WHEN AND HOW THE FIGHTING STOPS: EXPLAINING THE DURATION AND OUTCOME OF CIVIL WARS

Abstract

Previous research has shown that the duration of a civil war is in part a function of how it ends: in government victory, rebel victory, or negotiated settlement. We present a model of how protagonists in a civil war choose to stop fighting. Hypotheses derived from this theory relate the duration of a civil war to its outcome as well as characteristics of the civil war and the civil war nation. Findings from a competing risk model reveal that the effects of predictors on duration vary according to whether the conflict ended in government victory, rebel victory, or negotiated settlement.

Keywords:

• Civil war
• Conflict resolution
• Duration
• Competing risks

JEL Codes:

• C41
• D74

Notes

¹ An alternative control for economic development would be to use a measure of income inequality, since this is often hypothesized to be a cause of civil conflict. There are at least ten different ways/combinations in which income inequality can be measured, and depending on which one is adopted, different trends in inequality may result (Wade, 2004). The Gini coefficient is one way to measure inequality, but it gives excessive weight to the changes around the middle of the income distribution, at the expense of changes at the extremes (Cowell, 1977; Kawachi and Kennedy, 1997; Wade, 2001). Because our concern is controlling for income inequality's impact on civil wars, we would not be as concerned with changes around the middle, given that it is perception of increasing inequality in relationship to others that often increases the dissatisfaction with the government of those at the bottom of the economic ladder. Therefore, the Gini coefficient is not an appropriate measure. Further, the best source of Gini coefficients that we have found (e.g. the University of Texas Inequality Project) only covers about half of the country‐years in out dataset. For the cases where we do have Gini coefficients, it correlates with our GDP per capita measure at −0.7.

³ Missing values for the military size or population variables were replaced with their most recently measured value for each missing civil‐war year.

⁴ We initially tried to separate out the interventions as pro‐government, pro‐rebel and neutral. However there is not enough variation in the intervention types across the war outcomes to make this strategy feasible in the analysis.

² We combined negotiated settlements and truces from Doyle and Sambanis (2000). Otherwise the estimation of the models would be complicated by small numbers of those two outcomes.

⁵ These ongoing civil war cases are treated as right censored in the event history models.

⁶ We estimated a multinomial (conditional) logit model with a log time variable for our data (not reported here) and it produces qualitatively similar results to the Cox regressions presented here.

⁷ While not a proper test for the pooling, the fact that the sum of the log‐likelihoods for each of the outcomes is more than that for the pooled model suggests that the competing risks model is preferable.

⁸ The odds or odds ratio is the percentage change in the hazard for a one‐unit change in the covariates. It is computed by taking the base e exponent of the Cox regression model coefficients or from the hazard ratios. An odds ratio of 0 means that there is a −100% change in the hazard; 1, means that there is no effect; 2 is a 100% increase in the hazard.

⁹ Ideally, we would like to have data on casualties or deaths from conflict by year, but such data are not available. We employed the COW data and Doyle and Sambanis (2000) for total deaths from the conflict, scaled by population.

¹⁰ For the variables with interaction coefficients, the significance is based on a likelihood ratio test for the joint significance of the coefficient and its time interaction.