When the levee breaks: A forecasting model of violent and nonviolent dissent

Figures & data

Figure 1. Trends in nonviolent and violent conflict over observation period.

Table 1. Forecasting model summary.

Estimation methods	Outcome variables	Independent variables
Logistic regression Ridge Lasso Elastic Net k-Nearest Neighbors Random Forest Gradient Boosting Machine Neural Network ENSEMBLE- Unweighted ENSEMBLE- Ridge ENSEMBLE- Lasso ENSEMBLE- Elastic Net ENSEMBLE- Bayes	Unarmed conflicts Only Armed conflict only Combined unarmed conflicts and armed conflict (labeled as conflict)	Structural variables only Structural variables + ICEWS events data Structural variables + lagged outcome variable

Figure 2. Steps to train algorithms and evaluate their out-of-sample predictive power.

Table 2. Conflict predicted probability in the MENA region in 2011.^a

Conflict probability	Unarmed conflict probability	Armed conflict probability	Observed conflict	State name
99.5%	97.5%	0.6%	Unarmed	Morocco
95.67%	0.02%	74.61%	Armed	Iraq
66.6%	0.1%	93.9%	Unarmed	Iran
85.6%	0%	74.4%	Armed	Turkey
91.8%	7.5%	3.3%	Armed	Algeria
95.2%	50.7%	0.2%	Unarmed	Egypt
15.0%	0%	0.09%	Unarmed	Syria
0.8%	0.04%	0.1%	Neither	Lebanon
0.5%	0%	3.1%	Neither	Israel
0.2%	0.3%	0%	Unarmed	Tunisia
0.1%	0%	0.2%	Neither	Kuwait
0.3%	0%	1.3%	Neither	Saudi Arabia
0.1%	0.2%	0.05%	Unarmed	Bahrain
0.2%	0.1%	0.03%	Unarmed	Jordan
0.04%	0%	0.2%	Neither	United Arab Emirates
0%	0%	0.01%	Neither	Qatar
0%	0%	0%	Neither	Oman
NA	NA	NA	Both	Yemen
NA	NA	NA	Armed	Libya

^aYemen and Libya missing predicted probabilities due to missing data. Note as well that each model’s predicted probability is calculated independently, with errors varying across models and cases. Thus, while the probability of conflict is theoretically related to the joint probability of armed and unarmed conflicts, in some cases, the predicted probability of conflict diverges significantly from the combined predicted probabilities of armed and unarmed conflicts (e.g., in Iran, where the conflict model gives a probability of roughly 94.48% while the unarmed and armed models give roughly 4.49% and 35.17% probabilities, respectively.

Figure 3. Comparing the AUC-PR and AUC-ROC of the prediction algorithms across different models and outcomes.

Figure 4. Predictive power measured by AUC-ROC across models (columns) and algorithms (rows).

Figure 5. Predictive power measured by AUC precision-recall across models (columns) and algorithms (rows).

Figure 6. Variable importance in GBM model.