Working PaperPDF Available

Inter-Rebel Alliances in the Shadow of Foreign Sponsors



Content may be subject to copyright.
Inter-Rebel Alliances in the Shadow of Foreign
Milos Popovic
Columbia University
twitter: @milos_agathon
From the Patriotic Front struggle against the minority rule in Rhodesia to the
seven-party mujaheddin alliance in Afghanistan, inter-rebel alliances make the armed
opposition more resilient and successful in the face of government repression. Why
then do some rebel groups cooperate with each other while others do not? Drawing
on principal-agent theory, I argue that the presence of foreign sponsors is likely to
encourage alliance formation in civil wars especially when two rebel outfits share a
state sponsor than if only one of them receives foreign support. Shared sponsors may
demand cooperation between their agents and credibly threaten to punish them for
non-compliance. They may also insist on the establishment of umbrella institutions
to improve their monitoring and sanctioning capacity, and to increase the legitimacy
of their agents. I test this argument using the UCDP Actor dataset with new data on
alliances between rebel groups. I find strong evidence that shared sponsors increase the
probability of inter-rebel alliance.
Word count: 10963
I thank Milada Vachudova, Erin Jenne, Juraj Medzihorsky, Vujo Ilic, Editor Michael Colaresi as well as
three anonymous reviewers for their comments. My gratitude goes to Juraj Medzihorsky for running the
analysis on powerful workstation and Levente Littvay for replicating the material on the CEU server. All the
remaining errors are my own.
Numerous Western appeals to Syrian rebels to rally against the government and ISIS rest
on an idea that coordination, shared resources, and joint efforts should increase the odds of
rebel victory. Indeed, the ZANU-ZAPU (Patriotic Front) alliance toppled down the white
minority government in Rhodesia, the TPLF’s web of alliances with other Ethiopian groups
brought down the Mengistu regime, and the seven-party mujahideen alliance ushered in
the Soviet withdrawal from Afghanistan. Confronted with a unified opponent, governments
must invest more in military operations than if the rebels were divided. Yet, among 1,000
rebel groups fighting in today’s Syria some form alliances on their own (e.g. former Jabhat
al-Nusra and Free Syrian Army in 2014/2015 against Assad and Hezbollah) or in the
shadow of foreign sponsors (e.g. US-sponsored Syrian Democratic Forces or Iran-backed
Shiite groups) while others like Al-Qaeda or Jaysh al-Islam foster few or no alliances.
If inter-rebel alliances empower rebel groups and increase their odds of victory, why do
some make them while others do not? Akcinaroglu (2012) finds that less than a half of
all civil conflicts since World War II featured cooperation among government opponents.
What factors encourage alliances between rebel groups? Why does foreign support in some
conflicts foster rebel cooperation, while in other conflicts leads to belligerent rebels?
Analyzing these questions is crucial for explaining how external governments can affect
civil war dynamics, an issue that has only recently incited more scholarly interest in conflict
studies. Previous research on civil war dynamics mostly focuses on how interactions among
multiple armed actors affect conflict duration (Cunningham 2011), violence and civilian
casualties (Asal and Rethemeyer 2008;Metelits 2009;Bakke et al. 2012;Horowitz and
Potter 2013) and various war outcomes (Cunningham 2011;Nilsson 2008;Cunningham
et al. 2009;Akcinaroglu 2012;Phillips 2014). An emerging stream of conflict literature
shifts the focus to interactions themselves, showing that rebel groups may cooperate or fight
each other (Furtado 2007;Bond 2010;Bapat and Bond 2012;Fjelde and Nilsson 2012;
Christia 2012;Nygård and Weintraub 2014). This research has focused extensively on
internal factors, including rebel objectives (Furtado 2007), balance of power and winning
coalitions (Christia 2012), shared identity (Bond 2010), and rebel capabilities (Bapat and
Bond 2012). Although we know more about the impact of balance of power on rebel
interactions, little empirical research exists on how different configurations of foreign
sponsors influence rebel propensity for cooperation and conflict. This is a serious omission
given that previous studies show that foreign sponsors can decisively affect the organization,
effectiveness and survival of rebel groups (Sinno 2008;Salehyan 2010;Fjelde and Nilsson
2012;Salehyan et al. 2014).
By introducing foreign sponsorship into the study of inter-rebel dynamics, I expand
current theoretical insights with the focus on how different configurations of sponsors
can influence cooperation among rebel groups. This approach contributes to an emerging
work on state sponsorship of rebel groups (Salehyan 2009;Salehyan et al. 2011,2014;
Popovic 2015a,b;Szekely 2016) as well as to the rich scholarship on interstate rivalry (e.g.
Akcinaroglu and Radziszewski 2005;Colaresi 2005;Maoz and San-Akca 2012) by linking
incentives of external states to the actual use of rebel groups to shape and shove civil
war dynamics. In doing so, I depart from more prominent studies on inter-rebel alliances,
including Bapat and Bond (2012) who argue that rebel capabilities take precedence over
international factors in explaining alliance-making. In contrast, I break new ground in this
emerging literature by arguing that sponsors make the first move by providing support for
the alliance between two rebel groups, irrespective of their individual strength.
I use principal-agent framework to show that when one or more sponsors act as common
"principals" of both partners ("agents") in any given dyad, the alliance will be buttressed
through the use of selection, monitoring, and sanctioning mechanisms. Shared principals
enjoy the monitoring and sanctioning capability greater than that of single principals
because they can credibly threaten to deny resources to both agents for underperformance
or transgression. If one agent underperforms relative to the other, it gets punished by the
principal. An agent may disapprove of this system, but it may lack alternatives particularly
when the other partner is rewarded for good performance. Shared sponsors are likely to
demand cooperation between their agents, and devote considerable resources to this goal
because alliance may allow them to buttress their control over the rebellion, and more
efficiently navigate a dyad rather than multiple groups. Unless sponsors are poised to stir
chaos, an alliance, therefore, decreases transaction costs associated with the monitoring,
control and sanctioning of separate groups.
I borrow the measure for rebel alliance from Bapat and Bond to show that type of
foreign sponsors matters more than other predictors,
but I arrive at different results due to
our differences in time-frame for analysis (starting from 1975 instead of 1946, and ending
in 2009 rather than 2001), proxies for foreign support and statistical model.
The findings in this article offer novel and promising avenues for conflict resolution,
especially in relation to the Syrian war where multiple rebel groups are fighting the
government. Foreign sponsors can play a decisive role in bringing rebel groups together
or deepening divisions between them. Given the previous findings that rebel alliances
may prolong civil wars or lead to rebel victory this article suggests that dealing with rebel
alliances in multiparty civil wars such as the Syrian conflict requires third parties to devote
attention to external governments who often foster inter-rebel cooperation. Third parties
and mediators sometimes assume that rebel groups operate either independently from each
other or from foreign governments. This article indicates that understanding civil conflicts
requires understanding the international dimension of intrastate wars. Understanding the
connection between state sponsors and militants provides important directions for conflict
resolution; the failure to account for these ties may lead policy makers to unintentionally
prolong the civilian suffering.
This article is divided into four parts. First, I use principal-agent framework to shape my
theory of rebel alliance and offer hypotheses on the relationship between state sponsors
and inter-rebel cooperation. Second, I discuss my data and methods. In particular, I focus
on a rebel dyad as a unit of analysis rather than on a group or conflict as a whole. After this,
I present my statistical results, including those from 10-fold cross-validation, indicating that
the model with shared sponsors has the strongest explanatory and forecasting power. I
conclude by exploring the implications of this study for future research and policy.
Rebel Alliances in Civil Conflicts
Civil wars are increasingly characterized by a web of inter-dependencies between multiple
armed actors. Multiparty conflicts last longer (Cunningham 2011), increase the number of
civilian casualties (Metelits 2009;Bakke et al. 2012) and produce various war outcomes
(Nilsson 2008;Cunningham et al. 2009). While armed confrontation between the host
government and rebels is often the most visible aspect of civil wars, rebel groups also form
violent or cooperative relationships with each other. Rebel groups clash over access to
recruits, resources and territory (Fjelde and Nilsson 2012;Nygård and Weintraub 2014), but
they sometimes cooperate. Rebel alliances
improve the chances of opposition survival and
victory (Akcinaroglu 2012;Phillips 2014), increase the lethality of rebel attacks (Asal and
Rethemeyer 2008;Horowitz and Potter 2013), and hamper conflict resolution (Cunningham
Despite such important effects, the phenomenon of inter-rebel alliances remains an un-
derstudied topic. There is a handful of studies exploring the existence of inter-rebel alliance.
For example, Furtado (2007) develops a typology of rebel groups based on their goals and
available resources to argue that alliance formation depends on the ability of groups to
credibly commit to cooperation, and on the magnitude of counterinsurgency. Similarly,
Bond (2010) argues that shared identity facilitates cooperation between armed groups,
while power considerations drive the outlook of alliances. In another study, Christia (2012)
draws on realist theory in IR to suggest that rebel groups are attempting to simultaneously
be on the winning side of a war while also gaining the greatest possible benefits in doing so
(the "minimum winning coalition"). This means that a driving factor in alliance formation
is the relative strength of the different alliance partners vis-a-vis one another and vis-a-vis
other alliances. Due to frequent shifts in relative strength, Christia concludes that alliances
may be preserved if there is an external party capable of enforcing cooperation. Following
this conclusion, Bapat and Bond (2012) argue that alliance formation is more likely in
conflicts where the militants are weaker than the government. To overcome distrust and
form an alliance, weaker groups need a foreign sponsor that can use material supplies to
enforce cooperation and deter defection. (Ibid, 11). Bapat and Bond find that foreign
support has an impact on alliance only in interaction with relative rebel strength.
While empirical results on rebel alliance shed light on how balance of power is associated
with inter-rebel cooperation, present studies are limited in examining the influence of
external actors. External actors serve as a supporting factor to relative capabilities in
understanding alliance-formation. The assumption is that balance of power considerations
dictate the search for foreign support even though foreign sponsors may approach rebel
groups beforehand.
Looking at foreign sponsors offers a new way to address rebel alliances. This approach
acknowledges that sponsors play a decisive role in rebel behavior, organization and survival
(Sinno 2008;Salehyan et al. 2014). As Salehyan et al. (2014) show, foreign support from
multiple government encourages rebel groups to be more violent toward civilians because
no single state can effectively restrain the organization. Foreign support is critical for
both weak and strong rebels because it increases their military power and cohesiveness
(Staniland 2014). The more sources of support, the more likely rebels are to survive
government repression (Sinno 2008, 290). While weaker rebels should be more easily
induced to cooperation, both weak and strong groups may need an external enforcer
because anarchy stimulates concerns about relative gains (Grieco 1988, 498). Because a
prospective partner may grow stronger from cooperation, rebel groups may choose to rely
on their own capabilities. Therefore, foreign support is likely to have an effect on alliance
making irrespective of rebel relative strength.
Accordingly, Chandler (1983) writes that the disdain for Vietnam’s occupation of Cam-
bodia in the 1970s brought together Maoist Khmer Rouge and royalist FUNCINPEC even
though the latter was much weaker militarily than its ally. There is likewise a number
of alliances between two capable groups receiving external support such as, for example,
the Tigrayan People’s Liberation Front (TPLF) and The Eritrean People’s Liberation Front
(EPLF), Somali National Movement (SNM) and Somali Patriotic Movement (SPM) or FAP
and GUNT in Chad or Hamas and Hezbollah in Israel. During the civil war in Croatia and
Bosnia, respectively, the government in Belgrade was instrumental in forging ties between
the Serbian statelets and irregular forces even though these rebel groups had a parity with
the host government.
Alliance-Making and Principal-Agent Theory in Civil Wars
Conventional wisdom suggests that anarchy is a distinguishing feature of international
politics (Waltz 2010). Under anarchy, the lack of central authority to oversee and enforce
prospective deals encourages suspicion and hampers alliance-making (Snyder 1984). Not
only do states incur costs from negotiating and maintaining the alliance, but they also face
the possibility of unilateral defection from the agreement. Partners could minimize defection
using formalized treaties with protective clauses and mechanisms under international law
(Leeds et al. 2009) even though opportunistic abrogation seems inevitable when there
is a shift in threat perception, power, values and institutions of alliance members (Walt
1990;Leeds and Savun 2007). Maintaining alliances requires each party to credibly commit
to not cheat the other (Walter 2002). But when actors expect benefits from defection,
prospective partners cannot commit credibly to fulfill agreements. A credible outside actor
with superior strength and abundant resources might bring them together by offering
rewards and punishments or by establishing institutions (Axelrod and Keohane 1985).
The need for outside arbiters is even more pressing in civil conflicts where rebel groups
cannot rely on international law to safeguard inked deals. One way to understand the role
of external arbiters in rebel alliance-making is through principal-agent framework.
At a
minimum, principal-agent framework includes a principal who delegates authority to an
agent in order to solve collective decision-making problems, profit from agent expertise
or credibly commit to certain policies. The principal can select, monitor, and punish its
agent by manipulating the provision of resources (McCubbins and Kiewiet 1991, 27–34).
In conflict studies, this usually translates into a government providing money, sanctuary,
weapons or other tangible resources to rebel groups in return for their cooperation over
goals, organization and tactics (Byman and Kreps 2010;Salehyan 2010;Salehyan et al.
2011,2014;Szekely 2016). The threat to withdraw support allows the sponsor to deter
disobedience or pressure problematic rebel groups into submission (Salehyan 2010;Popovic
2015a,b). If this logic holds, then the principal could also induce its agent to cooperate
with other agents. Existing evidence shows that sponsors have pursued this path, including
the efforts to unite the Afghan mujaheddin, Kashmir insurgents, and, recently, the Syrian
opposition in order to increase their effectiveness against the incumbent government.
However, single principals may have interests that do not necessarily favor effectiveness.
One common interest is to inflict damage on enduring rivals without necessarily committing
to forging inter-rebel ties (Maoz and San-Akca 2012;Salehyan et al. 2011). Rivals can,
for example, exploit ethnic and ideological connections with rebels for domestic purposes
(Saideman 2002;Byman and Kreps 2010), or provide support for weakly organized rebels
to stir instability (Salehyan et al. 2011). Ultimately, sponsors may funnel resources to their
agents to fight other groups whose interests are seen as hostile to the sponsor’s political
goals (Fjelde and Nilsson 2012). Hafiz Assad’s Syria, for instance, supported Palestinian
organizations such as Popular Democratic Front for the Liberation of Palestine (PDFLP),
al-Saiqa, and the PFLP-GC against Fatah to check Arafat’s influence (Byman and Kreps
2010, 11). Pakistan supervised the rise and fall of the Kashmir insurgency, pitting Hizbul
Mujahideen against the Jammu and Kashmir Liberation Front (JKLF), and later Lashkar-
e-Taiba against Hizbul fearing that a dominant Kashmiri organization could take on a life
of its own and make a compromise with India (Haqqani 2010, 290). Therefore, single
principals, especially those involved in enduring rivalry, may not necessarily be inclined to
unconditionally support alliance formation but rather fuel instability that can ultimately
lead to international war and recurring conflict (Salehyan 2009;Colaresi 2014).
While single principals may lack interest in forging ties among rebels, the presence of
two agents serving different principals generates collective action problems at the top of the
delegation chain. Frequently multiple sponsors have widely different agendas (for instance,
Sudan and France in the Chadian civil war), which may have detrimental effect on the
ability of groups to cooperate.
Sponsors must agree that the cooperation between their
agents is desirable, find a mutually acceptable framework, and work together toward the
alliance formation. Assuming that sponsors agree on the goals and means of prospective
alliance, they are faced with the division of labor problem—who should incur greater costs
of monitoring, supplying and sanctioning the agents. This lack of unity is particularly
exacerbated with the increase in the number of principals who can veto coordination efforts.
As a consequence, disunited principals can send contradictory signals to their respective
agents, which in turn hampers agents’ alliance proclivity.
If the lack of common supervision fuels uncertainty between the prospective partners,
and sows the division between the principals, then alliance formation might benefit from
two agents sharing a principal. Shared principals may come into being in two ways.
In the ideal case, shared sponsors offer assistance simultaneously to both groups under
the condition that they cooperate against a third-party (e.g. U.S. support for the Syrian
Democratic Forces composed of Kurdish and Syriac groups). This implies that a sponsor
appears for the first time in a conflict without much history of interference in the target
country. Yet, this seems implausible because civil conflicts frequently attract neighbors
who either seek to exploit the turmoil or prevent war diffusion across their own borders
(Salehyan 2009;Kathman 2010). Another, more plausible, possibility is that a sponsor aims
to transform its role from single or different to shared principal. In this case, the sponsor
may seek to increase the effectiveness of its proxy forces and legitimize the opposition
struggle in the eyes of the international community. To do so, the sponsor must first rein
in its protege prior to soliciting other rebel groups. One approach is to grant a sanctuary
to potential partners as Sudan did with the Chadian rebels in 2000s. This leaves rebels
without much choice but to comply with the sponsor’s demands. Another possibility is that
the sponsor systematically appoints loyal cadres to the protege’s leadership. For example,
Pakistan solicited cooperation between Lashkar-e-Taiba and Hizbul Mujahideen in the
1990s by appointing "Pakistani militants and foreign mercenaries as commanders of [...]
the Hizbul Mujahideen" (Tribune 1998).
Once in control of the partners, shared principals enjoy leverage greater than that
of single principals because they can credibly threaten to deny resources to both agents
for underperformance or transgression.
When the sponsor supports one rebel group, the
insurgents have considerable leverage in the relationship, and can ignore the demands of the
principal. But when the support simultaneously flows to another rebel group, the leverage
of the first protege is weakened. Foreign sponsors benefit from finding an additional protege
so as to increase their leverage over their agents. If one rebel group is disobedient relative to
the other, the sponsor may divert resources to the more loyal rebel group, thereby reducing
the relative strength of the problematic agent. For example, if one rebel group fights more
often against its partner than the government, it is likely to lose the sponsor’s favor. Yousaf
and Adkin (1992, 150) demonstrate this mechanism in the case of the Afghan mujahedeen
who received arms and supplies from Pakistan only if they attacked Soviet troops; passive
alliance members were temporarily denied support if they failed to spend ammunition in
combat. If sponsors offer support sequentially, then it may behoove them to reduce the
moral hazard of offering support to only one rebel group through finding an additional
dissident organization to support. Once the patron is sponsoring two rebel groups, the
principle has the leverage to push both rebel groups to enter into an alliance. An agent
may dislike this scheme, but it often lacks viable alternatives, especially in the case where
the other partner reaps the rewards for good performance. Thus, this relative performance
evaluation should mitigate the fear of exploitation that would otherwise exist in schemes
with single and different principals.
There are also pitfalls for alliance formation even in the shadow of shared sponsors.
Shared sponsors may unknowingly incite competition between their agents. Cooperation
with foreign sponsors in Chad, Eritrea, or Lebanon generated significant divisions within
rebel groups. In particular, Libyan willingness to support FAN and FAP against the govern-
ment led to serious disagreements and ultimately spelled the end of the Second Liberation
Front in Chad.
Even worse, once external support helps rebels ascend to power, ties to
shared sponsors may compromise externally-backed groups in the post-conflict environment
in which the public may frame them as "traitors" or agents of foreign powers (Colaresi
2014). Less benevolent shared principals may pit their agents against one another in order
to increase their control (e.g. Pakistan pitting Pakistani groups against indigenous Kashmiri
groups). This should boost the agent’s fear of exploitation from both the principal and
partner and weaken the alliance in the long run. Further complicating things, multiple
shared principals can issue conflicting orders to the alliance partners, damaging the alliance
in two ways. For example, Free Syria Army’s (FSA) key sponsors pursue different policies
regarding Moscow-backed Astana talks on Syria: while Turkey participates in the talks,
the United States has opted out. The agents may play the principals against one another
to increase their autonomy. While the agents decrease the fear of exploitation from the
principals, they simultaneously become more vulnerable to each other as the external
arbiter is unable to sanction their behavior. This leads to a scenario in which despite shared
sponsors rebels end up fighting each other. For instance, skirmishes broke out between
CIA-sponsored Fursan al Haq, by and Pentagon-backed Syrian Democratic near Aleppo
last year.
Simultaneously, benevolent agents may fall prey to the conflict between their
principals. Because the agents may be confused whose orders to follow, their alliance might
run the risk of rupture.
The potential costs of foreign support may lead rebels to seek cooperation without
external guidance. Indeed, civil wars feature inter-rebel alliances absent sponsors (e.g.
Guatemala, El Salvador, Myanmar). One alternative is that similar ethnic background may
bring groups together. However, co-ethnic groups might see each other as competitors for
popular support and territorial control, and engage in outbidding (Bloom 2004). Eliminating
a rival co-ethnic group may be more beneficial because it allows the winner to attract the
membership of the defeated. Failing to do so poisons the relationship between co-ethnic
rebels and may lead to vicious circle of rivalry and violence. In this case, shared sponsors
could serve as a barrier to internecine fighting while single or different sponsors may
fuel outbidding. A more viable alternative is the ideological compatibility. Communist
insurgencies in Latin America during the Cold War and, more recently, jihadi coalitions in
Syria and Iraq suggest that shared ideology may act as catalyst for alliance. In some cases,
rebel groups embrace certain ideologies to appeal to specific sponsors (San-Akca 2016).
For example, leftist ideology served groups to make themselves look favorable to USSR
during the Cold War period. In this respect, it may be that rebel ideology depends on the
sponsor’s ideology, and that this, in turn, depends on access to foreign support. It is unclear
whether shared sponsors also prefer ideologically congruent agents beyond the Cold War
(e.g. Iran’s support for alliance between Shia Hezbollah and Sunni Hamas). Thus, it may
be that ideology can have both independent and intervening effect on alliance-making,
depending on the absence or presence of foreign support.8
In sum, these examples suggest that while shared sponsors is not without imperfections,
it promises to have direct effect on inter-rebel alliances, more so than shared ethnicity or
Foreign Sponsors and Inter-Rebel Alliance
If sponsors can delegate authority to individual rebel groups (Byman and Kreps 2010;
Salehyan 2010;Salehyan et al. 2014;Szekely 2016), then they should theoretically be able
and willing to foster relationships, belligerent or cooperative, between two rebel groups.
Existing research on rebel alliance hints at this possibility but either develops no specific
mechanisms that would link sponsors and inter-rebel alliance (Christia 2012) or doubts that
sponsors can mitigate the imbalance of power between the prospective alliance partners
(Bapat and Bond 2012). In contrast, this article links insights from the literature on foreign
sponsorship with the emerging literature on rebel alliances to argue that alliance-making
will depend on the structure of the relationship between sponsors and rebels. In this article,
there are three possible configurations: 1) sponsors may serve as a principal of a single
rebel group; 2) two groups may act as agents of different sponsors; 3) two groups may be
agents of a shared sponsor.
The first form is likely to offer narrow prospects for alliance formation due to the
imbalance of power between two prospective partners. The externally-backed rebel group
may be uninterested in the cooperation with a rebel group lacking external backing because
the presence of foreign support may boost the capabilities and self-confidence of the
former. Consequently, the externally-backed group may view its bargaining position as more
favorable and decide to dictate preconditions for cooperation. Ultimately, the externally-
backed group may choose belligerence over cooperation in an attempt to eliminate the
competition. For example, this logic corresponds with the observation of the British liaison
officer Captain Hudson in the Chetnik headquarters in occupied-Yugoslavia during World
War II. Analyzing the failure of the Partisans and Chetniks to form a viable alliance against
the German forces in 1941 Hudson notes that
The British promise of support had the effect of worsening Chetnik-Partisan
relations. When I first arrived at Ravna Gora and Uzice, at the end of October,
1941, before Chetnik-Partisan hostilities, Mihajlovic already knew by telegram
that he would get British support. He felt rightly that no one outside the country
knew about the Partisans or that he alone was not responsible for the revolt
(Maclean 1957, 126).
A similar pattern is visible in the Sri Lankan civil war where the Tamil Tigers used Indian
support to wipe out their competitors. Another possibility is that the sponsor may be
disinterested in fostering cooperation. If the sponsor desired genuine cooperation between
its agents, it would have provided support to both groups. Instead, the empowered
agent may be unleashed against other groups. For example, Pakistan encouraged Hizbul
Mujahideen to initiate fratricidal attacks against its former agent, JKLF. Similarly, Fjelde
and Nilsson (2012) find that such sponsors are more associated with inter-rebel violence
than cooperation. This implies that sponsors favoring one rebel group over the other may
hamper their potential for cooperation. This leads to the first hypothesis:
H1: Ceteris paribus, foreign support will have no effect on alliance formation if
it is directed to only one group.
Another possibility is that both prospective partners receive support, but from different
sponsors. While the imbalance of power becomes less of an issue, assuming that each
sponsor equally contributes to rebels’ capabilities, multiple principals face difficulties
synchronizing their policies. Given their diverging preferences, multiple sponsors lack
common standing toward their agents (Salehyan 2010). This, in turn, may lead principals
to issue contradictory directions to their agents. At a maximum, sponsors may impose
their own preferences on each other, preventing their agents from cooperation. One such
example is the failure of two Congolese rebel groups, Rally for Congolese Democracy (RCD)
and Movement for the Liberation of the Congo (MLC) to preserve their alliance once their
respective sponsors, Rwanda and Uganda, turned against each other over the spoils of war
in the eastern part of the country. Another example is the inability of Gulf countries to put
together an anti-Assad alliance of their fragmented agents. Sponsors must synchronize their
policies to make their agents cooperate. Due to collective action problems, this undertaking
is ultimately very costly, and sponsors often end up issuing contradictory directives to their
respective agents. This produces the second hypothesis:
H2: Ceteris paribus, foreign support will have no effect on alliance formation
when two groups receive it from different sponsors.
The final possibility is that two groups receive support from the same sponsor or
sponsors. Shared sponsors are likely to demand cooperation between their agents, and
may devote considerable resources to this goal than sponsors of a single group. Reasonably,
shared sponsors may provide support to multiple rebels to instigate chaos or maintain their
influence in the target country without a commitment to their cause. But the creation of
a rebel alliance may signal the sponsor’s resolve to topple down the target government.
The presence of shared principal with superior monitoring and sanctioning capabilities
should minimize the fear of cheating and exploitation inherent in the anarchic nature of
civil conflicts. Shared sponsors might favor rebel cooperation either because it corresponds
with their preferences or offers the possibility to manipulate alliance partners. For example,
Iran fostered cooperation between Hamas and Hezbollah throughout the 1990s and 2000s
as a part of their shared "resistance" against Israel and the West (Byman and Kreps 2010).
Shared sponsors can also use monitoring and sanctions to increase the cost of unilateral
defection "by offering material inducements to make alignment more attractive or by
threatening to punish disloyal regimes" (Walt 1997, 164). For instance, in his recollection
of Pakistan’s relationship with the Afghan mujahideen (Yousaf and Adkin 1992, 150–151)
specify how ISI officers manipulated the supply of weapons and ammunition to the seven
For planning purposes we worked on a rough percentage basis for each Party.
These were not permanently fixed; they varied slightly for operational reasons,
and sometimes they were deliberately reduced if a Party was seen not to be
pulling its weight in the field. Such reductions were normally gradual and
followed a verbal warning to the Leader. [...] If my officers reported a warehouse
was always full, sometimes for months, it meant that the Party was less than
enthusiastic at prosecuting the war, and as such never qualified for an increased
share of arms.
Second, the shared sponsor’s commitment may also signal its resolve to consolidate
control over insurgency. Shared sponsors may foster the establishment of umbrella institu-
tions to improve their monitoring and sanctioning capacity because it is easier to navigate a
collection of groups rather than multiple outfits. Umbrella organizations also ensure that
no single rebel outfit can negotiate separately with the incumbent government without the
consensus of the sponsor. These include, for example, Pakistan’s creation of the United
Jihad Council, an umbrella organization of jihadists based in Pakistan, Seven-Party alliance
in Afghanistan and, the Arab-sponsored Syrian National Council. Shared sponsors should
be most committed to forging cooperation between their agents when they have strategic
interests in the conflict-ridden country, such as the acquisition of the territory, resources or
population as well as weakening of their rival (Maoz and San-Akca 2012). For example,
the long-desired acquisition of Kashmir was the driving force behind Pakistan’s decision
to support the alliance between its agents, Lashkar-e-Taiba and Hizbul Mujahideen, and
to later establish the umbrella institution for Pakistani-based jihadi outfits. The shared
sponsor’s influence on alliance formation may be buttressed by common ties with its agents
such as ideology, ethnicity or religion. Under such circumstances, shared sponsors can
combine material support with legitimacy to foster cooperation between their agents.
While alliance formation may serve the shared sponsor’s interests, rebels receiving
external support are also likely to benefit from cooperation. Rebel groups may anticipate
valuable resources such as weapons, funding or sanctuary. For instance, Yousaf and Adkin
(1992, 150–155) portray how the ill-equipped Afghan mujahideen largely toned down their
differences to receive external support. Without such a support the mujahideen would have
risked an uncertain future against a stronger foe. Another advantage is that the shared
sponsor may guarantee that alliance partners will not exploit each other even if one of them
becomes much stronger. Shared sponsors can threaten to withdraw resources or punish an
agent for disobedience. This guarantee minimizes fear and distrust that would otherwise
deter cooperation under anarchy. Therefore, this leads to the third hypothesis:
H3: Ceteris paribus, when two rebel groups share a sponsor, they are more likely
to form alliance.
Data and Research Design
To examine these hypotheses, I have assembled an original dataset of all multiparty civil
conflicts for the period 1975–2009 in which rebel groups may or may not cooperate against
the government or other rebel groups. I begin with the existing UCDP Armed Conflict
Dataset (Gleditsch et al. 2002), which includes civil conflicts with two or more non-state
actors that are fighting against a government. Next, I code those civil conflicts where there
were at least two rebel groups active for any observed year. After selecting those conflicts, I
then arrange the dataset into dyad-years where a dyad includes two rebel groups. Following
Bapat and Bond (2012) a dyad is coded only if two rebel groups were active in the same
territory and year. For instance, Hamas and Hezbollah were both militarily active from
1987 until 1996, and MILF and MNLF were active throughout the 1980s, which makes
them potential partners in a respective conflict and time period. In this case, I analyze the
dyadic relationship between Hamas and Hezbollah, and between MILF and MNLF for the
period in which they were active. In most instances, the government is fighting more than
two rebel groups so I create annual dyads from every possible combination of the active
groups. If two groups were fighting within a country in two different territorial conflicts
(e.g. Kashmir insurgents and the Naxalites) they were not considered as potential dyads. In
rare instances, (i.e. ELN and FARC in Colombia, MCC and PWG in India, and SPM and SNM
in Somalia) the alliance between two groups collapsed only to be re-established within one
to two years.
The dataset covers the post-1975 period because the UCDP data on external support—
used to measure my main independent variable—records information only for this period.
To my knowledge, there are no datasets with similar time-sensitive and robust information
on the identity of foreign sponsors. This limits my ability to fully evaluate my argument
and competing explanations in the pre-1975 period. In total, the dataset includes 165 rebel
dyads nested within 985 dyad-years.
Alliance is a formal or informal cooperation between two rebel groups in which they share
resources or coordinate attacks against the government or other rebel groups. It denotes
whether there is alliance or not. I borrow this variable from Bapat and Bond who measure
alliance as "resource-sharing or tactical co-ordination between the groups at some time
during a year" (2012, 19). Their dataset includes information on alliance-making for the
period 1946–2001 or 1,318 observations with 429 occurrences of alliance (33 percent). In
the first stage of coding, I excluded from their dataset cases in which one of the potential
partners is an alliance (e.g. UIFSA in Afghanistan), military faction or nameless group
of organizations labeled "various insurgents" (these include, for instance, Chadian rebels
in the 1970s, Myanmar’s Shan insurgents, Lebanon’s sectarian organizations, and other
non-PLO groups in Israel).
Since my dataset extends to 2009, I then searched for additional evidence on alliance-
making in the UCDP External Support in Armed Conflict Dataset (Högbladh et al. 2011).
Similar to Bapat and Bond (2012), this dataset codes alliance as the provision of warring
(i.e. troops) or non-warring support (e.g. money, logistics or training) support as well
as the coordination of policies, including information on the identity of partners on an
annual basis. Additionally, I searched for mergers in the UCDP Non-State Actor dataset
since Bapat and Bond also record them.
Where there was evidence of one group providing
support to or merging with its prospective partner, alliance was coded 1, and 0 otherwise.
Following Bapat and Bond, I considered only cases where two rebel groups cooperated in
the same territory and year because transnational alliances may entail different alliance
dynamics and foreign sponsors may not hold the same influence on rebel groups. Therefore,
alliances with militant movements outside a conflict (e.g. MILF and ASG with Jemayyah
Islamiyya) or with transnational militant movements (e.g. ARS/UIC with Al-Qaeda) were
not considered. In sum, the data include roughly 21 percent of dyad-years with alliance
(206 observations) and 79 percent of dyad-years without an alliance (779 observations),
yielding a total of 985 observations.
Shared Sponsors
The central argument of this article is that sponsors are likely to boost alliance formation
if they are shared by both members of a dyad. To test this argument, I draw on UCDP
External Support in Armed Conflict Dataset (Högbladh et al. 2011). The key advantage of
this dataset is that it provides information on the identity of the sponsor and year of support
for every rebel group that fought against the government from 1975 to 2009. This allows
me to identify not only who were recipients of foreign support in a dyad, but also whether
the potential partners shared a sponsor for any given year. UCDP defines sponsor as a
government of an internationally recognized country that provides warring or non-warring
assistance to a party in an ongoing civil conflict (Ibid). This support can take a form of the
provision of weapons, funding, sanctuary, logistics, training, intelligence, regular troops and
other types of material support. In this article, I consider all these forms of support together.
The resulting variable, "Sponsor", measures whether any member of a dyad receives foreign
support in any given year or not.
Next, I determine whether one or both groups have sponsors to test the hypotheses.
This variable is an upgraded version of the previous in that it displays the source of foreign
support in the following way. If no member of a dyad received support in a given year,
the predictor takes the value of 0; if only one receives external backing the variable takes
the value of 1; support for both partners from different sponsors is coded 2; and if both
received support from the same sponsor it equals 3. It is worth noting that the value of 0
(no support) is regarded as a baseline category in the subsequent analyses.
Figure 1 here
The frequency plot in Figure 1 shows that no support and shared support are prevalent
in the data, while observations with different and single sponsors are rarer. Put simply, this
implies that foreign support more often takes the form of shared delegation than that of
single or competing sponsors. When these categories are compared with the occurrence
and non-occurrence of alliance, visible is a large discrepancy in observations related to
non-support, and a balance regarding shared sponsorship.
Other Predictors
Beyond testing the main hypotheses, this article also engages two main explanations in
the literature. The first is advanced by Bapat and Bond (2012) and includes an interaction
between relative rebel strength and foreign support. Relative rebel strength ("Weak Dyad")
denotes whether the dyad members are weaker than the host government. This variable
was coded following Bapat and Bond, in that I use the weaker of the Non-State Actor (NSA)
dataset figures (Cunningham et al. 2013) for both groups to denote their ability to resist
the government’s repression. The variable is coded 1 when both groups are weaker than
the government, and 0 if they are a match to or stronger than the government.
The second explanation comes from Christia (2012) who argues that rebel groups are
more likely to cooperate with groups of similar strengths. "Ratio" measures the balance
of power within any given dyad as a range of values from 0 (extreme imbalance) to 1
(balance). Drawing on the number of troops from the NSA (Cunningham et al. 2013), I
calculate "Ratio" by dividing the number of troops of group A by number of troops of group
B. If this argument holds, then "Ratio" should be positively associated with the probability
of alliance.
To ensure that my analysis does not simply reflect the impact of other predictors potentially
associated with both alliance formation and with the main variables of interest, I include a
number of dyad-, conflict-, and country-level controls. In particular, I include the variables
suggested by Bapat and Bond (2012) to control for the impact of environment.
"GDP per capita" is used to denote the government’s ability to control its territory.
Countries with lower GDP per capita should stimulate rebel groups to cooperate against
the government. I draw on Gleditsch (2002) for the measure of this variable, which is
log-transformed for the purpose of this article. Another proxy for the government’s absolute
capacity to deal with the insurgency is military spending ("Expenditure"). An increase in
military spending should signal the lack of capacity to tackle the insurgency. Thus, with
every unit increase in spending, there should be an increase in the likelihood of alliance.
This variable is borrowed from the COW dataset (Singer 1988), and log-transformed.
Another possibility is that the central government is recently formed and that multiple
rebel groups may seize this opportunity to join hands in toppling down the incumbent
regime. This predictor is taken from Gurr et al. (2010), and logged. Other controls include
non-contiguity—denoting countries, like Indonesia or the Philippines whose capitals are
physically separated from the rest of the territory—duration, which controls for temporal
dependence in the data, and ethnic and religious fractionalization.
Analysis and Discussion
The data is composed of dyad-year observations, where each dyad-year is nested within a
dyad. This implies that observations are not independent of each other given that there are
multiple rebel groups who operate within the same conflict, and may indirectly interact
with each other. Violating the assumption of independence of observations can lead to
biased estimates of coefficients and their standard errors (Barcikowski 1981). Accordingly,
I use multilevel logit regression, which is found to mitigate this issue (Gelman and Hill
2006). The multilevel model allows coefficients to vary across several levels even where
observations are non-independent, correctly modeling correlated error. In this article, I
cluster observations by government (country) level.10
Figure 2 reports the direction of each of the predictors on alliance formation, using
the coefficient estimates and confidence intervals.
The vertical dotted line represents no
effect; positive coefficients (associated with alliance) are represented by point estimates to
the right of the dotted line, while negative coefficients (associated with no alliance) are
represented by point estimates to the left of the dotted line. The horizontal solid lines
represent the 95 percent confidence intervals. These intervals display the range of values in
which one can be 95 percent certain that the true value of the parameter lies. The observed
relationship is regarded as above the conventional critical values when the interval bars do
not include 0. The effect is thus present if 0 lies outside the intervals, and unclear if 0 is
included in the intervals.
Figure 2 here
Figure 2 includes three models of alliance-making.12 I begin my analysis by presenting
Model 1, in which I test H1 and H2 that single and different sponsors will have no effect
on alliance-formation between two given rebel groups, whereas shared sponsors should
increase the probability of cooperation, as H3 suggests. As envisaged, the findings show that
the presence of either single or different sponsors has no effect on alliance-formation given
that their confidence intervals include the line of no effect. This squares with the wider
expectations in civil war literature that foreign support may not necessarily have a positive
effect on inter-rebel cooperation (Fjelde and Nilsson 2012). In fact, the external backing
may bolster power asymmetry between potential partners, encouraging competition and
fratricide rather than cooperation. In contrast, confidence intervals for shared principals
are positive exclude "zero", indicating that the effect of shared sponsors has a practical
significance for understanding the onset of alliance-formation. These results suggest that
there is more space for inter-rebel cooperation if the potential partners are agent to the
same foreign sponsor.
Model 2 tests the explanation advanced by Bapat and Bond (2012) that foreign support
("Sponsor") interacted with the strength of the dyad relative to the government ("Weak
Dyad") is more likely to lead to alliance formation in civil wars. I find no support for this
argument as the interaction effect ("Sponsor x Weak Dyad") displays unclear effect on
alliance. In contrast, the interaction term for foreign support ("Sponsor") has a positive
and considerable effect on the probability of alliance, indicating that sponsors of a strong
dyad may be more likely to lead to alliance formation. The failure to replicate findings
in Bapat and Bond (2012) is, first, due to a difference in time-period covered in our
respective analysis because I omit 236 observations from the pre-1975 period, and include
97 observations for the post-2001 period. Even when I constrain the time period to 1975–
2001 the results do not change.
Another reason is that Bapat and Bond use their own
measure of sponsorship, whereas I borrow the measure from the UCDP. But the UCDP
measure covers the post-1975 period. Their measure does not include information on
identity of sponsors, and that precludes me from using their proxy to test my argument.
Finally, the difference in statistical models might be driving the outcome. While Bapat and
Bond employ probit model, I use multilevel logistic model, which accounts for the fact that
rebel dyads are nested within conflicts.
On the other hand, the findings in Model 3 lend support to Christia’s argument that a
balance of power facilitates cooperation between rebel groups as the coefficient estimate
for balance ("Ratio") has a large and positive effect on alliance. Therefore, the shadow of
anarchy constraints cooperation as rebel groups fear exploitation from stronger partners.
Regarding control variables, the models offer two findings. First, an inter-rebel alliance
is likely to take place when potential partners are facing a durable incumbent regime.
Even though this may run against some expectations in the civil war literature, the logic
follows that of balance-of-threat theory where actors are likely to resolve their disputes if
their common foe is perceived as more dangerous. This dynamic is not fully captured in
coefficients for GDP per capita and expenditure, as neither have a clear effect on alliance
formation. Second, I find that alliance is more likely to take place in the early years of
conflict. Perhaps rebel capabilities are largely even at the outset of the conflict or the goal
of toppling down the government resonates well across the opposition spectrum. Either
way, this finding offers important implications for Syria where multiple groups have failed
to form alliance after six years of combat.
Now I turn to Model 1 and Model 3 in detail given that shared sponsors and Christia’s
troop ratio demonstrate practical significance for alliance-formation. Using the estimates
from Model 1 and Model 3, respectively, Figure 3 shows the predicted probability of alliance
by types of foreign sponsors (left-hand side boxplot) and troop ratio (right-hand-side line
plot with 95 percent confidence intervals) while holding other predictors constant. This
includes a prototype case with the following characteristics:
The country is contiguous and polarized along the ethnic and religious lines;
The incumbent regime has been in power for a decade, and spends more than USD 3
million annually on military with a GDP per capita of approximately USD 300;
The conflict has been active for two decades.
Figure 3 shows the substantive impact of shared sponsors on alliance-making as well as
a substantive variation in the effects of sponsor types. The probability of alliance-making
is at its lowest when neither of the prospective allies receives foreign support (less than
0.1). There is more probability of alliance for those dyads in which at least one group
receives foreign support (around 0.15). However, that is much lower compared to cases
in which both groups receive backing from different sponsors (0.25). Interestingly, this
finding suggests that cooperation is far more likely among groups with external backing
than those that rely solely on domestic support. Shared sponsors boost the probability of
alliance more than all other types of support taken individually. The probability of alliance
is this case is double than that of different sponsors (0.5 vs. 0.25), and nearly triple than
that of single sponsors (0.5 vs. 0.18). Shared sponsors are, therefore, a crucial factor in
predicting alliance-formation in future civil conflicts.
Figure 3 here
Moving to the right-hand side part of Figure 3, there is a clear upward trend in predicted
probability of alliance as ratio inches closer to 1, i.e. balance of power. As the balance be-
tween two groups increases the probability of alliance grows nearly five times. Although the
effect is impressive the 95 percent intervals are extremely wide, decreasing the confidence
in the result.
While the results and the substantive effects show the strong effect of shared sponsors
on the presence of inter-rebel alliance, it is equally important to explore the predictive
capacity of the models. I do so using 10-fold cross-validation, which is a machine learning
technique used to address the model underperformance and overfitting through a random
partition and analysis of data (Colaresi and Mahmood 2017).
I first determine predictive
performance of the models using Receiver Operating Characteristic (ROC) curves.
curve is used in applications in which data are class imbalanced to indicate the true and
false positive rates for a classifier.
In this article, the ROC curve shows the extent to which the model correctly classifies
"alliance" and "non-alliance". The ROC graph is summarized by the Area Under Curve
(AUC), which is the probability that the model correctly ranks positive cases ("alliance")
versus negative cases ("no alliance"), and that one has the greater probability than the other.
Models with greater predictive capacity gravitate toward the upper left corner of the plot,
and have higher AUC scores. This indicates the true positive rate against the false positive
rate for the different possible cut-points of a diagnostic test. The closer the curve to the
diagonal line, the more the prediction resembles a coin-flip (.5); the closer the curve to the
upper left corner, the more accurate the model—a perfect fit would have the curve hugging
the top-left corner (1).
Figure 4 here
AUC scores of 0.70 are regarded as fair, while AUC scores equal to or higher than 0.80
are considered good. The predictive performance in Figure 4 varies from 0.85 to 0.87.
Model 1 has the highest AUC of 0.87, while Model 2 and Model 3 have slightly lower
AUC scores of 0.85 and 0.85 each. I conclude that my model is slightly more capable in
predicting alliance-making.
These results show the overall performance of my model, but it is also important to
understand how to improve its predictive power. I, therefore, identify observations that
generate lower performance using the model criticism plot (Colaresi and Mahmood 2017).
The model criticism plot shows the distance between actual and predicted values by plotting
the latter for each observation on x-axis. Observations are then colored according to
their observed value (in this case, non-alliance is blue, while alliance is red). The plot
then ranks the predicted values in descending order on the y-axis. Extremely inconsistent
positive values (alliance is observed, but the model predicts low probability of alliance)
are colored in red and gravitate toward the southwest, whereas highly inconsistent zero
values (non-alliance is observed, but the model predicts high probability of alliance) appear
in blue toward northeast. Observations appear on the vertical separation plot on the right
y-axis, with most discrepant positive and zero observations being colored in intense red and
blue, respectively. These observations are labeled and connected by lines to their respective
points on the y-axis on the left.
Figure 5 here
Figure 5 displays the model criticism plot for Model 1.
Similar to ROC scores, Model 1
demonstrates a solid predictive performance as well as good separation of negative and
positive values given that there are only a few extremely inconsistent positive values in
the southeast. The top ten discrepant cases of alliance (red) are: SSDF–SNM in 1991
(Somalia), NRA–UPM in 1986 (Uganda), JEM–SLM in 2007 (Chad), Amal–LNM in 1985–
1986 (Lebanon), SPM–SNM in 1989, and EPDM–TPLF in 1989 (Ethiopia), FAR-PGT and FAR–
ORPA, all in 1979 (Guatemala), and ELN–FARC in 1991 (Colombia). Foreign sponsorship
is present in one instance, in the JEM–SLM dyad. The most common denominators for
the remaining nine observations are weaker capabilities relative to the government and a
rather long time period during which the dyads entered into alliance (within a decade of
their existence). Interestingly, most of these observations belong to the Cold War period
where ideological polarization played a critical role in intra-conflict dynamics as well as
in relation to foreign sponsors. Using Non-State Armed Groups (NAG) dataset (San-Akca
2016), I check whether the groups in those dyads shared one of the possible ideologies: left-
wing, nationalist, religious, right-wing. I find that eight out of ten dyads are ideologically
compatible. This suggests that adding shared ideology could be useful for exploring whether
it improves the performance of the model regarding alliance.18
I notice that highly discrepant cases of non-alliance are clustered in the Israeli–Palestinian
conflict where ethnic outbidding and foreign support among the Palestinian outfits may
have prevented the formation of cooperation. This reinforces my previous point that foreign
support may not necessarily contribute to inter-rebel alliance unless it comes from shared
This article builds on principal-agent framework to argue that an inter-rebel alliance
in civil conflict is more likely when any two potential partners share a foreign sponsor.
This relationship provides two-way benefits: sponsors can use alliance to manipulate the
dynamics of the conflict while rebels receive material incentives such as weapons, funding
or sanctuary. Moreover, shared sponsors are likely to invest effort in preventing potential
defections through controlling and monitoring mechanisms. The empirical analysis, using
novel panel data on rebel alliances, suggests considerable support for this argument. I find
that alliance formation is more likely when both potential partners receive external backing
from the same sponsor or sponsors. Moreover, I find that single and different sponsors have
an unclear effect on alliance formation and that no foreign support appears to offer the
worst prospects for cooperation.
Future research on civil war should take the foreign sponsorship of inter-rebel dynamics
more seriously. With a few exceptions, previous quantitative studies have largely explored
the relationship between the incumbent government and rebel groups. As this study
demonstrates, rebel groups develop ties with external governments, which can significantly
influence their strategies. Bringing together data on foreign sponsors and rebel groups
can shed more light on the blurry line between external and domestic actors in civil war.
Research on political violence may also benefit from this merger by examining how links to
sponsors affect rebel propensity for violence against civilians.
These insights may better inform both counter-insurgency and conflict resolution. While
governments may deal with separate groups, the alliance makes rebels a more formidable
opponent. Counter-insurgent strategies should factor in the presence of external ties, and
invest more efforts into resolving issues with foreign sponsors by diplomatic means. Conflict
resolution professionals should also take into account the preferences of foreign actors
before investing considerable resources into mediation efforts. As the Syrian conflict testifies,
there is no shortcut to peace, let alone cease-fire, when multiple foreign governments
interfere in insurgent interactions.
1I am thankful to Kanisha Bond for sharing the dataset.
I understand rebel alliance as a "formal or informal arrangement for security cooperation" (Walt 1990,
12) between two rebel groups.
Principal-agent theory originates from political economy, management, and law, but it has also found
its application in political science (McCubbins and Kiewiet 1991;Pollack 1997;Nielson and Tierney 2003;
Hawkins et al. 2006).
4I thank a reviewer for pointing out this issue and example.
5I thank a reviewer for highlighting leverage as a distinguishing feature of shared principals.
6I thank a reviewer for bringing up this example.
The model in Table 2 in the Appendix features shared ethnicity, ideology as alternative control variables.
9Mergers for the post-2001 period include MCC–PWG, UIFSA, and MJP–MPCI–MPIGO. Mergers from the
Bapat and Bond include PF, SRRC, URNG, and FMLN.
The multilevel logit analysis was conducted using a marginal likelihood estimator, and a logit link function
as implemented in the
package (Fournier et al. 2012) for the
language (Venables and Smith
2014). Missing data was corrected with multiple imputation using R’s
package (Honaker et al. 2011),
creating 500 imputations. This number is considered very high as the literature recommends only 10 to
50 imputations. But for many of the estimates, the rate of missing information was too high when a lower
number of imputations was used. Therefore, I increased the number of imputations to get the rate of missing
information below 5 percent for all estimates. The coefficient estimates and standard errors from 500 models
were pooled following "Rubin’s rule" (Little and Rubin 2014, 86–87).
Following American Statistical Association’s (ASA) suggestion to avoid p-values in favor of other ap-
proaches (Wasserstein and Lazar 2016), I choose confidence intervals to present my findings. Conventional
regression output with coefficient estimates, standard errors and p-values is in Table 1 in the Appendix.
I separated the latter from my main predictor because both relative rebel strength and troop ratio are
theoretically endogenous to foreign support.
13See Figure 4 in the Appendix.
K-fold cross validation (CV) is a way to analyze how the results of a model apply to an independent
sample, i.e. predictive accuracy of the model. CV randomly partitions the original data into similar folds of
subsets, and then performs analysis on a single subset ("training dataset"), while validating the analysis on
the other ("testing dataset"). In my case, I partition the dataset into 10 folds of 98 or 99 observations, carry
out CV on 500 multiple imputations, and pool predicted values for testing data.
I provide additional CV diagnostics such as the precision-recall (PR) plot and confusion-matrix in the
I use R package
developed by Zuhaib Mahmood. Package source:
17The model criticism plots for Model 2 and Model 3 are available in the Appendix.
Shared ideology turns out to be a "significant" predictor of alliance when included in a model with shared
ideology and a number of other alternative dyad-level variables (see Table 2 in the Appendix). I have also
included a robustness test for the model with shared ideology/ethnicity and other rebel-level variables by
interacting the Cold War dummy variable with shared ideology as well as including this variable as a predictor.
Ideological proximity is robust to the inclusion of the Cold War variable, which itself demonstrates no effect
on alliance-formation. These results are in Figure 8 in the appendix.
Akcinaroglu, S. (2012). Rebel interdependencies and civil war outcomes. Journal of Conflict
Resolution 56(5), 879–903.
Akcinaroglu, S. and E. Radziszewski (2005). Expectations, rivalries, and civil war duration.
International Interactions 31(4), 349–374.
Asal, V. and R. K. Rethemeyer (2008). The Nature of the Beast: Organizational Structures
and the Lethality of Terrorist Attacks. The Journal of Politics 70(02), 437–449.
Axelrod, R. and R. O. Keohane (1985). Achieving cooperation under anarchy: Strategies
and institutions. World politics 38(01), 226–254.
Bakke, K. M., K. G. Cunningham, and L. J. Seymour (2012). A plague of initials: Fragmen-
tation, cohesion, and infighting in civil wars. Perspectives on Politics 10(02), 265–283.
Bapat, N. A. and K. D. Bond (2012). Alliances between militant groups. British Journal of
Political Science 42(04), 793–824.
Barcikowski, R. S. (1981). Statistical Power With Group Mean as the Unit of Analysis.
Journal of Educational and Behavioral Statistics 6(3), 267–285.
Bloom, M. M. (2004). Palestinian Suicide Bombing: Public Support, Market Share, and
Outbidding. Political Science Quarterly 119(1), 61–88.
Bond, K. D. (2010). Power, Identity, Credibility & Cooperation: Examining the Development
of Cooperative Arrangements Among Violent Non-state Actors. Ph. D. thesis, Pennsylvania
State University.
Byman, D. and S. E. Kreps (2010). Agents of destruction? applying principal-agent analysis
to state-sponsored terrorism. International Studies Perspectives 11(1), 1–18.
Chandler, D. P. (1983). A History of Cambodia. Westview Press Boulder, CO.
Christia, F. (2012). Alliance formation in civil wars. Cambridge University Press.
Colaresi, M. (2014). With friends like these, who needs democracy? the effect of
transnational support from rivals on post-conflict democratization. Journal of Peace
Research 51(1), 65–79.
Colaresi, M. and Z. Mahmood (2017). Do the robot: Lessons from machine learning to
improve conflict forecasting. Journal of Peace Research 54(2), 193–214.
Colaresi, M. P. (2005). Scare tactics: The politics of international rivalry. Syracuse University
Cunningham, D. E. (2011). Barriers to peace in civil war. Cambridge University Press.
Cunningham, D. E., K. S. Gleditsch, and I. Salehyan (2009). It takes two: A dyadic analysis
of civil war duration and outcome. Journal of Conflict Resolution.
Cunningham, D. E., K. S. Gleditsch, and I. Salehyan (2013). Non-state actors in civil wars:
A new dataset. Conflict Management and Peace Science 30(5), 516–531.
Fjelde, H. and D. Nilsson (2012). Rebels against rebels explaining violence between rebel
groups. Journal of Conflict Resolution 56(4), 604–628.
Fournier, D. A., H. J. Skaug, J. Ancheta, J. Ianelli, A. Magnusson, M. N. Maunder, A. Nielsen,
and J. Sibert (2012). Ad model builder: using automatic differentiation for statistical
inference of highly parameterized complex nonlinear models. Optimization Methods and
Software 27(2), 233–249.
Furtado, C. S. (2007). Inter-rebel Group Dynamics: Cooperation Or Competition. The Case of
South Asia.
Gelman, A. and J. Hill (2006). Data analysis using regression and multilevel/hierarchical
models. Cambridge University Press.
Gleditsch, K. S. (2002). Expanded trade and gdp data. Journal of Conflict Resolution 46(5),
Gleditsch, N. P., P. Wallensteen, M. Eriksson, M. Sollenberg, and H. Strand (2002). Armed
conflict 1946-2001: A new dataset. Journal of peace research 39(5), 615–637.
Grieco, J. M. (1988). Anarchy and the limits of cooperation: a realist critique of the newest
liberal institutionalism. International organization 42(03), 485–507.
Gurr, T. R., M. G. Marshall, and K. Jaggers (2010). Polity iv project: Political regime
characteristics and transitions, 1800-2009. Center for International Development and
Conflict Management at the University of Maryland College Park.
Haqqani, H. (2010). Pakistan: Between Mosque and Military. Carnegie Endowment.
Hawkins, D. G., D. A. Lake, D. L. Nielson, and M. J. Tierney (2006). Delegation and Agency
in International Organizations. Cambridge University Press.
Högbladh, S., T. Pettersson, and L. Themnér (2011). External Support in Armed Conflict
1975–2009. Presenting New Data. In 52nd Annual International Studies Association
Convention, Montreal, Canada, March, pp. 16–19.
Honaker, J., G. King, M. Blackwell, et al. (2011). Amelia ii: A program for missing data.
Journal of statistical software 45(7), 1–47.
Horowitz, M. C. and P. B. Potter (2013). Allying to kill terrorist intergroup cooperation and
the consequences for lethality. Journal of Conflict Resolution, 199–225.
Kathman, J. D. (2010). Civil war contagion and neighboring interventions1. International
Studies Quarterly 54(4), 989–1012.
Leeds, B. A., M. Mattes, and J. S. Vogel (2009). Interests, institutions, and the reliability of
international commitments. American Journal of Political Science 53(2), 461–476.
Leeds, B. A. and B. Savun (2007). Terminating alliances: Why do states abrogate agree-
ments? The Journal of Politics 69(4), 1118–1132.
Little, R. J. and D. B. Rubin (2014). Statistical analysis with missing data. John Wiley &
Maclean, F. (1957). The Heretic: The Life and Times of Josip Broz-Tito. Harper.
Maoz, Z. and B. San-Akca (2012). Rivalry and state support of non-state armed groups
(nags), 1946–2001. International Studies Quarterly 56(4), 720–734.
McCubbins, M. D. and R. Kiewiet (1991). The Logic of Delegation. University of Chicago
Metelits, C. (2009). Inside Insurgency: Violence, Civilians, and Revolutionary Group Behavior.
New York: New York University Press.
Nielson, D. L. and M. J. Tierney (2003). Delegation to international organizations: Agency
theory and world bank environmental reform. International Organization 57(02), 241–
Nilsson, D. (2008). Partial peace: Rebel groups inside and outside of civil war settlements.
Journal of Peace Research 45(4), 479–495.
Nygård, H. M. and M. Weintraub (2014). Bargaining between rebel groups and the outside
option of violence. Terrorism and Political Violence (ahead-of-print), 1–24.
Phillips, B. J. (2014). Terrorist group cooperation and longevity. International Studies
Quarterly 58(2), 336–347.
Pollack, M. A. (1997). Delegation, agency, and agenda setting in the european community.
International Organization 51(01), 99–134.
Popovic, M. (2015a). Fragile Proxies: Explaining Rebel Defection Against Their State
Sponsors. Terrorism and Political Violence, 1–21.
Popovic, M. (2015b). The Perils of Weak Organization: Explaining Loyalty and Defection of
Militant Organizations Toward Pakistan. Studies in Conflict & Terrorism.
Saideman, S. M. (2002). Discrimination in international relations: Analyzing external
support for ethnic groups. Journal of Peace Research 39(1), 27–50.
Salehyan, I. (2009). Rebels Without Borders: Transnational Insurgencies in World Politics.
Cornell University Press.
Salehyan, I. (2010). The delegation of war to rebel organizations. Journal of Conflict
Salehyan, I., K. S. Gleditsch, and D. E. Cunningham (2011). Explaining external support
for insurgent groups. International Organization 65(04), 709–744.
Salehyan, I., D. Siroky, and R. M. Wood (2014). External Rebel Sponsorship and Civilian
Abuse: A Principal-Agent Analysis of Wartime Atrocities. International Organization,
San-Akca, B. (2016). States in Disguise: Causes of State Support for Rebel Groups. Oxford
University Press.
Singer, J. D. (1988). Reconstructing the correlates of war dataset on material capabilities of
states, 1816–1985. International Interactions 14(2), 115–132.
Sinno, A. H. (2008). Organizations at War in Afghanistan and Beyond. Cornell University
Snyder, G. H. (1984). The security dilemma in alliance politics. World politics 36(4),
Staniland, P. (2014). Networks of Rebellion: Explaining Insurgent Cohesion and Collapse.
Ithaca, NY: Cornell University Press.
Szekely, O. (2016). A friend in need: The impact of the syrian civil war on syria’s clients (a
principal–agent approach). Foreign Policy Analysis 12(3), 450–468.
Tribune (1998). ISI Sacks Hizbul Chief.
Venables, W. and D. Smith (2014). the r core team. An Introduction to R. Notes on R: A
Programming Environment for Data Analysis and Graphics Version 3(1), 07–10.
Walt, S. M. (1990). The Origins of Alliance. Cornell University Press.
Walt, S. M. (1997). Why Alliances Endure or Collapse. Survival 39(1), 156–179.
Walter, B. F. (2002). Committing to peace: The successful settlement of civil wars. Princeton
University Press.
Waltz, K. N. (2010). Theory of International Politics. Waveland Press.
Wasserstein, R. L. and N. A. Lazar (2016). The ASA’s Statement on P-Values: Context,
Process, and Purpose. The American Statistician.
Yousaf, M. and M. Adkin (1992). The Bear Trap: Afghanistan’s Untold Story. London: L.
None Single Different Shared
Number of observations
0 100 200 300 400 500 600
85 62
Figure 1: Distribution of Type of Sponsors by Alliance
Durability (ln)
Ethnic frac.
Religious frac.
Duration (ln)
Expenditure (ln)
GDP p.c. (ln)
Shared sponsor
Different sponsor
Single sponsor
−5.0 −2.5 0.0 2.5 5.0
Coefficient Estimates
Model 1: Alliance in the Shadow
of Sponsor
Durability (ln)
Ethnic frac.
Religious frac.
Duration (ln)
Expenditure (ln)
GDP p.c. (ln)
Sponsor x
Weak Dyad
Weak Dyad
−4 0 4
Coefficient Estimates
Model 2: Sponsor and Weak Dyad
(Bapat and Bond)
Durability (ln)
Ethnic frac.
Religious frac.
Duration (ln)
Expenditure (ln)
GDP p.c. (ln)
−4 0 4
Coefficient Estimates
Model 3: Relative Strength (Christia)
Figure 2: Pooled Models of Alliance-Making in Civil War, 1975–2009
Figure 3: Marginal Effects on Predicted Alliance based on Model 1 (left) and
Model 3 (right, including 95 percent confidence intervals in grey). GDP p.c., ratio,
military expenditure, conflict duration, religious and ethnic fractionalization, and
regime durability are held at mean values. The remaining control variables are set
to their modal values.
Model 1
False positive rate
True positive rate
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
AUC = 0.87
Model 2
False positive rate
True positive rate
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
AUC = 0.85
Model 3
False positive rate
True positive rate
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
AUC = 0.85
Figure 4: ROC Plots for Pooled Cross-Validated Models of Alliance-Making
AMAL & LNM−1986
AMAL & LNM−1985
FAR & PGT−1979
NRA & UPM−1986
EPDM & TPLF−1989
ELN & FARC−1991
SPM & SNM−1989
FAR & ORPA−1979
JEM & SLM−2007
SSDF & SNM−1991
PUK & KDP−1975
PIJ & PLO (FATAH)−2002
Observation (ordered by f)
0.00 0.25 0.50 0.75 1.00
Forecast Value
Figure 5: Model Criticism Plot for Model 1
... Group strength may have both first-order and second-order effects, e.g., stronger groups form more conflict ties, and they are especially attractive opponents for equally matched or weaker groups. Existing studies find support for the proposition that geographical location, territorial control and foreign sponsorship affect absolute and relative power of groups during war (Buhaug, Gates, and Lujala, 2009;de la Calle and Sánchez-Cuenca, 2015;Popovic, 2018). ...
Full-text available
p>Existing scholarship ignores relational interdependencies when attempting to understand the behaviour of non-state armed groups during civil war. This paper investigates the in- terconnected web of alliances and rivalries in the Yemen Civil War to answer the following question: why do armed groups fight each other? We employ a network approach to investigate determinants of intergroup violence. This emphasises the role of identity, arguing that operating in salient cleavages necessitates that groups align or distinguish themselves from each other. We further argue that informal cooperation incentivises violence the longer the war continues. These arguments are tested using pooled Exponential Random Graph Models to account for endogenous structures over time. Results indicate that shared identity is a significant driver of hostilities, moderating cooperation and amplifying the effects of group attributes. Robustness checks and simulations demonstrate that network models more accurately capture the underlying mechanisms to predict fighting in this case.</p
... Group strength may have both first-order and second-order effects, e.g., stronger groups form more conflict ties, and they are especially attractive opponents for equally matched or weaker groups. Existing studies find support for the proposition that geographical location, territorial control and foreign sponsorship affect absolute and relative power of groups during war (Buhaug, Gates, and Lujala, 2009;de la Calle and Sánchez-Cuenca, 2015;Popovic, 2018). ...
Full-text available
p>Existing scholarship ignores relational interdependencies when attempting to understand the behaviour of non-state armed groups during civil war. This paper investigates the in- terconnected web of alliances and rivalries in the Yemen Civil War to answer the following question: why do armed groups fight each other? We employ a network approach to investigate determinants of intergroup violence. This emphasises the role of identity, arguing that operating in salient cleavages necessitates that groups align or distinguish themselves from each other. We further argue that informal cooperation incentivises violence the longer the war continues. These arguments are tested using pooled Exponential Random Graph Models to account for endogenous structures over time. Results indicate that shared identity is a significant driver of hostilities, moderating cooperation and amplifying the effects of group attributes. Robustness checks and simulations demonstrate that network models more accurately capture the underlying mechanisms to predict fighting in this case.</p
The literature on delegated rebellion has treated principals (external states) and their agents (rebel groups) as the main factors in the inception of rebellion. Intriguingly, no attention has been paid to subnational elites as a separate, third actor. This article takes a novel perspective on delegated rebellion by ascribing agency to subnational elites. It introduces the theoretical concept of strategic entrapment, which shows that even subnational elites unwilling to follow the path of rebel violence may be trapped between the incipient rebel groups and a principal. As a result, subnational elites are sidelined and replaced by the principal’s rebel proxies.
Cooperation among militant organizations contributes to capability but also presents security risks. This is particularly the case when organizations face substantial repression from the state. As a consequence, for cooperation to emerge and persist when it is most valuable, militant groups must have means of committing to cooperation even when the incentives to defect are high. We posit that shared ideology plays this role by providing community monitoring, authority structures, trust, and transnational networks. We test this theory using new, expansive, time-series data on relationships between militant organizations from 1950 to 2016, which we introduce here. We find that when groups share an ideology, and especially a religion, they are more likely to sustain material cooperation in the face of state repression. These findings contextualize and expand upon research demonstrating that connections between violent nonstate actors strongly shape their tactical and strategic behavior.
Rebellion, insurgency, civil war-conflict within a society is customarily treated as a matter of domestic politics and analysts generally focus their attention on local causes. Yet fighting between governments and opposition groups is rarely confined to the domestic arena. "Internal" wars often spill across national boundaries, rebel organizations frequently find sanctuaries in neighboring countries, and insurgencies give rise to disputes between states. In Rebels without Borders, which will appeal to students of international and civil war and those developing policies to contain the regional diffusion of conflict, Idean Salehyan examines transnational rebel organizations in civil conflicts, utilizing cross-national datasets as well as in-depth case studies. He shows how external Contra bases in Honduras and Costa Rica facilitated the Nicaraguan civil war and how the Rwandan civil war spilled over into the Democratic Republic of the Congo, fostering a regional war. He also looks at other cross-border insurgencies, such as those of the Kurdish PKK and Taliban fighters in Pakistan. Salehyan reveals that external sanctuaries feature in the political history of more than half of the world's armed insurgencies since 1945, and are also important in fostering state-to-state conflicts. Rebels who are unable to challenge the state on its own turf look for mobilization opportunities abroad. Neighboring states that are too weak to prevent rebel access, states that wish to foster instability in their rivals, and large refugee diasporas provide important opportunities for insurgent groups to establish external bases. Such sanctuaries complicate intelligence gathering, counterinsurgency operations, and efforts at peacemaking. States that host rebels intrude into negotiations between governments and opposition movements and can block progress toward peace when they pursue their own agendas.
Increasingly, scholars interested in understanding conflict processes have turned to evaluating out-of-sample forecasts to judge and compare the usefulness of their models. Research in this vein has made significant progress in identifying and avoiding the problem of overfitting sample data. Yet there has been less research providing strategies and tools to practically improve the out-of-sample performance of existing models and connect forecasting improvement to the goal of theory development in conflict studies. In this article, we fill this void by building on lessons from machine learning research. We highlight a set of iterative tasks, which David Blei terms ‘Box’s loop’, that can be summarized as build, compute, critique, and think. While the initial steps of Box’s loop will be familiar to researchers, the underutilized process of model criticism allows researchers to iteratively learn more useful representations of the data generation process from the discrepancies between the trained model and held-out data. To benefit from iterative model criticism, we advise researchers not only to split their available data into separate training and test sets, but also sample from their training data to allow for iterative model development, as is common in machine learning applications. Since practical tools for model criticism in particular are underdeveloped, we also provide software for new visualizations that build upon already existing tools. We use models of civil war onset to provide an illustration of how our machine learning-inspired research design can simultaneously improve out-of-sample forecasting performance and identify useful theoretical contributions. We believe these research strategies can complement existing designs to accelerate innovations across conflict processes.
Many sources of economic data cover only a limited set of states at any given point in time. Data are often systematically missing for some states over certain time periods. In the context of conflict studies, economic data are frequently unavailable for states involved in conflicts, undermining the ability to draw inferences of linkages between economic and political interactions. For example, simply using available data in a study of trade and conflict and disregarding observations with missing data on economic variables excludes key conflicts such as the Berlin crisis, the Korean War, the Cuban Missile Crisis, and the Gulf War from the sample. A set of procedures are presented to create additional estimates to remedy some of the coverage problems for data on gross domestic product, population, and bilateral trade flows.
Some of the most brutal and long-lasting civil wars of our time – those in Afghanistan, Bosnia, Lebanon, and Iraq, among others – involve the rapid formation and disintegration of alliances among warring groups, as well as fractionalization within them. It would be natural to suppose thatwarring groups form alliances based on shared identity considerations – such as Christian groups allying with other Christian groups, or Muslim groups with their fellow co-religionists – but this is not what we see. Two groups that identify themselves as bitter foes one day, on the basis of some identity narrative, might be allies the next day and vice versa. Nor is any group, however homogeneous, safe from internal fractionalization. Rather, looking closely at the civil wars in Afghanistan and Bosnia and testing against the broader universe of fifty-three cases of multiparty civil wars, Fotini Christia finds that the relative power distribution between and within various warring groups is the primary driving force behind alliance formation, alliance changes, group splits, and internal group takeovers.