The Effect of Child Soldiers on Rebel Violence against Civilians

ABSTRACT

Existing work describes child soldiers as very violent towards civilians. Challenging this, I posit that children’s effect on group behaviour is conditioned by rebels’ civilian support. Because they have weak pre-existing norms, children are both prone to normalize violence and susceptible to rebel efforts to control their use of violence. They should thus closely follow group rules in their behaviour towards civilians, implying a moderating effect of these rules. I expect that child soldiering increases civilian victimization only for groups who lack incentives to show restraint towards civilians because they receive no support from them. Empirical tests support this expectation.

Acknowledgments

I thank Tobias Böhmelt, Ursula Daxecker, Christoph Dworschak, Tatiana Gelvez, Kristian Skrede Gleditsch, Martin Steinwand, Paul W. Thurner, two anonymous reviewers, the editors, as well as audiences at the Jan Tinbergen European Peace Science Conference 2019 and at the University of Essex for very helpful comments on this paper.

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Notes

1. One exception are studies that differentiate between recruits motivated by material and non-material incentives and examine how these individuals’ decision to desert are affected by political training (Weinstein 2005, 2006, Oppenheim et al. 2015).

2. While these studies of rebels all rely on evidence that is anecdotal or based on small convenience samples, common experiences of violence have generally been found to instil in-group cohesion and out-group antagonism that may be detrimental in post-conflict contexts but highly suited to create an effective fighting force (See e.g. Miguel et al. 2011, Voors et al. 2012, Cecchi et al. 2016).

3. This allows me to count smaller-scale violence against civilians and to differentiate between acts of one-sided violence committed inside and outside the conflict country. My main analysis employs a variable counting all violence against civilians. Results are robust to employing dependent variables counting violence in the conflict country only or using the UCDP 25 death threshold.

4. While the one-sided violence data is organised in actor-years, many of my control variables have a dyadic format and the sample is defined as groups that achieve 25 battle-related deaths while fighting one specific government. Groups active in multiple states can thus fight in multiple dyads in one year.

5. There is empirical support for using ZINBs instead of standard Negative Binomials (Greene 2012, p. 861ff.; Hilbe 2011, p. 371–379). In both main models, Vuong tests result in large and positive z-values, favouring the ZINB (Vuong 1989, Greene 2012, p. 863). As the Vuong test has been argued to be inappropriate for comparing overlapping models (Santos Silva et al. 2015, Wilson 2015), I also use alternative HPC tests (Santos Silva et al. 2015), which favour the ZINB.

6. Haer and Böhmelt also provide an ordinal indicator which codes child soldiering as non-existent (0), intermediate with children comprising less than 50% of a group’s forces (1), or high with children outnumbering adults (2). However, this more differentiated variable appears to suffer from substantial coding issues (Haer and Böhmelt 2017). Models using this alternative variable are presented in the appendix and mirror my main models in terms of their substantive results. This dataset has consequently also been expanded (Haer et al. 2019). However, this was done by differentiating by whether groups used forced or voluntary recruitment to enlist child soldiers, not by extending the period of observation. I hence use the original child soldiering dataset provided by Haer and Böhmelt.

7. It is crucial that this cooperation and support is voluntary and not the result of coercion. To ensure that mobilisation does not capture coercive actions by rebel groups, I present models in the appendix that explicitly control for rebels’ use of forced recruitment.

8. This observation is relevant as separatist groups are, all else equal, less likely to recruit children in the first place (Lasley and Thyne 2015).

9. See e.g. Faulkner et al. (2019), Haer et al. (2019), Lasley and Thyne (2015), Tynes and Early (2015), and Vargas and Restrepo-Jaramillo (2016).

10. Summary statistics are reported in appendix A where the controls are also discussed in more detail.

11. Models 2 and 4 also include some statistically significant results for control variables, most prominently mobilisation. I refrain from interpreting these results because these model specifications were chosen with only the effect of child soldiers in mind. The control variables were hence selected to control only for confounders regarding the relationship between child soldiering and violence against civilians. The coefficient estimates of anything but child soldiering and the interaction term may hence result from omitted variable bias as no efforts were made to achieve unconfoundedness for these variables, making them substantively mostly meaningless (See Cinelli and Hazlett 2020, p. 44–45, Hünermund and Louw 2020).

12. In addition, Figure 1 also indicates that the number of civilians killed by groups that use child soldiers and can mobilise local support is lower and statistically distinguishable so from the number of civilians killed by groups who employ children but have no such support (Loftus and Masson 1994).

13. See the appendix.

Additional information

Notes on contributors

Marius Mehrl

Marius Mehrl is a Postdoctoral Researcher at the Geschwister Scholl Institute of Political Science. He received his PhD in Government from the University of Essex in 2020. In his research, he uses quantitative methods to study questions relating to armed conflict, civil-military relations, and the international arms trade. His work on these topics has been published in journals including Conflict Management and Peace Science, International Interactions, and the Journal of Conflict Resolution.