Final Project - Executive Summary
Ava Cheevers
3/9/2021
Executive Summary
For events as destabilizing and dangerous as they are, coups are surprisingly frequent. The Cline Center at UIUC has compiled a comprehensive dataset of nearly 1000 coup conspiracies, coup attempts, and realized coups that have occurred worldwide since 1945.
I am using this data, supplemented with temporal nation-level data, to answer the research question of can we predict a coup attempt’s outcome based on attributes surrounding the coup and the nation in which it is occurring?
The Cline Center dataset contains many dummy variables on the attributes of the coup or coup instigators. Seen below, the distributions of coup outcomes seems to vary significantly based on the type coup, so these variables will be used for prediction.
- The Cline Center dataset also contains useful temporal data. Below, we see that the frequency of different event types varies by year. Thus, both year and month (which sees similar variation) will also be used for prediction.
Additional data has been joined to form a picture of the political and economic state of a country in the year that the coup event occurs, as well as to add in geographic data. These predictors will also be used in the model. The data sources are cited below.
Towards the goal of creating the best model for this data, the pre-processing steps in the recipe include removing predictors with near-zero variance, median imputing missing values, one hot encoding a categorical variable, BoxCox transforming skewed predictors, and normalizing all predictors.
Random forest, boosted tree, k nearest neighbors, and elastic net models were fit and tuned. The random forest model with mtry = 6 and min_n = 11 achieved the highest ROC AUC score, and thus was selected.
The selected model performed well on the testing data, achieving a ROC AUC score of about 0.81. From the ROC curve, we see that it was best at identifying realized coups.
- Features from all three data sources turned out to be of importance, as seen in the chart below. Whether a coup was
dissident, or initiated by a small group of discontents which may include ex-military leaders, religious leaders, ex-government officials or others, turned out to be the most important predictor of outcome.
While this model performed well, it could likely be improved with more training data, or with more political data or granular geographic data.
Future research questions include whether the occurrence of coup events can be predicted based on country and year, and whether coup events within a region could be modeled as Poisson or Hawkes self-exciting processes.
Data Citations
Peyton, Buddy, Joseph Bajjalieh, Dan Shalmon, Michael Martin, and Jonathan Bonaguro. 2020. Cline Center Coup D’état Project Dataset. Cline Center for Advanced Social Research. V.2.0.0. November 16. University of Illinois Urbana-Champaign. doi: 10.13012/B2IDB-9651987_V2
Enterline, Andrew, Michael Greig, Resat Bayer, Diane Dutka, et. al. 2017. National Material Capabilities Dataset. The Correlates of War Project. V.5.0.0. February 1. University of North Texas.
Jennifer Bryan (NA). gapminder: Data from Gapminder. https://github.com/jennybc/gapminder, http://www.gapminder.org/data/, https://doi.org/10.5281/zenodo.594018.