Targeting and Resistance: Reassessing the Effect of External Support on the Duration and Outcome of Armed Conflict

ABSTRACT

This article draws a distinction between external support which primarily serves to enhance rebel capacity to offensively target vital state interests and support which primarily increases rebel capacity to defensively resist state repression. Targeting support increases a rebel group’s incentive to behave aggressively, and is found to be associated with a shorter conflict duration when given to strong groups and a higher probability of a decisive conflict outcome. Resistance support increases a rebel group’s incentive to prioritise survival, and is found to be associated with a longer conflict duration.

Acknowledgments

I would like to thank the following people for their useful comments and helpful guidance in developing this paper: Kyle Beardsley, Laia Balcells, Idaean Salehyan, and Patricia Sullivan.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1. While some prior studies have considered direct intervention to be qualitatively different from support, here direct intervention is treated as merely one form of support. While it is likely the most costly form a support a sponsor can provide and tends to have the largest impact on conflict, there is no reason to treat direct intervention as conceptually distinct from other forms of support. Additionally, while the inclusion of this support category makes theoretical sense, its exclusion does not change the empirical results of the tests below.

2. Cunningham (2010) separates instances of interventions by the motivation of the intervener (independent or not independent of the supported combatant), and Regan and Meachum (2013) separate interventions into military, economic, and diplomatic. However, neither of these studies differentiated external military support by content. The only other study to have made such a differentiation and use econometric techniques to demonstrate the utility of such a differentiation, Sawyer et al. (2015), is discussed below.

3. It is also important to clarify that the targeting-resistance distinction is different from the notion of active and passive sponsors of terrorism present in Byman (2005). All rebel support considered here is actively provided. For example, when the provision of foreign sanctuary is considered below, only cases in which safe haven was intentionally provided are taken into account.

4. Additionally, these two forms of support are operationalised in a way that ensure that neither form of support reflects a higher level of commitment than the other. While it could have been plausible to measure these concepts in a way that meant targeting was associated with higher levels of commitment than resistance support, this is avoided by ensuring that only the intentional provision of safe haven in counted in resistance support. Furthermore, targeting support comes in the form of weapons much more often than in the form of direct intervention. While the magnitude of support is an important factor worth developing measures for, that is not within the scope of this article.

5. While it is said here that the effect of targeting support provided to weak groups is indeterminate, it is perhaps more accurate to say that it is dependent on the magnitude of the support received. Measuring such a magnitude is beyond the scope of this study, and therefore the effect of targeting support for weak groups is considered to be indeterminate for the purpose of the empirical tests conducted in this article.

6. Since Kalyvas and Balcells (2010) do not code technologies of rebellion for all civil conflicts, but only for full-scale civil wars, their categorisation can not be directly used in the analysis below without losing the majority of observations.

7. Given the small number of rebel victories since the end of the Cold War (6), it is not feasible to analyse the conditions leading to rebel or state victory independently. They are pooled here since the argument being tested has identical expectations for which support types affect the probability of all decisive outcomes.

8. Similar results are obtained if coups are excluded from the analysis entirely.

9. A test of the proportional hazards assumption showing that this model is appropriate can be found in Table 1 of the Appendix.

10. Separate analyses are conducted for duration and outcome to allow maximum comparability between the results and those in Cunningham et al. (2009). A competing risks approach (which combines the modelling of duration and outcome) has been used in some of the previous literature, but may not be entirely justifiable in this case. A competing risks approach assumes the independence of irrelevant alternatives (IIA), meaning that the probability of any failure event at a given time is not affected by the probability of any other failure event at that time. For example, the likelihood of rebel victory at time t must be independent from the likelihood of rebel defeat at t, which seems like an unsafe assumption to make. For this reason, as well as comparability with the results in CGS, duration and outcome are modelled separately instead of with a competing risks approach.

11. For example, in a three-year conflict-spell in which targeting support is provided in years one and two, but not year three, the rebel group will be recorded as having received 1, 2, and 3 years of support.

12. The rebel strength variable used here is strength independent of external backing. The correlation between a group being weak in a given year and receiving targeting support in that year is < |.1|.}.

13. An additional model was fitted to ensure that civil conflict in neighbouring states did not alter the effects of targeting or resistance support, and the results of this model are presented in Table 2 of the Appendix, available on the author's website.

14. More formally, if a multivariate normal distribution is defined as N(μ, ∑), a vector of 100,000 new coefficient values ($\tilde{\beta }$) are drawn from the distribution N ($\hat{\beta }$, cov($\hat{\beta }$)), where $\hat{\beta }$ is the vector of coefficients estimated by the model and cov($\hat{\beta }$) is the variance-covariance matrix from the fitted model.

15. While this type of simulation is usually used to simulate predicted values of the dependent variable, the CPH model is semi-parametric (the baseline hazard ratio is left unspecified), which restricts the analysis to the simulation of hazard ratios. This means that the relative risk of the failure event can be estimated, but not the absolute risk.

16. Full results from the model are in Table 3 of the Appendix. A ‘spaghetti plot’ showing the uncertainty around the marginal effects for this interaction are in Figure 3 of the Appendix.

17. Additional models were fitted to ensure that civil conflict in neighbouring states did not alter the effects of targeting or resistance support, and the results of these models are presented in Table 4 of the Appendix.}.

Additional information

Notes on contributors

Jordan Roberts

Jordan Roberts is a PhD Candidate at Duke University and a Pre-Doctoral Fellow at the University of Texas at Austin.