Study
Mohler and colleagues (2015) conducted a single-blind, randomized, controlled trial in three divisions of the Los Angeles Police Department (LAPD) in California to assess the effects of a fully automated statistical algorithm for predictive policing, the epidemic-type aftershock sequence (ETAS) model, on crime volume in dynamic hot spots.
The field trials were conducted in three LAPD divisions on the following dates: Foothill Division from Nov. 7, 2011, to April 27, 2012; North Hollywood Division from March 31, 2012, to Sept. 14, 2012; and Southwest Division from May 16, 2012, to Jan. 10, 2013. Rather than dividing patrol areas into partially matched blocks, patrol “mission maps” were generated independently by a crime analyst (the control condition) and the ETAS algorithm (the treatment condition) each day. Mission maps were identical in outward appearance, except for the exact placement of prediction boxes. Only after each set of predictions was generated was it randomly determined, using a Bernoulli random number generator, whether the control or treatment condition would be deployed to the field for the next 24-hour period. Hot spot locations dynamically changed each day, therefore days were randomly assigned to the treatment or control conditions. Command staff, supervisors, and patrol officers in each of the deployment areas were not aware of the distinction between treatment and control mission maps, and the police officers involved were the same on both control and treatment days. Field-deployed missions consisted of twenty 150-by-150-meter prediction boxes (which is the size of a city block in Foothill and held constant across the experimental regions) for each 12-hour shift per division. Patrol officers were directed to use available time to “get in the box” and police what they saw; therefore, control and treatment conditions compare directed patrol patterns in space and time, not differences in field tactics. A total of 62 control days and 62 treatment days in Foothill, 82 control days and 70 treatment days in North Hollywood, and 117 control and 117 treatment days in Southwest were randomly assigned during the trial, for a total of 510 days.
In the treatment condition, an ETAS crime forecast consisted of the top 20 prediction boxes in rank order displayed on a map and available for directed police patrol. The ETAS model was implemented as a fully automated, cloud-based machine learning system. Software was installed on the police records management system server that encrypted and sent crime report data (address, date, time, crime type) once every hour to cloud-based servers, where the data were geocoded using Google application programming interfaces and stored for later use. Once a day at 4:00 a.m. the ETAS parameters were re-estimated using the previous 365 days of crime data (up to the latest records) and boxes were ranked by intensity. Patrol commanders logged on with a web interface before each patrol shift and printed out a report that was then distributed to officers. The software also allowed the analyst on control days to place 20 boxes (identical in size and appearance to ETAS boxes) on the Google map instead of ETAS–generated boxes. The report contained a map of hot spot locations and a list of nearest cross-streets of the hot spot locations. Missions were set daily at 5:00 a.m. for two sequential shifts from 6:00 a.m. to 6:00 p.m. and 6:00 p.m. to 6:00 a.m. In Foothill, mission predictions were made for the 24-hour period starting at 4 p.m.
For the control condition, crime analysts placed a fixed number of 150-by-150-meter boxes within their operational environment in specified time windows. The goal of box placement was to identify a small set of locations where the analysts expected crime most likely to occur. In Los Angeles, three crime analysts acting independently were responsible for placing prediction boxes for Foothill, North Hollywood, and Southwest Divisions. Crime analysts were free to use any information and methods at their disposal to generate predictions. The LAPD followed a COMPSTAT policing model focused on the analysis of 7-day crime maps supplemented with ad hoc street-level intelligence. The LAPD analysts emphasized small clusters of recent crimes as signaling emerging problems in need of short-term response to interrupt formation of crime hot spots. An easy-to-use web interface was developed to allow analysts to choose prediction locations by clicking on a Google map. Analysts spent considerable time engaged in analysis of crime and intelligence data before committing to predictions. Each analyst spent about 1.5 to 2.0 hours each day inspecting crime maps and recent crime reports before placing their prediction boxes. In the Foothill Division, 20 predictions were generated for each 24-hour period beginning at 4:00 p.m. each day. In both North Hollywood and Southwest Divisions, 40 prediction boxes were generated each day, evenly divided between two shifts (6:00 a.m. to 6:00 p.m. and 6:00 p.m. to 6:00 a.m.).
Officer time on mission in control and treatment prediction boxes was used as the primary measure of officer activity. Officers generated a time-stamped call event on their in-car mobile data terminal upon entering and leaving any mission box. The total amount of time spent within prediction boxes was the officer time on mission, and each incident of an officer spending time within a box was a mission stop. Reported crimes were tabulated according to whether they occurred strictly inside or outside active control and treatment boxes. A crime occurring strictly inside an active prediction box (which is a box generated before the start of a patrol shift and held constant in that location for the duration of the shift) is considered a successful prediction. Regression models were constructed to assess the relationship between daily crime volume and patrol time on mission. Changes in crime volume (specifically burglary, car theft, and burglary-theft from vehicle) were assessed for all three LAPD divisions as a whole. No subgroup analysis was conducted.
