On the prediction of intrastate conflicts
This is a hopeful blog: can we predict when civil wars are bound to occur ? If not, can we stop them ? Such events are profoundly destructive for the citizens of these countries, and their effects are now inevitably globalized. Predicting these conflicts early on presents a major challenge for policymakers in an effort for peace. By successfully doing so, the identification of potential causal mechanisms would unravel a world of possibilities when it comes to understanding and defusing these catastrophes.
Economical crisis, infant mortality, anocracy… When a civil war occurs in a country, retrospectively, one can say “of course this is happening, this country had so many major issues !”. But which issues in particular lead to civil wars ? Are some more influential than others ? Are there patterns triggering civil wars, some sort of evil secret sauce ? Can we avoid civil wars by predicting them, and acting before the irreparable ? That’s what we will try to find out.
This decade has seen the increasing use of machine learning techniques, deep learning in particular, in a broad range of areas. This renewed enthusiasm, in part triggered by the dramatic increase in data and computing power, has implications in the domain of the political sciences. In 2015, Muchlinski and colleagues released a study on the use of Random Forests, a machine learning technique well suited for the prediction of rare events: Comparing Random Forest with Logistic Regression for Predicting Class-Imbalanced Civil War Onset Data. One of their findings was that this machine learning technique could rival with classical statistical methods in predicting civil war onsets, and brought some new insights into the emergence of these conflicts by analyzing the predictive power of the variables used.
In a similar fashion, our study aims at pushing this comparative analysis further, in the realm of Artificial Neural Networks, and Multi Layer Perceptrons (MLPs) in particular. Since Random Forest algorithms are powerful at handling severe imbalance in the data to answer such questions, we wish to study if MLPs can do the same. If proven useful, MLPs could, for instance, enhance the prediction of rare events in conflict data by leveraging their high flexibility in input modality: be it temporal, linguistic or visual data. To motivate their use, we will focus on four different interrogations about these conflicts:
This study presents some of the tools one can use when dealing with the prediction of rare events. We will go through the pipeline and analytical work needed to derive conclusions from predictive models, and raise some important questions as to the validity of these conclusions. But before diving into the core of this study, let us ask ourselves:
Mark Gersovitz and Norma Kiger define a civil war as “a politically organized, large-scale, sustained, physically violent conflict that occurs within a country principally among large/numerically important groups of its inhabitants or citizens over the monopoly of physical force within the country”.
Between 1945 and 2000, no less than 116 civil wars occurred across the world. The map below shows the number of years a country has been in civil war. As we can see, Southern countries are the most impacted, in Asia, the Middle East, Eastern Africa and South America. Some countries, such as Colombia, have been in civil war for decades !
In this project, we were particularly interested in the onset of civil wars, which are presented in the map below: