Program Profile: Philadelphia (Pa.) Predictive Policing Experiment (3PE)

This strategy sought to reduce crime in Philadelphia by testing three different patrol strategies. The program is rated Ineffective. Relative to control areas, there was no statistically significant difference in violent or property crimes in areas using one of two of the patrol strategies. Property crimes in an area using one of the strategies saw a statistically significant decrease, while there was a statistically significant increase in violent crime in areas using two of the strategies.

An Ineffective rating implies that implementing the program is unlikely to result in the intended outcome(s).

This program's rating is based on evidence that includes at least one high-quality randomized controlled trial.

An Ineffective rating implies that implementing the program is unlikely to result in the intended outcome(s).

This program's rating is based on evidence that includes at least one high-quality randomized controlled trial.

Program Goals

Predictive policing is a tactic that relies on software forecasts (or predictions) of potential locations of criminal events or crime hot spots. Often, law enforcement agencies use predictive policing to identify expected high-crime areas or targets (such as individuals who commit crimes) to implement a police response or intervention. In 2015 the Philadelphia (Pa.) Police Department sought to implement an intervention that incorporated predictive policing and could address high rates of crime, as the city had a homicide rate of 17.8 per 100,000 people and a violent crime rate of 1,029 per 100,000 people, compared with national homicide and violent crime rates (4.9 and 383 per 100,000 people, respectively). As a result, the Philadelphia Police Department implemented the Philadelphia Predictive Policing Experiment (known as 3PE).

3PE was a place-based experiment designed to reduce different types of violent and property crimes (such as homicides and burglary) by testing different (but operationally realistic) patrol strategies based on the use of crime predictions (estimated by a predictive policing software program) across all police districts in Philadelphia.

Program Components/Target Sites

3PE consisted of three specific police activities (or patrol responses) that were implemented in two phases: 1) the property crime phase and 2) the violent crime phase. Each phase was implemented daily for 14 weeks (or about 3 months). During the property crime phase (which ran from June 1 through Aug. 25, 2015), officers patrolled predicted crime areas from 8:00 a.m. to 4:00 p.m. and focused primarily on the following property crimes: residential and commercial burglary, and vehicle theft (including theft from stolen vehicles). During the violent crime phase (which ran from Nov. 1, 2015, through Jan. 31, 2016), officers patrolled predicted crime areas from 6:00 p.m. to 2:00 a.m. and focused primarily on the following violent crimes: homicide, rape, robbery, and aggravated assault.

In both the property and violent crime phases, patrol officers were assigned to one of the three patrol responses during their daily 8-hour shifts:

• Awareness patrol. In this response tactic, patrol officers were informed at roll call of the predicted crime areas and asked to pay attention to the areas when they were able. Patrol officers were still responsible for performing typical operational duties.
• Marked car patrol. In this response tactic, usually two patrol officers (though occasionally one patrol officer) were assigned to a single marked vehicle dedicated to patrolling the predicted crime areas. Patrol officers were not assigned calls for service or other operational roles; however, they could support other officers attending calls for service in their prediction areas if the call was related to a property or violent crime (the only exception was if a fellow officer required immediate assistance).
• Unmarked car patrol. In this response tactic, patrol officers were assigned to a single unmarked vehicle (and were often dressed in plain clothes), dedicated to patrolling the predicted crime areas. Similar to the marked car patrol tactic, patrol officers were not assigned calls for service or other operational roles; however, they could support other officers attending calls for service in their prediction areas if the call was related to a property or violent crime (the only exception was if a fellow officer required immediate assistance).

The experiment relied on a predictive policing software called HunchLab, which was a web-based predictive policing software that accessed historical crime data from the Philadelphia Police Department and used machine learning techniques to produce crime forecasts or predictions for the city (Ratcliffe and Taylor 2017). Consistent with the design of the study, the software was used to generate predicted crime areas for each of the three patrol responses).

Similar to other predictive systems, potential crime locations in HunchLab were defined by grid cells, which were 500 feet by 500 feet in size, laid over the city in a fishnet pattern. HunchLab’s algorithm used historical crime data to compute expected crime counts for each 500-foot by 500-foot cell. HunchLab then selected the cells with the highest expected counts and displayed them as likely crime grid cells on a map, called "mission grids". The treatment area for this study included these mission grids and contiguous cells that were adjacent to the mission grids. For additional information on the formation of mission grids, see the reviewed study by Ratcliffe and colleagues (2020a, 6–9).

Although the predictive policing algorithm calculated three predicted crime (or target) areas for each shift, these areas were not available for district personnel not involved in the study to view.

Study

Ratcliffe and colleagues (2020a) conducted a randomized controlled trial to determine the impact of the Philadelphia Predictive Policing Experiment (3PE) on reducing violent and property crimes across districts in the city. At the time of the study, the Philadelphia Police Department consisted of 6,200 sworn officers and 800 civilian personnel. According to the 2016 Census, Philadelphia had a population of 1.5 million people, with the majority from a White (45 percent) or Black (44 percent) racial background. In regard to ethnicity, 14 percent of people identified as Hispanic or Latino.

This experiment consisted of testing three patrol response strategies (i.e., the intervention conditions): 1) awareness patrol, 2) marked car patrol, and 3) unmarked car patrol. As a part of the experiment, each patrol strategy relied on a predictive policing software (called HunchLab), which used historical crime data to create a spatiotemporal forecast (or prediction) of areas of criminality or crime hot spots. During the experiment, data were collected from multiple sources, including the American Community Survey, Philadelphia's Open Street Map, the Philadelphia Zoning Authority, and the Philadelphia Police Department.

During the intervention, the three patrol responses were implemented in two phases, which included the strategies delivered in the property crime phase (June 1 to Aug. 25, 2015) and the violent crime phase (Nov. 1, 2015, to Jan. 31, 2016). Following the property crime phase, districts were re-randomized for the violent crime phase. The study authors separated the experiment into two phases for two reasons: 1) the Philadelphia Police Department had to prepare for an event that was being hosted in Philadelphia in late September (a visit from Pope Francis), and 2) to allow crime patterns to return to normalcy for the districts.

Philadelphia had 22 police districts; however, the study authors excluded district 77 (which contained the airport) for all phases of the study. The authors also excluded the district with the lowest crime rate during each phase, resulting in district 7 being excluded in the property crime phase and district 5 during the violent crime phase. Thus, there were 20 districts included in the study sample and randomly assigned (using blocked randomization with a 1:1:1:1 ratio) to the following conditions: 1) awareness patrol (n = 5); 2) marked car patrol (n = 5); 3) unmarked car patrol (n = 5); 4) business-as-usual “control” condition (n = 5).

Because of the design of the study, which aimed to examine whether the three different patrol responses had direct effects on reducing property and violent crimes, this CrimeSolutions review separated the three patrol responses as three discrete studies: Study 1 pertains to results for the awareness patrol response relative to the business-as-usual control condition. Study 2 (described below) pertains to results for the marked car patrol response relative to the business-as-usual control condition. Study 3 (also described below) pertains to results for the unmarked patrol response relative to the business-as-usual control condition.

In the awareness patrol strategy, districts were patrolled by officers who were made aware of predicted target areas at the start of their 8-hour shifts and were asked to pay attention to these district areas, within limits of their typical operational duties. Meanwhile, in business-as-usual “control” districts, patrol officers did not have access to the crime prediction software. At the pre-intervention period, there were no statically significant differences on expected crime patterns between districts in the awareness patrol intervention districts and the business-as-usual districts.

Crime data were collected for pre-intervention and post-intervention periods. Outcomes of interest included property and violent crimes. Property crimes were defined as residential and commercial burglary, motor vehicle theft, and theft from vehicles. Violent crimes were defined as shootings, homicides, aggravated assaults, and robberies. Property and violent crimes were measured using official data from the Philadelphia Police Department’s geolocated incident database.

Bayesian models with Markov chain Monte Carlo estimation were used to determine the difference between districts in the awareness intervention condition and the business-as-usual condition, at posttreatment. This study did not conduct subgroup analyses.

Study

Ratcliffe and colleagues (2020b) conducted a randomized controlled trial to determine the impact of the second patrol response strategy (dedicated marked car patrol), as a part of the Philadelphia Predictive Policing Experiment. As in Study 1 (Ratcliffe et al. 2020a), Study 2 used the same study sample of police districts in the city (n = 20). This study reports findings on districts randomized to the marked car patrol strategy (n = 5) relative to business-as-usual districts (n = 5).

For districts randomly assigned to the marked patrol condition, this intervention strategy consisted of patrol officers who were assigned to a single marked vehicle that was dedicated to patrolling the predicted crime areas identified by HunchLab. Patrol officers in this intervention condition were not assigned calls for service or other operational roles. Marked car patrol districts were relative to the same business-as-usual districts from Study 1. At the pre-intervention period, there were no statistically significant differences in expected crime patterns between marked patrol car intervention districts and business-as-usual districts.

Crime data were collected for pre-intervention and post-intervention periods. Outcomes of interest included property and violent crimes. Property and violent crimes were measured using official data from the Philadelphia Police Department’s geolocated incident database. Bayesian models with Markov chain Monte Carlo estimation was used to determine the difference between districts in the marked car patrol intervention strategy and the business-as-usual condition, at posttreatment. This study did not conduct subgroup analyses.

Study

Ratcliffe and colleagues (2020c) conducted a randomized controlled trial to determine the impact of the third patrol strategy (dedicated unmarked car patrol), as a part of the Philadelphia Predictive Policing Experiment. As in Study 1 and Study 2 (Ratcliffe et al. 2020a; Ratcliffe et al. 2020b), Study 3 used the same study sample of police districts in the city (n = 20). This study reports findings on districts randomized to the unmarked car patrol strategy (n = 5), relative to business-as-usual districts (n = 5).

For districts randomly assigned to the unmarked patrol condition, patrol officers drove a single unmarked vehicle that was dedicated to patrolling the predicted crime areas identified by HunchLab. Patrol officers in this intervention condition were not in uniform and were not assigned calls for service or other operational roles. Unmarked patrol districts (i.e., the intervention) were relative to the same business-as-usual districts from Study 1 and Study 2. At the pre-intervention period, there were no statistically significant differences in expected crime patterns between unmarked patrol car intervention districts and business-as-usual districts.

Crime data were collected for pre-intervention and post-intervention periods. Outcomes of interest included property and violent crimes. Property and violent crimes were measured using official data from the Philadelphia Police Department’s geolocated incident database. Bayesian models with Markov chain Monte Carlo estimation were used to determine the difference between districts in the marked car patrol intervention strategy and the business-as-usual condition, at posttreatment. This study did not conduct subgroup analyses.

In 2015, Temple University’s Center for Security and Crime Science partnered with the Philadelphia (Pa.) Police Department to investigate the impact of the Philadelphia Predictive Policing Experiment (Ratcliffe et al. 2020a; Ratcliffe et al. 2020b; Ratcliffe et al. 2020c). This was a 2-year project funded by the National Institute of Justice (Award No. 2014–R2–CX–0002), in which Temple University’s Center for Security and Crime Science worked with the Philadelphia Police Department to reduce violent and property crimes, by testing different (but operationally realistic) patrol strategies that could better optimize the use of crime predictions estimated by a predictive policing software program (i.e., HunchLab, now known as ShotSpotter Mission).

Before implementing the intervention, a briefing was held for all available midlevel and senior district–level leaders in the Philadelphia Police Department. In addition, instructional information was sent to each district and included two short instructional videos for crime analysts (i.e., those responsible for accessing crime predictions) and patrol officers. Further, the study authors were available to assist with technical issues (such as districts having trouble accessing the online predictive policing software) and provided support by fielding phone calls from districts and attending to districts in person to answer questions (Ratcliffe et al. 2018).

To assess the implementation of the different patrol strategies and monitor treatment integrity, field observations were conducted by trained observers (i.e., graduate research assistants and primary investigators). Trained observers collected observation data through ride-alongs with patrol officers assigned to the intervention areas. Field observations consisted of two elements: 1) structured and systematic observations of officers’ patrol behavior or activity, documented by the observer on a form every 15 minutes, and 2) open-ended ethnographic field notes. For the observation form, observers were able to choose from multiple options to describe where specific police activity occurred (such as at a station, a car outside grids, a car inside grids, a car chat with citizens, on break, an incident within the grid, an incident responded to outside the grid, “other”). In addition, patrol officers assigned to the marked and unmarked patrol conditions were required to complete written daily logs. Daily logs were collected regularly and cross-referenced for accuracy with administrative data (which linked officers to their pedestrian stops, arrests, and other indicators of activity). These logs included details of officer activity in the predicted target areas, amount of time on site, officer observations, and other pertinent information.

These sources were used in the development of the program profile:

These sources were used in the development of the program profile:

Brantingham, Patricia L., and Paul Jeffrey Brantingham. 1982. “Mobility, Notoriety, and Crime: A Study in the Crime Patterns of Urban Nodal Points.” Journal of Environmental Systems 11(1):89–99.

Cohen, Lawrence E., and Marcus K. Felson. 1979. “Social Change and Crime Rate Trends: A Routine Activity Approach.” American Sociological Review (44):588–608.

Cornish, Derek Blaikie, and Ronald V.G. Clarke. 1986. The Reasoning Criminal: Rational Choice Perspectives on Offending. New York, N.Y.: Springer.

Durlauf, Steven N., and Daniel S. Nagin. 2011. “Imprisonment and Crime: Can Both Be Reduced?” Criminology and Public Policy 10(1):13–54.

Ratcliffe, Jerry H., and Ralph B. Taylor. 2017. The Philadelphia Predictive Policing Experiment: Summary of the Experimental Design. Philadelphia, Pa.: Temple University.

Ratcliffe, Jerry H., Ralph B. Taylor, and Amber Perenzin Askey. 2017. The Philadelphia Predictive Policing Experiment: Effectiveness of the Prediction Models. Philadelphia, Pa.: Temple University.

Ratcliffe, Jerry H., Ralph B. Taylor, Amber Perenzin Askey, John Grasso, and Ryan Fisher. 2017. The Philadelphia Predictive Policing Experiment: Impacts of Police Cars Assigned to High-Crime Grids. Philadelphia, Pa.: Temple University.

Ratcliffe, Jerry H., Ralph B. Taylor, Amber Perenzin Askey, Ryan Fisher, and Josh Koehnlein. 2018. The Philadelphia Predictive Policing Experiment: Final Report. Washington, D.C.: Department of Justice, National Institute of Justice.

Ratcliffe, Jerry H., Ralph B. Taylor, and Ryan Fisher. 2020. “Conflicts and Congruencies Between Predictive Policing and the Patrol Officer’s Craft.” Policing and Society 30(6):639–55.

Ratcliffe, Jerry H., Ralph B. Taylor, Kevin Thomas, John Grasso, Kevin J. Bethel, Ryan Fisher, and Josh Koehnlein. 2019. The Philadelphia Predictive Policing Experiment: Summary of Experiment and Findings. Philadelphia, Pa.: Temple University.

Taylor, Ralph B., and Jerry H. Ratcliffe. 2020. “Was the Pope to Blame? Statistical Powerlessness and the Predictive Policing of Micro-Scale Randomized Control Trials.” Criminology & Public Policy 19(3):965–96.

Following are CrimeSolutions-rated programs that are related to this practice:

Hot Spots Policing

Hot spots policing strategies focus on small geographic areas or places, usually in urban settings, where crime is concentrated. Through hot spots policing strategies, law enforcement agencies can focus limited resources in areas where crime is most likely to occur. This practice is rated Effective for reducing overall crime and rated Promising for reducing violent, property, public order, and drug and alcohol offenses.

Evidence Ratings for Outcomes

	Crime & Delinquency - Multiple crime/offense types
	Crime & Delinquency - Violent offenses
	Crime & Delinquency - Property offenses
	Crime & Delinquency - Public order offenses
	Crime & Delinquency - Drug and alcohol offenses

Problem-Oriented Policing

These analytic methods are used by police to develop crime prevention and reduction strategies. The practice is rated Promising and led to a significant decline in crime and disorder.

Evidence Ratings for Outcomes

Crime & Delinquency - Multiple crime/offense types

Why might a practice's outcome ratings differ from the ratings of specific programs encompassed by that practice?