Exploring the Determinants of Crime-Terror Cooperation using Machine Learning

Abstract

Objectives

This study seeks to further strengthen extant knowledge regarding terrorist group involvement in organized criminal activity through two means. First, it measures a set of environmental and organizational characteristics for a sample of well-known terrorist organizations based on the crime-terror literature. Second, it illustrates the utility of inductive research designs for examining patterns in the criminal behavior of terrorist groups for theory building and the potential risk classification of new terrorist organizations in the future.

Methods

The authors utilize a random forest classification algorithm to examine three sources of information about a broad set of environmental and organizational factors determined to be of potential importance in predicting when a terrorist organization will engaged in organized criminal behavior. First, it examines out-of-sample accuracy through bootstrap cross-validation estimation. Second, it quantifies the predictive efficacy/importance of each measured factor. Finally, it utilizes partial dependence functions to examine the relational trend between the most important predictive factors and variation in the presence of organized criminal behavior.

Results

The study finds three results. First, predictive accuracy using readily quantifiable factors about the criminal behavior of terrorist organizations is good but could be improved upon. Second, organizational factors such as group size, ideology and attack behavior out perform environmental factors in terms of predictive performance. Third, it finds that the most important predictor variables have a predominately non-linear relationship with whether the algorithm would classify a group as engaging in organized criminal behavior or not.

Conclusions

The study finds that theory building should seek to examine temporal variation in the organizational structure of terrorist groups as a fruitful way forward for further understanding when a group is likely to engage in organized criminal behavior. It also suggests that scholars should seek to engage more critically with concepts surrounding the potential non-linear pathways in which groups end up engaging in organized crime. Finally, the results illustrate the utility of modern machine learning algorithms and inductive research processes for both academic and practitioner needs alike. Especially when dealing with a complex phenomenon with imperfect data.

Appendices

Appendix A: Logistic Regression Coefficients

Three logistic regression models were fit to examine variable significance and importance. Table N below displays the results. The first model uses every variable outlined in the data and operationalization section. The second model uses only organizational factors, and the third model uses only environmental factors. Comparing model log likelihood, the organizational model explains a greater amount of variation when compared with the environmental factor model. The AIC statistics also indicates that organizational variables are substantively better than environmental variables for explaining variation concerning criminal activity.

Examining the fully fitted model, the logistic regression results in three statistically significant predictors. The centralized group structure dummy variable is significant at the 10% level. Substantively, the logit coefficient indicates a positive relationship between centralized organizational structure and involvement in criminal activity. The coefficient estimates a 1.82 increase in the odds of a terrorist organization being classified as having involvement in criminal activity when it has a centralized organizational structure. This finding is interesting as it contradicts theoretical and anecdotal expectations about the relationship between decentralized organizational structure and criminal involvement. Namely the expectations concerning the strategic incentive of centralized structures to avoid high-risk activities.

Group size is statistically significant and conforms to the expectations held in the literature. Group size has a positive relationship with involvement in criminal activity. The regression coefficient indicates an average classification odds increase of 1.44 as organizations increase in group size. Finally, the religious ideology dummy variable is statistically significant and indicates a strong positive relationship with criminal involvement, with religious groups having a 14.8 odds increase in being classified as having criminal involvement. This finding is difficult to reconcile with both intuition and theoretical expectations. One would expect religious groups to shy away from illicit activities, yet this finding would suggest the opposite. While not statistically significant, the nationalist ideology dummy indicates the expected relationship as found in the literature. Interestingly, none of the environmental variables are statistically significant in the full model, but military spending and logged GDP are significant in the environment model (Table 3).

Table 3 Logistic regression

Appendix B: Distributional Comparison Between Panel Measures and Aggregated Cross-sections

See Table 4.

Table 4 Mean and variance scores across measure types

Appendix C: Tenfold Cross Validation Estimation of Out of Sample Accuracy

Equation used for estimating tenfold cross validation accuracy. K is for each fold, while Err represents the classification error rate for each fold. The results corroborate the comparison in the main text (bootstrap validation). The random forest (75.2%) is roughly 6% more accurate in classifying out of sample cases when compared to the logistic regression (68.9%) (Figs. 5, 6).

Fig. 5

Tenfold cross validation accuracy for random forest classifier

Fig. 6

Tenfold cross validation accuracy for logistic regression

$$CV_{k} = \frac{1}{k} \mathop \sum \limits_{i = 1}^{k} 1 - Err_{i}$$