The spillover of war in time and space: exploring some open issues

ABSTRACT

In spite of a voluminous empirical literature, the diffusion of civil war in time and space is still not fully understood and several issues remain open for debate. This article sheds new light on some of these issues. First, we assess the robustness of results to changes in the definition of neighbourhood (spatial dimension) and in the lag structure (temporal dimension). Second, we investigate the factors that determine the strength of civil war diffusion, focusing in particular on (i) the type of war (e.g. its intensity and scope), (ii) the quality of domestic polity, and (iii) interaction opportunities (e.g. ethnic ties and refugee flows). Third, we look at how the occurrence of war in a neighbouring country affects the duration (and not just the onset) of domestic civil war.

KEYWORDS:

• Spatio-temporal models
• civil war
• diffusion
• duration analysis

JEL CLASSIFICATION:

• C23
• C25
• D74
• N40
• O11

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ In the rest of this article we use the terms spillover, diffusion, and contagion interchangeably to indicate the spread of civil war from one country to its neighbour(s).

² In which case, one must decide between adopting a weighting system or counting each neighbour at war as one (Buhaug and Gleditsch 2008).

³ Colonial wars are no longer fought. Nevertheless the test of this hypothesis is still relevant to assess how their inclusion/exclusion affects the empirical results. Moreover, there are new forms of extra-systemic wars (e.g. those related to regime change or the pursuit of terrorist groups beyond state’s border) that are likely to become more frequent in the future and that share some common features with colonial wars.

⁴ Collier and Hoeffler (2001) is the working paper version of Collier, Hoeffler, and Söderbom (2004). Interestingly, the neighbourhood variable appears in the working paper version, but not in the published version.

⁵ Our preference for a random effect specification stems from two considerations. First, theoretical work shows that (i) the fixed-effects estimator is biased even when the time series dimension of the panel is fairly large and (ii) in a probit model, ignoring unobservable individual heterogeneity is not necessarily worse than using the fixed effects specification to account for it (Greene 2004). Second, the fixed effects estimator requires all countries that never experienced a civil war be dropped, drastically reducing the number of observations available for estimation and could actually generate a form of sample selection bias.

⁶ When a 100-km radius is used to define the neighbourhood, results are very similar to those obtained for a definition of neighbourhood based on land borders. These additional results are available upon request.

⁷ Each subsample includes only a fraction of the 4733 observations available for estimation in the full sample. Thus the standard errors of the estimated βs in Figure 1 are necessarily larger than those reported in Table 1.

⁸ If t_nB > t_n, then the variable would be equal to the time between t₀ and t_n; that is, the duration of the domestic civil war. If t_0B > t_nA (so that between t_nA and t_0B there is no war in the neighbourhood), then the variable is equal to the time between t₀ and t_nA plus the time between t_0B and whichever is smaller between t_n and t_nB.