1. Introduction
The fragmentation of militant movements and cooperation among them are central to the dynamics of multiparty conflicts (Seymour et al., Reference Seymour, Bakke and Cunningham2016). Inter-group cooperation improves organizational survival and military effectiveness (Horowitz and Potter, Reference Horowitz and Potter2014; Phillips, Reference Phillips2014). For example, the Taliban’s rapid takeover of Afghanistan after the U.S. withdrawal was facilitated by the alliances Taliban forged with former rivals (Giustozzi, Reference Giustozzi2021).
Most research on militant alliances has asked why groups cooperate, with the common answer being the desire to aggregate military capacity (Christia, Reference Christia2012; Phillips, Reference Phillips2014). Recent work has turned to with whom groups cooperate, reflecting the complexity of conflict systems with dozens of actors (Bapat and Bond, Reference Bapat and Bond2012; Asal et al., Reference Asal, Park, Rethemeyer and Ackerman2016; Bacon, Reference Bacon2018; Popovic, Reference Popovic2018; Gade et al., Reference Gade, Gabbay, Hafez and Kelly2019). Yet a key question remains underexplored. How do groups cooperate: what form do alliances take, and what strategic logics shape their design? (Balcells et al., Reference Balcells, Chen and Pischedda2022; Blair et al., Reference Blair, Chenoweth, Horowitz, Perkoski and Potter2022; Steinwand and Metternich, Reference Steinwand and Metternich2022). While scholars increasingly note that inter-militant group cooperation varies in depth and form (Christia, Reference Christia2012; Balcells et al., Reference Balcells, Chen and Pischedda2022; Blair et al., Reference Blair, Chenoweth, Horowitz, Perkoski and Potter2022), the strategic logic behind deeper, operationally more consequential alliances remains poorly understood.
Armed self-determination disputes offer an important yet underexamined setting to study these questions. Much of the existing theorizing on alliance formation has been developed for contexts in which insurgent groups contest control of the central government, such as Christia’s (Reference Christia2012) application of a minimum-winning-coalition framework to the Afghan Mujahideen, or models in which members of a winning coalition inevitably compete until a single actor captures the prize (Tan and Wang, Reference Tan and Wang2010). Conflicts over self-determination present a different political environment. Although contests for national power are absent, these struggles are nonetheless marked by chronic fragmentation and rivalry (Cunningham, Reference Cunningham2014). Yet armed self-determination conflicts also exhibit extensive cooperation among groups operating in close proximity.
Consider the alliances of the United Liberation Front of Asom (ULFA), a major rebel group pursuing self-determination in Northeast India. ULFA formed a high-end partnership with the Kamtapur Liberation Organisation (KLO), offering training in explosives and weaponry in exchange for sanctuary during Bhutan’s crackdown on ULFA bases (Banerjee, Reference Banerjee1999, Reference Banerjee2002). This high-end cooperation entailed the transfer of knowledge and infrastructure. In contrast, ULFA’s long-standing cooperation with the All-Tripura Tiger Force (ATTF) involved only weapons sales (Kalita, Reference Kalita2011). Under what conditions do militant groups in self-determination conflicts engage in high-end cooperation with one another?
I define high-end cooperation as the transfer of sensitive operational know-how, such as training, intelligence, and logistical assistance, from a militant group that possesses specialized capabilities to one that does not. These partnerships differ from symbolic cooperation (e.g., rhetorical endorsements or pledges of allegiance) and from transactional cooperation (e.g., arms sales). These alliances deserve closer attention because they require sharing sensitive assets, which are unevenly distributed across groups operating in the same conflict ecosystem, with autonomous partners, making these alliances far costlier and riskier than symbolic or transactional ones.
To explain such cooperation, I develop a principal–agent framework. High-end cooperation entails one group (the principal), which possesses the operational know-how, tactical expertise, and logistical networks required to provide high-end support, empowering another (the agent), which lacks such resources, to act on its behalf. By providing sensitive operational know-how, the principal seeks to pursue objectives it cannot achieve directly, such as diversifying tactics, projecting influence into a contested area, or accessing remote battle zones by delegating to an agent.
Yet delegation also exposes both actors to certain risks. For principals, the primary vulnerabilities are: first, the military/strategic risk that agents may repurpose these sensitive assets in ways that diverge from the principal’s aims (e.g., classic agency slack); and second, the political risk that empowered agents may siphon civilian support and emerge as political rivals within the broader militant field. For agents, the risks mirror these concerns: the military/strategic risk that such cooperation will further augment the operational advantages of the already dominant group within the conflict ecosystem; and the political risk that entering a principal–agent relationship will reinforce the agent’s politically subordinate position.
To mitigate these risks, principals strategically select agents, and agents strategically decide whether to accept high-end support. First, principals favor tactically complementary groups (e.g., those with distinct attack portfolios) because such partnerships maximize the gains from delegation. Second, they avoid ideologically proximate or politically substitutable allies (e.g., those who compete for the same civilian constituencies) as a buffer against downstream political rivalry. Agents partner with principals despite military/strategic and political risks because the benefits of receiving high-end support often outweigh these concerns. In fragmented conflict ecosystems where weaker or less capable groups face threats from both the state and rival factions, high-end support can enhance agents’ survival prospects and bolster political credibility. High-end cooperation thus reflects asymmetric alliance-building, wherein principals expand their capabilities and influence while preserving military and political dominance, and agents acquire military and political assets they cannot generate alone.
To evaluate my theory’s observable implications, I employ temporal exponential random graph models (TERGMs) on an original, time-series, directional network dataset covering 53 ethnonationalist militant groups active in Northeast India from 1981 to 2021. Existing datasets often exclude weaker groups, conflate different forms of cooperation, and lack information on directionality. My dataset addresses these limitations by coding eight discrete types of cooperation and specifying directionality for each dyadic tie. Northeast India offers rich variation in cooperation types, groups’ military capacity, and their claimed civilian constituency, making it an ideal setting to test the observable implications of my theory. TERGMs allow me to estimate the effects of exogenous covariates while accounting for network dependencies.
The results support my theory’s expectations. First, high-end cooperation is significantly more likely between groups with dissimilar attack portfolios, consistent with the logic of tactical complementarity. Second, the analysis also finds evidence supporting the proposed mechanism related to political threat perception: high-end cooperation is less likely between groups that appeal to the same civilian constituency. Placebo tests on arms-for-cash transactions and rhetorical endorsements reveal different patterns, reinforcing the claim that high-end cooperation follows a distinct strategic logic.
The argument advanced here is most applicable to domestically rooted militant alliances featuring self-determination groups, where the principal operates autonomously and both actors share the same conflict environment. Three scope conditions follow. First, in center-seeking insurgencies, the logic of alliance formation may differ. Second, the theory assumes no overriding external principal. If the principal is itself subordinate to a state sponsor, for instance, it may lack the freedom to choose its agents. Third, the theory assumes local embeddedness wherein both parties are native to the conflict zone. In transnational alliance networks, such as the Islamic State’s or al-Qaeda’s relationships with ideologically proximate groups in foreign theaters, cooperation likely follows a franchising logic. There, the principal’s broader geographic reach may reduce its vulnerability to political competition from local agents.
This study contributes to the literature by shifting the focus from why groups cooperate to how they do so. By identifying high-end cooperation as a distinct form, I offer a more nuanced view of militant alliances. My theory also reconceptualizes delegation risk in militant principal–agent relations as one containing political rivalry, and therefore, recasts militant alliances as products of competitive threat perception. Empirically, the findings help reconcile contradictory results in the literature. While some studies link shared constituencies to cooperation (Bacon, Reference Bacon2018; Gade et al., Reference Gade, Gabbay, Hafez and Kelly2019; Balcells et al., Reference Balcells, Chen and Pischedda2022; Blair et al., Reference Blair, Chenoweth, Horowitz, Perkoski and Potter2022), others associate them with infighting (Pischedda, Reference Pischedda2018, Reference Pischedda2020; Phillips, Reference Phillips2019). This study shows that shared constituencies discourage high-end cooperation, even as they may facilitate other forms of cooperation.
2. High-end vs. low-end cooperation
In one of the earliest discussions of how militant groups cooperate, Moghadam (Reference Moghadam2015) distinguishes between high-end and low-end forms. He envisions cooperation as a continuum of depth. At the high end are (a) mergers and (b) what he calls “strategic alliances,” which are relationships marked by extensive collaboration in training, intelligence-sharing, personnel, or operational know-how. In contrast, low-end cooperation includes tactical arrangements that are narrower in scope, or transactional ties such as arms sales or rhetorical endorsements, which involve limited exchanges.Footnote 1
While informed by Moghadam’s framework, I depart from it in two ways. First, I exclude mergers. Once groups merge, they form a new entity and no longer engage in inter-group cooperation in the conventional sense. Second, although recent work has begun to theorize and document mergers (Topal, Reference Topal2025), we know far less about high-end cooperation between autonomous groups, or what Moghadam calls “strategic alliances.” These are enduring, operationally consequential partnerships that do not involve organizational fusion. This study addresses that gap. For the remainder of the manuscript, I use the term high-end cooperation to refer to these non-merger alliances in which one actor transfers sensitive operational capabilities and know-how to another. Because these forms of know-how are not evenly distributed across armed organizations in a conflict ecosystem, such cooperation is typically directional: one group provides capabilities that another lacks.
Beyond Moghadam’s distinctions, I argue that high-end cooperation is fundamentally distinct from transactional or rhetorical ties because it imposes significantly greater costs. Like states, militant groups operate in anarchic environments that lack central enforcement mechanisms. Although foreign state sponsors may attempt to oversee the interactions among their militant proxies (Popovic, Reference Popovic2018), they rarely exert full control (Salehyan et al., Reference Salehyan, Gleditsch and Cunningham2011). As a result, militant groups must initiate, negotiate, and sustain cooperation under conditions of uncertainty, which heightens the cooperation costs of alliance formation (Bapat and Bond, Reference Bapat and Bond2012; Christia, Reference Christia2012; Zeigler, Reference Zeigler2016; Bacon, Reference Bacon2017).
Importantly, not all forms of cooperation entail equal costs. Rhetorical endorsements or one-off material exchanges (e.g., arms or funds) require minimal oversight. High-end cooperation, by contrast, creates deeper vulnerabilities. Chief among these are redistributive consequences for operational capacity. The internal distribution of operational capabilities within militant alliances shapes both the incentives for cooperation and the risks associated with sustaining it over time. Groups lacking such capabilities may seek out more capable partners to enhance their operational infrastructure, gain access to new resources, or elevate their standing in a competitive militant landscape (Asal and Rethemeyer, Reference Asal and Rethemeyer2008; Horowitz and Potter, Reference Horowitz and Potter2014; Blair et al., Reference Blair, Chenoweth, Horowitz, Perkoski and Potter2022). But prospective partners must also consider the costs of transferring resources, negotiating the terms of cooperation, and managing the long-term implications of altering the distribution of capabilities within the alliance (Balcells et al., Reference Balcells, Chen and Pischedda2022). High-end cooperation, thus, exposes groups to the risk of unintentionally shifting the dyadic balance of capabilities in ways that may weaken their own position.
3. A principal–agent theory of high-end cooperation in armed self-determination conflicts
If high-end cooperation can shift the distribution of operational capabilities, why would groups enter cooperative arrangements that risk augmenting the capabilities of another group operating in the same conflict ecosystem? I argue that high-end cooperation reflects a principal–agent relationship, in which one group (the principal), possessing specialized operational capabilities, transfers those skills to another (the agent) to carry out tasks the principal cannot perform directly due to geographic constraints, tactical limitations, or strategic risk (Horowitz, Reference Horowitz2010a). The alliance constitutes a form of delegation, “the process by which the principal offers a conditional grant of authority to an agent to act on their behalf” (Byman and Kreps, Reference Byman and Kreps2010, p. 3). The principal thus invests in the agent’s capacity, expecting that the agent will pursue objectives aligned with its own.
3.1. Why delegate?
Delegation serves as a cost-saving device in war (Salehyan, Reference Salehyan2011). Militant principals often lack the time, or task-specific expertise to carry out every operational objective they might have, prompting them to assign responsibilities to agents better suited for certain tasks. In other words, a principal delegates because another group holds a tactical niche. After all, “without some gains from specialization, there is little reason to delegate anything to anybody” (Hawkins et al., Reference Hawkins, Lake, Nielson, Tierney, Hawkins, Lake, Nielson and Tierney2006, p. 13). For instance, a group adept at conventional warfare might delegate Improvised Explosive Device (IED) bombings or assassinations to a smaller group with greater expertise or agility in these tactics. Delegation thus follows a logic of comparative advantage (Byman and Kreps, Reference Byman and Kreps2010) and enables principals to expand their tactical repertoire by leveraging the specialized strengths of others.
In self-determination conflicts that are almost by definition geographically confined to where an ethnic or identity group resides, comparative advantage also stems from differences in geographic embeddedness. For instance, a locally rooted agent may have better access, intelligence networks, or legitimacy in a particular area (Salehyan, Reference Salehyan2011). In these cases, delegation becomes a way to project influence into inaccessible or hostile spaces by relying on a better locally embedded partner with stronger reach in that locality. Therefore, delegation can also serve as a mechanism of spatial extension, allowing principals to operate across a broader geographic canvas without overstretching their own organization.
Finally, delegation may yield benefits beyond comparative advantage, namely, reputational insulation and plausible deniability. Some tactics like civilian-targeted violence or suicide bombings may offer strategic returns but carry significant reputational costs (Abrahms and Conrad, Reference Abrahms and Conrad2017). Principals may avoid undertaking such acts directly and instead delegate them to less reputationally sensitive agents (Byman and Kreps, Reference Byman and Kreps2010). This would help the principal avoid public backlash (Abrahms and Conrad, Reference Abrahms and Conrad2017) and punitive state responses (Onder, Reference Onder2023).
3.2. Why accept delegation?
Agents enter high-end cooperative relationships because receiving training, intelligence, or logistical support from a more capable principal can generate military, organizational, and political gains they cannot obtain on their own. Partnering with a militarily more capable organization provides access to tactical expertise, operational infrastructure, and intelligence networks that may substantially enhance the agent’s capacity to survive and operate.
First, in multiparty conflict ecosystems, weaker groups face threats from both the state and rival factions. A tightly knit alliance with a principal can serve as a deterrent against attacks from competing groups or state reprisals, enabling the agent to function with greater security and operational freedom (Morrow, Reference Morrow1991). Moreover, through high-end support, agents can gain access to capabilities, such as explosives training, that might otherwise remain out of reach.
Second, high-end cooperation can elevate the agent’s political standing. Access to specialized training and enhanced operational capacity can translate into more effective or more visible violent operations, which in turn can increase the agent’s credibility within the broader insurgent field. Such gains may improve recruitment prospects among radical constituencies, strengthen claims to represent their community, and enhance their bargaining position vis-à-vis other armed actors (Krause and Singer, Reference Krause, Singer, Reiter and Gärtner2001). Accepting delegation, thus, can help agents consolidate both military relevance and political legitimacy within a crowded conflict ecosystem.
3.3. The perils of delegation
Despite its strategic advantages, delegation in high-end militant cooperation exposes both principals and agents to serious risks. For principals, the classic concern is agency slack, where the agent shirks duties or pursues actions contrary to the principal’s goals. In extreme cases, the agent may even turn its newly gained resources against the principal itself (Salehyan, Reference Salehyan2011, p. 502). For agents, the danger takes the opposite form. Entering a high-end cooperative relationship with a militarily dominant principal can further consolidate the principal’s superiority within the conflict ecosystem, reinforcing rather than alleviating the agent’s subordinate position.
These risks are acute in militant alliances for three reasons. First, unlike state alliances, which can be formalized through binding contracts, militant partnerships are rarely codified due to their clandestine nature (Byman and Kreps, Reference Byman and Kreps2010; Mattes, Reference Mattes2012). This heightens informational asymmetries (Shapiro and Siegel, Reference Shapiro and Siegel2007) and lowers the barriers to shirking. Second, because high-end cooperation preserves the agent’s independent command structure, monitoring and enforcement mechanisms are absent (Byman and Kreps, Reference Byman and Kreps2010). This lack of oversight increases the likelihood that the agent will use its new capabilities to pursue its own goals rather than those of the principal.
Third, and most consequentially, high-end cooperation redistributes capabilities within the alliance, generating strategic vulnerabilities for both actors. For principals, enhancing the agent’s capabilities amplifies the danger that a newly empowered partner may become a future adversary. While similar concerns exist in state–militant proxy relations (Salehyan, Reference Salehyan2011), the problem is arguably more pronounced in militant-to-militant ties, where principals do not possess the monitoring and enforcement tools to restrain agents.
For agents, the same absence of monitoring and formal safeguards means they cannot constrain how the principal leverages the cooperation. Accepting high-end support gives the principal discretion over how much capability to transfer, when to transfer it, and how to use the resulting relationship. Without formal safeguards, agents have little ability to prevent a capability-wise dominant principal from using the cooperation to further expand its own agenda in ways that deepen the agent’s subordination within the conflict ecosystem. Thus, capability redistribution in either direction constitutes a military/strategic risk that both actors must weigh when entering high-end cooperative arrangements.
Beyond this military/strategic risk, high-end cooperation introduces an often-overlooked peril: political competition. This is largely absent from canonical principal–agent frameworks, for instance, those focused on state delegation to non-state actors, where differences in international status (sovereign actor vs. unrecognized, or proscribed non-state actor) reduce the risk of direct political rivalry. In militant alliances, however, transferring operational capabilities can also transfer political capital. Empowering an agent can reshape political hierarchies within the broader conflict ecosystem, altering how civilians, as potential providers of support to militants, evaluate the relative standing of each group.
For principals, the danger lies in elevating the agent’s political profile in the eyes of civilians. Armed groups can attract greater civilian support by demonstrating superior violence capacity (Kydd and Walter, Reference Kydd and Walter2006; Conrad and Greene, Reference Conrad and Greene2015; Belgioioso, Reference Belgioioso2018). Training, intelligence, and logistical support, and the consequent boost in violent capacity, can enhance the agent’s perceived competence, making it a credible alternative to the principal in the eyes of civilian supporters. This threat is especially consequential in self-determination conflicts when the agent is identity-wise substitutable with the principal, that is, when both groups appeal to overlapping constituencies or draw from the same support base (Pischedda, Reference Pischedda2018, Reference Pischedda2020; Phillips, Reference Phillips2019). In such cases, the agent’s political gains may come at the principal’s expense, potentially siphoning the principal’s supporters, volunteers, informants, and recruits. The principal may thus be cultivating a political rival.
For agents, the political risk takes the opposite form. Entering a high-end alliance with a dominant principal may reinforce the principal’s political preeminence within the conflict ecosystem. If civilians view the principal as the true source of operational capacity, cooperation can lock the agent into a subordinate role, narrowing its independent political appeal and limiting its ability to develop an autonomous support base. In this sense, the same cooperation that risks elevating the agent above the principal also carries the risk—for the agent—of entrenching a political hierarchy that favors the principal.
4. Partner selection in high-end cooperation
In canonical principal–agent theories, the standard response to agency slack is to design enforcement mechanisms that constrain agent behavior. As discussed above, such solutions are largely unworkable in militant alliances. Instead, I argue that militant principals and agents engaging in high-end cooperation rely on anticipatory strategies when selecting their partners. Principals strategically choose whom to empower in ways that maximize potential gains while mitigating the perils of delegation by hedging against downstream rivalry. Agents strategically accept such arrangements when the transferred capabilities enhance their own survival prospects and political standing.
In what follows, I first introduce a plausible empirical assumption check to validate the logic underlying this principal–agent framework. I then outline two testable hypotheses about partner selection, reflecting strategic calculations related to tactical complementarity and political differentiation, each serving to maximize both the principal’s and the agent’s gains while insulating them from the political risks of delegation.
A core implication of this principal–agent framework is that high-end cooperation presupposes an organizational capability differential that underpins the logic of delegation. The principal must possess operational skills, infrastructure, or specialized know-how that the agent lacks. It also clarifies why symmetric cooperation is especially perilous when sensitive knowledge, intelligence, or operational infrastructure is exchanged. If two groups are evenly matched in their capabilities, transferring high-end capabilities risks destabilizing the balance of power within the dyad, potentially leaving the original principal disadvantaged should the alliance collapse in the future (Morrow, Reference Morrow1991; Mattes, Reference Mattes2012). High-end cooperation, thus, should empirically manifest as cooperation between more capable organizations and less capable ones. To validate this assumption of the theory, I verify whether cooperating dyads exhibit the expected organizational capacity differential.
Organizational Capacity Differential Proposition: High-end cooperation occurs in dyads with a greater organizational capacity differential.
One way to maximize the gains from high-end cooperation is through tactical complementarity: selecting partners whose capabilities differ from, but complement, one another. This follows directly from the comparative-advantage logic discussed earlier; that is, groups benefit when they combine distinct operational strengths rather than duplicate existing ones (Byman and Kreps, Reference Byman and Kreps2010). Research on inter-organizational alliances shows that partnerships emerge when actors pool dissimilar yet mutually beneficial assets, such as technological innovation and market access in inter-firm alliances (Soda and Furlotti, Reference Soda and Furlotti2017), or distinct military assets and geostrategic positioning in inter-state alliances (Murdoch and Sandler, Reference Murdoch and Sandler1984).
This logic should apply to militant alliances. Armed organizations vary in their violent tactical repertoires and develop expertise at certain tasks (Horowitz, Reference Horowitz2010b, p. 39). Some excel in conventional warfare, others in guerrilla tactics or urban terrorism. Tactical complementarity suggests that, for example, a rural guerrilla group that lacks urban access might benefit from cooperating with a clandestine urban cell skilled in suicide bombings. Delegation, then, becomes a tool for expanding tactical diversity through cooperative outsourcing. For principals, partnering with agents whose attack portfolios diverge from their own expands their tactical repertoire without having to develop those skills in-house. For agents, accepting high-end support from a tactically distinct principal provides access to violent skills they are unlikely to develop independently. Hence,
Complementarity Hypothesis: The greater the dissimilarity in the attack profiles of two militant groups, the more likely they are to engage in high-end cooperation.
Beyond maximizing the tactical diversification gains from high-end cooperation, both principals and agents must manage the political risks of delegation by hedging against downstream rivalry. High-end cooperation can enhance an agent’s operational capacity and, by extension, its political stature within the broader militant landscape. These gains plausibly extend beyond the battlefield. When a principal empowers an agent, this may also raise the agent’s profile among civilian audiences, increasing its appeal as a potential political rival. This risk is particularly acute when both groups appeal to the same ethnic, religious, or ideological constituency. When one’s ally can function as a “substitute” for one’s organization (i.e., when fighters could plausibly join the ally after abandoning their current group (Hauenstein, Reference Hauenstein2023)), it can undermine the principal’s political position by co-opting its social support base (Pischedda, Reference Pischedda2020).
Recent scholarship argues that the cooperation-inducing effects of ideational alignment are offset by competitive dynamics among organizations with compatible political aspirations (Balcells et al., Reference Balcells, Chen and Pischedda2022). Groups with compatible aspirations can more easily co-opt each other’s support bases (Pischedda, Reference Pischedda2020). Cooperating with a potential competitor that can prey on one’s social support base can threaten one’s prospects for political success (Onder, Reference Onder2025). For principals, providing high-end support to an organization that draws on the same constituency heightens the danger of empowering a political competitor. Consequently, principals should prefer partners that appeal to distinct constituencies, thereby leveraging the benefits of cooperation without eroding their own political standing. Agents should likewise be more willing to accept high-end support from principals with whom they do not share a constituency, since, compared to accepting support from a principal that shares their constituency, doing so reduces the risk of becoming politically overshadowed. When constituencies differ, both parties can reap the benefits of cooperation without directly threatening one another’s political base.
Political Differentiation Hypothesis: High-end cooperation is less likely among militant groups that appeal to the same civilian constituency than among those that appeal to distinct constituencies.
4.1. Scope conditions
This theory of high-end cooperation applies most directly to militant alliances formed in conflict ecosystems featuring self-determination groups, where the principal retains autonomy over alliance decisions. Three scope conditions follow. First, in conflict ecosystems where militant organizations vie for control of the central state, the incentives and risks surrounding alliance formation may differ. As Christia (Reference Christia2012) demonstrates, for instance, ethnic or ideological affinities played little decisive role in shaping coalition dynamics among Afghan Mujahideen factions.
Second, the theory assumes the absence of an overriding principal in a “dual delegation” arrangement (Karlén and Rauta, Reference Karlén and Rauta2023). When a militant group is subordinate to an external state sponsor, it may be pressured to form alliances based on its state patron’s preferences rather than its own strategic calculus.
Finally, the theory assumes that principal and agent are embedded in a shared domestic conflict, where principals and agents vie for resources, audiences, and symbolic leadership in overlapping geographic spaces, and therefore downstream political competition constrains alliance formation. Delegation in transnational militant networks, such as al-Qaeda’s or the Islamic State’s relationships with ideologically proximate groups in foreign theaters, likely reflects a logic of franchising (Mendelsohn, Reference Mendelsohn2015), where the transnational principal faces less risk of political competition from local agents.
5. Research design
I employ Temporal Exponential Random Graph Models (TERGMs) on an original, disaggregated, time-series network dataset of 53 ethnonationalist militant groups active in Northeast India from 1981 through 2021. This dataset was constructed to address key limitations in existing publicly available data. Existing datasets tend to overlook smaller, weaker, and less lethal actors.Footnote 2 These groups may not produce high-casualty events but can still shape conflict dynamics through their cooperation with larger organizations.Footnote 3
Existing datasets also frequently over-aggregate different forms of cooperation. My dataset addresses this by coding eight discrete forms of cooperation. While some recent efforts, such as the Militant Group Alliances and Rivalries dataset (Blair et al., Reference Blair, Chenoweth, Horowitz, Perkoski and Potter2022), distinguish between categories such as financial, material, training, and operational support, they are not directional, and hence not suitable for testing implications derived from a principal–agent theory. In contrast, my dataset is directional; for each dyadic tie, I identify the sender and receiver of support.
Northeast India offers a fertile empirical setting for this analysis. Its long history of militant activity, ethnic fragmentation, rugged terrain, and porous international borders has given rise to a dense web of inter-group relations. The region is home to dozens of insurgent groups pursuing self-determination claims and regularly interacting with one another, providing natural variation in the depth of cooperation. Combined with rich variation in group capabilities and claimed constituencies, the region creates ideal conditions for evaluating the observable implications of my theory.
5.1. Data collection
I consider all armed non-state actors, including those typically categorized as rebels, insurgents, or terrorists. By adopting an inclusive approach, the dataset offers a more comprehensive view of inter-group militant relations than studies that restrict attention to either rebel groups (Bapat and Bond, Reference Bapat and Bond2012; Christia, Reference Christia2012) or terrorist organizations (Asal and Rethemeyer, Reference Asal and Rethemeyer2008).
To mitigate actor-selection bias, the dataset encompasses 53 groups, at least half of which have not been covered in previous network data collection efforts. These organizations were engaged in conflicts over independence, autonomy, or broader self-determination claims across the seven Northeastern states of India. The initial list of actors was drawn from the 26 Northeast Indian groups identified in the Uppsala Conflict Data Program/Peace Research Institute Oslo (UCDP/PRIO) Armed Conflict Dataset, version 21.1. Coders then gathered data on the basic characteristics of these groups from secondary sources. As these sources were reviewed, additional groups mentioned in relation to the UCDP/PRIO actors were recorded. Cross-referencing and triangulation were used to verify the existence of previously unlisted groups. The final inclusion criteria required that groups (a) operated in Northeast India between 1981 and 2021; (b) publicly identified themselves with a group name; and (c) used armed force in pursuit of a political goal.Footnote 4 Supplementary Appendix 2 includes the group list.
The dataset captures temporal variation in militant networks across four decades, allowing me to track actor entry and exit, the evolution of ties, and time-variant predictors of cooperation (Dorff et al., Reference Dorff, Gallop and Minhas2020). It identifies eight distinct forms of cooperation: (1) joint operations, (2) training support, (3) provision of arms and funds, (4) intelligence-sharing and logistical support, (5) joint planning or leadership meetings, (6) joint public statements, (7) participation in umbrella organizations, and (8) rhetorical support.Footnote 5 The codebook in Supplementary Appendix 1 details the coding rules.
Because the theoretical framework centers on strategic delegation, it is essential to distinguish between senders and receivers of support. Accordingly, for each recorded instance of cooperation in the dataset, I identified which group offered the support and which group received it. This allows for empirical evaluation of the theorized principal–agent dynamics. The resulting dataset includes 24,944 directed dyad-years.
5.2. Empirical strategy
I analyze a directed temporal network of high-end cooperation among 53 militant groups. This network is constructed for every year between 1981 and 2021 and captures annual ties in which one group provides another with training, intelligence, or logistical support.Footnote 6 A directed tie is recorded in a given year if Group A provided such support to Group B, with directionality identifying the principal (sender) and the agent (receiver). In total, 40 of the 53 groups (75%) engaged in at least one instance of high-end cooperation over the 40-year period.
Figure 1 illustrates the temporal evolution of the high-end cooperation network. Early on, the network is sparsely connected, with only a handful of dyads engaging in cooperation. From the early 2000s onward, the network densifies, reflecting an expansion in the number of groups participating in cooperation. A core-periphery structure emerges, in which a subset of groups serves as central hubs providing training, intelligence, and logistical assistance to multiple partners.
Evolution of the high-end cooperation network.
Figure 1 Long description
The image contains a grid of network plots, each labeled with a year ranging from 1982 to 2021. Each plot displays a circular arrangement of hollow nodes along the outer edge, with a smaller number of filled nodes connected by lines positioned toward the center or interior of the circle. In the earliest years, from 1982 through the late 1980s, only one or two filled nodes appear, with very few or no connecting lines between them. The network is sparse, with most nodes remaining hollow and unconnected. From the early 1990s onward, the number of filled nodes increases gradually. Short connecting lines begin to appear between a small cluster of filled nodes. By the mid-1990s, a small but visible grouping of connected filled nodes is present near the center of several plots. From 2000 onward, the number of filled nodes and the number of connecting lines between them grows more noticeably with each successive year. By the mid-2000s, a denser cluster of interconnected filled nodes is visible, with some nodes connecting to multiple others. From 2010 onward, the cluster of filled and connected nodes becomes more prominent. By 2015 to 2021, the connected structure occupies a larger portion of the interior space, with several nodes linking to multiple neighbors, forming a more complex connected arrangement. The hollow nodes arranged along the outer ring remain largely unconnected throughout the full timeline. The filled nodes and their connecting lines are concentrated in the interior region of each plot. No axis labels or units are present in any of the individual plots. No legend is visible within the image.
To illustrate the distinct network structure that emerges under high-end cooperation, Figure 2 shows the temporally aggregated visualizations of three networks: high-end cooperation (training, intelligence, and logistical support), transactional cooperation (arms-for-cash exchanges), and rhetorical support. The high-end cooperation network is relatively denser and more centralized. While the transactional cooperation network also exhibits some centralization, it is less densely connected and involves greater interaction among peripheral actors. The rhetorical support network is more diffuse still, with at least two separate hubs and no dominant central actors.
A network comparison of different forms of cooperation. (a) High-end cooperation. (b) Transactional cooperation. (c) Rhetorical cooperation.
Figure 2 Long description
The image shows three network visualizations labeled as (a), (b) and (c). Each network consists of nodes connected by lines, representing different forms of cooperation. Image (a) depicts a dense and centralized network, likely representing high-end cooperation. Image (b) shows a less dense network with some centralization, indicating transactional cooperation. Image (c) illustrates a more diffuse network with separate hubs, representing rhetorical support. The size of the nodes varies, suggesting different levels of influence or connectivity within each network.
Conventional statistical techniques rely on the assumption that dyad-year observations are independent, for instance, that Group A’s decision to support Group B has no bearing on its relationship with Group C. This assumption is problematic in alliance networks, where tie-formation decisions are often shaped by broader relational structures (Cranmer et al., Reference Cranmer, Desmarais and Kirkland2012). To address this, I employ TERGMs, which extend the ERGM framework to handle longitudinal network data (Leifeld et al., Reference Leifeld, Cranmer and Desmarais2018).
Compared to alternative approaches, such as latent space models, which capture dependence through latent positions in an abstract space, TERGMs define network statistics globally and explicitly represent structural dependencies like popularity and transitivity (Block et al., Reference Block, Stadtfeld and Snijders2019). Stochastic actor-oriented models, on the other hand, simulate the sequential decision-making of individual actors. The theory developed in this study, by contrast, is explicitly relational (e.g., dyadic characteristics such as complementarity and constituency overlap). TERGMs are better suited for testing hypotheses grounded in a dyadic tie-oriented model of alliance formation than actor-centered approaches (Block et al., Reference Block, Stadtfeld and Snijders2019).
5.3. Key predictors
Three key dyadic predictors are organizational capacity differential, attack portfolio dissimilarity (Complementarity Hypothesis), and shared constituency (Political Differentiation Hypothesis).
Organizational capacity differential operationalizes the assumption check derived from the principal–agent logic advanced earlier. Following George (Reference George2018), I use the logistical complexity of attacks as a proxy for organizational capacity. Specifically, I code assassinations, bombings, hijackings, and hostage-takings as logistically complex.Footnote 7 Using the Global Terrorism Database (GTD)’s attacktype variables, I compute the weighted proportion of a group’s logistically complex attacks in each year. This generates a continuous, time-varying indicator of capacity ranging from 0 to 1. Organizational capacity differential is measured as the absolute difference between two groups’ scores in a dyad. Higher values indicate greater capability differentials.
Attack portfolio dissimilarity operationalizes the logic of the Complementarity Hypothesis. To capture variation in groups’ tactical repertoires, I draw on the GTD’s attacktype variables. For each group-year, I calculate the proportion of attacks falling into each of the GTD’s eight predefined categories: armed assault, assassination, bombing/explosion, facility/infrastructure attack, hijacking, hostage-taking (barricade), hostage-taking (kidnapping), and unarmed assault. I then use a maximum-likelihood factor analysis to reduce these proportions to a unidimensional portfolio score per group-year. Attack portfolio dissimilarity measures the absolute difference between two groups’ scores in a dyad. Higher values indicate greater tactical dissimilarity, and thus greater complementarity.
Shared constituency captures the political risks highlighted in the Political Differentiation Hypothesis. Drawing on secondary sources, I identify the ethnic, tribal, or religious community each group claims to represent, using the definition of constituency as “the broad social group on whose behalf [militants] claim to fight” (Balcells et al., Reference Balcells, Chen and Pischedda2022). Shared constituency is a dichotomous variable coded 1 if both groups in a dyad claim to fight on behalf of the same ethnic/tribal or religious group, and 0 otherwise.
To account for endogenous features of the network, I incorporate two structural dependency terms in all TERGM specifications.Footnote 8 First, sender popularity is modeled using the ostar statistic,Footnote 9 which captures the tendency for well-connected actors to attract additional ties. This reflects reputational dynamics in militant alliance formation, where groups with visible histories of cooperation may be more attractive partners (Cranmer et al., Reference Cranmer, Desmarais and Kirkland2012; Bacon, Reference Bacon2017, Reference Bacon2018). Second, transitivity is modeled using the gwesp term (geometrically weighted edgewise shared partners), with a decay parameter of 1.Footnote 10 This term captures the likelihood of triadic closure (i.e., that groups with common allies are more likely to cooperate directly), reflecting alliance patterns shaped by shared enemies or trusted third-party brokers (Asal et al., Reference Asal, Park, Rethemeyer and Ackerman2016; Bacon, Reference Bacon2017).
5.4. Accounting for confounding
I include a set of control variables at both the group- and dyad-levels. At the group-level, I include a time-varying measure of groups’ capacity, constructed identically to the specification used for computing organizational capacity differential. Second, I include a binary indicator for whether a group is a splinter group, derived from my original data collection. Splinter factions (e.g., groups that emerged through organizational rupture from preexisting militant actors) may be seen as less credible alliance partners due to their history of defection.
At the dyadic level, I control for whether groups in a dyad share a common foreign state sponsor. Shared patrons can help mitigate alliance breakdowns by serving as external guarantors (Bapat and Bond, Reference Bapat and Bond2012; Popovic, Reference Popovic2018). This binary variable, derived from my original data collection, is coded 1 if both organizations received material backing from the same state sponsor, and 0 otherwise.
I include two controls for non-delegatory cooperation within the dyad: transactional and rhetorical cooperation, both derived from my original data collection. Transactional cooperation is a binary variable coded 1 if the dyad engaged in arms-for-cash exchanges during a given year. Rhetorical cooperation is a binary variable coded 1 if one group publicly praised, or declared loyalty to the other in the same year. These controls account for the possibility that other forms of cooperation are correlated with high-end cooperation, either as precursors or complementary activities.
6. Results
Table 1 reports the results from TERGMs, each modeling the likelihood of a directed high-end cooperation tie between a pair of militant organizations in a given year. Model 1 presents a naïve specification. Model 2 adds structural network dependency terms. Model 3 introduces a set of control variables capturing group-level traits and dyadic features. Finally, Model 4 adds controls for other forms of cooperation (e.g., transactional and rhetorical). Across all four specifications, the three key dyadic predictors, organizational capacity differential, attack portfolio dissimilarity, and shared constituency, are statistically significant at conventional levels and in the expected directions.
TERGMs of high-end cooperation among militant groups in Northeast India, 1981–2021
Table 1 Long description
The table reports four statistical models estimating factors associated with high-end cooperation among militant groups in Northeast India from 1981 to 2021, with coefficients shown alongside uncertainty intervals and significance markers. Across all four models, organizational capacity differential is positive and statistically significant, with estimates around 0.40 to 0.43. Attack portfolio dissimilarity is also consistently positive and significant, around 0.20 to 0.25. Shared constituency is consistently negative and significant, roughly minus 0.48 to minus 0.59, indicating lower cooperation when groups draw on the same constituency. Sender capability is negative and becomes statistically significant in the models that include network dependency and controls, while receiver capability is negative and significant in all models. When network terms are included, transitivity and sender node popularity are both positive and significant, suggesting cooperation is more likely in clustered networks and for more popular senders. Additional controls show transactional cooperation has a large positive and significant association, while rhetorical support is not statistically significant and joint foreign supporter has very wide uncertainty. The baseline edges term is negative and significant in every model, and all models use 24,944 observations; results should be interpreted with attention to the reported uncertainty intervals and model specification differences.
Notes: Bootstrapped pseudolikelihood estimates, as described in Desmarais and Cranmer (Reference Desmarais and Cranmer2012), reported. Temporal bootstrapping is used to correct the standard errors. Standard errors are based on 1000 network-year bootstrap iterations. Asterisks indicate that the coefficient is statistically significant at or beyond the traditional 0.05 level.
To begin with, the results support the theoretical expectation embedded in the Organizational Capacity Differential Proposition. The coefficient for organizational capacity differential is positive and significant across all models. High-end cooperation is more likely in dyads characterized by greater disparities in organizational capacity. This is consistent with a principal–agent framework: high-end cooperation rests on a capability differential logic that makes delegation meaningful, and empirically, ties do indeed form between more capable groups and less capable partners, which confirms an important assumption underlying my theory.
The results further support the two hypotheses of the study. First, the hypothesis that principals are more likely to pursue high-end cooperation with agents who possess distinct tactical specializations (Complementarity Hypothesis) is supported. The coefficient for attack portfolio dissimilarity is positive and statistically significant across all four model specifications. As the difference between groups’ tactical repertoires increases, so does the probability of high-end cooperation. This finding is consistent with the argument that principals and agents both benefit when they combine distinct violent skillsets, allowing each actor to expand operational capabilities without developing new tactics internally.
Empirical patterns of cooperation are also consistent with the Political Differentiation Hypothesis. Across specifications, the coefficient for shared constituency is negative and statistically significant. Groups that claim to represent the same ethnic/tribal or religious civilian base are less likely to engage in high-end cooperation. This aligns with the idea that both principals and agents seek to avoid political arrangements that heighten political substitutability, increase downstream rivalry, or risk political overshadowing.
To illustrate the substantive implications of the three predictors, Figure 3 plots the predicted probabilities for high-end cooperation ties across dyad-years, derived from the second TERGM in Table 1. Panel A shows density plots of the distribution of predicted probabilities across different values of key predictors. Across the top two plots, a clear rightward shift in the density curves for high values (orange) relative to low values (blue) is visible. This suggests that dyads with high attack portfolio dissimilarity and high organizational capacity differential (orange densities) are associated with greater predicted probabilities of high-end cooperation than dyads with low dissimilarity or symmetry (blue densities). The bottom plot shows the reverse pattern: dyads that do not share a constituency (orange) are more likely to high-end cooperate than those that do (blue), supporting the Political Differentiation Hypothesis.
Predicted probabilities of high-end cooperation. (a) Distribution of probabilities. (b) Micro-level interpretation.
Figure 3 Long description
The image A showing three density plots of predicted probability by group. First density plot text: Org. Capacity Diff. Legend text: High, Low. Horizontal axis label: Predicted Probability. Horizontal axis tick labels: 0.00, 0.01, 0.02, 0.03, 0.04, 0.05. Vertical axis label: Density. Vertical axis tick labels: 0, 100, 200. The density for Low is concentrated near 0.00 to 0.01 with two narrow peaks. The density for High extends farther along the horizontal axis, with visible mass beyond 0.01 and a long tail reaching toward 0.05. Second density plot text: Portfolio Dissimilarity. Legend text: High, Low. Horizontal axis label: Predicted Probability. Horizontal axis tick labels: 0.00, 0.01, 0.02, 0.03, 0.04, 0.05. Vertical axis label: Density. Vertical axis tick labels: 0, 100, 200. The density for Low is concentrated near 0.00 to 0.01 with two narrow peaks. The density for High extends farther along the horizontal axis, with visible mass beyond 0.01 and a long tail reaching toward 0.05. Third density plot text: Shared Constituency. Legend text: No, Yes. Horizontal axis label: Predicted Probability. Horizontal axis tick labels: 0.00, 0.01, 0.02, 0.03, 0.04, 0.05. Vertical axis label: Density. Vertical axis tick labels: 0, 100, 200. The density for Yes is concentrated near 0.00 to 0.01 with two narrow peaks. The density for No is shifted to higher predicted probability values, with a peak around the 0.01 region and a tail extending toward 0.05. The image B showing three point plots of predicted probability with interval lines. First point plot title text: Organizational Capacity Diff. Text: Thick equals 95 percent CI. Text: Thin equals 99 percent CI. Horizontal axis tick labels: Low, High. Vertical axis label: Predicted Probability. Vertical axis tick labels: 0.000, 0.005, 0.010, 0.015, 0.020, 0.025, 0.030. The point at Low is near 0.010. The point at High is near 0.020. Both points have vertical interval lines. Second point plot title text: Attack Portfolio Dissimilarity. Text: Thick equals 95 percent CI. Text: Thin equals 99 percent CI. Horizontal axis tick labels: Low, High. Vertical axis label: Predicted Probability. Vertical axis tick labels: 0.000, 0.005, 0.010, 0.015, 0.020, 0.025, 0.030. The point at Low is near 0.010. The point at High is near 0.020. Both points have vertical interval lines. Third point plot title text: Shared Constituency. Text: Thick equals 95 percent CI. Text: Thin equals 99 percent CI. Horizontal axis tick labels: Yes, No. Vertical axis label: Predicted Probability. Vertical axis tick labels: 0.000, 0.005, 0.010, 0.015, 0.020, 0.025, 0.030. The point at Yes is near 0.010. The point at No is near 0.018. Both points have vertical interval lines.
Panel B adopts the interpretive strategy proposed by Leifeld, Cranmer, and Desmarais (Reference Leifeld, Cranmer and Desmarais2018) for analyzing TERGM predictions at the micro-level. For each dyad-year in the sample, I compute predicted probabilities of high-end cooperation, stratify dyads by unique values of the three key predictors, and calculate the median predicted probability and associated confidence intervals within each stratum. The results are consistent: dyads with high attack portfolio dissimilarity and high organizational capacity differential are substantially more likely to engage in high-end cooperation, with predicted probabilities that are nearly double those of their low-dissimilarity or low-asymmetry counterparts. Dyads that share a civilian constituency are considerably less likely to form ties, consistent with my political rivalry mechanism.
6.1. Additional analyses
I conduct several additional analyses. First, I assess the endogenous goodness-of-fit (GOF) of the TERGMs. I simulate 100 networks at each time step and evaluate how well these simulated networks reproduce key structural properties of the observed network, such as degree distributions. The GOF assessments, visualized in Supplementary Appendix 4, suggest that the models fit the observed data well.
Second, I consider additional group-level covariates that may shape both a principal’s and an agent’s willingness to enter into high-end cooperative arrangements: a count of surrender events by each group to the government side, and binary indicators for whether a group operates in a region with an international border and espouse a leftist or religious ideology. All variables are derived from my original data collection. As shown in Supplementary Appendix 5, the main findings hold.
Third, to rule out the possibility that the main results reflect broader alliance formation tendencies rather than patterns specific to high-end cooperation, I conduct placebo tests using alternative network outcomes. I re-estimate TERGMs on networks of transactional cooperation (arms-for-cash exchanges) and rhetorical cooperation. As shown in Supplementary Appendix 6, these placebo models return null results. None of the key predictors are in the same direction or statistically significant in the rhetorical cooperation network, and only attack portfolio dissimilarity is significant in the transactional network. Moreover, the direction of some coefficients, such as organizational capacity differential, also reverses in the transactional cooperation models.
Finally, I conduct a battery of robustness checks using a wide array of network metrics. As reported in Supplementary Appendices 7 through 11, the main results are robust.
7. Discussion and conclusion
This study developed a theory of high-end cooperation rooted in the logic of principal–agent delegation to reorient existing understandings of militant alliances by showing how fragmentation in the militant field shapes not only why and with whom groups cooperate but also how.
The empirical results strongly support the theory’s core expectations. The empirical assumption that underpins the principal–agent logic is supported: ties overwhelmingly form between more capable and less capable actors. As for the hypotheses I advance, dyads characterized by high attack portfolio dissimilarity are significantly more likely to engage in high-end cooperation (i.e., complementarity). Finally, cooperation is less likely between groups that appeal to the same civilian constituency, consistent with the expectation that political substitutability heightens downstream rivalry. Moreover, placebo tests relying on transactional and rhetorical cooperation suggest these calculations are specific to high-end cooperation.
These findings suggest that high-stakes cooperation among armed groups is the outcome of strategic design. Principals construct alliances in a way that minimizes threats to their dominant position, and agents accept high-end support when doing so does not risk political overshadowing by a dominant partner. This has several broader implications for the study of civil war. First, it challenges the assumption that cooperation among armed groups requires ideological alignment, demonstrating instead that tactical complementarity and political non-substitutability are the key drivers of high-end cooperation. Second, while much of the literature treats civilian support as a function of a group’s behavior in isolation, my theory implicates that armed actor legitimacy is also relational: it is being challenged and renegotiated as armed groups cooperate, delegate, and compete. This opens new directions for thinking about rebel political legitimacy as a positional achievement. Third, this study suggests that shifts in the distribution of capabilities, even absent ideological splits, can destabilize conflict systems, offering a novel account of actor fragmentation in civil wars rooted in shifting power dynamics rather than preference divergence. Finally, it pushes the literature on delegation in civil wars beyond the conventional state–proxy dyad, calling for more attention to the inter-rebel forms of principal–agent delegation.
There are, nonetheless, caveats to be mentioned. First, like most work on armed groups, the analysis relies on secondary reporting, which may underreport less-central actors. Second, although TERGMs address network dependencies, the analysis remains observational and cannot fully eliminate the possibility of omitted variable bias or endogeneity. Finally, the geographic and organizational scope of the analysis focusing on ethnonationalist actors in Northeast India may limit the generalizability of the results to conflict systems dominated by Islamist militant groups, such as those in Pakistan, Afghanistan, Syria, or Nigeria, where well-resourced transnational jihadist actors seem to exert outsized influence over alliance networks.
Despite these limitations, the study offers both empirical and substantive contributions. Empirically, it introduces a new, directionally coded, longitudinal dataset that captures different forms of cooperation. Methodologically, it applies TERGMs to model network dependencies, something rarely done in the literature on armed group alliances. Substantively, it reframes how scholars understand militant cooperation: as a problem of delegation and risk management. Future research can build on these insights by examining how external interventions, militant leadership changes, or state repression reshape the delegation calculus over time.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/psrm.2026.10110.
Acknowledgements
I thank James A. Piazza, Cyanne E. Loyle, Douglas Lemke, Ore Koren, and Shahryar Minhas for detailed comments, and the participants of ISA 2024 Early Career Workshop and ISA 2025 Networks in Peace and Conflict panel for helpful feedback.
Funding statement
This work was supported by dissertation support grants from The Pennsylvania State University, the Minerva Research Initiative, and the United States Institute of Peace.
Open access
