2

Place-Based Predictive Policing

Andrew Ferguson, American University Washington College of Law and planning committee member, provided the opening presentation for a session focused on theoretical underpinnings, examples of use, and evidence of effectiveness for place-based predictive policing approaches. Place-based predictive policing, he explained, is based on theories in the field of environmental criminology. These theories posit that crime is not equally distributed across society and that “hot spots” of certain crimes can be identified. Analysis of environmental factors and historical data can reveal patterns of heightened risk for future crimes; these patterns can then be used to direct police resources in the hopes of deterring crime. Key messages from this session included the following:

• Place-based predictive policing is based on theories of environmental criminology, which suggest that crime is concentrated in specific areas influenced by environmental factors, creating hot spots. (Ferguson)
• Effective predictive policing (place-based or person-based) requires both accurate predictions and well-implemented interventions. (Ratcliffe)
• Rigorous research, particularly randomized controlled trials, could help validate place-based predictive models and ensure effective implementation in the field. (Lee)
• Future artificial intelligence (AI)-driven place-based predictive policing tools could become integral to law enforcement software. (Bueermann)

HISTORY OF PLACE-BASED PREDICTIVE POLICING

The idea that crime hot spots could be predicted through algorithms arose around 2009, said Ferguson. Predictive policing was presented as a promising path forward—an objective, data-driven policing method that could help police departments manage resources, manage officers, and address community concerns related to racial justice. Early predictive policing technologies used past crime data and other variables to predict future criminal risk in particular places. Deterrence was the primary intervention, he said, generally in the form of sending additional officers to identified areas to discourage crime. Use of these technologies quickly spread across the country. Ferguson described three of the early technologies. One tool was inspired by seismology and relied on research showing that certain property crimes are “contagious” and will lead to “aftershocks” of similar offenses close to the time and place of the original crime. Contagion is thought to occur either because the same group commits the crimes or because of an environmental vulnerability that encourages further crime, Ferguson explained. The algorithm for this tool used data from incident calls, calls for service, and crime type, time, and place, but it did not use arrest data. These data were used to generate maps of 500-feet-by-500-feet areas with elevated risk of a particular crime. Police were given risk maps and advised to patrol high-risk areas or put a police car in the area to deter new crimes. The tool could also track and report the amount of time officers spent in a targeted area, said Ferguson.

Another early place-based predictive policing technology was described as a “patrol management system” that forecasted areas of risk and suggested ways police could respond. This model used data on past crimes as well as information on other factors that could predict crimes, such as seasonality, day of the week, holidays, and sporting events. Machine learning technologies analyzed these data and produced color-coded crime zones, which were used to allocate patrols and to inform officers about which types of crimes might be expected in certain areas. The third technology examined the risk environment, to determine why crimes were occurring in given areas. This model looked at the physical environment of a city as a “terrain of overlapping risks” and identified areas with multiple risks in close range as areas with a heightened likelihood of future crime. For example, said Ferguson, the model might identify gas stations that are open late at night as areas of particular vulnerability because people tend to gather in these areas and get into altercations. Based on these modeled risks, the program proposed strategies to enhance positive spatial influences—that is, ways to improve the physical environment to reduce crime. The idea behind the technology, said Ferguson, was that police organizations are risk-management agencies that can address vulnerabilities in the community and use strategies that go beyond specific deterrence.

Early models were primarily aimed at deterring crime through increased police presence or structural changes to the environment (e.g., lighting, cameras). These place-based predictive policing tools relied on crime data, said Ferguson, noting that the specific type of crime data affects the output. For example, a report could be based on calls for service, arrests, or all reported crimes; thus, reports could differ due to the demographics of individuals who are more likely to report crimes, the types of crimes most likely to be reported, when arrests are made, biases that may impact crime reports, and other factors. Initial predictive efforts focused on burglary and theft but expanded to violent crimes such as gun crimes, robberies, and assaults. Ferguson noted that he is unaware of examples of predictive policing used to monitor crimes such as fraud, white-collar crime, or interfamily crimes.

Early technologies raised several concerns, said Ferguson, including concerns that data were used as justification for over-policing in disenfranchised areas. Other concerns involved diversion of funds from supporting communities to surveilling communities, and the possibility that police would focus their resources on measurable crimes rather than all crimes. Finally, Ferguson noted concern that predictions of high-crime areas would justify increased police stops, undermining community members’ constitutional rights against unreasonable search and seizure. Ferguson explained that in the years since early place-based predictive policing tools were implemented, many of the initial programs ended or were redesigned. However, using data to predict crime has continued, and predictive analytics are now commonly embedded in everyday systems of policing.

EVIDENCE ON PLACE-BASED PREDICTIVE POLICING

Jerry Ratcliffe, Temple University, began his presentation by emphasizing that predictive policing is made up of both predictive analytics and, even more essentially, a police response to those predictions. Ratcliffe joined other speakers in encouraging particular attention to the policing element of predictive policing. “Good policing,” he argued, is effective, proportionate, and procedurally just, and it prevents crime more effectively than does unfocused, reactive policing. For example, said Ratcliffe, it is not particularly impactful to assign one officer to each area of a city regardless of crime rates, or to send officers out only after a crime has been committed.

Substantial evidence suggests that effective, proportionate, and procedurally just policing prevents crime, said Ratcliffe (see Box 2-1). Putting police at the right place at the right time has been shown to reduce crime, he explained, and using data to predict crime can help establish police presence where it is needed. Ratcliffe suggested that both current and long-term efforts could help with crime reduction, namely long-term efforts to address

poverty and improve education and, more immediately, predictive policing to better prevent violent crime.

Ratcliffe described his early work in place-based predictive policing in Philadelphia, where he and his colleagues used near-repeat analysis and related statistical methods to analyze shootings in the city. They found that in the first two weeks after a shooting event, the likelihood of a repeat shooting within 400 feet was 33% higher than expected (Ratcliffe & Rengert, 2008). Over half of shooting victims in Philadelphia were shot

within two blocks of their homes, and over 90% of victims were non-White. These insights are important, he said, because they suggest that police have an opportunity to address violence in communities that have historically not been well served. Ratcliffe acknowledged that this particular project was unable to prevent repeat shootings, but this was mainly due to limited police resources.

More recently, Ratcliffe and colleagues collaborated on the Philadelphia Predictive Policing Experiment (Ratcliffe et al., 2021). This project, said Ratcliffe, involved an experiment in which small city zones were randomized to three interventions or a control. The interventions were (a) making officers aware of crime predictions, (b) putting a marked police car in the target area, or (c) putting an unmarked car in the target area (Ratcliffe et al., 2021). Business-as-usual served as the control. Researchers found no difference in violent crime rates but found a 31% reduction in property crime in areas with marked cars (Ratcliffe et al., 2021). Ratcliffe emphasized, however, that the actual number of crimes prevented was relatively small: The reduction represented three crimes over three months in three patrolled grids. The inputs used were victim-generated data on property crime and violent crime.

Discussing the algorithm used in this experiment, Ratcliffe noted the attention paid to ensuring the algorithm was accurate and responsible, and he emphasized the importance of careful selection of inputs. One algorithm that he worked with predicted victimization more accurately if race was included as a measure. However, he noted that race variables are often a proxy measure for a history of discrimination and unequal access to resources. To encourage conscientious use of race as an input, the algorithm was designed so that the user had to explicitly choose to use race as a measure; Ratcliffe was not aware of anyone activating this feature. In addition, Ratcliffe and colleagues were careful not to use police-generated variables such as arrest data in the victimization algorithm.

A wide variety of data can be used as inputs for predictive algorithms, said Ratcliffe, but sometimes a simple approach works as well or better than a more complicated one. He gave an example of a study from Sweden comparing the predictive accuracy of various data inputs, in isolation and combined (Doyle & Gerell, 2024). While combining variables (e.g., place attributes, social characteristics) increased the accuracy of predictions, using data on past crimes alone, per the article’s authors, “still does a really good job,” said Ratcliffe. He explained that place-based prediction tends to be accurate because certain areas have higher opportunity for crime than others; crime levels will stay high unless underlying opportunity structures are changed. He noted that while disparities in education, poverty, unemployment, and other variables may impact crime, other influences are at play. To demonstrate this point, Ratcliffe pointed out that over the past

decade, the levels of violence in some major cities have shifted dramatically, but the underlying levels of education, poverty, and unemployment have remained largely the same. Ratcliffe expressed his support of long-term investments to reduce poverty and improve education, stating that these non-policing solutions play a role in reducing crime. In the meantime, he argued, predictive policing could help prevent some of the thousands of shootings that occur each year, for example. He noted, however, that crime prediction is not the end of the analysis—it is the first step to determining why crime is happening in certain areas and what can be done to address it.

When discussing the costs and benefits of predictive policing, he emphasized, it is important to compare predictive policing with actions police are currently taking rather than an ideal hypothetical. He drew a comparison to self-driving cars—while it is true that self-driving cars get into accidents, their accident rate is far lower than that of human drivers. “The comparison point is not no policing,” Ratcliffe stated. Police have always used some form of prediction to make choices about where to focus resources, he said. Historically, such prediction has taken the form of rudimentary predictive policing, in which a sergeant sent extra officers to areas thought to be crime prone. This approach had inherent biases and lacked transparency, Ratcliffe noted. While newer algorithms for predicting crime may not be perfect, he said, they improve upon the former opaque, non-data-driven approach. He encouraged stakeholders to have realistic goals for predictive policing and to focus on improving on the baseline of current police actions.

Findings from a Systematic Review

Youngsub Lee, University College London, and his colleagues conducted a systematic review examining the effectiveness of data-driven place-based predictive policing (Lee et al., 2024). The review focused on evidence supporting effective crime reduction, he noted, and did not directly discuss normative or ethical aspects of predictive policing.

Lee and colleagues screened the literature and found 161 studies, which they classified into four categories based on the strength of the evidence. Only studies that presented numerical outcomes (e.g., reduction in crime rates) were included. Studies that utilized randomization and real-world testing, and that considered displacement effects, were classified as having the strongest evidence and placed in the Type 1 category. Studies that utilized randomization and real-world testing but did not account for displacement effects were classified as Type 2. Retrospective studies that tested algorithms on past data and applied predictions to past crime data were classified as Types 3 and 4, with lower evidentiary value. Six studies were classified as Type 1 or Type 2 (including a Philadelphia study described in Ratcliffe et al., 2021); these studies demonstrated no, limited, or moderate effectiveness (Table 2-1).

TABLE 2-1 Type 1 and 2 Studies (n = 6)

SOURCE: Presentation by Youngsub Lee on June 24, 2024, derived from Lee et al., 2024.

The remaining 155 studies were retrospective studies with limited utility in demonstrating effectiveness, said Lee. These studies primarily generated a prediction rate for crimes and did not consider whether and which police interventions could reduce the crimes.

Notably, the six strongest studies suggest that matching prediction to intervention is a key element of the effectiveness of predictive policing, said Lee. This means that it is important to pair even the best prediction models with robust, carefully designed efforts to implement appropriate interventions in the field. While the retrospective studies do not offer much evidence in terms of effectiveness, they do provide a wealth of research on prediction models for future applications in this area, he said.

Overall, the review found evidence that the use of predictive policing models can effectively reduce crime, said Lee. However, given the valid concerns about place-based predictive policing, he expressed his perspective that evidence is not yet strong enough to offset concerns. More research is needed to ensure that place-based and person-based predictive policing models are effective and are implemented in ways that address valid community concerns. Specifically, prospective randomized controlled trials could test the impact of clearly specified interventions on target crimes. Additionally, it is important for these studies to test displacement effects, in which crime moves elsewhere in response to prevention efforts. Furthermore, said Lee, given the importance of proper field implementation, training officers to carry out appropriate interventions is critical. In addition, Lee called for more research on the best type of data to use for inputs in predictive policing algorithms. As Ratcliffe noted, crime data alone are fairly effective in predicting future crimes, said Lee. If police want to use data from non-police sources in addition to criminal data, evidence needs to show that the use of multiple data sources leads to significant efficiencies in crime prediction over the use of police-derived data alone, Lee stated.

Lee explained that this systematic review is part of an overall project designed to investigate trust in technology-enabled policing. Governments have an obligation to transparently explain policing and its tools to citizens, said Lee. If effectiveness is the basis for implementing predictive policing tools, sharing the strengths and limitations of such technologies with citizens is critical. Only when all parties are equally informed about the issue under discussion can society engage in democratic decision making regarding the use of predictive policing, Lee said.

Law Enforcement Perspective

Jim Bueermann, Future Policing Institute, shared his perspective on predictive policing approaches as a long-time police chief. Based on his experiences, Bueermann suggested that predictive policing tools have several

benefits. The mandate of policing is community safety, and predictive policing approaches could help fulfill this mandate by enabling law enforcement to predict and prevent crime, he said. By identifying where or toward whom police should be directing their time and energy, predictive policing approaches could also mitigate the impact of the ongoing recruiting and hiring crisis and help control costs through increasing the efficiency of policing. Finally, said Bueermann, predictive policing approaches could increase the public’s trust and confidence in policing.

On the other hand, Bueermann said, some people see predictive policing approaches as rife with problems. From this perspective, particular applications of predictive policing may exacerbate bias in policing, deepen existing issues with wrongful arrests and convictions, damage community perceptions of the police, and remove the human element from aspects of police decision making. Proponents and critics both have legitimate arguments, said Bueermann, so decisions about whether and how to use predictive policing tools require a collective discussion and a co-production of knowledge around the development and use of these technologies and their costs and benefits.

NEXT-GENERATION PLACE-BASED PREDICTIVE POLICING

Several speakers noted that the term “predictive policing” is no longer widely used. Bueermann agreed and said that rather than focusing attention on a specific technology or approach, it is important for decision makers to consider how the expansion of AI will impact policing. AI is becoming ubiquitous in every field and its use is being rapidly normalized, he noted. While the term “predictive policing” may be on its way out, similar data-driven approaches, increasingly relying on AI, could be built into broader technology ecosystems for police use. Such ecosystems could include multiple tools and technologies that rely on data, machine learning, and AI, Bueermann hypothesized.

Bueermann cautioned that while AI tools are growing exponentially, policies and guidelines around AI use have lagged. Use of predictive technologies and AI-based tools could dramatically improve policing or invite a cascade of problems, depending on how society responds. Improving knowledge about AI and developing and implementing policies and principles for its safe and responsible use are necessary at the federal, state, and local levels, he said. Community involvement in decisions about whether and how to acquire and use AI tools is valuable, as is national coherence around guiding principles for safe and responsible use of police AI, he stated. Bueermann also highlighted the value of police technologists—professionals who understand advanced law enforcement technologies and their underlying criminological theories—for operationalization

and implications. Technologists could also help law enforcement agencies navigate engagement with commercial technology developers, said Bueermann.