Competitive Alliances and Civil War Recurrence

Abstract

Why do internal wars start anew after they apparently end? I argue that rebel arrangements made for strategic reasons during wars sometimes create provocative effects even after conflicts end; coalitions formed between opposing groups during conflicts often precipitate disruptive commitment problems at the end of wars. This competition can abet the conflict renewal process, especially after wars terminating in military outcomes. Using new data on competitive militant alliances in civil wars, this study probes how rebel rivalries impact peace duration after wars. The evidence indicates that rivalry within coalitions shortens the period of postconflict peace. Wars ending in military victories give way, as many scholars argue, to lengthier periods of peace. But this effect reverses in the context of conflicts characterized by competitive alliances. Indeed, the combination of competitive alliances and a military victory strongly precipitates a resumption of hostilities. This perpetuation of the “conflict trap” proves especially pronounced when rebels win wars. My study implies that peacekeeping or third party forces may find the least local consent for their presence precisely where they matter most in post-conflict environments. As former work has shown, victorious rebels, having wrested power by force from vanquished governments, have relatively little desire for outside interference.

These people are temporary allies and must be treated with respect as long as they are of use to us. There is no question that they are enemies in the end.

— Daniel Ortega, 1979 ¹

Introduction

Factions in civil wars rarely create and consummate alliances in an orderly fashion. More often, alliances derive from expedience—as forced marriages of necessity between groups with a history of animosity and little trust in one another. Consider a few examples. A report from the University of Pennsylvania on civil war in Africa describes the dynamics between the Ninja, Cobra, and Cocoye militias as follows: “The politics of Congo-Brazzaville are triangular.…  Two sides of the triangle become aligned against the third, and those alliances are constantly shifting.” ² In Zimbabwe in the early 1980s, two former allies of the Patriotic Front, the Zimbabwe African National Union - Patriotic Front (ZANU-PF) and the Zimbabwe African People's Union (ZAPU), fought a bitter war between themselves immediately after their successful armed struggle for independence. And a tenuous alliance between deposed head of state Prince Sihanouk and the communist Khmer Rouge resulted from Lon Nol’s power grab in a 1970 Cambodian coup. Before being forced from power, Prince Sihanouk had continually denounced the Cambodian communists as “worse than common bandits.” ³ Unsurprisingly, the Khmer leadership turned on Sihanouk as early as 1973, electing to even liquidate some of its own members who had collaborated with the former prince. These examples notwithstanding, the study of inter-faction alliance dynamics remains relatively underdeveloped within the growing literature on civil wars.

Indeed, overall scholarly interest in internal conflicts continues to expand. Wars within states are now more common than wars between them. In fact, the prevalence of civil war points to an empirical regularity scholars describe as the “conflict trap” (Collier et al. 2003 ). Once a country falls into a civil war, its likelihood of relapsing into a new war significantly exceeds the odds for most countries to experience any civil conflict at all. This war recidivism phenomenon means that there are more civil wars than there are countries that have had them. The uneven distribution of internal conflicts across states features in many civil-war datasets. The 108 cases of civil war from 1944 to 1997 in the Correlates of War dataset span only 54 nations, and as few as 26 countries in the dataset experienced only a single war. Furthermore, the 124 civil wars listed in Doyle and Sambanis ( 2000 ) occurred in 69 nations. ⁴ Despite this consistent pattern across multiple data sources, civil war recurrence and conflict transformation remain open issues.

The field of civil-conflict resolution now profits from disaggregating conflicts into the multiple actors that often comprise them. Civil wars often exhibit rival factions and multiple challengers competing for common and sometimes conflicting goals. The Peace Research Institute of Oslo (PRIO)/Uppsala Armed Conflict database identifies 288 incidents of internal war since 1945, of which 30 percent include more than two combatants; some contain up to 10 participants. ⁵ The fact that multiple actors frequently participate in civil conflicts is more than an empirical artifact of the PRIO-Uppsala data. Combining figures from Cunningham ( 2006 ) on veto players in civil wars with the Doyle and Sambanis ( 2000 ) dataset reveals that just shy of 40 percent of the civil wars coded by Doyle and Sambanis involve three or more veto players, as identified by Cunningham ( 2006 ). For example, civil wars in Liberia, former Yugoslavia, Lebanon, Sudan, and the Democratic Republic of Congo were all multiparty affairs. Ongoing hostilities in Syria, Libya, and Yemen also exhibit a diversity of competing participants.

Thus, a growing body of scholarship relaxes the assumptions that civil wars are two-player arrangements and that rebels are unitary actors. By doing so, this work illuminates important aspects of civil wars. ⁶ But it so far largely neglects the role alliances may play in generating conflict recurrence. I argue that high degrees of competition between allied rebel parties produce a security dilemma internal to coalitions, thereby exacerbating commitment problems. This competitive dynamic greatly increases the chances of war relapse, even, and especially, after conflicts reach their military conclusion. I present newly collected data on civil war alliances. My dataset specifically identifies both instances of alliance formation among warring rebel groups and also internal rivalries within such alliances. Using the new data, I probe whether, and how, competitive rebel alliances impact both the likelihood and timing of war recurrence. ⁷ My findings suggest that conflictual alliances significantly precipitate new wars. They also result in shorter spells of peace in postconflict settings. Additionally, the results refine the developing consensus that military outcomes, and in particular rebel victories, yield relative stability and less war relapse. Postconflict peace in the wake of government defeat tends to emerge primarily in the absence of fractious alliances among victorious rebels.

Prior Explanations for War Recurrence

The puzzle of civil war recurrence naturally evolved from earlier work on civil war onset (Collier and Hoeffler 1998 , 2004 ; Fearon and Laitin 2003 ). Factors critical for spawning original conflicts—such as levels of economic development or weak political institutions—no doubt condition the sustainability of peace and the likelihood of war recurrence. But scholars now look beyond these factors in order to understand how long conflicts last and how they end (e.g., DeRouen and Sobek 2004 ; Mason, Weingarten, and Fett 1999 ). However, much work on civil war recurrence has tended to emphasize the characteristics of war termination.

Most analysts agree that how civil wars end impacts the likelihood of war recurrence in the postconflict setting. However, scholars continue to debate the nature of postwar conditions and the mechanisms linking them to persistent stability and renewed violence. Early efforts argued that negotiated settlements create greater odds of renewed conflict than outright military victories by one side. ⁸ Warring sides generally preserve their ability to resume hostilities in the wake of negotiated settlements. For instance, Wagner ( 1993 ) argues that settlements allow parties to maintain organizational capacity and identity after the war, even if they disarm as part of the peace process. On the other hand, a military victory results in the defeat, disarming, and sometimes the outright destruction, of the enemy. This significantly curtails the capacity of the losers to renew combat under such circumstances, especially when the winning force maintains its power to repress (see also Licklider 1995 ; DeRouen and Bercovitch 2008 ). However, Pearson et al. ( 2006 ) find no evidence of greater postwar stability after military outcomes. Doyle and Sambanis ( 2006 ) similarly discount this claim in their own analysis.

Nested within the military outcomes issue sits a related question: Which type of military victory—rebel or government—proves more stable, and why? Some of the work addressing this query indicates that victories achieved by rebels are more likely to yield sustainable peace than those by incumbent governments. Toft ( 2010 ) supports this assertion. She argues that rebel victories constitute the ‘best case' outcome in terms of avoiding civil-war recurrence and in terms of generating the greatest levels of democratization in the long-term, postwar environment. Research by Quinn, Mason, and Gurses ( 2007 ) posits that rebel triumphs are more likely to produce durable peace because they tend to completely delegitimize government claims to power; by contrast, when rebels lose they do not suffer similar losses in legitimacy. Mason, Gurses, Brandt, and Quinn ( 2011 ) refine this argument, suggesting that it may take time for rebels to consolidate their victory. But once they do, peace will likely prove more durable than when the government defeats a rebellion. ⁹ However, Kreutz ( 2010 ) finds that government victories reduce the incidence of war recurrence. In short, we lack a clear consensus concerning the matter.

Moreover, the field has begun to turn away from a more narrow approach that focuses simply on war and peace as outcomes. Joshi and Mason ( 2011 ) relate the type of conflict termination to the size of the resulting governing coalition that follows the war. Their research reveals that negotiated settlements are more likely than military victories to produce broader governing coalitions after wars. But when conflicts do end in military victories, victorious governments prove more likely than rebels to expand the size of postwar coalitions. And notably, their work also indicates that the size of such governing coalitions positively affects peace duration in the aftermath of wars. A related effort by Mukherjee ( 2006 ) suggests that the sustainability of peace after decisive military victories by either side improves when the vanquished party receives power-sharing concessions. Critically, the interactive effect between power-sharing agreements and decisive military victories significantly increases the duration of peace, rather than the independent effect of each variable.

Findings on external intervention remain central to this body of research. The growing literature on peacekeeping and negotiated settlements highlights how commitment problems regularly bedevil efforts to successfully end wars and prevent their renewal. Walter ( 1997 ) demonstrates that attempts at negotiated settlements often fail to bring peace to fruition in civil wars because the combatants are unable to make credible guarantees to the other parties. Negotiations fall apart because civil-war antagonists are asked to do “what they consider unthinkable”—to demobilize, disarm, and disengage their forces. They therefore risk rendering themselves vulnerable to surprise attacks or reneging by their opponents. ¹⁰ Expanding upon this insight, Walter ( 2002 ) finds that central to the sustainment of peace-promoting negotiations is outside intervention—specifically, third parties capable of enforcing, verifying, and mediating the implementation of posttreaty transitions among vulnerable parties. ¹¹ Adding to this line of inquiry, Doyle and Sambanis’ ( 2006 ) investigation into peace building suggests that international assistance is critical for stable peace, especially when wars have been particularly violent and have seriously eroded local capacities. The authors find that peace operations tailored to the task can actually “trump military victories.” ¹² That is, comprehensive peace agreements have better success rates than military victories for delivering peace.

Fortna ( 2008 ) also sheds light on how and where peacekeepers contribute to sustained conflict resolutions. ¹³ Her work demonstrates that international peacekeeping forces go where they are in most need and where the chances of success are most grim. ¹⁴ This includes multiparty conflicts marked by weak governments, high lethality, and large population displacement. Moreover, peacekeepers should be in relatively higher demand in wars marked by weak governments and strong rebel forces but not in cases where rebels or governments achieve outright victories. ¹⁵ Fortna further argues that peacekeepers disrupt pathways back to war by altering the incentives of belligerents, lowering uncertainty about participant intentions, promoting inclusion, and lessening the risk that skirmishes will reignite hostilities. More recently, Call ( 2012 ) emphasizes the totemic role of political participation in peace-building efforts. His work suggests that the political exclusion of former warring parties plays a decisive part in many instances of war recurrence.

Although a qualified consensus holds that external interventions tend to improve the chances of durable peace after internal wars, disagreement persists about which types of missions are most successful and also how to measure success. ¹⁶ Rates of success therefore tend to vary from study to study, and we find little consistency about what levels of violence properly constitute a return to war. ¹⁷ Recent studies also call into question peacekeepers’ capacity to reduce violence (Costalli 2014 ), the means by which interveners typically pursue peace-building (Autesserre 2010 , 2014 ), and the potential for open-ended international commitments to distort domestic politics (De Waal 2009 ). This debate demonstrates that peacekeeping can help consummate desirable outcomes but that interventions often fail to reach their full potential.

In summary, the literature on sustaining peace after civil wars contains several related branches. Work dedicated to negotiated settlements emphasizes the confounding nature of commitment problems and the capacity of third parties to mitigate such issues. Research focusing on how wars end has begun to evolve to reflect the interactive nature of various war outcomes with other conflict attributes. This article combines insights from previous scholarship, and by focusing on competitive alliance dynamics, helps identify the conditions under which a given outcome is more or less likely to generate stable postconflict scenarios.

Alliances and the Shadow of Rivalry

The role of alliances in internal conflict represents a relatively new area of scholarly interest. To date, few works bring insights from the study of alliance formation at the international level to the study of intrastate conflict. ¹⁸ But as already highlighted, many civil wars clearly involve shifting coalitions. In this section, I attempt to fold this empirical reality into a parsimonious account of coalition dynamics within an internal conflict setting. I argue that the shadow of rivalry between groups can exacerbate commitment problems and thereby engender repeat conflicts.

International relations scholars generally think of alliances as a means for at least two actors to augment or pool their defensive capabilities and resources against a common external threat. As Snyder ( 1997 , 4) writes, “Alliances are formal associations of states for the use (or nonuse) of military force, in specified circumstances against states outside their own membership.” ¹⁹ We might find it tempting simply to transfer this view from the international arena to the domestic realm. Thus, we might view civil-war alliances as, in essence, simply efforts by rebel groups to augment and enhance their overall fighting capacity. We would expect, all things being equal, that alliances would increase the success of rebels. Unfortunately, matters prove far more complicated: Coalitions spawned in the nettle of rebellion generally grow into less-formalized arrangements and often involve significant ambiguity. Worse still, in the highly anarchic environment characterizing many internal wars, alliance dynamics and the outcomes they produce seldom take a well-ordered form. ²⁰

Alliances, moreover, need not imply cooperation. This inference animates both interstate and civil-war coalitions. Uncertainty about alliance commitment means that allying parties—be they states or rebelling factions—possess opportunistic incentives to improve their security at the expense of their allies (e.g., Snyder 1984 ; Bapat and Bond 2012 ). Furthermore, an alliance seldom operates with the efficiency of a unitary actor. Members of an alliance may share a common objective to deter or expropriate, but they likely will disagree on how best to achieve this goal. The internal and sometimes divisive relations of the alliance itself condition the choices allies make. The presence of a common enemy may initially drive groups together, but external threats do not operate independently of the discord often present within alliances. The two forces combine and interact, leaving actors no choice but to deal with both simultaneously. Allying parties must deal with antagonistic allies as well as enemies (Weitsman 1997 ). The result? Very often, and even in the presence of common danger, the ability among allies to pool common strengths is eclipsed by even greater division, animus, and mutual fear between them. These dynamics, while often characteristics of alliances between states, serve to also confound the traditional pursuit of revolt at the domestic level. ²¹

Conventional wisdom holds that the “rebel’s dilemma” essentially redounds to a problem of collective action and collective dissent (Lichbach 1998 ). But inserting alliance calculations into this array complicates an already difficult problem for groups pursuing the task of rebellion. Clearly, alliance politics present an instance of the collective action problem. This stems from the very public goods nature of defense efforts and deterrence. Shared defense is a burden especially prone to the kinds of difficulties associated with common-resource issues. Many assume that defense contributions within an alliance are perfectly substitutable and hence, benefit all members of the coalition (e.g., Sandler and Hartley 2001 ). Free riding and exploitation prove to be common features of alliance arrangements, varying on threat levels and the relative strength of alliance members. And in the case of civil war, exploitation means more than mere shirking or abandonment. Where friction and uncertainty reign supreme, pure survival presents the more pressing dilemma for rebel groups formulating their alliance strategies. As a consequence, they face a security dilemma of various degrees that stems from factors entirely internal to the alliance.

A few examples help illustrate this point. Some coalitions exhibit only minimal internal discord and rivalry. The National Convention (NC) fighting for Namibian independence was a fairly cohesive coalition between the South West Africa People’s Organization (SWAPO), the South West African National Union (SWANU), and the National Unity Democratic Organisation (NUDO) elements. By contrast, the Sandinista National Liberation Front (FSLN) fighting in Nicaragua’s civil war against the Somoza dynasty was a coalition beset with rivalry. Strategic and tactical differences over how to achieve power escalated to such a point during the 1970s that they became less internal disputes than outright conflicts between opposing movements. The fighting among the three branches of Sandinismo became so bitter that factions only reestablished unity under extreme pressure from Cuban leaders in 1979. ²² Similarly, infighting among allied factions marks Chad’s lengthy history of internal conflict. For example, the National Liberation Front (FroLiNat) was a highly disputatious coalition of eleven loosely organized rebel forces. When FroLiNat finally overthrew the government in 1979, its various factions immediately fell into fighting among themselves, as one leader after another evicted his predecessor only to meet with further rebellion (Zartman 1993 ). These internal disputes produced a paradoxical effect: The factions could not rule together, but no one of them could rule alone (Zartman 1986 ).

Fear of exploitation implies that rival groups, even when they come together in a common cause, appreciate a distinct possibility that they may fight one another—and not infrequently, they do. This was the case in East Timor in 1975 when Revolutionary Front for an Independent East Timor (FRETILIN) forces fought a mini civil war against the Timorese Democratic Union (UDT), its former coalition partner. The two groups would later re-coalesce into a united front in the 1980s (Pinto, Jardine, and Nairn 1997 ). Similarly, Kurdish groups in Iraq maneuvered to eliminate each other in the late 1970s. By 1980, a de facto war between the Patriotic Union of Kurdistan (PUK) and the Kurdistan Democratic Party (KDP) emerged, with various other factions also involved. ²³ However, by 1986, while the KDP and PUK continued to denounce each other, there was growing recognition that they could no longer afford such internecine conflict. In 1987, the factions announced a formal alliance and fashioned the Iraqi Kurdistan Front (IKF) later the same year.

And at the most extreme level, the ongoing war in Somalia reveals that alliance loyalties in civil war are not permanent and cannot be taken for granted. This failed state has shown precisely the kind of susceptibility to the war trap—where various groups and clans come together, break apart, remove the government, and then fight anew to replace it. The fighting against the Barre regime in the late 1980s devolved into a broader war between the government and a loose coalition of rebel organizations. The factions generally fell into two main groups: North Somali movements, headed by the Somali National Movement (SNM); and Central Somali movements, led by the United Somali Congress (USC). The war escalated until the Somali army eventually collapsed in defeat. However, instability and rivalry within the victorious coalition hastened a new round of fighting and insurrection among the tribal factions that continues to this day. ²⁴ These instances demonstrate that civil war alliances are resilient constructs, capable of accommodating rivalry and revenge dynamics—at least temporarily, but seldom permanently. ²⁵

These examples also make clear that fear of reprisal in largely anarchic conditions means allying parties must contend with a security dilemma. ²⁶ Once rival groups elect to form an alliance against a common enemy, they face what Snyder ( 1984 ) describes as a “secondary alliance dilemma.” Defection is of paramount concern to members of a coalition. And defection is always among the options within an alliance marked by at least a minimal degree of internal discord. But in civil wars, defection implies more than abrogation of alliance obligations; it may invoke the wholesale liquidation of former partners. Thus the security dilemma brings about its own consequences for allying parties in civil wars. The first deals with levels of preparation and fighting capacity. Caustic and unstable coalitions mean allied groups are highly incentivized to respond in kind to any increases in strength or capability on the part of their partners. Should a coalition collapse, being caught unprepared in an “every man for himself” situation is a state of affairs that rebel groups can ill afford. The security dilemma therefore dictates that groups cannot sit by idly. When one party increases its defensive provisions, the other is inclined to replicate this action for fear of being abandoned or attacked.

The somewhat unexpected upshot of this is to dissuade free riding by making it a costly option. In the absence of such internal competition, each faction preserves a distinct incentive to free ride on its ally’s capabilities. However, the specter of the security dilemma means free riding is not optimal. As the Somalia episode reveals in its starkest form, free riding in the presence of a rivalrous ally is not a costless strategy. In fact, it can be suicidal. It is precisely under such unpredictable and capricious conditions that the costs of nonparticipation and free riding may equal those of participation (Kalyvas and Kocher, 2007 ). The extended result of the security dilemma and the penalty it imposes on free riding in a highly charged atmosphere is to render coalitions and their respective factions devoid of any inducement toward reconciliation or even disarmament, even in the wake of military successes. On the contrary, the defeat of one foe may transform latent intraparty feuds into diabolical engagements. ²⁷

Some conclusions follow from this. In the first place, it is improper to assume that every alliance is the same in its capacity to promote new wars. The more protracted are the differences within any alliance, the more severe is the security dilemma its members face. In as much as competition between allies dissuades free riding, it should also increase and preserve the potentiality for these factions to renew conflict as the conditions of the war evolve. This leads to the counterintuitive result that enemies may engender highly effective coalitions in civil conflicts. But very often such arrangements, while they may precipitate the end of one war, can generate others.

Hypothesis 1: Competitive coalitions increase the likelihood of civil war recurrence .

To extend and clarify this discourse, it is useful to relate it to the generally accepted wisdom regarding military outcomes and civil-war recurrence. As previously cited, the Wagner hypothesis posits that military victories are less likely than negotiated settlements to facilitate new wars. But the dynamics described above suggest that competing allies may upset this otherwise staid result. When an alliance environment is hostile, commitment and defection problems may transcend each party’s capacity for consigning to peace. Defeated allies may blame one another for their losses in an effort to resume the charge of rebellion in the future. Or, in victory, triumph over a common enemy may exacerbate problems held in abeyance during original conflicts. These difficulties may now resurface with a vengeance.

Hypothesis 2: Competitive coalitions increase the likelihood of civil war recurrence after wars ending in military victories .

Hypotheses 1 and 2 reflect the overall instability characterizing civil conflicts marked by competing and antagonistic allies. And the latter is an extension of the Wagner hypothesis. We may further extend the discourse by considering how such an environment amplifies commitment problems common to internal wars.

Alliances necessitate commitments from all parties (e.g., not to attack another faction, to defend an allied group against elimination, or to transfer aid and other vital resources). And most importantly, a rebel alliance implies a commitment to acceptable power-sharing arrangements in the aftermath of a potential victory. But all factions must manage such commitments with an eye on future bargaining power in the postwar regime. Under such conditions, each party has a distinct incentive to renege or to resist postconflict demobilization efforts. And the more alliance arrangements impact future bargaining leverage, the worse the problem becomes. Therefore, alliance guarantees made prior to or during the course of a conflict become less credible once a common enemy is eliminated. So what may appear as violence begetting violence is in many ways a consequence of commitment quandaries intensified by an environment of hostility and mistrust.

Ex-ante rivalry between allies interacts with and intensifies these commitment problems. Clarity regarding how commitment problems originate permits an even more finely tuned prediction concerning war outcomes and competitive rebel alliances. A commitment problem is generally the product of a shift in the balance of power ( $\Delta p$ ) between two or more competing parties (Fearon 1995 ; Powell 2006 ). ²⁸ When such changes are sufficiently large (or expected to be so), parties cannot reliably pledge not to take advantage of shifts in their favor (or to prevent those that do not favor them). A military victory is precisely the type of occurrence that will engender large changes in the balance of power, changes obviously benefiting unbeaten sides. Under such dynamic circumstances, agreements that were formally mutually acceptable to each allied faction may no longer be sustainable after military successes against governments. That is to say, the $\Delta p$ vis-à-vis competing allies may be sizable after defeating the government, especially if one group inherits or earns a majority of the state power apparatus. The shift in the balance of power is generally reflective of a change in one ally’s likelihood of eliminating another at some point in the future. In the absence of a common enemy, larger groups within successful competitive alliances no longer have a cause to exercise restraint with less enfranchised allies. They may no longer be willing to honor the previctory status quo. More specifically, $\Delta p$ entices allied parties to renege and alter the equilibrium in a manner more reflective of the new balance of power in postconflict settings. Consequently, after rebel victories wary allies are less able to credibly convey assurances they will not take advantage of more vulnerable partners, or even vanquished governments. Table 1 characterizes the expected outcomes.

Table 1.

Predicted outcomes

	No rebel victory	Rebel victory
No competitive alliance	–	Less new war
Competitive alliance	More new war	More new war

Victorious governments do not face the same kind and magnitude of commitment issues highlighted here. Without doubt, a government victory will induce a change in the balance of power ( $\Delta p$ ) vis-à-vis the rebels. But, given that rebels are almost always the weaker party, the magnitude of such a change should be relatively less than after a rebel win. The following hypothesis reflects the discussion above.

Hypothesis 3: Competitive coalitions are more likely to increase the chances of civil war recurrence after wars ending in rebel victories than after conflicts ending in government victories .

Research Design and Data

Civil War Data

The aim of this study is to investigate and present an understanding of civil war recurrence that highlights the importance of alliances and internal competition. It does so via a series of survival models using newly collected information on alliances in civil wars. For the purposes of case identification, this study employs two well known datasets on civil war: the Doyle and Sambanis ( 2000 ) database on peacekeeping and civil wars, with appropriate updates from the 2006 version (D&S), and the Uppsala Conflict Data Program / Peace Research Institute Oslo (UCDP/PRIO) Armed Conflict Dataset (ACD). ²⁹ These two civil war datasets are used because their respective units of observation differ in important ways for the study of war recurrence, thereby offering two distinct settings within which to test the premises set forth above.

The D&S data treat each conflict observation in an episodic fashion defined in accordance with the following criteria: (a) the war caused more than one thousand battle deaths, (b) the war represented a challenge to the sovereignty of an internationally recognized state, (c) the war occurred within the recognized boundary of that state, (d) the war involved the state as one of the principal combatants, and (e) the rebels were able to mount an organized military opposition to the state and to inflict significant casualties on the state. All conflicts in the dataset end in a military victory, a political settlement/treaty, or a ceasefire/truce. Moreover, because the D&S dataset emphasizes peacekeeping processes, it includes more short-lived cease-fires and is sensitive to government turnover. By contrast, the ACD dataset is strictly based on levels of violence for each conflict dyad every year. More precisely, UCDP defines a conflict dyad as “a contested incompatibility that concerns government and/or territory where the use of armed force between two parties, of which at least one is the government of a state, results in at least 25 battle-related deaths.” Therefore, in the absence of a military outcome or some agreement between the parties, an episode ends in a given year if during the follow-on year the conflict is inactive.

These conceptual and definitional distinctions between the two datasets imply slightly altered interpretations of war recurrence . War may begin anew in the former dataset only after it has ended in one of three distinct ways: settlements, military victories, or truces. By contrast, the ACD dataset emphasizes phases of conflict according to levels of violence in each dyad. Owing to the strict battle death criterion in the ACD database, the resumption of a war observation may take place in a given year merely on account of the conflict having drifted into a period of low activity and then picking up again. To present a more complete test of the propositions expounded above, both datasets are included in the statistical estimations below.

The modified D&S dataset employed here contains 125 episodes of war from 1944 to 2001. However, wars that were ongoing in 2001 and ended in the next decade were updated accordingly. The data were also changed to reflect instances where the authors revised their original coding because they deemed it to be wrong based on further research. The ACD dataset was modified in the following manner. If the intensity of a conflict never eventually reached the war threshold of one thousand battle deaths, then the observation was dropped. This coding criterion is commensurate with the theory developed above, which was not motivated by minor or low-intensity conflicts that may exhibit differing dynamics, especially regarding alliances. Extra-systemic and interstate conflicts are also excluded. Consecutive yearly dyads (of each conflict) are converted into a single conflict episode, the length of which was determined by the start date of the first yearly episode and the end date of the last episode. These modifications result in 193 conflict episodes from 1946 to 2012 in the ACD database.

New Data and Coding the Independent Variable(s)

In examining the propositions outlined above, this study brings to bear new information about alliances during civil wars. More specifically, it identifies not only instances of military alliances during the internal conflicts identified in the data, but extant rivalries and divisions within them. Because the ACD dataset is dyadic according to each rebel faction, identifying particular alliances involving “side B” of the dyad was straightforward, usually from reading an account of each conflict episode in the UCDP Conflict Encyclopedia. ³⁰ The episodic D&S data, on the other hand, only indicate the number of factions active in each conflict. Therefore, an initial task involved identifying the multiple actors and parties involved in each dispute and then ascertaining potential coalitions. For this reason, every case in the data was cross-referenced with data from Cunningham ( 2006 ), which lists the multiple players involved in each war. In this manner, a detailed list of actors for each case was compiled as a useful point of departure for identifying alliances. ³¹

The following coding rules were employed to classify coalitions in both the D&S and the ACD datasets. First, an alliance was coded for a particular war episode if there was a formal declaration of joint support between two or more groups. Because many civil war alliances lack such overt and official sanctioning, further identification mechanisms were required. Second, an alliance was coded if there was documented evidence of resource sharing between groups. And third, in the absence of either of the first two criteria, an alliance was coded if there was evidence that two or more factions simultaneously engaged the same enemy. The ‘simultaneous engagement’ rule does not pertain to separate movements fighting the same government in differing regions (often over distinct issues). To qualify as an alliance under this rule, the factions involved had to engage in fighting at the same time and same place against the same third party (government). And finally, the coding effort is sensitive to the timing of coalition formation and conflict episodes. All alliances had to be in meaningful existence sometime within the precise timeframe of each conflict episode.

Because the theory above is predicated on rivalry and competition, a final task was to determine which alliances were distinctly conflictual in nature. Toward this end, at least one of the two following criteria had to be satisfied. First, if the allied groups fought one another either before or during their alliance, their respective coalition was categorized as competitive. This infighting is strictly limited to episodes occurring before the end date of the civil war. Imposing this requirement helps to avoid endogeneity problems, whereby poor war outcomes may presumably lead to internal tensions. Second, if factions were known adversaries for ideological or other contentious reasons, the alliances they formed were also categorized as competitive arrangements. It is important to point out that alliance rivalry or competition is never coded on the sole basis of ethnic differences between allies. If either the alliance or the competitive nature of that alliance is what may be considered a borderline case, it is labeled ambiguous and dropped in a second set of analyses. This coding rule creates two versions of the alliance data—version ‘a’ includes all cases, and version ‘b’ includes only unambiguous observations. By these coding rules, the study identifies 57 instances of alliances, 42 of which are marked by some degree of competition in the D&S data. When ambiguous cases are dropped, the total number of alliances falls to 41, of which 31 are characterized as competitive. ³² For the ACD data, 92 alliances are identified, 71 of which are competitive; when ambiguous cases are removed, the numbers fall to 76 and 57, respectively.

According to the coding schema above, the three categories of alliances are as follows: (1) unified forces not involving any alliances, (2) alliances marked by little or no internal rivalry, and (3) competitive alliances characterized by internal rivalries. These categories are operationalized in the following fashion. The variable competitive ally = 1 if a coalition was characterized by internal rivalry according to the criteria set out above and is otherwise set to 0. The competitive ally covariate is thus a dummy variable reflecting the most contentious coalitions. The next category, ally , is a proxy for alliances not marked by internal rivalry and competition. The variable ally = 1 for any alliance not identified as competitive and is 0 otherwise. The omitted category in this coding arrangement is no ally . ³³

The Dependent Variables

The analytics below make use of Cox proportional hazards models, a generalized form of survival analysis. Given this model specification, the event of interest, also known as the failure variable, is war recurrence . The time variable is the number of days between the end of a conflict episode and (when appropriate) the start of another. The time component specifically indicates the duration of peace. The war recurrence and, when relevant, time to recurrence information for the D&S dataset are adapted from Fortna ( 2004 ), which identifies the time between the termination of fighting and the start of another war in the Doyle and Sambanis ( 2000 ) dataset. All nonrecurring conflicts are right-censored as of 30 July 2008. ³⁴ And importantly, according to Fortna’s coding criteria, the failure of peace is never attributed to a subsequent conflict in the same country between substantially different actors. Of the 125 D&S conflict episodes, 51 recurred. ³⁵

As noted, all observations in the ACD dataset are based on battle death criteria each year. The start and end dates of each episode allow users to determine the duration of each observation and if the conflict recurred. However, because highly unrelated conflicts within the same country often share the same conflict identification number ( ID ), it is necessary to categorize two recurrence variables. ³⁶ The first identifies recurrence of an episode any time a conflict (based on the conflict ID ) ended and started again, irrespective of the parties identified in the subsequent conflict ( recur_any ). Based purely on conflict ID in the UCDP dataset, 96 of the 193 episodes recur. A second dependent variable restricts recurrence episodes to only those conflicts where there is a sufficiency of linkage between the two conflict episodes ( recur ). ³⁷ This essentially requires that at least some part of the antigovernment combatants from the first episode also participate in the follow-on war. Based on this criterion, 80 of the ACD episodes recur. ³⁸ The time between repeat conflicts is determined according to the start and end dates for each episode. All nonrecurring conflicts in the ACD data are right-censored as of 30 July 2012.

Control Variables

To the theoretical insight developed here, a reasonable objection may be raised that political environments that cause social and organizational fragmentation also lead to war recurrence. There is undoubtedly a kernel of truth to this claim. Environments rife with social, political, and economic grievances are also where one may expect to observe internecine coalition rivalries. For this reason, it is imperative to control for weak state structures, potential class cleavages, meddling international actors, and other sociopolitical factors. Therefore, control variables in the models presented here include levels of democracy ( Polity ), peacekeeping operations ( pko ), military participation by major powers, ³⁹ and life expectancy measures, along with an ethnic heterogeneity index. ⁴⁰ The models also control for the length of each conflict ( duration ). Unless otherwise noted, all control variables are measured at the beginning of each conflict episode. ⁴¹ A war type category further distinguishes between the war aims of the rebelling side(s). It separates the conflicts into those fought over autonomy or separatist aims from those targeted specifically at control of the state apparatus (1 = state war , 0 = autonomy/self-determination ). ⁴² Additionally, the number of veto players is included to account for the dynamics associated with multiparty conflicts. ⁴³ As many of the covariates listed here are colinear, they are not all folded into the same model simultaneously.

While rebel strength is an obvious influence on conflict outcomes and recurrence, it also is a prospective factor in determining if a rebelling faction will form an alliance, making it a critical control variable for this study. To account for this characteristic, a control for the fighting capacity of rebels, vis-à-vis the government, is employed. This rebel-specific indicator, taken from the nonstate actor dataset (Cunningham, Gleditsch, and Salehyan 2013), is an ordinal measure of military strength relative to the government. Because this variable is originally coded in a dyadic fashion (from the Uppsala-PRIO dataset), each rebel group was matched to specific conflict episodes in the D&S data. The variable ( rebel strength ) rates the relative strength of the rebels as strong (3), at parity (2), or weak (1), in terms of their ability to wage conflict. In the D&S data, if more than a single rebel group was identified within a specific conflict episode, the final coding reflects the highest rating any one of the groups received.

Additional covariates reflect conflict outcomes. Doyle and Sambanis ( 2000 ; 2006 ) characterize four distinct outcomes for any given war: government victory, rebel victory, settlement, and truce/ceasefire. Similar to Cunningham et al. ( 2009 ), this work combines settlement and truce outcome into a single category of agreement . This is the omitted category, as it allows a direct comparison between military and nonmilitary outcomes pertinent to the discussion above. The outcome variables in the ACD dataset are taken from the UCDP Conflict Termination dataset (Kreutz, 2010 ), based on the following categories: government victory , rebel victory , agreement , and low activity/other . ⁴⁴ Again, the agreement category includes both peace and ceasefire agreements and unless otherwise noted serves as the omitted category. The low activity outcome indicates that the episode did not attain the battle deaths threshold the following year. Table 2 summarizes the various conflict outcomes for the two datasets. ⁴⁵

Table 2.

Conflict outcomes

	D&S		ACD
Outcome	N	%	N	%
Government victory	41	32.8	23	11.9
Rebel victory	30	24.0	20	10.4
Agreement	48	38.4	50	25.9
Low activity / other	—	—	76	39.4
(Ongoing)	6	4.8	24	12.4
Total	125	100	193	100

Note : ACD, Armed Conflict Dataset; D&S, Doyle, and Sambanis ( 2000 ) dataset.

Results

Table 3 displays the results of Cox proportional hazards models from the D&S data. Table 4 shows similar models for the ACD data. The models estimate the effects of each covariate on the “hazard” of war recurrence in a given time period, provided that peace has lasted up to that period. Time units are measured in days. The tables report the hazard ratios and the robust standard errors, clustered by conflict episode, for each of the covariates. ⁴⁶ Hazard ratios are interpreted relative to 1. A hazard ratio of greater than 1 indicates that the event of war recurrence or, equivalently, peace failure is more likely, and by implication, the duration of peace should be shorter. A hazard ratio of less than 1 indicates peace failure is less likely and that peace spells should be longer. All ongoing conflicts are not included in the analyses.

Table 3.

Cox proportional hazards models: duration of peace (D&S)

	Model 1 recur (a)	Model 2 recur (a)	Model 3 recur (b)	Model 4recur (b)
Ally	1.135 (0.587)	1.083 (0.563)	2.106 (1.174)	3.798 ** (2.014)
Competitive ally	5.389 *** (1.725)	4.530 *** (1.773)	6.148 *** (1.980)	5.883 *** (2.400)
Veto		0.825 (0.120)		0.728 ** (0.100)
Victory gov.		0.310 ** (0.170)		0.482 (0.271)
Victory rebel		0.540 (0.227)		0.766 (0.329)
Rebel strength		0.934 (0.291)		1.025 (0.345)
War type		1.715 (0.778)		2.436 ** (1.035)
Major		1.169 (0.415)		1.162 (0.375)
Duration (ln)		0.883 (0.158)		1.004 (0.169)
Polity		0.934 *** (0.014)		0.930 *** (0.013)
Life expectancy		0.937 *** (0.017)		0.942 *** (0.017)
Ethnic heterogeneity		1.010 *** (0.004)		1.012 *** (0.005)
Failures	51	51	51	51
Log pseudo likelihood	-210.356	-187.748	-210.140	-186.346

Note : Hazard ratios are reported. Robust standard errors in parentheses. **p < 0.05, ***p < 0.01. N = 119.

Table 4.

Cox proportional hazards models: duration of peace (ACD)

	Model 5 recur_any (a)	Model 6 recur_any (a)	Model 7 recur_any (b)	Model 8 recur_any (b)	Model 9recur (a)	Model 10recur (b)
Ally	0.410 * (0.206)	0.388 * (0.223)	0.617 (0.302)	0.764 (0.447)	0.473 (0.301)	0.741 (0.472)
Competitive ally	3.035 *** (0.709)	2.821 *** (0.765)	3.830 *** (0.941)	4.165 *** (1.243)	3.749 *** (1.105)	5.044 *** (1.621)
Veto		1.087 (0.120)		1.024 (0.118)	1.069 (0.129)	1.021 (0.131)
Victory gov.		0.494 * (0.209)		0.530 (0.249)	0.273 * (0.204)	0.270 (0.215)
Victory rebel		1.003 (0.386)		1.016 (0.406)	0.734 (0.296)	0.711 (0.312)
Low/other		1.322 (0.349)		1.537 (0.456)	1.364 (0.418)	1.605 (0.567)
Rebel strength		0.779 (0.180)		0.849 (0.187)	0.682 (0.224)	0.703 (0.237)
War type		1.080 (0.246)		1.184 (0.266)	0.814 (0.207)	0.969 (0.235)
PKO		0.348 ** (0.157)		0.378 ** (0.174)	0.447 (0.222)	0.482 (0.241)
Duration (ln)		1.017 (0.051)		0.995 (0.054)	1.053 (0.060)	1.041 (0.067)
Polity		0.997 (0.004)		0.999 (0.004)	0.998 (0.004)	0.999 (0.004)
Ethno-linguistic fractionalization		2.789 (1.329)		1.828 (1.138)	1.787 (1.092)	1.321 (0.876)
Failures	96	96	96	96	80	80
Log pseudo likelihood	−427.643	−414.619	−444.454	−413.840	−345.173	−343.734

Note : Hazard ratios are reported. Robust standard errors in parentheses. *p<0.1, **p < 0.05, ***p < 0.01. N = 169.

The first pair of models in Table 3 focuses on the ‘a’ versions of the D&S data and the latter two models use the ‘b’ versions. The results of these models offer good evidence that divisive rebel alliances meaningfully alter the risk of peace failure after wars. In the absence of any control variables (models 1 and 3), competitive alliances increase the hazard of war recurrence. Adding a whole array of potentially confounding covariates (models 2 and 4) does not greatly diminish the positive and statistically significant effect rivalrous coalitions have on promoting repeat wars. Moreover, the substantive impact of competitive alliances on conflict renewal is quite strong. The risk of a new war increases fourfold in model 2 and more than five times in model 4 on account of competitive alliances. The noncompetitive alliance covariate ( ally ) has a similar influence only in model 4.

Table 4 includes the results of six estimated models using the ACD data. Models 5 through 8 determine peace to have failed whenever recur_any = 1, and employ both the ‘a’ and ‘b’ variants of the alliance variables. Models 9 and 10 define a failure of peace according to the recur variable, again using both versions of the alliance data. As is evident in Table 4 , the effect of the competitive alliance covariate on the hazard of war recurrence in the ACD dataset is very much in line with its impact in the D&S data, even though the definitional unit of observation differs between the two datasets. Competitive alliances increase the risk of peace failure significantly in all six models. In sum, Tables 3 and 4 provide confirmatory evidence in support of hypothesis 1. ⁴⁷

It is worth pointing out that recurrence is not a mere proxy for wars of many players. Although the veto variable and the alliance indicators are highly correlated, ⁴⁸ the number of veto players in an original war never significantly increases the hazard rate of peace failure in any of the 10 models displayed in Tables 3 and 4 . On the contrary, the number of veto players significantly lowers the risk of war renewal in the D&S models ( Table 3 ). In fact, the hazard ratio of 0.728 in model 4 suggests that a unit increase in the number of veto parties lowers the risk of renewed conflict by approximately 27 percent. In the ACD models ( Table 4 ), the veto variable has no discernable effect on conflict renewal. The hazard ratios associated with the veto variable are never statistically indistinguishable from 1.

In order to more fully flesh out the relationship between alliances, multiparty conflicts, and peace duration, additional models are presented. Using the D&S data, model 11 in Table 5 includes only the number of veto -players; models 12 and 13 add to this the respective ‘a’ and ‘b’ variants of the alliance variables. Table 6 repeats this analysis with the ACD data. In isolation, the veto -player variable has a positive estimated influence on the rate of war recurrence, and is statistically significant in model 11. This increase is evident in both the D&S and ACD datasets (models 11 and 14). Furthermore, the anticipated impact is nearly identical in both datasets—upwards of 20 percent. These results alone suggest the overall number of parties involved in conflicts should complicate efforts to preserve peace. The outcome is in line with Cunningham’s ( 2006 ) original finding that wars characterized by multiple actors tend to be more intractable than traditional two-party affairs. However, the peace-confounding effect of many actors is eliminated when the alliance measures are incorporated into the respective models, as evident in models 12 and 13, as well as 15 and 16. The reversal effect of the veto parameter is particularly pronounced in the three models in Table 6 . In sum, the conflict recurrence results do not simply reflect higher incidence of war renewal in multiparty wars.

Table 5.

Cox proportional hazards models: duration of peace (D&S)

	Model 11 recur	Model 12 recur (a)	Model 13 recur (b)
Ally		1.106 (0.611)	2.110 (1.225)
Competitive ally		5.260*** (1.767)	6.161*** (2.312)
Veto	1.206* (0.120)	1.024 (0.113)	0.998 (0.109)
Failures	51	51	51
Log pseudo likelihood	−224.716	−210.333	−210.139

Note : Hazard ratios are reported. Standard errors in parentheses. *p<0.1, ***p < 0.01. N = 119.

Table 6.

Cox proportional hazards models: duration of peace (ACD)

	Model 14 recur	Model 15 recur (a)	Model 16 recur (b)
Ally		0.501 (0.297)	0.692 (0.399)
Competitive ally		4.487 *** (1.350)	5.373 *** (1.605)
Veto	1.162 (0.115)	0.914 (0.099)	0.910 (0.103)
Failures	80	80	80
Log pseudo likelihood	−380.794	−359.718	−357.867

Note : Hazard ratios are reported. Standard errors in parentheses. ***p < 0.01. N = 169.

As indicated in previous sections, the primary aim of this work is not to adjudicate the Wagner hypothesis, but to refine it by embedding it within a more nuanced theory of coalitions. As such, interactive models are required—those able to account for combined effects. The following models probe the potentially interactive nature of competitive coalitions over various war outcomes. These tests are designed to more specifically address the premise that rivalrous alliances may upset the stability normally associated with military victory or the capitulation of losing sides. More precisely, the extended models test the proposition that a decrease in the likelihood of war recurrence is associated with military victories only in the absence of competitive alliance arrangements. Only the ‘b’ versions of the data are employed, meaning they include no instances of ambiguous coding. Table 7 presents three additional Cox models employing the D&S data. Model 17 includes the interaction between competitive ally and military outcome ; model 18 presents the interaction between competitive ally and government victory ; and model 19 incorporates an interaction between competitive ally and rebel victory . Table 8 presents similar analyses using the ACD dataset in conjunction with the recur variable.

Table 7.

Cox proportional hazards models with interaction: duration of peace (D&S)

	Model 17	Model 18	Model 19
Ally	2.581 (1.529)	3.836 ** (2.062)	2.795 * (1.575)
Competitive ally	2.312 (1.212)	5.999 *** (2.533)	2.594 * (1.426)
Veto	0.757 ** (0.102)	0.729 ** (0.101)	0.785 * (0.110)
Victory gov.		0.495 (0.296)	0.388 * (0.212)
Victory rebel		0.767 (0.330)	0.194 * (0.168)
Military victory	0.286 ** (0.156)
Military victory * Competitive ally	5.507 *** (3.477)
Gov. victory * Competitive ally		0.838 (1.123)
Rebel victory * Competitive ally			8.691 ** (8.340)
Rebel strength	1.058 (0.370)	1.025 (0.345)	1.075 (0.365)
War type	2.748 ** (1.235)	2.404 ** (1.025)	2.386 * (1.066)
Major	1.096 (0.373)	1.157 (0.377)	1.044 (0.358)
Duration (ln)	1.068 (0.172)	1.003 (0.169)	1.078 (0.173)
Polity	0.922 *** (0.014)	0.930 *** (0.013)	0.924 *** (0.013)
Life expectancy	0.935 *** (0.017)	0.942 *** (0.018)	0.942 *** (0.016)
Ethnic heterogeneity	1.012 *** (0.005)	1.012 *** (0.005)	1.012 ** (0.005)
Failures	51	51	51
Log pseudo likelihood	−183.579	−186.334	−183.053

Note: Hazard ratios are reported. Robust standard errors in parentheses. *p<0.1, **p < 0.05, ***p < 0.01. N = 119.

Table 8.

Cox proportional hazards models with interaction: duration of peace (ACD)

	Model 20	Model 21	Model 22	Model 23
Ally	0.733 (0.461)	0.743 (0.475)	0.712 (0.449)	0.712 (0.449)
Competitive ally	4.757 *** (1.562)	5.023 *** (1.632)	4.752 *** (1.542)	4.752 *** (1.542)
Veto	0.979 (0.121)	1.018 (0.132)	1.007 (0.121)	1.007 (0.121)
Victory gov.		0.254 * (0.203)	0.255 * (0.205)	0.158 *** 0.113
Victory rebel		0.713 (0.313)	0.197 (0.225)	0.122 ** (0.127)
Military victory	0.217 ** (0.158)
Low/other	1.651 (0.583)	1.606 (0.567)	1.617 (0.566)
Agreement				0.618 (0.216)
Military victory * Competitive ally	4.101 * (3.091)
Gov. victory * Competitive ally		1.232 (1.154)
Rebel victory * Competitive ally			5.350 (6.168)	5.350 (6.168)
Rebel strength	0.885 (0.298)	0.705 (0.237)	0.853 (0.294)	0.853 (0.294)
War type	0.960 (0.231)	0.967 (0.237)	0.980 (0.233)	0.980 (0.233)
PKO	0.508 (0.256)	0.485 (0.244)	0.483 (0.241)	0.483 (0.241)
Duration (ln)	1.005 (0.067)	1.039 (0.067)	1.007 (0.067)	1.007 (0.067)
Polity	1.001 (0.004)	1.000 (0.004)	1.000 (0.004)	1.000 (0.004)
Ethno-linguistic fractionalization	1.308 (0.821)	1.335 (0.901)	1.147 (0.728)	1.147 (0.728)
Failures	80	80	80	80
Log pseudo likelihood	−343.126	−343.721	−342.585	−342.585

Note : Hazard ratios are reported. Robust standard errors in parentheses. *p<0.1, **p < 0.05, ***p < 0.01. N = 169.

As is evident in models 17 and 20, conflicts terminating in military victories exhibit a substantially lower incidence of war recurrence. In both models, the hazard ratios for this single covariate are less than 1 and statistically significant. This result, paralleling those in DeRouen and Bercovitch ( 2008 ), presents strong confirmatory evidence in favor of the Wagner hypothesis. Decisive military victories substantially reduce the risk of conflict resurgence. However, this outcome is irrespective of alliances that may or may not be active at the conclusion of conflicts. Fortunately, the binary nature of both the variables comprising the interactive term facilitates interpretation of the conditional results. The hazard ratio associated with the interaction term ( military victory x competitive ally ) is greater than 1 and significant in each of the two models: just over 5.5 in the D&S data (model 17) and more than 4.1 in the ACD data (model 20). The interactive outcome offers support for hypothesis 2. This is an interesting and important modification to the Wagner hypothesis. Absent a competitive alliance, military victories generally favor the chances of peace. However, the arc of rivalry among allies hastens a return to conflict after military victories.

The models in the remaining columns of Tables 7 and 8 unpack the military outcome even further by disaggregating the event into government and rebel victories. In these models, the competitive alliance variable interacts with the government victory and rebel victory covariates, respectively. Focusing first on government victory , we notice that the direction of influence for this covariate is the same across the two datasets, but statistically significant only in the ACD data (model 21). The hazard ratios for this outcome are considerably less than 1, indicating that war relapse is less likely in the wake of government successes. This outcome replicates the findings from Kreutz ( 2010 ). However, the interaction term ( government victory x competitive ally ) is not statistically significant in either the D&S model or the ACD model (models 18 and 21). An obvious disruption to the postconflict setting associated with government victory is not evident in these interaction results. Moreover, the estimated direction of the interactive effect ( government victory x competitive ally ) is positive in the ACD dataset (model 21) and negative in the D&S data (model 18). The destabilizing dynamics shown to be associated with competitive alliances in postconflict settings do not meaningfully interact with government victories to precipitate follow-on conflicts.

The same cannot be said for wars terminating in rebel successes. Models 19 and 22 in Tables 7 and 8 examine the interaction between competitive alliance and rebel victory . In both the D&S and ACD datasets, the independent impact of rebel success is to diminish the danger of reigniting domestic conflagrations. The hazard ratios associated with rebel victory are less than 1 and statistically significant in model 19. This general result is in line with Toft ( 2010 ) and others demonstrating the relative stability of victory achieved at rebel hands. But this outcome demonstrates the effect of rebel victory in the absence of competitive alliances. The interaction terms suggest that an opposite pattern emerges when rebel victories are also characterized by contested coalitions. The respective hazard ratios of the interaction variables in models 19 and 22 are 8.69 and 5.35. Both ratios are appreciably greater than 1 and the former is statistically significant. ⁴⁹ Finally, model 23 replicates model 22, but changes the omitted category of the outcome variable to low activity . The substantive results of model 23 match those of 22. Indeed, the peace-inducing impact of rebel victories is estimated to be even stronger, while the interactive results again demonstrate a reversal of this trend.

The inference of this analysis is that competing allies reverse the standard effect of rebel victory on conflict recurrence. This outcome represents a departure from Quinn et al. ( 2007 ), Toft ( 2010 ), and other studies reflecting the general stability associated with rebel victories. Allowing for the conditional nature of the conflict outcome is the primary reason for the difference in findings. ⁵⁰ It is even more remarkable that such a consistent pattern of reversal emerges from both datasets, especially in light of the relative paucity of military victories in the UCDP Conflict Termination database. ⁵¹ But it is also for this reason that this analysis should be interpreted with caution. Some of the key interactive results from the models using the ACD data remain at the margins of conventional levels of significance. But rather than fixate on p -value and data limitations, it is perhaps more useful to focus on the similar tendencies in the hazard ratios associated with the interactive covariates and their independent counterparts in both datasets. As the section below demonstrates, these parallel trends translate into highly similar predicted probabilities in the survival of peace across the two data sources.

A final comparison of the predicted percentages of peace survival under designated scenarios offers an indication of what would otherwise be called marginal effects. ⁵²Figure 1 presents a visual interpretation of the relative changes in war recurrence across the independent and interactive variables. It portrays the survival functions (of peace) inclusive of three cases: rebel victory , the absence of rebel victory , and rebel victory x competitive ally . ⁵³ The x-axis is time (measured in days). The left side of the figure uses the D&S data, and the right side uses the ACD dataset. All unlisted variables are set to their mean or median values. The two images on both sides of the figure show a striking semblance. While the survival percentages differ, the relative patterns between the evaluated cases look identical. Examined in isolation, rebel victories appear highly stable. The predicted survival rate remains well above 80 percent for conflicts ending in rebel victories (as indicated by the dotted line) in both datasets. The dashed line signifies that the absence of a rebel victory is generally less sustaining of peace. However, a great shift is evident in the survivor function for the interaction term. The solid line reveals a substantial reversal in the stability of wars ending in rebel wins. The survival function drops precipitously when conflicts are marked by competitive alliances and won by rebels.

Figure 1.

Cox proportional hazards models: predicted survival functions

Discussion

This examination of alliances and civil war recurrence informs our understanding of the causes of conflict transformation and renewal in a number of ways. First, it refines our traditional understanding of the relationship between military victories and renewed war. It strongly suggests that not all military successes are equally likely to yield sustainable peace. These results prove robust across two datasets on civil conflicts and differing conceptualizations of conflict recurrence . They help move the scholarly discourse beyond Atlas and Licklider’s ( 1999 , 36) keen but under-examined observation that one obstacle to sustaining peace after civil wars is “often a breakdown in relations among former allies, not former foes.” This process plays a critical role in producing renewed violence, even after those conflicts that end in clear military victories.

This study also improves our knowledge of the consequences of rebel victories in civil wars. It highlights the highly conditional aspect of rebel successes. Examined in isolation, rebel victories appear to produce relatively stable outcomes. But the general stability associated with rebel successes in civil wars wavers under the perverse influence of contentious allies. Competitive alliances make renewed conflict more likely by rendering parties less capable of overcoming commitment problems common to civil wars—especially after rebel victories. Thus, postconflict stability after civil wars won by rebels is to a high degree conditional on the absence of competitive or tenuous coalitions, as well as possibly on other confounding factors.

These findings highlight a critical issue in the study of civil wars. Scholars should not focus exclusively on types of outcomes. Contextual conditions surrounding the termination of wars shape postconflict possibilities. Future studies should aim to uncover how the different ways that civil wars end render them more or less susceptible to the specific process through which peace ‘breaks down’ that I focus on here. In broader terms, the scholarly debate should therefore move beyond the peculiarity of attributing postconflict stability solely to various types of war resolutions.

My study also adds to a growing list of works on the topic of alliances and alliance behavior in civil conflict. This article joins that line of scholarship in emphasizing the importance of strategy and politics. Alliances in civil conflicts sometimes transcend ethnic, racial, religious, linguistic, and other identity-based explanations for coalition behavior. Thus, existing work suggests that parties form alliances for strategic reasons. Christia ( 2012 , 33) notes that warring groups confronted with alliance choices during civil wars face two competing considerations: winning the war and maximizing postwar political control. Bapat and Bond ( 2012 , 14) note that allies, while gaining from mutual assistance, face strong incentives to prevent their partners from becoming too powerful. This tension implies serious commitment problems. And Akcinaroglu ( 2012 , 887) argues that variation in the degree of accommodation among allies shapes their willingness to sacrifice long-term benefits to realize short-term gains at one another’s expense. My study provides further evidence for how civil war coalitions create opportunities and dangers for their members.

My work also raises unsettling policy implications. On one hand, “giving war a chance” presents a less straightforward prospect than advocates maintain. Waiting for peace to take hold could entail countenancing decades of conflict phases of various intensities. On the other hand, recommendations regarding third-party involvement in civil conflicts must take into consideration existing intergroup antagonisms. When rivals seek out mutual accommodation, the outcome is usually short-lived at best and fratricidal at worst. Success in war, moreover, does not tend to ameliorate differences between hostile rebels. On the contrary, it may reopen the door to violence. Can third parties and peacekeeping forces assist competing factions to mitigate the difficulties outlined here? In theory, the answer is yes . Research cited here suggests that peacekeeping can serve as an effective instrument for precluding new wars. The more intervening forces appreciate divisions among parties, the more they should be able to take steps to address them. In practice, however, success may prove difficult, if not elusive. Prior studies demonstrate that the belligerents themselves often retain the strongest voices for or against the presence of peacekeepers. Triumphant parties, moreover, possess little incentive to concede to international overtures for peacekeeping forces (Fortna 2008 ). This means that postconflict situations most in need of peacekeepers—ones in which rivalrous coalitions upend governing regimes—are least likely to see them.

The literature on protest and rebellion emphasizes how grievances against governments drive conflict. However, my argument highlights how grievances and the strategic interplay that rebel groups have with one another provide a source of volatility and instability. Indeed, expediting the overthrow of regimes may lead to even wider conflicts—especially when participating coalitions are highly factionalized. As such conditions obviously operate in a variety of contemporary conflicts, we ignore them at our peril.

Author’s note: Thanks are owed to Karen Remmer, Jan Pierskalla, Arturas Rozenas, Emerson Niou, William Keech, Stephen Gent, Andrew Radin, Navin Bapat, Kathleen Cunningham, “grant-writer-z” as well as workshop participants at the University of Maryland and the University of North Carolina at Chapel Hill. I am also indebted to the three anonymous reviewers and the editor at ISQ for their valuable critiques of this manuscript. All remaining errors and shortcomings are mine alone.

¹ As quoted in Miranda and Ratliff ( 1993 , 4).

² http://www.africa.upenn.edu/Hornet/irin_21799.html (accessed June 25, 2015).

³ See Etcheson ( 1984 , 128).

⁴ Reported in Quinn et al. ( 2007 ). A similar pattern is evident in the Fearon and Laitin ( 2003 ) civil war dataset.

⁵ Reported in Cunningham ( 2006 ).

⁶ This list includes conflict duration (Cunningham 2006 ), war outcomes (Nilsson 2008 ; Cunningham et al. 2009 ) rebel fragmentation (Cunningham et al. 2012 ), inter-rebel violence (Fjelde and Nilsson 2012 ), and spoiler effects (Kydd and Walter 2002 ). Additionally—and important to this project—a small subset of this work seeks to unpack both the conditions for alliance formation among rebel groups (Bapat and Bond 2012 ) and how alliances affect conflict transformation after negotiated settlements (Atlas and Licklider 1999 ). Most recently, Akcinaroglu ( 2012 ) examines how alliances and interdependencies among rebel groups impact the likelihood of rebel success against governments. Christia ( 2012 ) probes how power distributions affect factors including alliance formation and alliance switching, as well as fragmentation.

⁷ A related study by Rudloff and Findley ( 2011 ) examines how rebel group fragmentation impacts recurrence.

⁸ Noted exceptions to this are Walter ( 2004 ), which focuses on micro-level motivations present at the end of wars, and more recently Walter ( 2014 ), which examines governance structures in postconflict settings.

⁹ More specifically, the authors note that the odds of peace failure following a rebel victory are relatively high in the immediate aftermath of a conflict, but this effect diminishes substantially after approximately three years.

¹⁰ Walter ( 1997 , 335–36).

¹¹ Work on external interventions has also spawned a robust literature on the nature and institutional design of peace settlements. See Hartzell ( 1999 ), Hartzell et al. ( 2001 ), Hartzell and Hoddie ( 2003 ) and Mattes and Savun ( 2009 ).

¹² Doyle and Sambanis ( 2006 , 5).

¹³ For work specifically addressing UN interventions, see Howard ( 2008 ).

¹⁴ See Fortna ( 2008 ) Chapter Two and Appendix B, as well as Fortna ( 2004 ).

¹⁵ Indeed, such a dynamic played out in the wake of Qaddafi’s fall from power in Libya in 2011, in which leaders on Libya’s National Transition Council objected to postconflict peacekeepers. The North Atlantic Treaty Organization (NATO) did not press the matter in the face of such aversion. See Chivvis and Martini ( 2014 , 5).

¹⁶ On this consensus, see Fortna and Howard ( 2008 ).

¹⁷ Autesserre ( 2014 , 22) discusses this in detail, noting ranges of 31%–85%.

¹⁸ A recent exception, however, is Christia ( 2012 ).

¹⁹ Emphasis added.

²⁰ The same may be said of alliances forged in the anarchic conditions of the international system. Alliance ambiguity may elicit real or perceived fears of abandonment on the part of one or more sides (Cha 1999 ).

²¹ A rich line of research within the subfield of security studies probes the issue of managing conflict within alliances. For more on this, see especially Schroeder ( 1976 ), Weitsman ( 1997 ), Krebs ( 1999 ), and Pressman ( 2008 ). Cha ( 1999 ) also treats the quasi alliance of antagonistic states beleaguered by historical animosities.

²² This summary is from Miranda and Ratliff ( 1993 , 12–13).

²³ In a curious twist of events, Iran openly supported the KDP in Iraq while engaged against the Kurdish Democratic Party of Iran (KDPI) in Iran (McDowall 2000 ).

²⁴ Shay ( 2008 ) discusses these Somali alliances and their respective factions in detail.

²⁵ See Balcells ( 2010 ) for a theory of rivalry and revenge dynamics in civil war.

²⁶ The classic treatment of the security dilemma is Herz ( 1950 ).

²⁷ Again, while not specifically under examination herein, to varying degrees the same may be true of wartime alliances between states.

²⁸ In the words of Powell ( 2006 , 183): “Shifts in the distribution of power are at the heart of Fearon’s [1995] three kinds of commitment problems.”

²⁹ Version 4-2012 from Gleditsch et al. ( 2002 ). Where appropriate, ongoing conflicts and recurrence episodes were updated with the subsequent version.

³⁰ Other sources were consulted as well. This includes Human Rights Watch reports, Minorities at Risk, and Onwar.com. All the identified alliances are listed in the dataset according to each episode.

³¹ This list was also double checked against other sources on civil war: Fearon and Laitin “Random Narratives,” and Toft ( 2010 ).

³² The few instances where more than a single alliance is observed for a particular conflict are not broken into multiple observations. Accounting for multiple alliances may be a logical extension to the analysis. Importantly, there are no instances of both competitive and non-competitive coalitions identified in any single conflict episode.

³³ Any alliance not classified as “competitive” is automatically placed into the ally category. The variable no ally = 1 if no alliance is coded for a particular episode of conflict and it is set to 0 otherwise.

³⁴ More specifically, Fortna ( 2004 ) identifies dates of ceasefire and renewed warfare. The ceasefire dates generally correspond to the D&S dates for the end of each conflict episode, but may not necessarily indicate the end of the war. For the few instances in which the ceasefire predated the end of war dates indicated by D&S, the conflict episode was coded as ending according to the D&S date. This preserves the episode of conflict (and duration) identified by Doyle and Sambanis ( 2000 , 2006 ).

³⁵ The longest time to recurrence is nearly 27 years for Rwanda’s conflict ending in 1964. The shortest is for the case of Afghanistan, in which the 1996 Taliban victory over Rabbani resulted in a new government in Kabul and a new conflict against the United Front almost immediately.

³⁶ For example, the civil war in Cameroon between the government and the Union of the Peoples of Cameroon (UPC) ended in 1961. The last remnants of the UPC were defeated in the 1970s, and its various leaders were the targets of successful assassination attempts. But according to the ACD, this conflict episode ID (158) recurs 23 years later when the forces of Ibrahim Sale attempted to overthrow the Cameroonian regime in 1984.

³⁷ On such linkages, see Sambanis and Schulhofer-Wohl ( 2009 ), which includes a very cogent discussion of war recurrence in the context of partition as a solution to civil conflicts.

³⁸ The longest gap between conflict episodes is just over 32 years for the Baluchistan conflict in Pakistan ending in 1977 and renewing in 2009. The shortest is 417 days for the Ugandan war ending in 1992.

³⁹ The variable major is not an indication of peacekeeping forces. It measures direct major power military participation or extensive political support for one or more of the parties to the conflict. Major powers are the five permanent members of the United Nations (UN) Security Council. If a major power only participates in a UN peace operation, this is not coded as a major power involvement.

⁴⁰ For the D&S data, the ethnic heterogeneity index ranges from 0 (minimum heterogeneity) to 144 (maximum heterogeneity). In the ACD dataset, the “ethnic, linguistic, and religious fractionalization” measure ( fractionalization ) is taken from Alesina et al. ( 2003 ). Missing data precluded incorporating a life expectancy measure into models using the ACD data.

⁴¹ Robustness tests for the D&S models also incorporated the following controls: GDP per capita, external intervention by third parties, UN involvement, measures of illiteracy , and a Gini index of inequality . A robustness check in the D&S models further controlled for war intensity by including a measure of total deaths and displacement. It is highly correlated with conflict duration . See the supplementary appendix.

⁴² Certainly, some wars exhibit characteristics of both. These are crude distinctions, as has been pointed out by Licklider ( 1995 ).

⁴³ This variable corresponds to Cunningham’s ( 2006 ) lenient veto players covariate, which includes the total number of parties involved in a conflict.

⁴⁴ Because the Kreutz ( 2010 ) dataset only covers war terminations through 2009, wars ending between 2009 and 30 July 2012 were coded by consulting current news sources, usually The Financial Times or British Broadcasting Corporation (BBC).

⁴⁵ The inclusion of smaller-scale conflicts would considerably increase the number of Government Victories from the Conflict Termination database.

⁴⁶ Clustering the errors by country yields comparable results.

⁴⁷ An additional check on each model in Tables 3 and 4 divided all conflict episodes into two groups—those marked by competitive coalitions and those without—and predicted the survival functions for each group. A test for the equality of the respective survival functions rejected the null ( p < 0.01) in every instance.

⁴⁸ The correlation coefficients between the veto measure and the alliance covariates are as high as 0.46. This is not at all surprising, given that the veto list was a key component in identifying the various alliances.

⁴⁹ The standard errors are appreciably large on account of the low percentage of conflicts ending in rebel wins. This paucity translates to a relatively low incidence of war terminations reflective of competitive alliances and rebel victories (21 in the D&S dataset and 6 in the ACD data).

⁵⁰ In Toft ( 2010 ), a given outcome is compared against an omitted category that clumps together all the other outcomes . Thus the various outcome variables are never included together in the same model. Effectively this means that the baseline category changes across each model. This modeling technique does not permit a comparison between the effects various war outcomes (e.g., rebel victory and government victory ) have on war recurrence.

⁵¹ There may be concern that wars ending in low activity might reflect lulls rather than meaningful terminations to conflicts—thereby overdetermining the actual number of war recurrences. To address this matter, a further robustness check restricted the war recurrence variable for this outcome to include only those conflicts with periods of little to no violence lasting at least two years. Replicating the analyses in Table 8 with this revised data generated similar results.

⁵² Estimating “traditional” marginal effects is not possible in a Cox model on account of the proportionality assumption. For this reason, the changes in the predicted survivor functions over various cases are used.

⁵³ To avoid cluttering the graphic, the evaluation and plotting of survivor functions for the two values of competitive ally were omitted from the figure. One may easily examine the interaction whereby rebel victory conditions the baseline effect of competitive ally . Doing so indicates that rebel victory further exacerbates the positive and strong effect competitive ally has on peace failure.