The Philadelphia predictive policing experiment

“Where items are applicable, our article follows a CONSORT framework for the reporting of randomized trials (Moher et al. 2010) and is organized as follows.”

“Blocked experimental designs have been around for roughly a century and are generally credited to statistician R. A. Fisher (Box 1980; Fisher 1935).”

“A substantial canon demonstrates that crime is unevenly distributed among places and victims (Felson 1987; Sherman et al. 1989; Weisburd and Eck 2004), and a number of spatial theories of crime can explain short-term changes in crime risk for small areas (Chainey and Ratcliffe 2005).”

“Previous randomized studies examining policing in crime hot spots, however, have relied on a year of crime prior to the experiment (Sherman and Weisburd 1995) or even longer (Ratcliffe et al. 2011; Taylor et al. 2011) to identify hot spot locations.”

“The crime prediction algorithm (Azavea 2014) uses machine learning techniques that incorporate components of risk terrain modeling (Caplan et al. 2011), near repeats (Ratcliffe and Rengert 2008), collective efficacy (Sampson 2012), and self-exciting point processes (see Mohler et al. 2011).”

“2 Hot spots policing is relevant because it continues to be the case that “focused police resources on crime hot spots provided the strongest collective evidence of police effectiveness that is now available” (National Research Council 2004, p. 250).”

“A substantial canon demonstrates that crime is unevenly distributed among places and victims (Felson 1987; Sherman et al. 1989; Weisburd and Eck 2004), and a number of spatial theories of crime can explain short-term changes in crime risk for small areas (Chainey and Ratcliffe 2005).”

“As Perry et al. (2013) point out, the theoretical foundations allowing for crime prediction are ably supported by routine activity theory (Cohen and Felson 1979), the rational choice perspective (Cor-nish and Clarke 1986), and crime pattern theory (Brantingham and Brantingham…”

“We used mixed effects models, with weeks (n = 14) nested within districts (n = 20), estimating full Bayesian models with Markov chain Monte Carlo (MCMC) estimation (Gelman et al. 2003; Gilks et al. 1996).”

“This is the best way to “obtain an unbiased estimate of the effect of selecting one treatment over another” (Detry and Lewis 2014, p. 85) in a randomized control trial.”

“A substantial canon demonstrates that crime is unevenly distributed among places and victims (Felson 1987; Sherman et al. 1989; Weisburd and Eck 2004), and a number of spatial theories of crime can explain short-term changes in crime risk for small areas (Chainey and Ratcliffe 2005).”

“Previous randomized studies examining policing in crime hot spots, however, have relied on a year of crime prior to the experiment (Sherman and Weisburd 1995) or even longer (Ratcliffe et al. 2011; Taylor et al. 2011) to identify hot spot locations.”

“Previous research has shown that street crime (Chainey et al. 2008) and property crime patterns (Gorr et al. 2003) are more predictable than other types of crime.”

“Concepts such as the Koper Curve (Koper 1995) or strategies such as focused foot patrol (Ratcliffe et al. 2011) may only be effective for long-term, chronic hot spots and prove ineffectual at addressing dynamic day-to-day crime projections, like those often identified by modern algorithms.”

“A substantial canon demonstrates that crime is unevenly distributed among places and victims (Felson 1987; Sherman et al. 1989; Weisburd and Eck 2004), and a number of spatial theories of crime can explain short-term changes in crime risk for small areas (Chainey and Ratcliffe 2005).”

“The crime prediction algorithm (Azavea 2014) uses machine learning techniques that incorporate components of risk terrain modeling (Caplan et al. 2011), near repeats (Ratcliffe and Rengert 2008), collective efficacy (Sampson 2012), and self-exciting point processes (see Mohler et al. 2011).”

“The crime prediction algorithm (Azavea 2014) uses machine learning techniques that incorporate components of risk terrain modeling (Caplan et al. 2011), near repeats (Ratcliffe and Rengert 2008), collective efficacy (Sampson 2012), and self-exciting point processes (see Mohler et al. 2011).”

“Constructing crime harm levels involves, in part, weighting crime occurrences by seriousness (Ignatans and Pease 2016; Ratcliffe 2015a, b; Sherman 2007; Sherman et al. 2016).”

“Predictive policing approaches such as risk terrain modeling (Caplan et al. 2011; Kennedy et al. 2011; Moreto et al. 2014) and techniques using short-term event patterns (Johnson et al. 2009a; Mohler et al. 2011) have emerged over the last decade or two with the promise of identifying short-term…”

“Previous randomized studies examining policing in crime hot spots, however, have relied on a year of crime prior to the experiment (Sherman and Weisburd 1995) or even longer (Ratcliffe et al. 2011; Taylor et al. 2011) to identify hot spot locations.”

“…approaches such as risk terrain modeling (Caplan et al. 2011; Kennedy et al. 2011; Moreto et al. 2014) and techniques using short-term event patterns (Johnson et al. 2009a; Mohler et al. 2011) have emerged over the last decade or two with the promise of identifying short-term crime hot spots.”

“Because each treatment district reverts to a business-as-usual state after eight treatment hours, it is possible to establish if there is a detectable increase in offenses during a subsequent nonexperimental time block, which would suggest temporal crime displacement (Bowers and Johnson 2003).”

“The Philadelphia predictive policing experiment At present, “there is little experimental evidence from the field demonstrating whether implementing an advanced analytics predictive model, along with a prevention strategy…works to reduce crime” (Saunders et al. 2016, p. 348).”

“…related to hot spots policing, then two recent observations from Weisburd and Telep are relevant: (1) there are numerous strategies that have not yet been rigorously tested, and (2) much more needs to be learned about the impact of new technology on policing effectiveness (Weisburd and Telep 2014).”

“Constructing crime harm levels involves, in part, weighting crime occurrences by seriousness (Ignatans and Pease 2016; Ratcliffe 2015a, b; Sherman 2007; Sherman et al. 2016).”

“Previous research has shown that street crime (Chainey et al. 2008) and property crime patterns (Gorr et al. 2003) are more predictable than other types of crime.”

“This is founded on evidence that offenders “forage” in nearby areas (Johnson et al. 2009b), and this acts in conjunction with a generalized risk heterogeneity model that highlights increased risk in certain areas (Johnson 2010; Tseloni and Pease 2003).”

“…colleagues’ comment of being “unaware of any field experiments conducted with patrol dosage randomly assigned to predictive policing hot spots” (Sher-man et al. 2014, p. 108) and “Currently, the policing aspect is mostly overlooked in evaluations of predictive policing” (Rummens et al. 2017, p. 7).”

“Arrest can reinforce “the capacity of the police to deter crime through the threat of apprehension” (Lum and Nagin 2017, p. 344).”

“This is founded on evidence that offenders “forage” in nearby areas (Johnson et al. 2009b), and this acts in conjunction with a generalized risk heterogeneity model that highlights increased risk in certain areas (Johnson 2010; Tseloni and Pease 2003).”

“Blocked experimental designs have been around for roughly a century and are generally credited to statistician R. A. Fisher (Box 1980; Fisher 1935).”

“While techniques such as aoristic analysis (Ratcliffe 2000; Ratcliffe and McCullagh 1998) can help, limitations with Philadelphia Police data reporting precluded using this approach here.”

“Constructing crime harm levels involves, in part, weighting crime occurrences by seriousness (Ignatans and Pease 2016; Ratcliffe 2015a, b; Sherman 2007; Sherman et al. 2016).”

“This requires any police department to be extremely responsive if a predictive intervention is to be successful (Haberman and Ratcliffe 2012; Ratcliffe and Rengert 2008; Wyant et al. 2012).”

“…approaches such as risk terrain modeling (Caplan et al. 2011; Kennedy et al. 2011; Moreto et al. 2014) and techniques using short-term event patterns (Johnson et al. 2009a; Mohler et al. 2011) have emerged over the last decade or two with the promise of identifying short-term crime hot spots.”

“First, while Drawve (2016) demonstrated that the PAI is more suited to comparisons between techniques or parameter choices, we report PAI mean and standard deviations for the five control districts across the property crime phases.”

“While techniques such as aoris-tic analysis (Ratcliffe 2000; Ratcliffe and McCullagh 1998) can help, limitations with Philadelphia Police data reporting precluded using this approach here.”

“If predictive policing is, in essence, an application of microlevel hot spots policing, then there is a need to consider explicitly the strategic translation of a place-specific microlevel policing tactic up to a district-wide or city-level implementation (Sherman et al. 2014).”

“Gorr and Lee (2015) have suggested that short-term, temporary hot spots, which may be identified by predictive policing methods, can be a more effective focus for crime prevention than large chronic hot spots and provide for greater equity of crime reduction resources.”

“This was estimated to be more than necessary given the short half-life of successes associated with hot spot policing interventions (for example, see Sorg et al. 2013).”

“Predictive policing approaches such as risk terrain modeling (Caplan et al. 2011; Kennedy et al. 2011; Moreto et al. 2014) and techniques using short-term event patterns (Johnson et al. 2009a; Mohler et al. 2011) have emerged over the last decade or two with the promise of identifying short-term…”

“We report the prediction accuracy index (PAI), a ratio measure comprising a numerator of the proportion of crime events falling within the mission grids for each district and a denominator of the percentage of the relevant police district covered by the mission grids (Chainey et al. 2008; Eck et al. 2017; Flaxman et a…”

“This requires any police department to be extremely responsive if a predictive intervention is to be successful (Haberman and Ratcliffe 2012; Ratcliffe and Rengert 2008; Wyant et al. 2012).”

“We report the prediction accuracy index (PAI), a ratio measure comprising a numerator of the proportion of crime events falling within the mission grids for each district and a denominator of the percentage of the relevant police district covered by the mission grids (Chainey et al. 2008; Eck et al. 2017; Flaxman et a…”

“Constructing crime harm levels involves, in part, weighting crime occurrences by seriousness (Ignatans and Pease 2016; Ratcliffe 2015a, b; Sherman 2007; Sherman et al. 2016).”

“…comprising a numer-ator of the proportion of crime events falling within the mission grids for each district and a denominator of the percentage of the relevant police district covered by the mission grids (Chainey et al. 2008; Eck et al. 2017; Flaxman et al. 2019; Tompson and Townsley 2010).”

“As Weisburd and colleagues note, “police and scholars must ask not just ‘what the strategy is’ that will be effective but also ‘how we gain enough dosage’ for such strategies to have large-scale crime prevention impacts across the city” (Weisburd et al. 2015, p. 384).”

“…(PEI) which indicates how well a forecasting algorithm performs compared to the best it could have done, given the available crime and distribution (Lee and Eck 2017), as shown: Where n* is the maximum number of crimes within the district that are forecastable within the area predicted (“ a ,” or…”

“…of historical data to create a spatiotemporal forecast of areas of criminality or crime hot spots that will be the basis for police resource allocation decisions with the expectation that having officers at the proposed place and time will deter or detect criminal activity” (Ratcliffe 2014, p. 4).”

“Second, block randomized designs may improve statistical power (Gill and Weisburd 2013).”

“When Summers and Rossmo (2019) provided over 130 offenders with a vignette featuring two police cars driving by in a short space of time, 63% said they would displace their offending to another time or place, and 23% said they would give up for the day.”

“…Copenhagen, Denmark (Attermann 2017) (62) Hot Spots Policing in Bogotá, Colombia (Blattman et al. 2017) (63) Philadelphia Predictive Policing Experiment (Ratcliffe et al. 2017) (64) Flint DDACTS Program (Rydberg et al. 2017) (65) Operation Strikeforce in Buffalo, New York (Wheeler and Phillips…”

“While prior short term hot spots approaches have yielded successful interventions (Mohler et al. 2015; Ratcliffe et al. 2020), the majority of the research on effective hot spot strategies on crime point to more long term, problem solving approaches (Braga et al. 2014).”

“American policing has its roots in slavery (Browne, 2015), with slave patrols serving as the first examples of organized policing units (Bass, 2001; Reichel, 1988; Silberman, 1978).”

“Ratcliffe et al. (2021) tested a predictive model in Philadelphia.”

“…between analysts and police officers may be unstable and issue-driven because officers believe their experiential knowledge provides a better rationale for resource allocation than the outputs of data-driven tools and algorithms (Meijer et al., 2021; Ratcliffe et al., 2020; B. Taylor et al., 2007).”

“Moreover, in their recent evaluation of Philadelphia’s predictive policing program, Ratcliffe and colleagues [57] also observed surprising effects related to violent crime in treatment areas as such events increased during their trial.”

“While this work focuses on long term crime forecasting, the application of using DBSCAN can potentially be extended to short term crime forecasting (Flaxman et al., 2019; Garnier et al., 2018; Lee et al., 2020; Mohler et al. 2015; Ratcliffe et al., 2020).”

“rows of crime incident for the period of 1st January 2003 to 13th May 2015 [32].”

“Neither is there an expert agreement with regards to the effectiveness of algorithmic policing (Meijer and Wessels 2019, Browning and Arrigo 2021, Ratcliffe et al. 2021).”

“The revelations from this research not only lay the groundwork for subsequent inquiries into language usage within law enforcement communication systems but also offer broader insights into the human consequences of such systems.”

“Finally, previous research has observed that—in addition to altering motorist behavior—marked police vehicles can reduce more traditional forms of crime and disorder (Ratcliffe et al., 2021), even when parked and unoccupied (Worrall et al., 2022).”

“Finally, Ratcliffe et al. (2020) performed a randomized experiment to explore the impact of different patrol strategies on violent and property crime in microscale predicted crime areas.”

“The lack of empirical evidence for such predictive programs, or the opacity of such evidence, have caused problems, along with the models and their subsequent predictions being accepted in the dayto-day life of the police, who are used to being guided by their experience and intuition rather than by the results of a "…”

“The findings of a recent randomized experiment [35] suggested that the use of predictive policing software can reduce certain types of crime but also highlighted the challenges of estimating and preventing crime in small areas.”

“Experiments study the effects of using predictive policing to allocate patrols compared to other approaches at the district-level, examining effects on crime (Mohler et al., 2015; Ratcliffe et al., 2020) and racial disparities in arrests (Brantingham et al., 2018).”

“…carefully consider prioritizing AI applications with relatively strong empirical support, such as geospatial predictive patrols, which have shown greater effectiveness than traditional methods in reducing crime when applied to high-risk areas (e.g., Mohler et al., 2015; Ratcliffe et al., 2021).”

“While predictive policing (see Rat-cliffe 2019; Ratcliffe et al. 2021; Berk 2021; Hunter et al. 2021; Sabol and Baumann 2022) is in its infancy, incorporating the thefts of commodities (e.g., metals, fuel, livestock) into predictive policing models may demonstrate their usefulness.”

“These models can then be used by law enforcement or other stake-holders to allocate resources for crime prevention [16], [19], community policing [1], collective efficacy initiatives [21], or other interventions.”

“This sort of automated predictive tool has not just been a product of criminological theory, but also a means to test and develop said theories in a range of fi eld experiments (Haberman and Ratcliffe, 2012; Mohler et al., 2015; Ratcliffe et al., 2021).”

“Related to the implementation of crime prevention strategies through Community Policing in Dialogic Patrols, data on Tanggon Kosala, Vol.14, No.1, 2025, pp. 45-53 e-ISSN 2775-8478.”

“The findings of a recent randomized experiment (Ratcliffe et al., 2020) suggested that use of predictive policing software can reduce certain types of crime but also highlighted the challenges of estimating and preventing crime in small areas.”