“…Given the long term predictability of crime at micro places however (Andresen et al 2017;Curman et al 2014;Weisburd et al 2004;Wheeler et al 2016), we believe such an approach to identify places that are long term crime hot spots, and decompose the effects to different local crime generators and demographic factors, is an approach that can be successfully incorporated into problem-oriented hot spots policing strategy. 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). Thus even if one can effectively predict crime at micro places, it may be that long term strategies to reduce crime at places is the more effective approach for police departments to take.…”
Objectives: We illustrate how a machine learning algorithm, Random Forests, can provide accurate long-term predictions of crime at micro places relative to other popular techniques. We also show how recent advances in model summaries can help to open the 'black box' of Random Forests, considerably improving their interpretability. Methods: We generate long-term crime forecasts for robberies in Dallas at 200 by 200 feet grid cells that allow spatially varying associations of crime generators and demographic factors across the study area. We then show how using interpretable model summaries facilitate understanding the model's inner workings. Results: We find that Random Forests greatly outperform Risk Terrain Models and Kernel Density Estimation in terms of forecasting future crimes using different measures of predictive accuracy, but only slightly outperform using prior counts of crime. We find different factors that predict crime are highly non-linear and vary over space. Conclusions: We show how using black-box machine learning models can provide accurate micro placed based crime predictions, but still be interpreted in a manner that fosters understanding of why a place is predicted to be risky.
“…Given the long term predictability of crime at micro places however (Andresen et al 2017;Curman et al 2014;Weisburd et al 2004;Wheeler et al 2016), we believe such an approach to identify places that are long term crime hot spots, and decompose the effects to different local crime generators and demographic factors, is an approach that can be successfully incorporated into problem-oriented hot spots policing strategy. 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). Thus even if one can effectively predict crime at micro places, it may be that long term strategies to reduce crime at places is the more effective approach for police departments to take.…”
Objectives: We illustrate how a machine learning algorithm, Random Forests, can provide accurate long-term predictions of crime at micro places relative to other popular techniques. We also show how recent advances in model summaries can help to open the 'black box' of Random Forests, considerably improving their interpretability. Methods: We generate long-term crime forecasts for robberies in Dallas at 200 by 200 feet grid cells that allow spatially varying associations of crime generators and demographic factors across the study area. We then show how using interpretable model summaries facilitate understanding the model's inner workings. Results: We find that Random Forests greatly outperform Risk Terrain Models and Kernel Density Estimation in terms of forecasting future crimes using different measures of predictive accuracy, but only slightly outperform using prior counts of crime. We find different factors that predict crime are highly non-linear and vary over space. Conclusions: We show how using black-box machine learning models can provide accurate micro placed based crime predictions, but still be interpreted in a manner that fosters understanding of why a place is predicted to be risky.
“…Districts were re‐randomized prior to the violent crime phase. A full suite of experimental results (Ratcliffe et al., 2020) and operational implementation challenges and lessons (Ratcliffe, Taylor, & Fisher, 2019) were reported elsewhere. The full final study report also is available (Ratcliffe, Taylor, Askey, Fisher, & Koehnlein, 2019).…”
Section: Focusmentioning
confidence: 99%
“…As described in Ratcliffe et al. (2020, Figures 2 and 3, 17/27): [O]fficers patrolled the treatment areas to varying levels throughout the shift, with officers getting to the treatment areas earlier for the 8 a.m. to 4 p.m. property crime phases, but officers patrolling treatment areas more extensively later in the shift for the violent crime phase (6 p.m. to 2 a.m.). At least 50% saturation of treatment areas was achieved for 3.5 and 3.75 hours, respectively, for property and violent crime phases.…”
Section: Focusmentioning
confidence: 99%
“…More specifically for property crime (Ratcliffe et al., 2020, Figure 2), for marked and unmarked cars considered together, using 15‐minute block coding by ride‐along observers in 79 observations, this means that in at least 50% of these observations observers coded activity occurring within the mission areas for 15 specific 15‐minute segments, or 3.75 hours.…”
Hinkle et al. (2013) highlighted a statistical powerlessness problem in hot-spots policing experiments in midsized cities with moderate property crime rates. The current work demonstrates that this problem is less readily resolved than previously suspected. It reviews results from a predictive policing randomized control trial in a large city with property crime rates higher than Chicago or Los Angeles. It reports, for the first time, a graphical analysis indicating the marked car patrol intervention, practically effective at the 500′ by 500′ (mission) grid level, three grids per shift, likely had a district-wide impact on reducing reported property crime. In addition, it reviews results of a series of thought experiments exploring statistical power impacts of four modified experimental designs. Only one alternative design, with spatially up-scaled predictive policing mission areas and concomitantly higher property crime prevalence rates, produced acceptable statistical power levels. Implications follow for current theoretical confusion in community criminology about concentration effects and units of analysis, and how models organize impacts across those different units. Implications follow for practice amid ongoing concerns about whether predictive policing works and, if it did, how to gauge its impacts and social justice costs. Policy Implications: The current work brings to the fore important questions beyond "does predictive
“…13 The analysis is based on an estimate that the team calls "predictive efficacy," which describes the extent to which the most harmful crimes are able to be predicted and deterred. Violent crimes like assault and homicide may be more difficult to predict than property crimes like burglary or larceny (e.g., Ratcliffe, Taylor, and Askey 2017), but "communities" may find it more valuable to prevent violent crimes because they are more harmful; efforts to thwart more easily predictable crimes may have greater success rates, but the payoff from deterring less predictable and more harmful crimes may be higher.…”
Predictive analytics and artificial intelligence are applied widely across law enforcement agencies and the criminal justice system. Despite criticism that such tools reinforce inequality and structural discrimination, proponents insist that they will nonetheless improve the equality and fairness of outcomes by countering humans’ biased or capricious decision-making. How can predictive analytics be understood simultaneously as a source of, and solution to, discrimination and bias in criminal justice and law enforcement? The article provides a framework for understanding the techno-political gambit of predictive policing as a mechanism of police reform—a discourse that I call “predictive policing for reform.” Focusing specifically on geospatial predictive policing systems, I argue that “predictive policing for reform” should be seen as a flawed attempt to rationalize police patrols through an algorithmic remediation of patrol geographies. The attempt is flawed because predictive systems operate on the sociotechnical practices of police patrols, which are themselves contradictory enactments of the state’s power to distribute safety and harm. The ambiguities and contradictions of the patrol are not resolved through algorithmic remediation. Instead, they lead to new indeterminacies, trade-offs, and experimentations based on unfalsifiable claims. I detail these through a discussion of predictive policing firm HunchLab’s use of predictive analytics to rationalize patrols and mitigate bias. Understanding how the “predictive policing for reform” discourse is operationalized as a series of technical fixes that rely on the production of indeterminacies allows for a more nuanced critique of predictive policing.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
