Replication Data for: 空间依赖与武装冲突预测

武装冲突的分布呈现空间依赖特征,具有较高概率扩散至其邻近地区。但是,既有的冲突研究,尤其是冲突预测研究并没有对空间依赖特征进行充分利用,更忽略了影响冲突的因素在更加精细的空间层次上的互动,这也制约着冲突预测向更加精细化的方向发展。在大数据时代,关于冲突的空间依赖理论不断发展,以及冲突相关空间数据与空间分析方法不断更新,都为研究者实现理论与实证的有机结合提供了机会,使他们可以更好地利用冲突的空间依赖特征为冲突预测服务。本文回顾了处理空间依赖特征的一般路径,阐述了在更加精细的时空单位捕捉空间依赖的一种路径,并基于缅甸的国内冲突案例(2010—2020年),借助机器学习的框架,通过分离总体持续期模型与集成贝叶斯模型平均方法,展示在大数据时代认真对待空间依赖性可以在更加精细的时空维度进一步提升冲突预测的准确率。本文的分析表明,通过充分利用研究对象本身的空间依赖性质进行模型建构,并辅以恰当的机器学习方法,即使模型中只有少量随时间变化的变量,也可以实现非常高的预测准确度。本文的研究路径因此对建立关于“一带一路”沿线国家的武装冲突的预警预测系统具有较大的政策启示。

The spatial distribution of armed conflicts exhibits spatial dependence, with a high probability of spreading to adjacent regions. However, existing conflict studies, especially conflict prediction research, have not fully exploited the spatial dependence characteristic, and have even neglected the interactions of conflict-influencing factors at finer spatial scales, which restricts the development of more refined conflict prediction. In the era of big data, the continuous development of spatial dependence theories related to conflicts, as well as the continuous updating of conflict-related spatial data and spatial analysis methods, provide researchers with opportunities to organically integrate theory and empirics, enabling them to better leverage the spatial dependence characteristic of conflicts for conflict prediction. This paper reviews the general approaches to addressing spatial dependence, elaborates on a method for capturing spatial dependence at finer spatiotemporal units, and based on the case of domestic armed conflicts in Myanmar (2010–2020), uses a machine learning framework combined with the split population duration model and ensemble Bayesian model averaging, to demonstrate that taking spatial dependence seriously in the big data era can further improve the accuracy of conflict prediction at finer spatiotemporal dimensions. The analysis in this paper shows that by fully leveraging the spatial dependence nature of the research object for model construction, supplemented by appropriate machine learning methods, even with only a small number of time-varying variables in the model, very high prediction accuracy can be achieved. The research approach of this paper thus has significant policy implications for establishing early warning and prediction systems for armed conflicts in countries along the "Belt and Road Initiative".