Model

The model consists of the three players previously introduced: militant group A, militant group B, and a government G.Footnote 18 While A and B are separate entities with distinct preference profiles, both consider G to be a common enemy. The primary goal of G is to disarm both A and B, thereby re-establishing its monopoly on force. By contrast, A and B seek to survive G's repressive attempts. Should A and B ally, both organizations improve their ability to survive by pooling resources and improving efficiency through co-operation.

Figure 1 presents the sequence of moves in the game, while Table 1 presents the interpretation of the model's parameters.Footnote 19 At the first move, A decides whether or not to propose an alliance between itself and B. If A does not propose an alliance, both A and B fight G separately, using only their own resources. However, if A makes an alliance proposal, and B accepts, both A and B develop security ties, and improve their efficiency and joint survival capability. However, since A and B are separate entities with variable preferences, forming such an alliance will to some degree dilute the original agendas.

Fig. 1

Bilateral game

Table 1

Parameters of Model

If an alliance forms, G begins the fulfilment subgame by offering a concession π, where π∈[0,1].Footnote 20 The concession is available to A and B if either breaks co-operation with their respective partner. If the concession is taken by either group, the group disarms and shares information with G. However, G only grants the concession to the defecting group if it requires the information to defeat the group's partner. In other words, if both A and B disarm in hopes of accepting the concession, G does not grant the concession to either party. Since both groups voluntarily disarm and surrender their political ambitions, G has little need to pay either of them to gain information.Footnote 21

The table beneath Figure 1 presents each of the possible outcomes of the game along with the payoffs to each player. We assume that G's top preference is to disarm both groups A and B. Therefore, G receives a payoff of 1 if it disarms A and B, either through violence or negotiation. If G is successful in disarming either group, the group receives a payoff of 0. By contrast, A and B receive payoffs of 1 should they independently withstand G's repression. However, since the alliance dilutes the agendas of both groups, A and B's payoffs are discounted by δ∈[0,1] if they survive as part of an alliance. The agendas of the two groups are more compatible as δ→1 and less compatible as δ→0.

The fulfilment subgame repeats infinitely until G successfully disarms both militant groups. We assume that G forcibly disarms A with probability p∈[0,1] and forcibly disarms B with probability q∈[0,1], which are assumed to be drawn from independent distributions. Since the game terminates for either player if they are disarmed, the groups’ probabilities of survival (p,q) serve as their respective discount factors.Footnote 22 We further assume that while G's power to repress is at its highest in the initial stages of the conflict G's ability to repress both A and B declines with time. This assumption is based on previous empirical studies indicate that the ability of governments to repress militant groups appears to be decreasing as the duration of the conflict increases.Footnote 23 These studies suggest that these groups grow more resilient with time due to their adaptation to government tactics, their ability to innovate, and their ability to signal their staying power to their target population.Footnote 24 We assume that G disarms both groups with probability p ^tq ^t, where t represents the number of rounds played. As is standard in models of conflict, each player (G, A, B) is assumed to pay a cost c to engage in armed conflict. For simplicity, we assume that the total cost for each player for the entire campaign is represented by C _{(G ,A ,B )}∈[c,∞).

Should A and B form an alliance, the efficiency gains from co-operation reduce G's probability of disarming both groups to θpq, where θ∈[0,1]. This indicates that if A and B ally, G's probability of success is less than its probability of disarming either A and B individually. If G fails to disarm either group, G receives a payoff of 0, minus the cost of prosecuting the conflict C_G. If either A and B accept G's concession, the group that accepts the concession receives a payoff of π and agrees to disarm. The information provided by the group that accepts the concession improves G's chance of forcefully disarming the other group by β∈[1,2].Footnote 25 If G successfully disarms one group through negotiation and the other through force, G receives a payoff of 1−π, which captures G's benefit for suppressing the militants minus the cost of the concession. If both groups disarm and attempt to exploit each other in hopes of getting G's concession, both groups receive payoffs of 0 since G reneges on its commitment to grant either side π.

Solution

We solve the game using the subgame perfect solution concept (SPE). While the complete formal solution is presented in the appendix, this section outlines the intuition behind the model's solution. We begin by examining the fulfilment subgame. We see that if 0 > 1−β(p,q)−c _{(A ,B )}, both groups will exploit each other. Since the price of exploitation exceeds the cost of disarming, both groups will disarm if they believe they will be exploited. This indicates that co-operation is increasingly unlikely if (p,q)→1, or as C _{(A ,B )} increases. We see that vulnerable militants who suffer greater costs are unlikely to sustain co-operation. If either A or B perceive the other group as vulnerable, the groups will anticipate opportunistic behaviour prior to forming the alliance, and will choose to fight individually at the beginning of the game. This suggests that the mutual fear of defection is a valuable tool for governments. Since both militant groups fear that the other will exploit, governments have an easier time repressing the groups.

If 0 > 1−β(p,q)−c _{(A ,B )} and G is willing to offer a concession, both groups have dominant strategies to exploit. However, if the government does not offer a concession and sets π=0, the strategy pairing (Co – operate, co – operate) is also a Nash solution. Assuming that militants choose the efficient equilibrium in multiple equilibria situations, the solution to the fulfilment subgame is (Co – operate, co – operate). In other words, if G refuses to negotiate, alliances will be sustained in the fulfilment subgame, which gives groups incentives to form alliances at the game's outset. The model illustrates a somewhat interesting prediction. Governments that are conciliatory can use the mutual distrust between militants to prevent alliances from forming. By contrast, governments that refuse to negotiate allow militants to develop credible alliances. Footnote 26

To induce A successfully to betray B, G must make an offer π with a value greater than or equal to A's value for sustaining long term co-operation with B. Should A co-operate with B indefinitely, A's payoff converges to a value of

$$\frac{\delta }{{1\,{\rm{ - }}\,(1\,{\rm{ - }}\,pq\theta )}}\,{\rm{ - }}\,{{C}_A}\, = \,\frac{\delta }{{pq\theta }}\,{\rm{ - }}\,{{C}_A}.$$

Similarly, B's value for extended co-operation with A converges to a value of (δ/pqθ) – C_B. We see that as pqθ→1, indicating that G has greater repressive power, the value for extended co-operation decreases for both A and B. This in turn decreases the value of π needed to induce either group into negotiation. Consequently, we see that if G has a high likelihood of disarming either A or B, both A and B are more likely to accept concessions. Formally, if the following condition is true, G offers a concession to A, which is also generalizable to B:

$$\beta q\left[ {1\,{\rm{ - }}\,\left( {\frac{\delta }{{pq\theta }}\,{\rm{ - }}\,{{C}_A}} \right)} \right]\, + \,(1\,{\rm{ - }}\,\beta q)({\rm{ - }}{{C}_G}) \gt pq\theta \,{\rm{ - }}\,{{C}_G}$$

Simplified:

$$q(\beta (1\, + \,{{C}_G}\, + \,{{C}_A})\,{\rm{ - }}\,p\theta )\,{\rm{ - }}\,\frac{{\beta \delta }}{{p\theta }}\, \gt \,0.$$

Figure 2 presents a graphical interpretation of Condition 2. We see that as {p,q}→1, G is more likely to negotiate. We also see that alliances between vulnerable militants are unstable since these groups are relatively easy to pay off. Facing at least one weaker group, G's strategy would be to find the weak link in the alliance and negotiate with that party. Since the weak link is either less likely to survive conflict or less resolute, the weak link offers G an opportunity to use a ‘divide and conquer’ strategy. To avoid this outcome, militants are only likely to co-operate with other groups that can both resist government repression and share some degree of affinity. Groups with compatible agendas that can withstand government repression increase the value of future co-operative interactions, thereby increasing the level of concessions that G must pay. If both groups are capable of surviving long into the future, the benefits for extended co-operation will exceed any offer that G makes. Consequently, we see that if G has a low probability of disarming the groups through force, both groups stand to gain from alliance co-operation.

Fig. 2

Government's utility for negotiating as a function of its ability to disarm militant groups A and B

We are able to see the extent of these gains from Figure 3, which plots the militants’ EU(Co – operation) – EU(∼Co – operation). We see that bilateral co-operation is only preferable if the government's repressive capacity is low. These results indicate that we should observe bilateral co-operation between militants to occur in cases where the government is unlikely to repress both of the groups, and where the groups share similarities in ideology.

Fig. 3

Utility to militant group A for bilateral co-operation as a function of the government's ability to disarm militant groups A and B

Implications

From the bilateral model, we see that if governments are likely to disarm their militants using violence, the groups’ value for extended co-operation falls. In these cases, the government can negotiate with the militants, rendering the groups unable to ally due to their mutual fears of exploitation. Similar to previous studies, these results indicate that the divide and conquer approach appears to be a very effective tool of counter-insurgency. However, since groups operating in weaker states are less likely to be disarmed forcibly, the value of extended co-operation to these groups will be higher. Governments may be unable to offer either group a concession that is more enticing than the prospect of extended co-operation, particularly if the groups are ideologically compatible. Consequently, we would expect that bilateral co-operation is only likely amongst ideologically similar groups operating in weaker states.

Empirically, we do see groups that are capable of resisting government repression banding together to form alliances. For example, Osama bin Laden's al-Qaeda and Ayman al- Zawahiri's Egyptian Islamic Jihad were both well-developed organizations with similar ideologies. According to the model, these groups should be capable of forming a credible alliance. Yet, although some anecdotal evidence exists in favour of the model, we also see many alliances that do not appear to match these predictions. For example, consider the alliance between the Zimbabwe African National Union (ZANU) and the Zimbabwe African Peoples’ Union (ZAPU) during the Rhodesian civil war. Despite facing a government with a powerful military force, and considerable rivalry between the parties, this alliance was sustained throughout the conflict. This observation, and several others, suggests a different type of alliance pattern between militants. While the current model allows us to understand bilateral co-operation, we next turn to a discussion of asymmetric alliances. This new specification makes the argument that groups facing powerful governments will seek alliances involving a state sponsor to constrain other groups from exploitative behaviour.

Model

To model the conditions under which asymmetric alliances form and sustain, we introduce the sponsor S into our previous game (see Figure 4). If S participates in an alliance with A and B, S alters the dynamic of co-operation. In its role as the dominant partner in the alliance, S agrees to absorb the costs of conflict for both parties (C_A, C_B), so long as both sides sustain co-operation. However, should either A or B defect from the alliance, S imposes a punishment α _{(A ,B )}∈[0,1]. As a result of S's participation, defection is no longer costless for either group, and this would be intended to deter both groups from betraying the alliance. In exchange, S assumes control over the alliance and further distorts the agenda of both groups by δ_S. In doing so, S gains some value V∈[0,1] for each period in which the alliance is sustained. However, due to both the cost of policing the militants and the risk of potential government retaliation, S must pay a governance cost C_s∈[C_A + C_B, ∞) to sustain the alliance. Assume that S must pay this cost up-front, regardless of whether or not A or B attempt to exploit their alliance partner. S's utility for intervening is represented by the following statement:

$${{U}_S}\,{\rm{ = }}\,{{V}^t} \,{\rm{ - }}\,{{C}_S}{\rm{.}}$$

Fig. 4

Asymmetric game

Since V∈[0,1] and is discounted by t, S's payoff converges to:

$$ {{U}_S}\, = \,\frac{1}{{1\,{\rm{ - }}\,V}}\,{\rm{ - }}\,{{C}_S}. $$

Let us assume that S's value for sponsoring the alliance V remains private information, and that the players do not have complete information as to its value. This indicates that both A and B are uncertain whether S will sustain the alliance indefinitely or whether, at some point, S will cease serving as a guarantor. Let us assume that the players have a flat prior belief as to the value of V. If this is the case, the probability that S will continue to participate in the alliance indefinitely is

$$\frac{{{{C}_S}{\rm{ - }}1}}{{{{C}_S}}}\, = \,V_{S}^{\rad } .$$

If $ {{V}_S}\, \gt \,V_{S}^{\rad } $, S participates indefinitely whereas if $ {{V}_S}\, \lt \,V_{S}^{\rad } $, the governance costs of the alliance are too high for S to sustain, and S will eventually abandon the alliance. Should this occur, the alliance between A and B transitions to the bilateral fulfilment game depicted in Figure 2. Substantively, this indicates that if S exits the alliance, A and B will continue in sustaining the alliance only if both groups trust that their counterpart will also sustain co-operation in the absence of external enforcement. Therefore, we see that for both A and B, allowing S to participate in the alliance contains some element of risk. On one hand, if the groups are unwilling to trust each other without an external enforcement mechanism, and there is some possibility that S will eventually abandon the alliance, there is some risk that both A and B will at some point become vulnerable to the other's exploitation. On the other hand, if both A and B are able to co-operate in the bilateral setting, S's decision to abandon the alliance does not alter the incentive to co-operate. However, since S's participation further distorts the agenda of the alliance, it is possible that these groups will prefer bilateral co-operation as opposed to asymmetric co-operation.

Solution

As in the previous model, the decision of the militants to become allies is determined by G's power to repress the groups. In the previous game, G's divide and conquer strategy effectively prevented relatively weaker militants from forming alliances, due to their mutual fear of betrayal. In this version of the game, however, G's power to divide and conquer is undermined by S's threat to punish defectors α. While both A and B always exploit if 0 > 1−β(p,q)−C _{(A,B )} in the bilateral game, A and B will only exploit in an asymmetric alliance if −α _{(A ,B )} > 1−β(p,q)−C _{(A ,B )}. We see that exploitation is more difficult for both A and B if S governs the alliance. Additionally, if S's punishment is sufficiently large, A and B will require additional concessions to offset the cost of punishment α, which raises the price G must pay to break the alliance. If S is part of the alliance, G negotiates with A if:

$$ \beta q\left[ {1\,{\rm{ - }}\,\left( {\frac{{\delta {{\delta }_s}}}{{1\,{\rm{ - }}\,\left( {1\,{\rm{ - }}\,\frac{{{{C}_S}\,{\rm{ - }}\,1}}{{{{C}_S}}}} \right)(pq\theta )}}\,{\rm{ - }}\,{{C}_A}\,{\rm{ - }}\,{{\alpha }_A}} \right)} \right]\, + \,(1\,{\rm{ - }}\,\beta q)({\rm{ - }}{{C}_G})\, \gt \,pq\theta \,{\rm{ - }}\,{{C}_G}. $$

Therefore, a credible alliance forms if:

$$ \beta q\left( {1\, + \,{{C}_G}\, + \,{{C}_A}\,{\rm{ - }}\,p\,{\rm{ - }}\,{{\alpha }_A}\, + \,\frac{{\delta {{\delta }_S}}}{{1\,{\rm{ - }}\,\left( {1\,{\rm{ - }}\,\frac{{{{C}_S}\,{\rm{ - }}\,1}}{{{{C}_S}}}} \right)(pq\theta )}}} \right)\, \gt \,0. $$

In bilateral conditions, the militants’ mutual fear of exploitation could prevent alliances from forming unless both groups were relatively more resistant to collapse. However, with a state sponsor, the threat of sponsor punishment increases the price governments must pay. As intuition suggests, the model predicts that as α _{(A ,B )}→1, indicating greater levels of punishment for defectors, both groups are more likely to sustain co-operation, indicating that the sponsor's threat to punish shores up the weak link in the alliance. In effect, the sponsor serves as an institution to constrain the groups from breaking contracts. Interestingly enough, this logic appears to resemble the traditional role of the mafia boss, which is intended to enforce contracts among illegitimate entities.Footnote 33 Similarly enough, the sponsor acts in the same capacity by constraining the weak link from behaving opportunistically. These results reflect the intuition that weaker groups should band together, while simultaneously addressing the concern that agreements must be credible. If a sponsor intervenes, militants become credible by building security institutions, not unlike states in the international system. While we typically think of militants as operating outside legal structures, the model seems to indicate that unless groups are very strong they should prefer asymmetric security agreements that will constrain their behaviour.

Yet, the introduction of a sponsor does not always improve the prospects for co-operation due to the uncertainty over the sponsor's willingness to sustain the alliance. From the bilateral game, we know that if A and B are capable of resisting G's efforts to suppress, they can sustain an alliance if Condition 2 is false. Substantively, this indicates that if the groups are facing a relatively weaker G, they are capable of sustaining an alliance with or without S's assistance. S's decision to exit the alliance may actually be beneficial to the two groups, given that S further distorts both groups’ agenda. However, if Condition 2 is true, this indicates that A and B cannot co-operate in the bilateral setting, and require S's assistance to sustain the alliance. In this case, G is relatively stronger, and both A and B have incentives to exploit each other. A and B can only be deterred from exploiting each other if S participates in the alliance and threatens to punish either group should they defect. Once S abandons the alliance, neither A nor B can credibly commit to sustaining the alliance, which will cause co-operation amongst the militant groups to collapse. We see that the decision to participate in an asymmetric alliance is risky for weaker A and B, and depends on both groups’ belief that S will sustain the alliance indefinitely.

Figure 5 replicates Figure 3 using the asymmetric conditions. Unlike the bilateral game, we see that asymmetric alliances are more likely to form if one or both of the alliance partners are weaker. If S's governance costs are high, S is increasingly likely to abandon the alliance at some point, which will eventually undermine the incentive to sustain co-operation with the other group. However, we see that if {p,q}→1, the groups are likely to co-operate with S, even if S is likely to abandon the alliance at some point in the future. On the one hand, since these groups are likely to disarmed by G, these weaker groups are increasingly willing to risk co-operation with S to give themselves some chance of survival. Substantively, this indicates that desperate militant groups facing stronger government adversaries are likely to accept S as a guarantor, even if doing so will only temporarily extend their alliance. On the other hand, we see that co-operation is also likely if (p,q}→0, indicating that G is less likely to repress the groups. In these cases, Condition 2 is likely to be false, indicating that both A and B can simply transition to bilateral co-operation if S abandons the alliance. These results suggest that S's participation has no effect on the probability that relatively stronger groups co-operate, but has a significant effect on the probability that weaker groups co-operate. While weaker groups are more likely to seek sponsorship than stronger groups, sponsors can free-ride if the groups are stronger. We know that if, on the one hand, A and B cannot sustain a bilateral alliance, the probability that S sustains co-operation indefinitely is:

$$V_{S}^{\ast } \, = \,\frac{{{{C}_S}\,{\rm{ - }}\,1}}{{{{C}_S}}}.$$

On the other hand, if A and B can sustain bilateral co-operation without sponsorship, the probability that S sustains co-operation throughout the alliance is:

$$ V_{S}^{{\ast\ast }} \, = \,\frac{{{{\delta }_S}\, + \,{{C}_S}\,{\rm{ - }}\,{\rm{1}}}}{{{{C}_S}}}. $$

Fig. 5

Utility to militant group A for asymmetric co-operation as a function of the government's ability to disarm militant groups A and B

For S to sustain co-operation, it must be true that $ {{V}}\, \gt \,V_{S}^{\ast } $ in the first case and $ {{V}_S}\, \gt \,V_{S}^{{{\rm{\ast}}\ast }} $ in the latter. Since $ V_{S}^{{{\rm{\ast}}\ast }} \, \gt \,V_{S}^{\ast } $, we see that it is more difficult for S to sustain an alliance indefinitely in the second case as opposed to the first. This demonstrates that S is more likely to sustain co-operation indefinitely when A and B are too weak to form an alliance on their own as opposed to cases where S can free-ride. Therefore, we see that sponsors are more likely to intervene in alliances if the militant groups that are facing governments with relatively greater repressive power. Further, we see that, as in the previous case, S is more likely to intervene as C_S→0, indicating that S prefers to intervene if the cost of governing the alliance between A and B decreases.

Implications

The asymmetric model gives us several insights into the alliance behaviour of militant groups. Militants facing governments with substantial repressive power have incentives to form alliances, yet these groups have little credibility in the eyes of potential alliance partners. To overcome this problem, militants have incentives to bind themselves to a sponsor, which signals credibility to possible alliance partners. Once vulnerable groups establish credibility, they are capable of co-operating with other groups, which in turn increases the prospects of resisting the government's repressive efforts. We also see an additional role typically not considered for sponsors of militant groups. Rather than just serving as a pool of potential resources, sponsors can serve a key function by constraining militants from negotiating with target governments, particularly when target governments are powerful. This may explain why governments are very quick to denounce states for supporting militants. Sponsors effectively eliminate the ability of governments to use negotiation for counter-insurgency, which appears to be one of their most effective weapons. We further see that sponsors are selective as to when they take on this role. In cases where the groups are already capable of forming alliances, sponsors will free-ride off the efforts of the groups rather than directly participate. However, if the militant groups are too vulnerable to co-operate credibly on their own, sponsors are more likely to intervene. Empirically, we observe this behaviour in several cases, including Iran's assistance in coalescing multiple groups into Hezbollah, American and Saudi co-ordination of elements of the mujahedeen, and Pakistani use of various Kashmiri groups. In addition to this anecdotal evidence, Asal et al. provide some evidence that sponsorship may increase the ability of groups to seek partners, though this analysis does not specify the conditions under which sponsorship can be either helpful or harmful.Footnote 34 We now present a research design to test the model's hypotheses.Footnote 35

Theoretical Variables

Government repressive power against the weak link

According to the theoretical model, the government's equilibrium behaviour is to co-opt the weak link, or the weaker of the two groups, in order to break the alliance. We therefore need some indicator that captures the ability of the weaker group to withstand the government's repression. We capture this variable using Cunningham et al.’s data on dyadic relations between violent non-state actors and governments.Footnote 45 The expanded non-state actor conflict data (NSA) contain a variable that dyadically estimates the power of various militant groups relative to their target governments. The scale ranges from much weaker, indicating the government is far more powerful and likely to disarm the group, to a stronger score, indicating that the government is unlikely to disarm the group. We compile the NSA strength score for both groups in each of the conflict dyads of the dataset. Since the theoretical prediction is that the government will offer concessions to the weaker of the two groups, we use the weaker of two NSA strength scores as our indicator to capture the government's repressive power against the weak link. If the dyadic pair contains a much weaker link, or a group that is quite vulnerable, our expectation is that an alliance will not form unless a sponsor is present. To illustrate the coding procedure, consider the case of the South African conflict during the 1980s. UCDP/PRIO identifies three major groups engaging in political violence against the state: the African National Congress (ANC), the Pan Africanist Conference (PAC) and the Azanian People's Organization (AZAPO). Both PAC and AZAPO are coded as being ‘much weaker’ than the South African state, whereas the ANC is coded as ‘weaker’. Therefore, in the dyad between PAC and AZAPO, the weak-link variable would be coded as ‘much weaker’, given that neither group was likely to resist efforts by South Africa to co-opt. Similarly, in the dyad consisting of the ANC and PAC, the weak-link variable is also coded as ‘much weaker’. While the ANC was relatively stronger than the PAC and, therefore, better able to resist the state's efforts to co-opt it, PAC was a weaker organization and was therefore vulnerable to becoming co-opted by the South African government. The weak-link variable captures the crucial component in predicting whether alliances form, and what type of alliances form. On the one hand, if the weak link in the dyad is much weaker than the state, we would expect these groups to be less likely to form bilateral alliances, but more likely to form asymmetric alliances with the help of a sponsor. On the other hand, if the weak link in the dyad is at parity or stronger than the state, this would suggest that the values of {p,q} are much lower, indicating that bilateral alliances are more likely to form and asymmetric alliances are less likely to form. When merged with the alliance data, the NSA data consist of groups with five categorizations of the relative power of the group to the state: much weaker, weaker, parity, stronger and much stronger. Using the dyadic pairings, our weak-link variable is distributed as follows:

We see that the weak-link variable is heavily skewed towards the weaker categories. Unfortunately, this scaling creates two categories that account for 97 per cent of the variation in the key independent variable. Therefore, we use three different measures of the weak-link variable in our analysis. First, we use the weak-link scale as it is, despite the fact that two of the categories consist of only 3 per cent of the observations combined. We then create a second, dichotomous weak-link indicator that simply combines categories (1, 2 and 3). This indicator essentially differentiates those groups that have a much weaker link from those whose weak links have a score that is > 0. We believe that these three indicators collectively will allow us to test both Hypotheses 1 and 2, since the indicator exhibits variation and best captures how the weak link affects the probability of an alliance.Footnote 46

Alliance Control Variables

We include several control variables to test both Hypotheses 1 and 2, and Hypothesis 3. To test the former set, we first include the state's per capita gross domestic product (GDP) score, which Fearon and Laitin use a proxy for state strength.Footnote 47 The authors argue that increased per capita GDP scores serve ‘as a proxy for the state's overall financial, administrative, police and military capabilities and … will mark more developed countries with terrain more “disciplined” by roads and rural society more penetrated by central administration’.Footnote 48 The authors further argue that lower per capita GDP scores may make joining an insurgency more appealing to individuals, given the poor prospects of high economic returns from the legitimate economy. The per capita GDP variable, therefore, serves as a continuous indicator of the government's ability to project power over its territory. While our key independent variable captures the strength of the groups relative to the government, we believe that we must also control for the government's absolute ability to control and police its territory. The mean GDP per capita score is 1.96, with a min of 0.196 and a max of 11.3.

Sponsorship Control Variables

To test Hypothesis 3, we also require several control variables to capture the dynamics between the potential sponsor and the target government. By using only Thompson rivals, we control for the level of hostility between the two states. However, we also require measures of several other concepts that are systematically correlated with hostile international interactions.Footnote 54 These variables are as follows.

Method

The data to test both Hypotheses 1 and 2 are organized as a cross-sectional time series using a binary dependent variable (BTSCS).Footnote 58 Since both the alliance and sponsorship variables are dichotomous, we test the hypotheses using a probit model. We first test Hypotheses 1 and 3 using the dataset on potential alliance formation. Hypotheses 1 and 2 argue that if the government's repressive capacity is low the militants are more likely to form bilateral alliances, whereas these groups require a sponsor to assist them when the government's repressive capacity is high. We test these hypotheses using three models. We first present the results using the first weak link score, followed by the results produced by the second. We then present a model that uses the second weak-link score and includes regional control variables for sub-Saharan Africa, North Africa/Middle East and Asia. Latin America is left as the excluded category.Footnote 59 Each model is estimated with robust standard errors clustered on the government.

We next test Hypothesis 3 using the set of potential sponsors. The key explanatory variable in this test is the government's repressive power against the alliance. Our expectation is that the potential sponsors are more likely to intervene if the government's repressive power is relatively high, but are more likely to free-ride (and therefore not intervene) if the government's repressive power is weaker. Unfortunately, the loss of observations resulting from the use of the weak-link indicator results in only eight observations of sponsorship, which are perfectly predicted by all three of the weak-link indicators. Therefore, we follow Fearon and Laitin's 2003 study and use per capita GDP as a proxy for the government's repressive power against the weak link. Again, each model is estimated with robust standard errors clustered on the government.