Predictive Policing
Predictive Policing
Predictive Policing
org
Predictive Policing
10.27.2015
SARAH BRAYNE, ALEX ROSENBLAT, and DANAH BOYD
Introduction1
Predictive policing refers to the use of analytical techniques by law enforcement to make statistical
predictions about potential criminal activity.2 Predictive policing can involve either predicting
events (i.e., forecasting when and where crimes are likely to occur) or people (i.e., individuals likely
to be victims or perpetrators of crimes). Instead of relying on an officer’s ‘hunch’ about an area,
“predictive policing uses the power of ‘big data’ to isolate patterns.”3 A 2013 RAND report offers
a taxonomy of predictive methods, identifying four categories of predictive policing:
This primer offers an overview of what is currently known about predictive policing and
highlights unanswered questions about the implications of predictive policing.
Context
Predictive policing is emblematic of a broader trend towards data-driven decision-making in a
wide range of fields. Within criminal justice, actuarial methods have long been an important
component of managing risk.5 In the 1920s, Ernest Burgess of the Chicago School of sociology
created the first parole prediction instrument, in which he calculated probability estimates of
reoffending. Although actuarial methods have existed in criminal justice for almost a century,
predictive analytics have only become systematically incorporated into law enforcement practices
over the past two decades. 6
Predictive Policing 2
The shift towards predictive policing occurred in the late 2000s. In the face of low crime rates,
departments were facing pressure to push crime rates even lower. Following the logic that the
‘low-hanging fruit had already been taken’ (i.e., the easy crimes had already been intercepted or
prevented), law enforcement needed a means by which to further reduce crime rates. Coupled
with pressure to allocate resources more efficiently in light of tight budgets, in 2008, Bratton, who
had moved on to become the Chief of the Los Angeles Police Department, began working with
federal agencies to assess the viability of a more predictive approach to policing.
The root causes of crime vary. Therefore, it is important to consider the different underlying
assumptions informing different predictive models. For example, Another participant who had
been working with criminal justice organizations to increase access to data on traffic stops, found
that models predicting violent crime sometimes identify people, rather than places, as their
outcomes.
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org
Predictive Policing 3
Another form of predictive policing is person-based. In such systems, law enforcement may
predict individuals or groups most likely to be involved in crimes, either as victims or offenders.
Person-based predictive policing could involve social network analysis or regression models using
risk factors.9 One of the civil liberties questions that arises is: what are the implications of the
police modeling risk for individuals who have no criminal record? For example, as the White
House report on Big Data documents, “In response to an epidemic of gang-related murders, the
city of Chicago conducted a pilot that shifts the focus of predictive policing from geographical
factors to identity. By drawing on police and other data and applying social network analysis, the
Chicago police department assembled a list of roughly 400 individuals identified by certain factors
as likely to be involved in violent crime. As a result, police have a heightened awareness of
particular individuals that might reflect factors beyond charges and convictions that are part of
the public record.”10
Once individuals are identified as high risk, police interventions could range from patrolling the
areas where individuals reside to more aggressive interventions – such as talking to identified
individuals or their family members.
Whether location or individual based, there are four key stages in the practice of predictive
policing:11
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org
Predictive Policing 4
In the first stage, data is collected. These data can range from basic crime data (i.e., when and
where historical crimes occurred), to more complex environmental data such as seasonality,
neighborhood composition, or risk factors (e.g., vacant lots, parks, ATMs). The second stage
involves data analysis, which yields predictions about future crime. When deciding what
predictive method to use, law enforcement needs to consider both the type of crime they want to
target and their department’s resources.
The third stage in the predictive cycle is police intervention. Usually police intervention involves
distributing crime forecasts to commanders who use them to make decisions about where to
deploy officers in the field. Patrol officers are also sometimes given reports at briefings to inform
them where to go while on shift. During uncommitted time (i.e., when they are not responding to
calls for service), patrol officers focus their time and resources to surveilling the people and places
models suggest are likely to be involved in future crime. It is important to note that the very act of
predictive policing creates new data as well. For example, police describe how they ‘enter and
clear predictive boxes,’ and they enter their status in the computer terminal in their car to notify
dispatch and keep records for future analysis.12
The fourth stage, target response, highlights that this predictive policing cycle, or, ‘battle rhythm’
as some law enforcement officials have called it, grows increasingly complex over time. Law
enforcement needs to account for individuals’ responses to police intervention. As mentioned
earlier, the intervention could serve as a deterrent, preventing crime from occurring, or could lead
to the displacement of crime to a different area.
Rhetoric
There is considerable hype in media and policy circles about predictive policing. TIME Magazine
named predictive policing as one of the 50 best inventions in 2011. With its emphasis on ‘big data’
analytics, predictive policing is touted as a means to improve both efficiency and equity.
In terms of efficiency, advocates of predictive policing say it can more accurately predict future
crimes than humans can.13 One possible explanation for this is that humans overestimate
meaningful patterns.14 Traditional crime analysts may possess exaggerated perceptions of crime
patterns and consequently overreact by directing a disproportionately high amount of police
resources to an area. Additionally, police departments face resource constraints, so predictive
policing’s ability to ‘do more with less’ can be appealing from a budgetary perspective.
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org
Predictive Policing 5
“standardize information across shifts and experience levels” ultimately “eliminating the concern
about adequate information sharing.”15 It is important to avoid false binaries such as “intuition-
driven” versus “data-driven” policing, because in practice, neither approach exists in isolation
from the other; each informs the other in consequential ways.
Just as people are excited about predictive policing, there are anxieties that have bubbled to the
surface since its implementation. Fears stem from civil rights groups, researchers, the communities
that are being policed, and law enforcement themselves. Civil rights advocates suggest predictive
policing will be used to profile and harass people who have not committed any crime, and that it
can do so under the patina of objectivity.
Researchers suggest that new analytic techniques may, to a certain extent, reproduce
conventional police practices, but under the guise of data science. Mirroring debates occurring
over discriminatory lending,16 some of the variables included in predictive models may simply be
a proxy for race or other protected categories.
Individuals in law enforcement argue predictive models should not substitute experience, ‘street
smarts,’ and officer intuition. It is worth noting that law enforcement is not uniform in their
opinion of predictive policing. Rather, there are divisions, with some officers viewing it as another
useful tool in their toolbox, others viewing it as an entrenchment of managerial control, and
others still viewing it as a subversion of their experiential knowledge.17
Some (but not all) concerns are driven by misconceptions. The major pervasive misconception is
the Minority Report myth. Minority Report, a book and subsequent film that depicts a dystopian
world in which individuals are arrested for crimes “precogs” foresee them committing in the
future, is relentlessly invoked in discourse around predictive policing. This is a misleading
characterization of predictive policing that distracts from the important issues at hand relevant to
having a productive conversation about the promises and perils of predictive policing moving
forward. Predictive policing is not akin to a crystal ball. It does not foretell the future. Rather,
when used correctly, it can give police officers probabilistic information about where to go and
who to police. Much of the time, predictive models serve to confirm police intuitions. However,
they may also be able to dispel officers’ individual perceptions of crime or crime rates that may be
inaccurate or outdated. The implications of predictive policing techniques have much more to do
with the bureaucratic processes that surround the technology than the technology itself. The same
tools can be used to redress existing inequities or amplify them.
Actors
There are a variety of relevant actors involved in predictive policing, including local, state and
federal governments and agencies, predictive platform designers, vendors, researchers, and
consultants.
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org
Predictive Policing 6
Most predictive policing pilot projects in local police departments are made possible by injections
of federal funding. For example, in 2011, the Los Angeles Police Department received a three
million dollar grant from the U.S. Department of Justice (DOJ) to conduct a multiyear analysis of
predictive policing. In 2015, the Miami Police Department received $600,000 to do the same.
Smaller cities tend to take up predictive tools after pilots are run in larger departments. Some of
the cities that are currently using predictive policing include: Chicago, IL, Memphis, TX, Los
Angeles, CA, Santa Cruz, CA, Minneapolis, MN, Palm Beach, FL, Dallas, TX, Vancouver
(British Columbia, Canada), Charlotte-Mecklenburg, NC, Nashville, TN, Glendale, AZ, East
Orange, NJ, Baltimore, MD, New York City, NY, Philadelphia, PA, Miami, FL, New Castle, DE,
Lincoln, NE, and Montevideo, MN.
One of the biggest predictive policing vendors is PredPol. Researchers Jeff Brantingham and
George Mohler, currently at the University of California Los Angeles and Santa Clara University,
respectively, suggested it was possible to predict certain crimes much like it was possible to
forecast the distribution of earthquake aftershocks. They co-founded PredPol, now headed by
Larry Samuels. Informed by predictive techniques used in research on counterinsurgency
operations, PredPol is run on a cloud-based SaaS platform and uses three types of data in its
proprietary algorithm – place, type, and time of crime. It includes three years of data, weighting
the more recent data more heavily. The algorithm generates 500 by 500 square foot predictive
boxes on maps, indicating areas where particular crimes are most likely to occur. PredPol is
currently used in almost 60 departments, the largest of which are the Los Angeles Police
Department and the Atlanta Police Department.
Another software is Hunchlab, designed by the GIS firm Azavea. Robert Cheetham, a former
crime analyst for Philadelphia Police Department is the President and CEO. Hunchlab’s
statistical models account for the interaction of social, behavioral and physical risk factors.
Hunchlab is currently used by the Philadelphia Police Department and the Miami Police
Department. In 2015, the New York City Police Department announced their plans to begin
testing Hunchlab software.18
A key contrast between PredPol and Hunchlab is their analytic strategy. Informed by near-repeat
theory, PredPol has a parsimonious model, only using historical crime data. Whereas Hunchlab
also uses near-repeats, it also uses risk-terrain modeling19 which involves a much wider range of
variables, such as seasonality, collective efficacy (the willingness of individuals to intervene), school
calendars, and environmental risk factors. An interesting conflict and break in the predictive
policing narrative is that advocates of PredPol preempt some of the criticism leveled at Hunchlab,
namely for including too many factors that could be considered discriminatory in their models.
Individuals at PredPol are careful to emphasize their software does not predict who commits
crimes, but rather focuses on what types of crimes are predicted to occur where and when.
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org
Predictive Policing 7
So, how is predictive policing different from conventional hot spot policing? There are three main
differences. First, although preliminary data suggests they actually yield substantively similar
predictions, hot spot policing uses density maps. In other words, hot spot policing involves simply
plotting crimes in order to visualize a geographic distribution of crime, and subsequently
deploying officers to the “hottest areas.” Although the predictive boxes generated by algorithms
often overlay conventional hot spots, predictive boxes are much smaller than hot spots and
sometimes generate boxes that do not match up with conventional hot spots. A second important
difference is that predictive algorithms are highly opaque, a phenomenon referred to as
“algorithmic secrecy.”20 Simply stated, police officers understand how heat maps are generated,
but the algorithmic process by which predictive boxes are generated is invisible and difficult to
interpret. Finally, whereas hotspot policing is retrospective in its analytic approach, predictive
policing is prospective. Hotspot policing directly maps where crimes occurred in the past, but
predictive policing could predict crime will occur in a location that there has not been a crime
before. The exact ways in which this would happen are often shielded by the proprietary nature of
predictive policing algorithms.
Understanding the actors and agencies involved in funding and adopting predictive policing
efforts is helpful as we transition into the next section on open questions about the practice.
Knowing the institutional architecture of predictive policing is integral to knowing where the
points of leverage are and understanding where and how research, measurement, and assessment
can occur.
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org
Predictive Policing 8
officers regardless of whether they use formal predictive methods or not) – may also explain the
drop in crime.
- How do decisions by law enforcement affect the validity of the data? What can be done to
correct biases in the data?
- How does the implementation of predictive policing affect the data that these techniques
rely on?
- Who can meaningfully assess the algorithms that are being implemented and what kind of
algorithmic oversight is appropriate?
- How does missing or biased data shape the predictions that these systems make?
Ongoing assessment of these systems will be critical, but is by no means guaranteed. We need to
ask what kinds of structures are necessary for evaluating the efficacy of predictive policing. One
possibility is to run randomized control trials (RCTs) across divisions. Such trials need to be
evaluated by researchers independent from the department, the designers of the strategy, and the
vendors. For example, we should not simply rely on assessments of a policing initiative by the
consulting firm that designed the strategy, which is currently the case in some departments.
How do different policing tools perform, relative to each other and relative to
earlier methods?
Although we know about comparative accuracy of crime prediction in patrol (e.g., we have data
on the predictive power of algorithms vs. analysts), we do not know how platforms measure up
against one another (e.g., Hunchlab vs. PredPol). Moreover, even less is known about the
application of predictive models in investigations.
- Are there more cases cleared by arrest when detectives employ predictive analytics?
- How do law enforcement officers treat the information they receive from each system?
- What bureaucratic dynamics influence how officers incorporate the information they
receive into their policing practices?
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org
Predictive Policing 9
It is also important to consider not only how efficacious a predictive technique is, but also how
transparent it is. As mentioned above, algorithms tend to be relatively inscrutable for police
officers, and research suggests that simple heuristics may be almost as effective as advanced
computational methods. For example, stop heuristics with simple scoring rules may have
comparable accuracy to fuller statistical models.21 Models that are easier to understand can
improve police buy-in and can be less expensive, making it possible for smaller departments with
fewer financial resources to adopt the approach.
How can Fourth Amendment protections be preserved in the context of these new
tools?
The way law enforcement uses predictive analytics challenges the traditional paradigm of Fourth
Amendment law because predictive information may be used to justify stops under the existing
Fourth Amendment precedent.2223 In a 2015 law review article, Ferguson asks whether a stop can
“be predicted on the aggregation of specific and individualized, but otherwise noncriminal,
factors.”24 He argues that otherwise noncriminal factors “might create a predictive composite that
satisfied the reasonable suspicion standard.”25 In other words, predictive analytics may effectively
make it easier to meet the reasonable suspicion standard in practice, thus justifying more police
stops. Ferguson suggests that if the police use big data to reach the threshold of reasonable
suspicion, the courts “should require a higher level of detail and correlation using the insights and
capabilities of big data.”26
- In the context of predictive policing, should we raise the standard of reasonable suspicion?
- Should being a high risk individual be grounds for police surveillance (with or without a
warrant)?
- How will all the new forms of data being generated through predictive policing be used?
What will be admissible under the exclusionary rule?
How will predictive policing affect the overall dynamic between police and the
communities they serve?
We must also consider questions about the implications of predictive policing in terms of civil
rights and social inequality.
- What kind of hit rate (i.e., the percentage of stops in which an officer makes an arrest or
issues a summons) constitutes defensible public policy?
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org
Predictive Policing 10
As Barocas and Selbst argue, discrimination may be an artifact of the data collection and analysis
process itself. Even with the best intentions, algorithmic decision-making can lead to
discriminatory practices and outcomes.27 Algorithmic decision procedures can “reproduce
existing patterns of discrimination, inherit the prejudice of prior decision makers, or simply reflect
the widespread biases that persist in society. It can even have the perverse result of exacerbating
existing inequalities by suggesting that historically disadvantaged groups actually deserve less
favorable treatment.”28 Each of the steps in data analysis can create “possibilities for a final result
that has a disproportionately adverse impact on protected classes, whether by specifying the
problem to be solved in ways that affect classes differently, failing to recognize or address
statistical biases, reproducing past prejudice, or considering an insufficiently rich set of factors.”29
Simply stated, we need to ask whether predictive policing reduces police discretion (and to what
extent we want to reduce it in the first place), and whether it serves to exacerbate or remedy
existing inequalities in police practices.
Finally, returning to an age-old question in policing literature: to what extent is the crime rate a
function of enforcement practices? More generally, to what extent is what we “know” about crime
conditional on how we know it? We need revisit these questions in the context of predictive
policing, because the ‘raw’ data inputted into predictive models is in fact not ‘raw’ at all, but
rather fundamentally social.
Having answers to these types of questions will help us determine to what extent predictive
policing is defensible public policy, and what the best practices of predictive policing are. It is an
anachronism to think that the police can (or should) stop using data or predictive models.
Consequently, we need to design and implement models with these questions in mind.
1 We are very grateful for the strong contributions and insights made by David Robinson, Harlan Yu, Corrine Yu, Patrick
Davison, and Angie Waller in the research and production of this primer.
2 Perry, W. L., McInnis, Brian, Price, C.C., Smith, S.C., and J.S. Hollywood. 2013. “Predictive Policing: The Role of
Crime Forecasting in Law Enforcement Operations.” RAND Safety and Justice Program. Available at
http://www.rand.org/content/dam/rand/pubs/research_reports/RR200/RR233/RAND_RR233.pdf.
3 Ferguson, A. G. 2012. “Predictive Policing and Reasonable Suspicion.” 62 Emory Law Journal 259: p.265.
4 Perry, W. L., McInnis, Brian, Price, C.C., Smith, S.C., and J.S. Hollywood. 2013. “Predictive Policing: The Role of
Crime Forecasting in Law Enforcement Operations.” RAND Safety and Justice Program. Available at
http://www.rand.org/content/dam/rand/pubs/research_reports/RR200/RR233/RAND_RR233.pdf.
5 Ericson, R. V., and Haggerty, K. D. 1997. Policing the Risk Society. Oxford University Press.; Lyon, D. 2003.
Surveillance as social sorting: Privacy, risk, and digital discrimination. Psychology Press.
6 Feeley, M. M. and Simon, J. 1992. “The New Penology: Notes on the Emerging Strategy of Corrections and Its
Implications.” Criminology, 30(4), 449-474.; Harcourt, B. E. 2007. Against Prediction: Profiling, Policing, and Punishing in an
Actuarial Age. University of Chicago Press.; Marx, G. T. 2005. “Seeing Hazily (but not Darkly) Through the Lens: Some Recent
Empirical Studies of Surveillance Technologies.” Law and Social Inquiry, 30(2), 339-399.
7 Ferguson, A. G. 2012. “Predictive Policing and Reasonable Suspicion.” 62 Emory Law Journal 259: p. 272.
8 Ferguson, A. G. 2012. “Predictive Policing and Reasonable Suspicion.” 62 Emory Law Journal 259.
9 Papachristos, A. 2009. "Murder by Structure: Dominance Relations and the Social Structure of Gang Homicide."
American Journal of Sociology 115(1): 74-128.
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org
Predictive Policing 11
10 Podesta, J., Pritzker, P., Moniz, E.J., Holdren, J., and Zients, J. 2014. “Big Data: Seizing Opportunities, Preserving
Values.” Executive Office of the President., p. 31.
https://www.whitehouse.gov/sites/default/files/docs/bigdataprivacyreportmay12014.pdf.; Moser, W. “The Small Social
Networks at the Heart of Chicago Violence,” Chicago Magazine, December 9, 2013, http://www.chicagomag.com/city-
life/December-2013/The-Small-Social-Networks-at-the-Heart-of-Chicago-Violence/.
11 Based on Price, M. et al., 2013. National Security and Local Police. New York: Brennan Center for Justice.
12 Some departments gather this data automatically by using vehicle locators on cars.
13 Uchida, C.D., and Swatt, M.L.. 2013. “Operation LASER and the Effectiveness of Hotspot Patrol: A Panel Analysis.”
Police Quarterly 16(3): 287-304.
14 For example, see research on cancer clusters: Trumbo, Craig W. 2000. “Public Requests for Cancer Cluster
Investigations: A Survey of State Health Departments.” American Journal of Public Health 90: 1300-1302; Robinson D. 2002.
“Cancer Clusters: Findings vs. Feelings.” MedGenMed 4: 16.
15 Friend, Z. 2013. “Predictive Policing: Using Technology to Reduce Crime.” FBI Law Enforcement Bulletin.
16 Squires, G.D. 2004. Why the Poor Pay More. Westport: Praeger.
17 Brayne, S. 2015. “Stratified Surveillance: Policing in the Age of Big Data.” Dissertation, Princeton University.
18 Other software, such as the analytic software designed by IBM and formerly used by the Memphis Police Department,
also exist. However, this primer focuses on PredPol and Hunchlab, as they are currently the largest operating in this space.
19 The NYPD is working with researchers at Rutgers University and John Jay College of Criminal Justice to evaluate the
efficacy of risk-terrain modeling for allocating police resources. See: “Effectiveness of Risk Terrain Modeling for Allocating Police
Resources.” National Institute of Justice. Accessed November 19, 2015. http://www.nij.gov/topics/law-
enforcement/strategies/predictive-policing/pages/risk-terrain-modeling-for-allocating-resources.aspx.
20 Pasquale, F. 2015. The Black Box Society: The Secret Algorithms that Control Money and Information. Harvard
University Press.
21 Goel, S. Rao, J.M., and Shroff, R. 2015. “Precinct or Prejudice? Understanding racial disparities in New York City’s
Stop-and-Frisk Policy.” SSRN Scholarly Paper. No. ID 2572718.
22 See Ferguson 2012; Ferguson, A. G. 2015. “Big Data and Predictive Reasonable Suspicion.” 163 University of
Pennsylvania Law Review 1137.
23 Joh, E. E. 2014. "Policing by Numbers: Big Data and the Fourth Amendment." 89 Washington Law Review 35.
24 Ferguson, A. G. 2012. “Predictive Policing and Reasonable Suspicion.” 62 Emory Law Journal 259: p.330.
25 Ferguson, A. G. 2012. “Predictive Policing and Reasonable Suspicion.” 62 Emory Law Journal 259: p. 335.
26 Ferguson, A. G. 2012. “Predictive Policing and Reasonable Suspicion.” 62 Emory Law Journal 259: p. 336.
27 Barocas, S. and Selbst, A. 2016. “Big Data’s Disparate Impact.” California Law Review. 104. SSRN Scholarly Paper.
28 Barocas, S. and Selbst, A. 2016. “Big Data’s Disparate Impact.” California Law Review. 104, p. 3.
29 Barocas, S. and Selbst, A. 2016. “Big Data’s Disparate Impact.” California Law Review. 104, p. 5.
DATA & CIVIL RIGHTS: A NEW ERA OF POLICING AND JUSTICE datacivilrights.org