shifting from Russian to Chechen soldiers. Taking the 452 cumulative attacks that preceded Russian sweeps as our baseline, we would find only 273 postsweep attacks (or between 345 and 205 attacks with a 95% confidence interval) if Chechens, not Russians, had conducted these operations.

TABLE 3.

Impact of Sweeper Ethnicity: Difference-in-Difference Estimation

Notes: 145 matched pairs.

Table 4 reports ordinary least squares (OLS) regression estimates of Treatment's effects both alone (Model 1) and with all 16 covariates added as control variables (Model 2). In each case, the decrease in mean attacks when shifting from Russian- to Chechen-staffed sweeps is substantively large and statistically significant at the p = .0001 level. We might worry, however, that this effect is driven largely by dynamics specific to Groznyy, Chechnya's much contested and highly symbolic capital city, thus obscuring broader patterns in the remaining populated settlements. I therefore reestimated Models 1 and 2 without observations from Groznyy's four districts (Models 3 and 4). Once again, Treatment remains both highly significant and substantively important, with a less than 10% attenuation in ethnicity's effect being reported in either model.Footnote ²⁰

TABLE 4.

Treatment Effects: OLS Regression

Notes: Robust standard errors clustered on individual populated settlements are listed in parentheses. **p < .01, ***p < .001.

Although the average treatment effect provides a useful summary measure, it is plausible that ethnicity's causal impact actually varies over time. That is, although the treated and control groups are matched on similar time windows to control for common temporal trends, it is plausible that ethnicity has a different impact in 2002 than, say, 2005, owing to endogenous developments in the war itself. I therefore reestimated differences in mean insurgent attacks for Russian and Chechen sweeps at six-month intervals for 2004 and 2005, a time period that encompasses 83% of all recorded Chechen-only sweeps. As Figure 2 illustrates, there is some variation in the mean number of postsweep attacks, but the overall trend remains remarkably consistent by sweeper ethnicity across time. A small increase in mean insurgent attacks is consistently noted after Russian sweeps, for example, while Chechen sweeps are uniformly marked by a decrease in mean insurgent attacks. In short, sweeper ethnicity has causal effects on insurgent violence at both the aggregate level (2000–5) and within finer-grained increments of time (2004–5).

FIGURE 2.

Difference in Mean Postsweep Attacks, by Sweeper Identity, 2004–5

Note: Nearly 83% of all Chechen-only sweeps occurred during 2004–5.

Timing of Insurgent Attacks: Estimating Hazard Rates

A second empirical strategy for estimating ethnicity's causal effects involves examining the timing of the insurgents' first postsweep attack. To date, the temporal dynamics of civil wars have largely been ignored. Yet we may uncover clues about the impact of ethnicity if we discover that the timing, and not simply the amount, of insurgent violence is conditional on the identity of the sweeping soldiers.

Do we observe differences in the timing of the first insurgent attack after Russian and Chechen sweeps? Simple Kaplan-Meier failure estimates reveal that the timing of attacks does in fact hinge on sweeper ethnicity. Insurgents are, for example, much more likely to attack after a Russian operation. At least one attack was recorded after 79% of Russian sweeps (115/145), whereas only 62% (90/145) of Chechen sweeps were followed by an insurgent attack.Footnote ²¹ As Figure 3's left panel illustrates, insurgent response times also vary considerably by soldier ethnicity. Insurgents typically attack much more quickly after Russian operations: it takes only three days for 25% of Russian-swept villages to record an attack, compared with nearly 13 days after Chechen operations to reach the same 25% “failure” rate. Similarly, half of all Russian-swept villages record at least one attack only 17 days after these operations; Chechen-swept villages, by contrast, take nearly 44 days to reach the same level. After 57 days, a full 75% of all Russian-swept villages have been the site of an insurgent attack, a failure rate never equaled by Chechen-swept villages.

FIGURE 3.

Timing and Hazard of First Postsweep Insurgent Attack

Note: 290 Observations, 2000–5. Hazard rates obtained using a Weibull survival regression with gamma frailty terms clustered on individual settlements (Model 6 in Table 5).

We can also test the relationship between sweeper ethnicity and the “hazard” rate of observing an insurgent attack using parametric survival regression. Weibull regression was chosen because it provides the smallest Akaike Information Criterion value, meaning that Weibull offers the most efficient estimation. All regressions were run with shared frailty terms centered on individual populated settlements to capture residual heterogeneity among subgroups where repeated observations are the norm. Essentially a form of random effects for duration models, the frailty function is particularly useful in the absence of randomization because it accounts for unobserved heterogeneity that may lead certain settlements to “fail” at rates systematically higher than other locations (Box-Steffensmeier and Jones Reference Box-Steffensmeier and Jones2004, 146–8). In these models, coefficients should be interpreted as the hazard ratio associated with a particular variable, with a 1 signifying no impact on the “hazard” of observing an insurgent attack and coefficients greater than or less than 1 indicating an increase or decrease in likelihood of an attack, respectively.

As Table 5 details, Treatment is associated with a substantial decrease in the odds of observing a postsweep attack. More specifically, a shift from Russian to Chechen soldiers is associated with a 60% decrease in the likelihood of a postsweep attack (or between 72% and 42% with a 95% confidence interval) in Model 6. Dropping Groznyy-specific observations leaves the estimated hazard rate largely changed at about a 55% decrease (or between 70% and 32% with a 95% confidence interval). Figure 3's right panel graphically illustrates the sharp variance in estimated hazard rates by sweeper ethnicity, with the most dramatic difference occurring in the initial three weeks after the sweep occurs.

TABLE 5.

Treatment and Insurgent Attack Hazard: 90-Day Treatment Windows

Notes: Weibull regression with gamma frailty terms centered on individual villages (standard errors in parentheses). ***p < .001.

Note that these findings are not artifacts of the decision to adopt 90-day treatment windows. As a robustness check, I reestimated Models 5 to 8 with treatment windows shortened to 14 and 30 days. Treatment continues to be highly significant and associated with a marked decrease in the probability of observing an insurgent attack. Model 6, for example, returns a hazard coefficient of 0.366 (p = .001) at the 14-day mark and 0.418 (p = .001) at the 30-day mark, suggesting a decrease in insurgent attack hazard rates of about 73% and 58%, respectively, when moving from Russian to Chechen soldiers.Footnote ²²

Additional Evidence: Joint Sweeps

We might wonder, however, what happens when Russian and Chechen units are paired together during the same operation. It is plausible, for example, that these combined units outperform their monoethnic counterparts because they combine Russian military power with Chechen linguistic skills and local knowledge (Felter Reference Felter2005). Ethnicity, in this view, plays only a secondary role in explaining insurgent violence, with greater weight placed on unit characteristics and their internal dynamics. Yet if the advantages of coethnicity are non-transferrable or, alternatively, are degraded by the presence of non-coethnics, then we are likely to observe joint units at or near Russian-only units' performance. This comparison helps locate ethnicity's causal effects by testing whether coethnic advantages are a function of specific Chechen soldier attributes or arise instead from interaction between soldiers and the population, in which case coethnic advantages may actually be attenuated by overt cooperation with non-coethnic forces.

Do insurgent attack patterns after joint sweeps resemble those of Russian-only sweeps? To answer this question, I repeated the matching procedure outlined previously using joint patrols as the treatment and Russian-only sweeps as control observations. Nearest neighbor matching with replacement was adopted, and no control village was used more than twice. Particular emphasis was placed on exact matching with covariates (Prior Attacks, Large-Scale Theft, and Killing) that might condition village selection by sweeping forces.

As Table 6 reveals, the balance across all covariates is excellent regardless of the specific equality of distribution test—standardized bias, Wilcoxon rank-sum, or Kolmogorov-Smirnov—employed. We are therefore left with 77 joint operations matched with Russian-only sweeps.Footnote ²³

TABLE 6.

Balance Summary Statistics and Tests: Russian and Joint Sweeps

Notes: 77 matched pairs.

This matching yields a surprising (non)finding: difference-in-difference estimation indicates that the treatment effect associated with a shift from Russian-only to joint operations is not statistically significant (Table 7). Although a modest reduction in the mean number of postsweep insurgent attacks is recorded, this difference fails to reach conventional levels of significance in any of the models estimated (Models 9–12). This finding is robust to both the inclusion of the full set of matched covariates and the dropping of all Groznyy-specific observations. Thus, while some improvement over Russian-only operations is observed, the magnitude of the estimated treatment effect pales when compared with the change associated with shifting from Russian-only to Chechen-only operations.

TABLE 7.

Treatment Effects, Joint Patrols: OLS Regression

Notes: Robust standard errors clustered on individual populated settlements are listed in parentheses. ***p < .001.

We can also estimate the impact of Treatment on hazard rates of insurgents' first attack in the 90 days following a sweep operation. No difference is observed in insurgent response rate using Kaplan-Meier failure estimates. We observe 55 attacks (failures) after the 77 Russian-only operations, whereas 63 are recorded for the 77 joint sweeps.Footnote ²⁴ A full 25% of swept villages experience an attack in the first four days following both Russian and joint operations, with the first divergence between failure rates coming at the 50% failure mark: it takes 18 days for Russian-swept villages to reach 50% failure versus 24 days for joint-swept villages. Roughly two months after the initial sweep, 75% of all joint-swept villages have experienced at least one attack. Russian-swept matches, by contrast, never reach this failure mark.

Repeating the survival analysis, Table 8 reports hazard estimates for Treatment using Weibull regression with gamma frailty terms clustered on individual villages. These reports are useful for underscoring a largely negative conclusion: the shift from Russian-only to joint operations is associated with a modest decrease in the hazard of observing an insurgent attack, but none of these coefficients reach conventional levels of statistical significance. Resetting the treatment windows at 14 and 30 days as a robustness check yields the same negative conclusion, namely, that the hazard rates associated with Russian and joint sweeps are broadly similar regardless of the length of the treatment windows.

TABLE 8.

Treatment and Insurgent Attack Hazard, Joint Sweeps: 90-Day Treatment Windows

Notes: Weibull regression with gamma frailty terms centered on individual villages (standard errors in parentheses). ***p < .001.

Simply put, joint operations are triggering broadly similar insurgent reactions as Russian-only sweeps. To be sure, there is a modest decrease in insurgent violence after joint sweeps relative to Russian-only operations. Yet rather than creating an ideal pairing of Chechen intelligence and Russian military might, these joint operations failed to match the performance of Chechen-only operations. It appears that the presence of Russian forces is sufficient to dissipate most of the advantages that accrue to Chechen soldiers.

CONDITIONAL EFFECTS OF VIOLENCE: IDENTIFYING THE MECHANISM(S) AT WORK

Substantial evidence thus exists to suggest that patterns of insurgent violence are at least partly conditional on soldier identity. Explaining why this particular relationship is observed is difficult, however. Indeed, there is no shortage of plausible mechanisms at work, a problem compounded by the fact that these mechanisms typically generate equivalent expectations about insurgent behavior. Although we cannot definitely rule in favor of a specific mechanism with available data, the following sections begin the task of eliminating some mechanisms while also providing limited evidence in favor of others.