The Risks and Rewards of Cooperation

Armed groups exist in acutely anarchic environments.Footnote ⁸ They can appeal to no authority, they are regularly defeated,Footnote ⁹ and their members are frequently killed. As a result, despite their typically long-term political objectives, they tend to operate with relatively short time horizonsFootnote ¹⁰ and overwhelmingly prioritize their own security.Footnote ¹¹ This immediacy has a profound impact on their behavior and results in competition for recruits, funding, and notoriety.Footnote ¹² It also renders organizations hesitant about any action, including cooperation with other groups, that exposes them to additional risk.

The risks of cooperation stem, in part, from the potential for the infiltration of or defection by alliance partners.Footnote ¹³ For instance, the alliance between the Revolutionary Armed Forces of Colombia (FARC) and Sendero Luminoso broke down in 2011 after documents seized in raids on FARC bases contributed to the capture of FARC commander Alfonso Cano and Sendero Luminoso commander Comrade Artemio.Footnote ¹⁴

The risk of defection and security breaches is exacerbated by the weak mechanisms for enforcement and monitoring that persist among militant organizations. These organizations cannot create institutions in the same way as states and consequently have fewer means by which to verify compliance or punish partners for defection. While armed groups can formalize cooperation and require reporting or oversight, these steps exacerbate security risks.Footnote ¹⁵ For instance, in the prior example, documentation and formalized cooperation between FARC and Sendero Luminoso contributed to the security breach that ultimately undermined the alliance.

The challenges of cooperation do not end once a relationship is initiated. Alliances are difficult to sustain in the face of pressure from the state for at least three reasons. First, constant engagement against state security forces can preoccupy militants and exhaust resources, leading alliance obligations to go unfulfilled. Consider the alliance between the Red Brigades and the Armed Proletarian Nuclei (NAP). Their alliance collapsed in 1978 after Italian police decimated NAP in 1977, raiding at least thirty-nine safe houses, seizing weapons and operatives, and ultimately killing NAP leader Antonio Lo Musico.Footnote ¹⁶ As the police campaign expanded, NAP was forced to withhold operatives from planned operations with the Red Brigades, which contributed to the breakdown of the relationship.

Second, state repression can undermine alliances, particularly material alliances, by inhibiting groups’ ability to interact. In general, material cooperation requires communication, contact, and transactions between group members. When state pressure increases, it limits these necessities of alliance maintenance. For example, repression exacerbated communication problems between AQ and its affiliates, prompting one AQ leader to ask: “How can we correspond with brothers in Algeria, Iraq, Yemen and Somalia? Sometimes there is no other means after taking precautions.”Footnote ¹⁷

Third, repression reduces trust. Consistent losses force a reappraisal of security. Allies, as outsiders to an organization, are typically the first to come under suspicion. For instance, cooperation between AQ and the Haqqani Network was strained in 2010–11 as AQ commanders feared that Haqqani operatives were passing information to Pakistan's Inter-Services Intelligence.Footnote ¹⁸ No evidence has emerged that this actually occurred, but AQ's paranoia illustrates how state pressure can impede cooperation and breed dissent among allies.

Why then do armed groups cooperate at all? A substantial body of evidence suggests that they do so because, under some conditions, the benefits can outweigh the risks.Footnote ¹⁹ Cross-national studies confirm that alliances can augment armed groups’ capabilities, lethality, and resilience.Footnote ²⁰ Taken together, these benefits motivate groups to accept the inherent risks and uncertainty of alliance commitments.

The first wave of research on alliance formation examined the role of entrepreneurial group members and their personal relationships.Footnote ²¹ More recently, some researchers have shifted their attention from individuals to organizations.Footnote ²² Much of this work emphasizes commonalities between potential partners—in terms of ideology, tactical capabilities, ethnicity, alliance portfolios, or foreign sponsors—that serve as a foundation on which cooperation can build.Footnote ²³ These shared features demonstrate trustworthiness, credibility, and dedication; they can also improve monitoring. With security fears looming large, this can facilitate credible commitments.Footnote ²⁴

Understanding cooperation between militant groups thus requires explaining how groups commit themselves in the face of repression. One established answer is that shared state sponsors can prevent militant groups from unilaterally defecting, particularly when groups are weak or repressed.Footnote ²⁵ However, many groups—even weak groups—that lack a shared sponsor still form durable alliances. Indeed, we find that groups share a sponsor in just 4.2 percent of the 5,918 alliance-years in our data. This rarity of shared sponsorship suggests that militant groups must devise additional ways to solve commitment problems.

In response, we focus on the role of shared ideology as a vehicle for developing and maintaining the trust and commitment required for alliance formation and durability.Footnote ²⁶

Ideological Commonality and the Dynamics of Cooperation

Following most recent work in this area,Footnote ²⁷ we use Sanín and Wood's definition of ideology as a “more or less systematic set of ideas that includes the identification of a referent group (a class, ethnic, or other social group), an enunciation of the grievances or challenges that the group confronts, the identification of objectives on behalf of that group (political change—or defense against its threat), and a (perhaps vaguely defined) program of action.”Footnote ²⁸

Work in this area indicates that, for many organizations, ideologies can be overlapping and imprecise but are relatively static. Change in ideology is relatively rare, in part because it generally occurs over long time frames. For example, the ideology that governed Basque Fatherland and Freedom had both leftist and nationalist-separatist elements. While leftism was more central to the group's ideology in the late 1960s and early 1970s as the group was debating how to inspire a lucha de masas (mass proletariat struggle) to achieve Basque independence, leftist economic thought was integral to the group's platform in later years as well.Footnote ²⁹ Moreover, consistent with nearly all existing and cutting-edge work in this area,Footnote ³⁰ our coding of ideology is time-invariant.Footnote ³¹

Research indicates that while group ideology is sticky and deeply linked to organizational identity, it is not immutable. For example, during their conflict with the Russian state, Chechen militants shifted from a primarily nationalist orientation to one more inflected with Salafism.Footnote ³² This was, in part, strategic since it facilitated relationships with some of the most capable terrorist organizations in the world, but it was only possible because most Chechens were Muslims. Similarly, the handful of Palestinian organizations that have survived for long periods have, while consistently hewing to nationalist objectives, generally transitioned from a leftist ideological perspective to a more religiously oriented one. Again, this had substantial strategic value as the Cold War drew to a close and Islamist organizations came to the fore, but it also reflected the changing orientation of their constituency.

We hold that ideological commonality offers four benefits that make cooperation more likely and durable: it lengthens the shadow of the future, facilitates monitoring and enforcement, provides access to common authority structures, and enhances trust.

First, as noted by scholars such as Bapat and Bond, shared ideology lengthens the “shadow of the future” because groups in the same ideological circles are more likely to share other allies, communicate with common peers, and interact repeatedly over time.Footnote ³³ Jemaah Islamiya, for instance, worked with AQ to expand its operational reach beyond Indonesia to Singapore, Thailand, Malaysia, and the Philippines. Yet, Jemaah Islamiya and AQ did not cooperate in isolation; they were embedded within a broader jihadist network that included Abu Sayyaf, the Moro Islamic Liberation Front, Jamaah Anshurat Tauhid, and others.Footnote ³⁴ Similar webs also existed among European and Latin American far-left groups.Footnote ³⁵ These ideological networks are important because they raise the stakes of alliance commitment. News of defection or abrogation can quickly disseminate throughout the ideological network and undermine a group's reputation, undermining its chances of forging other alliances in the future.

Second, shared ideology can facilitate monitoring and enforcement by granting access to a wider community of nonmilitant co-ideologues. Whether diasporas,Footnote ³⁶ religious adherents, or simply foreign supporters, armed groups can leverage external supporters—who are typically under less immediate pressure from the state—to detect and sanction violations of alliance obligations. This works particularly well when nonmilitant co-ideologues are hard-liners who view the militant struggle as a mission to resist infringements on their core ideological values.Footnote ³⁷ For example, throughout the 1980s and 1990s, as the Kurdistan Workers’ Party (PKK) suffered setbacks amid a Turkish military crackdown, Kurdish populations in Germany, Belgium, and France engaged in advocacy and fundraising to sustain the PKK's relationship with its European affiliates.Footnote ³⁸ More recently, AQ leveraged the charity organizations of nonmilitant co-ideologues to transfer money to its affiliates when more direct methods became unavailable.

Third, shared ideology provides access to common authority and social structures that reduce uncertainty and promote cohesion.Footnote ³⁹ For instance, in Syria, fraught alliances between Ahrar al-Sham, Suqur al-Sham, Jabhat al-Nusra, and Jaish al-Islam were reinforced after Yusuf al-Qaradawi (considered “the world's most influential Sunni cleric”) publicly hailed Ahrar al-Sham efforts as “a legitimate jihad.”Footnote ⁴⁰ Although each of these groups received support from state sponsors, they did not share the same sponsor. Rather, Saudi Arabia, Qatar, and Turkey supported different factions, and largely sought to disrupt alliances between their sponsorees and groups supported by other states. Salafi clerics urged cohesion in spite of competition between state sponsors. This role for ideological authority figures is not unique to Salafist groups. Appeals to leading Leninist and Çayanist thinkers were important in sustaining the alliance between the Turkish People's Liberation Front (THKP-C), the Mahir Çayan Suicide Group, and the Revolutionary Youth Federation in Turkey.Footnote ⁴¹ Ideological authorities can also incentivize cooperation by heralding partnership with groups deemed ideologically pure and sanctioning groups deemed illegitimate. These labels carry reputational costs that subsequently affect groups’ ability to attract recruits, financing, and political support.Footnote ⁴²

Fourth, when two groups are ideologically aligned, their common ideals and values enhance mutual trust.Footnote ⁴³ Ideological congruence makes it easier for groups to have confidence in the motivations of potential and existing partners. This, combined with the other benefits described earlier, suggests that groups sharing ideologies have unique advantages that facilitate alliance commitment. This leads to our first expectation:

Once cooperation takes hold between militant organizations, it is frequently targeted and challenged by states. Shared ideology facilitates sustained cooperation in the face of state repression through these same four mechanisms, and it is in this context that its effect should be most clearly visible. Indeed, as ideologically aligned organizations perceive a greater threat, their monitoring, enforcement, and commitment may even strengthen in response because research shows that in-group resolve increases in response to out-group pressure.Footnote ⁴⁴ In cases of cross-ideological collaboration, by contrast, cooperation under repression becomes increasingly difficult to sustain. Thus, we expect that:

The Critical Role of Religion

Not all ideologies are equivalent. We have argued that shared ideology offers relationship benefits because it lengthens the shadow of the future, facilitates monitoring and enforcement, provides access to common authority structures, and enhances trust. While most shared ideologies have some combination of these attributes, shared religion is particularly robust on these dimensions when compared to alternatives such as leftism or nationalism.Footnote ⁴⁵

Networks of religiously motivated organizations are some of the most established and wide reaching. This makes the shadow of the future even more consequential by exacerbating the negative consequences of alliance defection. These well-developed religious networks also make it more likely that groups can obtain useful information on alliance partners even before collaborating.

Shared religion also facilitates commitment because external religious constituencies consist of large, devoted audiences and developed, robust authority structures. This means that religiously oriented militant organizations have more nonmilitant co-ideologues than other groups, and hence have more potential sympathizers who can reward cooperation, detect shirking, and punish defection. Perhaps most significantly, religions can provide access to nonmilitant authority structures that are resistant to state repression.Footnote ⁴⁶ As Walter argues, the role of religious authorities is central to commitment because these figures hold cheap means of rewarding cooperation and punishing defection.Footnote ⁴⁷

For example, some religious figures leveraged promises of holy reward and punishment to help smooth over alliance tensions between AQ and al-Qaeda in Iraq, as well as various Salafist factions in the Syrian civil war. Even in Northern Ireland, preachers like Ian Paisley helped cement cooperation between loyalist groups like the Ulster Volunteer Force and the Ulster Defence Association.Footnote ⁴⁸ Strengthening this effect, states tend to hesitate to crack down on religious leaders and organizations, even those with apparent ties to militancy, for fear of alienating the broader public and escalating a conflict.Footnote ⁴⁹

Of course, while religion—and specifically extremist variants of Salafism—is the dominant shared ideology among militant groups today, this was not always the case. Leftist organizations were ubiquitous throughout the 1970s and 1980s, drawing on a large, devoted, international mass of adherents at the ideology's peak during the Cold War. Consequently, their networks may have similarly benefited from some of the mechanisms described here. For instance, communist parties and activists around the world helped coordinate transnational revolutionary efforts. Likewise, leftism, more so than other nonreligious ideologies like ethnonationalism or right-wing extremism,Footnote ⁵⁰ featured clear authority structures. Theorists and revolutionaries such as Marx, Lenin, Trotsky, Mao, Castro, and Guevara formed focal points through which militant groups could claim ideological credibility.Footnote ⁵¹ These individuals facilitated commitment because they could decry defection as deviation from the proper, revolutionary path. For example, Turkish communist intellectual Mahir Cayan, founder of the THKP-C, successfully leveraged his Marxist credentials to denounce Dev-Genc, another Turkish communist group, for “right-deviationism” after Dev-Genc strayed from THKP-C and endorsed a Kemalist faction of the Workers’ Party of Turkey.Footnote ⁵² Subsequently, THKP-C attracted new support from the People's Liberation Army of Turkey for its ideological stance.

Taken together, we anticipate that religious ideology is most closely associated with the formation of militant cooperation and durability in the face of repression. But this derives from religions’ institutions, followers, and dense contemporary networks, rather than theology or the nature of religious belief or practice.Footnote ⁵³ We anticipate that our hypothesized relationships will be most pronounced for organizations that share a religious ideology, and that similar but attenuated effects will exist for secular ideologies with substantial numbers of adherents and international institutionalization (as with leftism in the Cold War era).

Rhetorical and Material Cooperation

Rhetorical and material relationships differ in their purpose, costs, and requirements. With few exceptions, rhetorical alliances are of value only if they are public. Because these alliances are about expressed support for another group and exclude material exchange, they are largely driven by a belief by one group in the merits of another group's cause. Moreover, rhetorical alliances are “cheaper” than material alliances that entail the expenditure of resources (e.g., guns, bombs, transportation, training) and effort (e.g., smuggling weapons, crossing borders).

Nonetheless, rhetorical alliances may pose substantial reputational risks. Because they are public, rhetorical declarations of support affect judgments about the ideological commitment and general credibility of both groups. Indeed, with their association clearly visible, groups may also be more concerned about the conduct of their public rhetorical allies as compared to private material allies. Groups risk backlash if their rhetorical allies suffer devastating setbacks or commit atrocities that sap public support. In both cases, rhetorical partnership with a malfeasant or weak ally can give the impression that a group is itself weak or mismanaged.

In contrast, material alliances usually operate best when they are clandestine. They are usually developed for the exchange of specific resources like arms, territory, training, operational plans, and funding to remedy operational deficits. The physical nature of this exchange makes these relationships more vulnerable to infiltration and defection, and the need for monitoring generates further security vulnerabilities. States also have a greater incentive to disrupt material relationships because of their immediate link to the capacity for violence. Finally, the higher physical costs and risks of material cooperation generate greater concerns over reciprocity.

Differences in their purposes, costs, and mechanics mean that rhetorical and material alliances are likely to respond differently to repression. As repression shortens time horizons, exacerbates resource deficits, and raises the likelihood of outright collapse, material alliances become more instrumentally valuable. Yet, concerns about security and reciprocity may be especially daunting when a group is facing a period of intense repression.

Thus, while repression increases the need for material support, it also makes it much harder to establish and maintain. In this context, organizations need mechanisms to overcome the concerns over security, reciprocity, and partner defection that gain salience in material cooperation. The result is that while rhetorical alliances can be conducted at arm's length with less immediate risk to security, this is infeasible in the context of material alliances. We therefore expect the cooperation-enhancing effects of shared ideology in the face of repression to be concentrated in the context of material cooperation.

While security concerns and commitment problems are less acute in the context of rhetorical cooperation, this does not imply that state repression has no impact on rhetorical relationships. Because the functional value of rhetorical support is contingent on both the performance of the ally vis-à-vis the state and the extent to which it imposes reputation costs, these relationships also become riskier as repression increases. Repression can reduce the selectivity of the violence that militant groups engage inFootnote ⁵⁴ and lead to operational setbacks or collapse.Footnote ⁵⁵ Increased repression, therefore, raises the risk that a group's allies will engage in counterproductive civilian victimization and/or collapse outright. These will concern groups involved in rhetorical alliances because of the greater reputational costs they entail. In sum, the risks of rhetorical cooperation grow with increased repression, but the benefits do not. Moreover, these risks are not conditioned by shared ideology because they are divorced from concerns over bilateral security and trust.Footnote ⁵⁶

The Militant Group Alliances and Relationships Data Set

To assess these expectations, we collected time-series data on all known militant relationships from 1950 through 2016, which we introduce here as the MGAR data set.Footnote ⁵⁷ We consider all independent, violent, nonstate organizations. This includes groups described as terrorists, rebels, and civil war actors around the world. As such, these data provide an important extension to those that focus on terroristFootnote ⁵⁸ or rebel groupsFootnote ⁵⁹ in isolation. The unit of analysis is the undirected dyad-year, which allows us to model evolving relationships among militant actors over time, including change and termination.

We began by creating a comprehensive list of militant organizations from existing data sets, and then mapping on the militant relationships identified in other data, including the Terrorist Organizational Profiles (TOPS) database,Footnote ⁶⁰ the Global Terrorism Database (GTD),Footnote ⁶¹ and the Armed Conflict Dataset (ACD).Footnote ⁶² From there, we investigated the alliances and either confirmed or discarded them; for those we confirmed, we added a time-series component and characterizations of the nature of the cooperation.

With existing data verified, we turned to the task of identifying previously unknown relationships. To do so, we assigned teams of researchers with relevant language skills to countries and gave them a list of all known militant groups ever active in that country. The teams coded relationships and also added organizations they uncovered that were missing from existing data sets.Footnote ⁶³

The process of documenting these relationships builds on a substantial body of prior research. We supplement and link together related data from TOPS, GTD, ACD, the Big Allied and Dangerous database,Footnote ⁶⁴ the Minorities at Risk Organizational Behavior data set,Footnote ⁶⁵ and the Mapping Militant Organizations project,Footnote ⁶⁶ among others. We have maintained identifiers from these data and matched our data with TORG crosswalk identifiersFootnote ⁶⁷ to allow researchers to link our data with other work in the field.Footnote ⁶⁸ The goal is to contribute to an aggregation of knowledge about militant networks that is open-source, adaptable, and widely available.Footnote ⁶⁹

The resulting MGAR data set consists of 7,402,203 undirected dyad-year observations, including 1,318,819 unique undirected dyads and 2,613 unique militant groups.Footnote ⁷⁰ The data set includes 7,409 undirected dyad-years in which a relationship exists between groups, of which 5,918 are cooperative and 1,491 are hostile.Footnote ⁷¹ Though not employed in this analysis, the full data set is even larger because it also includes alliances between militant groups and states, diaspora communities, political parties, and nonviolent organizations. The sheer number of ties identified in our data belies the notion that militant organizations primarily operate in isolation to preserve secrecy and security—there is a dense web of relationships, which we outline in the data description that follows.

MGAR captures when each relationship began and ended, and what it entailed. We provide a categorization of the nature of the relationship between groups, ranging from allies (the strongest form of cooperation) to associates, supporters, fans (verbal but not material support), hosts (primarily states), rivals, and competitors (the strongest form of conflict). For each cooperative dyad-year we also code whether the relationship included operational (e.g., shared membership, joint operations, tactical advising), material (i.e., arms transfers), territorial (e.g., shared bases), training, or financial support (i.e., cash transfers). Both the categorical and content variables are time variant.

These distinctions in the data allow us to isolate material from purely rhetorical cooperation. Material alliances are those in which at least one of the following is exchanged: operational, material, territorial, training, or financial support. Rhetorical alliances are defined as those in which groups are cooperative but none of these forms of physical support is exchanged.Footnote ⁷²

Descriptive Overview

Data from MGAR reveal important patterns in militant cooperation. The absolute number of alliances between militant groups is substantial (Figure 1). In general, alliances were relatively more common between 1965 and 1980, before declining from 1980 to 1990, and then increasing steadily again from 1990. The degree of militant cooperation—both initiation and durability—peaked in relative terms in 1970. These trends in militant cooperation reflect well-known political dynamics. From 1965 to 1980, the systemic political environment was characterized by decolonization struggles, global ideological contestation, and substantial state sponsorship of militant organizations, all of which facilitated alliance formation and maintenance by injecting militant groups with resources and revolutionary ideological currency.Footnote ⁷³

FIGURE 1.

Militant alliances by type

As the decolonization concluded and the Cold War waned from 1980 to 1990, militant cooperation declined in absolute and relative terms. The dissolution of multiethnic empires and the proliferation of weak states gave rise to the creation of many new militant groups (i.e., potential alliance partners), but waning ideological contestation and state sponsorship reduced the ideological and material resources vital for alliance formation and maintenance.Footnote ⁷⁴ Intuitively, Figure 2 shows that the relative decrease in cooperation in this era is driven by a decline in alliances between leftist groups, precisely those most affected by the decline of ideological contestation and state sponsorship in the late Cold War.

FIGURE 2.

Militant alliances by shared ideology

Since 1990, militant cooperation has increased steadily. This increase is driven by the rise of alliances between groups sharing a religious ideology. As religious ideologies, especially Salafism, supplanted communism as the dominant global, militant ideological currency after the Cold War, religiously oriented cooperation proliferated.Footnote ⁷⁵ By 2016, the global proportion of militant dyads that were allied rebounded to its 1980 level. As of 2016, there were forty rhetorical and 294 material alliances between militant groups, including twenty rhetorical and sixteen material alliance initiations. These figures amount to cooperation in more than one-tenth of 1 percent of all militant undirected dyads worldwide.

Alliance initiation has remained relatively constant since the mid-1980s, and has not increased along with the total increase in militant cooperation since 1990. This suggests that the total increase in militant cooperation is driven in large measure by the growing durability of alliances, not by an increased ability of groups to form alliances in the post–Cold War era. This insight suggests the possibility that alliance initiation and maintenance have different causes, and particularly that shared ideology (e.g., co-religion) may be more important for sustaining than for initiating cooperation under repression.

Apart from broad trends in the degree of militant cooperation, Figure 1 also reveals an important difference in patterns of rhetorical and material cooperation. In particular, material alliances are far more frequent than rhetorical alliances, accounting for about 84 percent of all alliance-years in the data. While the degree of rhetorical cooperation is relatively constant over time, material cooperation is subject to far more temporal variation, with peaks in the periods from 1965 to 1980 and from 1990 to 2016. Looking at alliance formation, rhetorical alliance initiation outpaces material alliance initiation in only two years: 1973 and 2016. In 1973, the stark increase in rhetorical alliance formation reflects the rise of nonmaterial cooperation between prominent, mostly leftist groups in the Middle East, Western Europe, and Latin America. Responding to increasing repression in the wake of events like the Munich massacre, the Yom Kippur War, and the Ezeiza massacre in Argentina, leftist/communist groups mobilized to rhetorically support ideological compatriots around the world.Footnote ⁷⁶ The 2016 surge in rhetorical alliance formation reflects a deliberate strategy by ISIS to legitimate its claim of a caliphate by extracting loyalty pledges from a network of global affiliates.Footnote ⁷⁷ Some of those initially rhetorical alliances have since matured into material alliances.

Many of the most lethal militant alliance networks, such as those of ISIS, stretch across continents; at the the same time, large militant networks also emerge within individual countries. Because the MGAR data capture relationships for a large, global sample of militant groups, they offer a unique opportunity to explore the dynamics of domestic versus transnational cooperation.Footnote ⁷⁸ Figure 3 distinguishes alliances between groups based in the same or neighboring countries (domestic) from alliances between groups based in noncontiguous countries (transnational).Footnote ⁷⁹ Domestic alliances are more common, accounting for about 68 percent of all alliance-years. Notably, the ebb and flow of domestic and transnational cooperation follow similar trends over time.

FIGURE 3.

Militant alliances by geographic scope

Table 1 provides summary statistics for core variables in our data for all dyad-years and for cooperative dyad-years. These offer a broader portrait of the MGAR data and echo many of the insights highlighted graphically in Figures 1 through 3. The alliance content variables, including the disaggregation of material and rhetorical alliances, correspond to the content variables described before, while the relationship-type variables correspond to the categorical coding we described. The data show that of the five types of support we code, operational support is most common, occurring in about 65 percent of all alliance dyad-years. Training and material support are also frequent, occurring in 57 percent and 56 percent, respectively; territorial and financial support are less common, occurring in 40 percent and 48 percent, respectively. The mean and median number of forms of support exchanged in an alliance dyad-year is 3, indicating that most militant cooperation yields the exchange of relatively numerous and robust material resources.

Table 1.

MGAR summary statistics

Ideology is shared in more than 91 percent of all alliance dyad-years. Shared religion is particularly significant: religion is shared in just 6 percent of all undirected dyad-years but in 37 percent of alliance dyad-years. Consistent with work on state sponsors, we find that shared sponsorship also occurs disproportionately in alliance dyad-years.Footnote ⁸⁰ Still, shared sponsorship is not present in about 95 percent of alliance dyad-years, so other commitment mechanisms must also exist to support the observed extent of militant cooperation. Finally, even in the summary statistics it is clear that most militant groups generally, and most alliances dyad-years specifically, operate in highly repressive contexts.

Empirical Strategy

To test our hypotheses, we start with the list of 2,613 unique militant groups in MGAR, and create an undirected dyadic data frame that links each militant group to every other contemporaneous militant group for each year of each group's existence. This procedure yields 7,402,203 observations, reflecting every undirected dyad-year in which a relationship could exist. Hence, our unit of analysis is the militant group undirected dyad-year.

The dependent variable captures whether an alliance exists between groups in an undirected dyad-year and, if so, the nature of the alliance: rhetorical or material. We distinguish between alliance initiation and durability (i.e., duration) to test Hypotheses 1 and 2, respectively. The dependent variable in our principal analyses is defined using the alliance content variables: operational, material, territorial, training, and financial support. If one or more of these forms of support is exchanged in a dyad-year we code the alliance as material. If an alliance exists but evidence suggests that none of these five forms of support was exchanged, we code the alliance as rhetorical. In dyad-years (n = 376) where data on the various forms of support are missing because insufficient information existed to make a determination, we code no alliance.Footnote ⁸¹

To establish the robustness of the finding we also assess a version of the dependent variable derived from a categorical variable in MGAR that captures relationship type: none, allies, associates, supporters, hosts, fans, rivals, or competitors. In this formulation, material alliances are defined as allies, associates, supporters, or hosts, and rhetorical alliances are defined as fans. We employ the content-based variable in our primary analyses because it is based on objective information about the repertoires of support within alliances; by contrast, the categorical variable is based on coders’ more subjective assessments of the relationship. Nevertheless, these measures are highly correlated (ρ = 0.973), and all results are robust to the choice of measure.Footnote ⁸²

The primary independent variables are shared ideology and repression. Shared ideology is a dichotomous indicator based on the MGAR group ideology variables. We coded whether each group was leftist/communist, nationalist/separatist, far-right-wing, or environmentalist, along with a categorical variable for each group's religious identity. As we have noted, there is relative consensus in the literature that ideologies can be overlapping but are generally slow moving. In response, our coding categories are not mutually exclusive but are time invariant. We define shared ideology dyads as those in which both constituent groups share a secular or religious ideology. For instance, Basque Fatherland and Freedom and the Irish Republican Army are coded as sharing ideology because both groups are nationalist/separatist.Footnote ⁸³

We measure repression using Fariss's latent human rights index.Footnote ⁸⁴ Fariss produces a dynamic latent variable model of physical integrity rights—individuals’ rights to protection from arbitrary bodily harm and coercion. The index combines information on torture, political imprisonment, disappearances, and extrajudicial killings. We reverse-scaled the indicator for ease of interpretation, so higher values mean more repression and less respect for human rights. We measure repression dyadically: our repression score is the maximum value of the reverse-scaled rights index across the countries in which the two groups are based. All results are substantively similar when we measure repression with the CIRI physical integrity rights index.Footnote ⁸⁵ Both the Fariss and CIRI measures vary across countries and over time.Footnote ⁸⁶ Because we are primarily interested in the effect of shared ideology on militant cooperation as repression increases, we interact the shared ideology indicator with our repression measure.

Estimation

We take two primary modeling strategies in our main analyses: a series of multinomial logistic regressions to test Hypothesis 1 (on alliance initiation) and a series of variance-corrected hazard models to test Hypothesis 2 (on alliance durability). We opt for multinomial logistic regression to model alliance initiation because our dependent variable is a three-point, categorical measure, where 0 means no alliance initiation, 1 means rhetorical alliance initiation, and 2 means material alliance initiation. Because we are interested in the differences between rhetorical and material alliance formation, 0 serves as our reference category. Thus, in the multinomial logistic approach, we estimate two equations: one for rhetorical alliances relative to no alliance and one for material alliances relative to no alliance. The multinomial logistic framework allows us to estimate the relative probabilities of our different outcomes—rhetorical or material alliance initiation—from the same sample, rather than from independent samples as in standard logistic regression. It also has the advantage that it makes no assumptions about proportional odds as in an ordered logistic regression.Footnote ⁹³

Our multinomial logistic models are expressed as

$$\eqalign{& Pr( Y_{d, t} = { {\rm \smallcaps ALLIANCE\;INITIATION\;CATEGORICAL}}) = {\rm Logit}( \alpha _0 + \beta ( {\rm \smallcaps SHARED\;IDEOLOGY}_{d, t}) + \cr & \gamma ( {\rm \smallcaps REPRESSION}_{d, t-1}) + \delta ( {\rm \smallcaps SHARED\;IDEOLOGY}_{d, t}{\rm ) ( \smallcaps REPRESSION}_{d, t-1}) + \theta ( X_{u, s}) + \epsilon _{d, t}) } $$

where the outcome is the initiation of a rhetorical or material alliance, d indexes undirected dyads, and t indexes years. α ₀ denotes the constant, and δ gives the effect of shared ideology as repression increases in undirected dyad d in year t = 1. θ are estimates from a vector of covariates, and ${\epsilon} _{d, t}$ are undirected dyad-clustered standard errors.

To test Hypothesis 2, on alliance durability, we turn to variance-corrected hazard models. This approach allows us to directly estimate the effect of shared ideology, conditioned on repression, on the time to militant alliance termination. In these analyses we subset the data to cooperative dyad-years and code termination when groups experience alliance discontinuation (i.e., an alliance in year t−1 and not in year t).

While not widely recognized in the literature on militant cooperation, repeated failures are common in the data. Many militant groups terminate alliances only to re-ally with former partners in the future. For instance, the Palestine Liberation Organization (PLO) and the Popular Front for the Liberation of Palestine (PFLP) first allied from 1968 to 1973. PLO–PFLP cooperation broke down in 1974 when the PFLP joined the Rejectionist Front in opposition to the PLO's adoption of the Ten-Point Plan. Nevertheless, the PLO–PFLP alliance reemerged in 1981 when the PFLP rejoined the PLO Executive Committee. It terminated a second time in 1999 amid lingering debate over the merits of conciliation with Israel.

In traditional hazard modeling approaches, event times are assumed to be conditionally independent. This assumption does not hold in our framework because the risk of failure is correlated within units. Clustering of terminations among dyads means that previous failures raise the risk of future failures. Given the threat of event dependence, we employ a variance-corrected gap-time Cox estimator.Footnote ⁹⁴ This model stratifies the risk set by the number of failures a dyad has experienced to control for event dependence and clusters standard errors by undirected dyad to account for within-dyad correlations between failures. The estimator uses gap time because we are interested in modeling the time since the last failure. By employing the Cox estimator, we make no additional parametric assumptions about the shape of the underlying hazard function.Footnote ⁹⁵

The hazard in this framework can be written as

$$ \eqalign{& \lambda _{d, s} (t) =\cr & \lambda _{0s}( t-t_{s-1}) e^{\beta ( { \rm \smallcaps SHARED\;IDEOLOGY}_{d, t}) + \gamma ( {\rm \smallcaps REPRESSION}_{d, t-1}) + \delta ( {\rm \smallcaps SHARED\;IDEOLOGY}_{d, t}{\rm ) ( \smallcaps REPRESSION}_{d, t-1}) \, + \, \theta ( X_{d, t}) }} $$

where the outcome is the termination of a rhetorical or material alliance, d indexes undirected dyads, and s denotes the cumulative number of alliance breakdowns within the dyad, which stratifies the risk set. λ _0s is the baseline hazard rate, and t − t _s−1 specifies a gap-time formulation, where the hazard is the risk of failure for alliance breakdown s since the occurrence of alliance breakdown s − 1. δ gives the effect of shared ideology as repression increases in undirected dyad d in year t − 1. θ are estimates from a vector of covariates. In all models, we cluster standard errors by undirected dyad and use Efron's method for ties.

Disaggregating Shared Ideology

The prior analyses assessed ideology broadly, pooling shared secular (i.e., nationalism/separatism, leftism/communism, right-wing, and environmentalism) and religious ideologies in a single indicator. There is, however, reason to anticipate heterogeneity among ideologies in their value for alliance commitment as repression grows. We specifically posit that shared religious ideology is an especially efficient mechanism for alliance commitment because religions offer large, dispersed networks of believers and authority figures who can monitor and sanction alliances between adherent groups. Similarly, religious ideologies offer cheap mechanisms for rewarding cooperation and punishing defection.Footnote ¹⁰⁴ To examine heterogeneity across ideologies, we replicate our main analyses of alliance initiation and durability, replacing the general indicator for shared ideology with specific indicators for shared religion and shared leftism/communism. These ideologies comprise most of the alliances in the data, and over time, religion and leftism have provided the most substantial ideological capital for militant groups.Footnote ¹⁰⁵

Our repression measure captures the overall climate in a country, but does not distinguish which groups are targeted. Because different ideologies may face different levels of repression at various points, we focus these disaggregated analyses on periods when the respective militant ideology was most salient, on the logic that these are times when state repression posed the greatest threat to cooperation. In the absence of granular data on which ideologies face the brunt of the repression, this periodization helps capture variation in the intensity of pressure on different ideological groups. We study shared religion during the post-9/11 period through 2016, when the Global War on Terror drove increasing pressure on jihadist groups; and shared leftism/communism during the Cold War, when US forces bolstered anti-communist forces around the world.

Shared Religion in the Post-2001 Era

Figure 6 tests our expectation that shared religion is uniquely important for generating alliance commitment. In the top panel we present results from a multinomial logit model of alliance initiation. In the bottom panel we depict the effect of shared ideology on material alliance durability. Models follow the same specification employed earlier, though we subset to the post-2001 period, when state repression against religious (and especially jihadist) groups was most intense.Footnote ¹⁰⁶

FIGURE 6.

The effects of shared religion on alliance initiation and durability

In the top panel, key differences emerge in the effect of shared religion compared to shared ideology in general. At low levels of repression, religion has no significant effect on the probability of rhetorical initiation (p = .424), but this effect becomes significantly positive as repression increases (p = .069). Shared religion increases the probability of material initiation at low levels of repression, but this effect is substantively modest and narrowly significant (p = .076). These findings are consistent with Bloom's conclusion that co-religionists are uniquely susceptible to outbidding dynamics, competing for recruits and resources when repression is low and clerics and adherents have reduced incentives to demand cooperation.Footnote ¹⁰⁷ In contrast, and in line with Hypothesis 1, shared religion is associated with a significant increase in the probability of both rhetorical and material initiation as repression increases. Moving from the tenth to the ninetieth percentile of repression increases the probability of rhetorical initiation 130 percent and the probability of material initiation 116 percent. In other words, repression has promoted cooperation between co-religionist militant groups in the post-2001 period. AQ's relationship with its network of Salafi affiliates is instructive of religion's unique commitment role for cooperation. As counterinsurgent pressure increased after 2001, AQ expanded its material relationships with ideologically sympathetic organizations around the world.

In the bottom panel of Figure 6 we study the effect of shared religion on the durability of material cooperation.Footnote ¹⁰⁸ Findings comport with Hypothesis 2. The results in the bottom-left panel show that, for groups that share a religious ideology, the risk of material alliance termination is positive and imprecisely estimated when repression is low, but the effect of shared religion strengthens considerably as repression increases.Footnote ¹⁰⁹ Groups that share religion are about 55 percent less likely to suffer material termination at the highest level of repression. This finding comports with Woldemariam's argument that reduced repression attenuates incentives for cooperation and makes militant fragmentation more likely, but that repression can lead to insurgent cohesion if commitment is high.Footnote ¹¹⁰ Equivalently, the results in the bottom-right panel show that the risk of material alliance breakdown is 150 percent greater for groups without shared religion at the highest level of repression. Overall, these findings help illustrate why global linkages between jihadist militant groups like AQ, ISIS, and their affiliates have endured despite two decades of massive counterterror pressure. Shared religion provides a particularly complete package of commitment devices that aid cooperation in the face of threat.

Although we focus on material alliances in Figure 6, we find substantively identical results for rhetorical alliances.Footnote ¹¹¹ Given the small number of co-religious rhetorical alliances in the post-2001 period, we hesitate to draw strong inferences about the relative importance of shared religion for generating commitment in rhetorical versus material alliances under repression. However, these results suggest that co-religious rhetorical alliances are significantly less likely to break down for all values of repression above the tenth percentile. This effect is largely driven by the network of rhetorical alliances ISIS built in 2015 and 2016 as a means of legitimizing its caliphal project.Footnote ¹¹²

Shared Leftism/Communism in the Cold War

We posit that the effectiveness of religious ideology as a commitment device derives from its built-in institutions and monitoring mechanisms, rather than theology. As such, secular ideologies with substantial numbers of adherents and broad networks of institutions should follow a similar pattern. In the twentieth century, leftism/communism represented the most prominent secular ideology, with a global network of ideologues and substantial state support for adherent militants. The end of the Cold War may have weakened the commitment value of the ideology, as communism's transnational networks and primary state supporter evaporated.Footnote ¹¹³ However, we expect, as with religion, that leftism/communism had a distinct positive effect on alliance commitment during the Cold War, when the ideology was a salient ideological force and anticommunist repression was intense. To investigate this possibility, we specify an additional model where we interact shared leftist ideology and repression, and subset to the Cold War era, defined as 1950 to 1989.Footnote ¹¹⁴

The top panel of Figure 7, shows the effect of shared leftism on rhetorical and material alliance initiation at low and high values of repression. Increasing repression reduces the probability of rhetorical alliance initiation between groups sharing leftist ideology by 100 percent as repression increases from its tenth to ninetieth percentiles. This effect differs from the effect of shared religion on rhetorical alliance initiation, where we observe an increasing probability of initiation as repression increases. Like shared religion in the post-2001 period, however, shared leftist dyads are 26 percent more likely to ally materially as repression increases. This result aligns with our expectation in H1. More broadly, these findings highlight a key difference between rhetorical and material alliances in their responsiveness to repression, highlighting the importance of our exercise in distinguishing them.

FIGURE 7.

The effects of shared leftism/communism on alliance initiation and durability

Turning to the bottom panel of Figure 7, results on material alliance durability also suggest that shared leftism during the Cold War plays a less important commitment role under repression than shared religion after 2001. Specifically, whereas we find that shared religion significantly reduces the probability of material alliance termination as repression increases, figures in the bottom panel suggest that shared leftism has a more modest effect. During the Cold War, shared leftism reduced the probability of material alliance termination as repression increased, but not significantly so. Rather, the graphs in the bottom panel show that the negative effect of shared leftism on alliance breakdown in the Cold War is mostly a product of the unique fractiousness of cooperation between leftist groups at low levels of repression. As Cronin argues, “The left-wing and anarachistic groups of the 1970s … were notorious for their inability to articulate a clear vision of their goals that could be handed down to successive generations after the first generation of radical leaders departed or were eliminated.”Footnote ¹¹⁵ As a result, leftist militant organizations throughout the Cold War were prone to disputes over doctrine, particularly in periods when repression was low and allies had time and space to debate revolutionary theory.Footnote ¹¹⁶ In total, shared leftism reduced the risk of material alliance breakdown during the Cold War, but not enough to generate significant alliance commitment. We find no evidence that shared leftism facilitated rhetorical alliance commitment.Footnote ¹¹⁷

Beyond shared religion and leftism, we analyze shared nationalism/separatism, right-wing, and environmentalist ideologies in the online appendix.Footnote ¹¹⁸ We do not have strong a priori expectations about the relative commitment value of these secular ideologies, but we do anticipate that their roles in facilitating alliance initiation and durability under repression will be less important than shared religion. In contrast to shared religion, shared nationalism/separatism only marginally increases the probability of material alliance initiation, and this effect attenuates considerably at high levels of repression. Similarly, compared to other ideologies, alliances between right-wing and environmentalist groups are less likely to be initiated, and more prone to be terminated, as repression increases. Overall, these findings point to the unique commitment value of shared religion for militant cooperation.

Robustness and Extensions

The chief threat to inference in our empirical framework stems from the endogeneity of state repression and militant cooperation. If governments increase repression because militant groups form material alliances, our results may capture the government response to militant cooperation, not the role of shared ideology as a commitment device. In our primary models, we use a one-year lag of repression to avoid temporal confounding, but this strategy is imperfect because governments may increase repression in expectation of militant cooperation.

This endogenous process argument cannot explain our finding that shared ideology reduces the probability of material alliance termination as repression increases. States have an unconditional incentive to respond to material cooperation between militant groups with repression. Hence, if reverse causality were solely driving our results—that is, if governments were increasing repression in response to militant cooperation or the threat thereof—we would not expect significant differences in the probability of termination between dyads with and without shared ideology.

To further address potential endogeneity between repression and cooperation in our models, we take two complementary approaches. First, we re-estimate our core specifications using a semi-parametric marginal structural model (MSM). The MSM is fitted in two stages. We estimate the probability that a group experiences its observed treatment history (in this case, we define treatment as high repression) in the first stage and use these estimates to construct inverse probability of treatment weights. Then we re-estimate the effect of shared ideology on alliance initiation and durability as repression increases those weights. In this way, inverse probability weighting controls for time-dependent confounding between repression and cooperation.Footnote ¹¹⁹ In line with our main results, the MSM shows that shared ideology unconditionally increases the probability of alliance initiation and significantly reduces the risk of material, but not rhetorical, alliance termination under repression.Footnote ¹²⁰

Second, we re-estimate our models using two-stage residual inclusion (2SRI). 2SRI is a control-function approach that addresses endogenous explanatory variables by estimating a reduced-form model of the endogenous variable (in our case, repression) and including first-stage residuals as regressors in the second stage. Because our measure of repression is continuous, we estimate the first-stage selection equation using linear regression with country-level covariates. In the second stage, we re-estimate our multinomial logistic and hazard models while including the residual term calculated in the first stage. 2SRI estimates also confirm that shared ideology unconditionally increases the probability of alliance initiation and significantly reduces the risk of material but not rhetorical alliance termination.Footnote ¹²¹ While neither 2SRI nor the MSM can fully mitigate endogeneity concerns, concordance between the estimates from these models and our core results builds confidence in our findings.

Next, we probe the sensitivity of our results to alternative estimation techniques. One of our primary modeling strategies is multinomial logistic regression, which does not assume proportional odds. Taking advantage of the ordered structure of our dependent variables, we show that ordered logit models yield substantively identical results.Footnote ¹²² Second, while we control for dyadic and country-level covariates, it is likely that there remains unobserved heterogeneity across undirected dyads and years. Fixed effects could absorb time-invariant confounding, but they are too computationally intensive to implement in our primary models. Linear probability models estimated with high-dimensional fixed effects for undirected dyads and years indicate that increases in repression are associated with rhetorical alliance initiation for groups that share ideology.Footnote ¹²³ The converse holds for groups that lack shared ideology.

Apart from alternative estimators, we re-examine the core results using different specifications, including alternative measures of cooperation and repression. First, we re-analyze our core models using the dependent variable based on relationship type, rather than on alliance content.Footnote ¹²⁴ Second, we use a different measure of repression based on CIRI rather than Fariss.Footnote ¹²⁵ Third, we re-estimate our models on a sample of observations for which both groups involved in a dyad appear in the GTD—these models drop observations with missingness in the capability ratio measure.Footnote ¹²⁶ Fourth, because AQ and ISIS are uniquely well connected and capable, and have been the target of aggressive and sustained counterterrorism operations, we show that our results are robust to excluding dyads involving these groups or their core affiliates.Footnote ¹²⁷

Finally, as an extension on our descriptive analyses, we examine heterogeneity in the value of shared ideology under repression for domestic versus transnational alliances.Footnote ¹²⁸ Transnational militant networks, such as those of AQ and ISIS, represent perhaps the most prolific and dangerous form of militant cooperation.Footnote ¹²⁹ However, the transaction costs of transnational alliances are far higher than those of domestic alliances, so commitment should be more difficult to form and sustain for groups cooperating across a broad geographic area versus within a single conflict ecosystem.Footnote ¹³⁰ Transnational alliances represent a hard test for the role of shared ideology in generating alliance commitment as repression increases.

To test for differences in the effect of shared ideology on alliance initiation and termination as repression increases, we rerun our core models on two distinct samples. The domestic sample includes all undirected dyad-years involving groups based in the same country or contiguous neighboring countries, and the transnational sample includes all observations involving groups based in noncontiguous neighboring countries. We find that the theorized effects are magnified in the expected direction for transnational material alliances, precisely where commitment is most challenging. As in Table 2, shared ideology increases the probability of rhetorical and material alliance initiation, and this effect is generally unconditioned on repression. A similar pattern emerges in the context of alliance durability. Both domestic and transnational alliances are less likely to terminate when they share ideology, particularly religion, as repression increases, but the magnitude of this effect is significantly greater for transnational than for domestic alliances. The predicted breakdown-reducing effect of shared ideology is 35 percentage points greater for transnational than for domestic material alliances at the highest value of repression. Combined, these tests build additional support for our theory and show that the effect of shared ideology on alliance formation and durability is largest for transnational alliances, precisely those in which commitment problems are most severe.