The Global Terrorism Database (GTD) provides information on terrorist attacks that have been documented in the news since 1970. Curated by The National Consortium for the Study of Terrorism and Responses to Terrorism (START), this database indicates suicide bombings are much less common than other types of terrorist attacks. Figure 1 shows a plot of attacks that were classified as suicide bombings vs. non-suicide bombings over time.

Figure 1. Suicide vs. Non-Suicide Bombings Over Time

Similarly, when we look at terrorist activity at the group level, we see that the number of groups that do engage in suicide bombings is low and, among those groups, the proportion of attacks that are suicide bombings (versus other types of attacks) is small. See Figure 2a, 2b, and 2c for illustrations of these trends, using GTD data from 2017.

Figure 2a. Proportion of Groups that Engaged in Suicide Bombing in 2017

Figure 2b. Among Groups that Engaged in Suicide Bombings in 2017, Proportion of Attacks that were Suicide Bombings

In order to see this pattern among groups that conducted relatively few attacks in 2017, we can zoom in on Figure 2b to groups that conducted 50 or fewer attacks (see Figure 2c).

Figure 2c. Among Groups that Engaged in 50 or Fewer Suicide Bombings in 2017, Proportion of Attacks that were Suicide Bombings

Suicide bombings tend to be concentrated in specific countries; most countries do not experience suicide bombings as a form of terrorism. Figure 3 displays a heatmap of suicide bombings that occurred in 2017.

Figure 3. Heatmap of Suicide Bombings in 2017

Nonetheless, suicide bombings are more deadly than other forms of terrorism; see Figure 4 for a plot of the number of people killed in suicide bombings versus non-suicide bombings in 2017 alone.

Figure 4. Log Number of Poeple Killed by Suicide vs. Non-Suicide Bombings in 2017

Note. Given that the data were heavily skewed right, the data were transformed by adding 1 and then taking the log of each datapoint.

• Understand characteristics of terrorist groups that predict whether they engage in suicide bombings
• Understand situations/contexts in which suicide bombings occur
• Build a model that accurately predicts whether or not a terrorist group engages in suicide bombings
  • Focus on Recall: Generate a model with very few false negatives
• Interpret my final model's feature importances with an eye toward predicting future suicide bombings

I used two databases in this project. The first, called the GTD, is an open-source database including information on terrorist events around the world from 1970 through 2017. Unlike many other event databases, the GTD includes systematic data on domestic as well as transnational and international terrorist incidents that have occurred during this time period and now includes more than 180,000 cases. For each GTD incident, information is available on the date and location of the incident, the weapons used and nature of the target, the number of casualties, and--when identifiable--the group or individual responsible.

I also used One Earth Future's (OEF) [Terrorist and Insurgent Organizations' Service Provision across time dataset (TIOS 2.0)] The TIOS 2.0 provides indicators associated with whether and how more than 400 violent non-state actors provide public goods and services across more than four decades (1969–2013). There are 6,659 group-years (i.e., rows) in the TIOS 2.0.

Both the GTD and the TIOS 2.0 create their databases by culling journal articles on terrorist activities and coding for their variables of interest. For instance, for the TIOS 2.0, OEF searched within new stories for words and phrases that correlate with service provisions (see Table 1 for example words).

Of relevance to the current project, the GTD contains one variable indicating whether a given terorrism event was due to a suicide bombing. My primary goal was to use machine learning to model features affiliated with suicide bombing using the GTD and TIOS 2.0. The TIOS 2.0 contains matching variables for the GTD in order to merge the two databases. Doing so allows one to observe individual terrorist group activities in a given year, including the number of suicide bombings groups engaged in.

This project builds on my first Capstone, which explored the TIOS 2.0 data, and my second Capstone, which modeled characteristics of suicide bombings using the GTD database. READMEs for both of those projects outline extensive EDA relating to the GTD and TIOS 2.0.

Recall that my goal was to build a model that would predict whether terrorist groups would engaged in suicide bombings. I opted to use features from the GTD that had feature importances > .04 in a Random Forest model outlined in my Capstone 2 project. I also included public goods and services (i.e., service provisions) provided by terrorist groups, available in the TIOS 2.0 database. Table 1 provides feature labels, names, and details.

Table 1. Feature Labels, Definitions, and Source

As illustrated in Figures 1-2, there is class imbalance in my target variable, which is the proportion of groups that engaged in suicide bombings. To deal with class imbalance, I upsampled my data, resulting in N = 1,476 (n = 738 for each group).

I generated training and testing samples for the upsampled dataset using the sklearn.model_selection.train_test_split method. I used a 75:25 (train:test) split on the data.

I first ran two Decision Tree models where features were either (a) a count of the number of times a feature occured in a group-year, or (b) a binary indicator of whether or not a feature occurred in a given group year (see Table 2). I observed Accuracy, Recall, and Precision in order to select the optimal model. I focused particularly on Recall, which is an important index for predicting suicide bombings. That is, it is important to "catch" groups that may engage in a suicide bombings in a given year, which is reflected by the Recall score.

Table 2. Decision Tree Statistics and Model Characteristics

I selected Model 1 for the next phase of model testing because this model had a higher recall score that Model 2. Next, I ran Model 1 through a series of algorithms (see Table 3 and Figure 5). I found that model fit statistics using a Random Forest model were slightly better than the remaining models.

Table 3. Model Statistics and Hyperparameters for Algorithms

Figure 5. ROC Curves for Five Algorithms Testing Model 3.

Using a Random Forest Classification, I created a confusion matrix (Figure 6) and displayed feature importances (Figure 7).

Figure 6. Confusion Matrix Using Random Forest Classification.

Figure 7. Feature Importances Using Random Forest Classification

Table 4. T tests Assessing Whether Groups that Engage in Suicide Bombings Differ in Feature Prevalence

Note. Independent samples t tests were conducted using a Bonferonni correction (adjusted critical p value = .0004).

We can use feature importances (see Figure 7) and directional effects observed from t tests (see Table 4) to make assumptions about groups that conduct suicide bombings. The most important feature in predicting whether groups engage in suicide bombing is their use of explosive vehicles to conduct attacks. That is, groups that engage in suicide bombings use explosive vehicles significnatly more often than groups that do not engage in suicide bombings.

Terrorist groups that use suicide bombing are also more likely to claim responsibility for their attacks than their counterparts, suggesting the former seek notoriety for their deeds. That is, these groups are forthcoming and tend not to hide in the shadows.

Finally, groups that engage in suicide bombings tend to provide significantly more services than groups that do not engage in suicide bombings. According to OEF researchers, some violent groups seek legitimacy and international support through service provision (see Heger & Jung, 2015).

Taken together, this work suggests authorities and individuals should be particularly vigilent in areas that have a high density of vehicles. Authorities should monitor terriost groups that have claimed responsibility for suicide bombings in the past -- particularly those that provide services for their communities.

Heger, Lindsay L. and Danielle F. Jung (2015) Negotiating with rebels: The effect of rebel service provision on conflict negotiations. Journal of Conflict Resolution, 61, 1203-1229.