We investigate the determinants of onset, duration and incidence of civil wars, and their sensitivity to different coding rules. Whatever the coding rule used, incidence of civil war is largely determined by poverty, country size, mountainous terrain and ethnic diversity. Poverty reduces the opportunity cost of rebellion, and mountainous terrain makes rebellions harder to defeat, while ethnic diversity creates potential conflict, since some linguistic groups may feel under-represented. Continuation and onset of civil war are non-overlapping subsets of incidence, depending on whether civil war occurred in the previous year. We exploit this aspect to test whether the determinants of continuation and onset are statistically significantly different. For four out of five coding rules, we cannot reject the null hypothesis of common determinants of civil war onset and continuation. This is the main contribution of the paper. Ethnic diversity is particularly high in Africa, where many civil wars occur, but we show that the significance of ethnic diversity is not just an unidentified Africa effect. Our estimated model works as well for Africa as for other areas. There is evidence that ethnic polarization matters as well as fractionalization. We show that polarization understates the propensity for conflict in societies with high ethnic diversity, where the polarization measure tends to be low, but otherwise outperforms fractionalization as a predictor of conflict. Mountainous terrain is generally not significant when ethnic polarization is included.