Democratization and civil war

Abstract

This article examines the impact of civil war on democratization, particularly focusing on whether civil war provides an opportunity for institutional reform. We investigate the impact of war termination in general, along with prolonged violence, rebel victory and international intervention on democratization. Using an unbalanced panel data set of 96 countries covering a 34-year period, our analysis suggests that civil war lowers democratization in the succeeding period. Our findings also suggest that United Nations intervention increases democratization, as do wars ending in stalemates. However, wars ending in rebel victories seem to reduce democratization. These findings appear robust to conditioning, different instrument sets, modelling techniques and the measurement of democracy.

Keywords:

• civil war
• democracy
• conflict
• democratization
• outcomes of war

JEL Classification:

• H56
• N40
• O11

Notes

¹ We obtained similar results examining rents as a share of population and excluding all but oil rents. These results are available upon request.

² We run Fisher’s test without and with a trend variable for democracy, log of GDP, openness to international trade and population, among others. Detailed test statistics are available upon request.

³ Comparing a two-way random-effects GLS estimator and a two-way within estimator, we reject the null hypothesis that the differences in the two sets of estimated coefficients are not systematic with a chi-squared test with 11 degrees of freedom and a resultant test statistic of 23.07.

⁴ We employ a Breusch–Pagan test and reject the null hypothesis of homoscedasticity with a chi-squared test with 1 degree of freedom and resultant test statistics of 13.95 and 56.89 for the within estimator without and with a lagged dependent variable, respectively.

⁵ We employ the Wooldridge test for autocorrelation in the panel data and reject the null hypothesis of no first-order autocorrelation with a F(1,87) test statistic of 39.802 and 92.771 for the within estimator without and with a lagged dependent variable, respectively.

⁶ GMM estimators with too many moment conditions can be subject to overfitting biases in small samples (Bond, 2002). We thus compare the unrestricted and restricted estimates, and the loss of information from deep lags is thought to be minimal.

⁷ We also estimate the one-step GMM estimator with the lag-limits set to three or greater; the one-step GMM estimator with the lag-limits set to three, and the one-step GMM estimator with collapsed, forward orthogonal instruments. These estimates are available upon request.

⁸ The full results, including estimated coefficients and SEs for the control variables, are available upon request.