Title page

Contents

Abstract 5

Executive summary 6

Method 6

Results 7

Discussion 7

Introduction 8

Outlaw motorcycle gangs 9

Gang databases 11

Risk assessment 12

Predicting high-harm offending among outlaw motorcycle gang members 13

Current study 13

Method 14

High-harm offending 14

Data 15

Analytic approach 16

Limitations 19

Results 20

National-level analysis 22

State-level analysis 23

Discussion 30

Model accuracy: The trade-offs of using suboptimal data 30

Implications and disruption opportunities 33

Data availability and access is pivotal for a path forward 35

The use of transparent machine learning in police settings 36

Conclusion 37

References 38

Appendix: Removing individuals with no prior recorded offending 45

Table 1. Error and accuracy calculations of the confusion matrix 18

Table 2. Descriptive statistics 20

Table 3. Confusion matrix for random forest model trained on high-harm offending at the national level 23

Table 4. Confusion matrix for random forest model trained on high-harm offending in four jurisdictions 25

Table 5. Feature importance for each model developed to predict high-harm offending 27

Table 6. Summary findings 29

Figure 1. Receiver operating characteristic (ROC) curves for random forest (grey) and logistic regression (green) predicting high-harm offending among OMCG... 22

Figure 2. Receiver operating characteristic (ROC) curve for random forest (grey) and logistic regression (green) predicting high-harm offending among OMCG... 24

Table A1. Confusion matrix for random forest model trained on high-harm offending using restricted sample 46

Figure A1. Receiver operating characteristic (ROC) curve for random forest (grey) and logistic regression (green) predicting high-harm offending among OMCG... 45