Predicting Arrests for Terry Traffic Stops in Seattle
Introduction
In 1968, the supreme court ruled in Terry v. Ohio that an officer can stop and search (better known by some states as stop and frisk) a person if the officer had reasonable suspicion as to their activity (the court does actually differentiate between reasonable suspicion and mere suspicion). Thus, Terry Stops were born. Terry Stops do further extend to traffic stops as well and officers are able to search vehicles on the same premise. With the recent rises of questionable police practices, it’s worth asking can we create a model that identifies that factors that increase the chances of the person in question being arrested?
Dataset
This dataset was taken from data.gov and can be explored here.
First Thing’s First, There’s Two Kinds of Arrests
What? How? Well there’s two arrest variables in the database:
- Arrest Flag, which notes whether the subject was physically arrested on scene
2. Stop Resolution, which reflects whether a terry stop was inputted into the system as the resolution to the stop
Furthermore is an interesting potential dilemma. Arrests in the system clearly outnumber physical arrests on-scene by about 7,000 so we know that not everyone that was arrested because of a terry stop was necessarily physically arrested at the stop. So, it’s a viable machine learning problem.
The other way around is more concerning. Nearly everyone (except for two people) that was physically arrested were nearly all inputted into the system as an arrest which question whether an arrest in the system is a machine learning question (hint: it should still be because it doesn’t account for the other 7,000 arrested people).
So, How Do We Choose Which One To Model?
Easy, we do both.
Arrest Flag, How Well Can We Predict Whether Someone Was Arrested On-Scene?
A gradient boosting model performed best on predicting whether someone was arrested on scene. In fact, it’s precision on correctly predicting whether someone had been physically arrested on scene was 99%. Here are the top 5 model features that influence this outcome:
- Year
- The strongest feature. The more recent the year, the higher the likelihood that the subject was physically arrested on scene (there is also an implication here that physical on-scene arrests have been increasing as time has gone on)
2. Was The Stop Resolution Also An Arrest?
- This didn’t guarantee that the person was physically arrested on-scene but it did increase the chances that they were
3. Month
- The later months in the year are more likely to lead to a physical on-scene arrest then earlier months
4. Juvenile Disturbances
- Juvenile disturbances were more likely to lead to a physical on-scene arrest then other types of calls
5. Auto Recovery
- Calls made in terms to a car needing to be towed was more likely to lead to a physical on-scene arrest then other types of calls
Stop Resolution, How Well Can We Predict If A Terry Stop Ultimately Resolved As An Arrest In The System?
This was much more difficult to predict but the best performing model for this issue was a Random Forest model which was able to predict whether a stop resulted in an arrest in the system with 86% precision. These were the most important features:
- Arrest Flag, Were They Physically Arrested On-Scene?
- Important and did increase chances of arrest being made in the system but didn’t guarantee it
2. Year
- Similar to the arrest flag model, later years have seen higher likelihoods of arrests from Terry Traffic stops
3. Hour
- Later hours (evening) made it more likely that a stop would result in an eventual arrest in the system
4. Call Type As A Suspicious Person/Suspicious Circumstances
- This raised the the chances of arrest in system but is unclear as to what clearly defines suspicion in this case.
5. Officer Age
- Younger officers were more likely to lead to an arrest in the system then older officers
Conclusion
While these models can give us a good sense as to what features can increase the likelihood of arrests in different contexts, it’s unclear as to whether this data can be generalized beyond Seattle. One especially interesting inclusion is the fact that year is a big feature in both models meaning that the closer we get to the present, the more likely a physical arrest and/or an arrest in the system. That definitely deserves a closer look. There are also other trends in Seattle Police behavior that would require a deeper dive in order to give an accurate assessment as to why these features are important (i.e., Why are older officers more likely to arrest then younger officers? Why do later months have more arrests then earlier months?).
