Introduction

2Interest in the scientific dimension of situational crime prevention should not be seen simply as a show of erudition. Equating this prevention model to legitimate scientific practice is an important condition for it to be disseminated, accepted and recognized (Benbouzid, 2011). Moreover, while discourse on crime prevention in England, the birthplace of situational crime prevention, has largely been shaped as discourse on technological and social science and innovation (Tilley, Laycock, 2007), it has rarely adopted a political perspective on the scientific controversies underpinning them.

3The qualitative analysis of this body of texts shows a process of mathematization (Israël, 1996) and a controversy that is often overlooked in the official history of the development of situational crime prevention: the controversy between Ken Pease and Tim Hope (and their respective colleagues). The former is known for having contributed decisively to situational crime prevention (SCP) by proposing the principle of prediction of repeat victimization (Laycock, Farrell, 2003; Pease, Tseloni, 2014), and the latter for having formalized community crime prevention (CCP) as a substitute model for situational crime prevention (Hope, 1995; Hope, Karstedt, 2003). The difference between these two prevention models is well-known in the criminology community: CCP considers the social organization of a city at all levels as both the cause and the solution of criminality issues, while SCP considers that the immediate environment in its simplest material forms directly affords the means for anticipation, surveillance and prevention of criminal acts. But the dividing line between these two models is not simply a conceptual difference between the SCP and CCP models – which in fact have often been shown to overlap in many ways. It was actually the similarities between the initial projects of Tim Hope and of Ken Pease that allowed for the full-blown controversy to arise; that is, their underlying critical analyses made situational crime prevention debatable and substantially amendable.

4The controversy revolves around research on the statistical modelling and the prediction of repeat victimization, on which Pease and Hope worked together in the 1990s. Their aim was to understand why a small number of victims suffer the most burglaries, and how processes of victimization are repeated in space and time. The controversy itself can be mapped as a non-dialogue around the statistical tests that model repeat victimization and produce a new form of time- and space-based situational crime prevention assisted by algorithms, known as predictive policing (Perry et al., 2013). As the empirical data piled up, interpretations of the repeat victimization phenomenon increasingly weighed towards local experiments on situational crime prevention. Despite Tim Hope’s many attempts to refute this interpretation, the idea that the best way of reducing delinquency was to use situational crime prevention techniques as emergency security measures to protect repeat victims gradually gained currency in public policies on security.

5The Pease/Hope controversy seems to be an ideal case study for investigating the scientific content of the black box of situational crime prevention. Most of the numerous critiques of situational crime prevention have not actually meticulously unpacked the theories and scientific facts that legitimize it. Only with an in-depth exploration of the controversy did I gradually discover that the scientific roots of predictive policing could be traced back to the history of research on the modelling of repeat victimization.

6To further this intuitive understanding that had become a conviction, and to back it up more soundly, I carried out a systematic analysis of the written sources and especially the scientific reports and articles on repeat victimization. My aim was to explore available sources in detail, in order to see how they corresponded and were articulated to one another. This investigation of “science in the making” required me to cross-compare written sources with information obtained from the individuals that I saw as being at the core of the controversy. I was not able to meet them all, but did have several long discussions with the two main protagonists, Tim Hope and Ken Pease. With Tim Hope I held two long interviews, in Vilnius and Prague, respectively in September 2011 and September 2014. He furthermore provided me with a long account of the key periods of his career and the scientific debates in which he had been engaged. In the case of Ken Pease, I was unfortunately unable to meet him face-to-face but he did agree to an intense exchange of emails between 2010 and 2014. This correspondence was based on my precise questions, with all the latitude that he considered appropriate in order to answer them. The aim was above all to obtain clarification on the stages of his own career in relation to his published work – which I had mapped out prior to our correspondence –, as well as the circumstances of his work on situational crime prevention, and the reorientation of his research towards repeat victimization phenomena. Based on these questions and this reorientation, I asked him to explain how he agreed and disagreed with Tim Hope.

7The analysis of the scientific dynamic of this controversy, from the 1990s to date, draws on the sociology of science and techniques (Pestre, 2006) and, in particular, that of the sociology of quantification (Armatte, 2010; Desrosières, 1993). I treat the scientific models “symmetrically”, without focusing more on one or the other of the two positions studied. In opening the black box (Latour, 1989) of the models, I consider them as the cognitive instruments that the protagonists in a battle use to support the arguments justifying their different ways of conceiving of and implementing crime prevention.

8As a sociologist of science, my aim is limited to ensuring that predictive policing (an avatar of situational crime prevention) remains “questionable”, by keeping myself outside of any debate on its effectiveness. I will have attained my objective if this article enables the reader to understand predictive policing above all as a problem that is both, and inextricably, political and scientific.

I - The quantitative criminology group and the statistical modelling of repeat victimization

11By no means did this proposition stand to reason. In fact it was contrary to the common sense of all the stakeholders of prevention and security policy, especially the police, who never imagined that a burglar would commit the same burglary several times. Pease nevertheless managed to persuade the actors on the ground to apply situational crime prevention techniques to people who had already been burgled once (patrols, security advice, organization of neighbourhood watches, etc.). The project very quickly appeared to be successful. There was not only a steep decline in the number of burglaries, but also real involvement by the actors on the ground. Research on repeat victimization consequently appeared to be a promising avenue for giving situational crime prevention new impetus.

12As soon as he had some success in the government sphere, Pease envisaged transforming this applied problem of repeat victimization into a full-blown research programme. Very simply, with the help of a few Canadian colleagues, he started by describing the statistical distribution of temporal series of repeat victimization, based on police data (Polvi et al., 1991). He showed that in the first month following a burglary, the risk of being a victim of another burglary was twelve times higher for homes having already been victims than for the others. After this one-month period, the risk declined considerably.

13This first statistical description of the temporal evolution of victimization was a cue for Pease to delve deeper into research on repeat victimization. He applied more sophisticated statistical skills by enrolling econometricians from Manchester University, specialized in financial modelling. Pease suggested that they analyze repeat victimization based on data from the large-scale British Crime Survey. This was the beginning of the creation of a research team called the Quantitative Criminology Group (QCG), initially comprised of the three econometricians Denise Osborn, Alan Trickett and Ellingworth, and of three criminologists: Ken Pease, Graham Farrell, a young psychologist who had just completed his PhD on repeat victimization, and Tim Hope, a young sociologist specialized in crime prevention (I will revert to Tim Hope further on in this paper).

1) Incidence, prevalence and concentration

14With funding from the Economic and Social Research Council, the QCG first embarked on a study of repeat victimization for prevention policy, but without a priori associating its research with any particular prevention model. The first step was to analyze the data from victimization surveys, mapping the distribution of the probability of repeat victimization in the population. The researchers used fairly sophisticated statistical models (laws of distribution and hypotheses on random variables) to obtain an approximate mathematical description of the data. At the outset, the QCG first envisaged an exploratory analysis in which it sought to describe the differences of incidence of victimization in different geographical areas. It published its analysis in What is different about high crime areas (Ticket et al., 1992), now considered as a seminal article in repeat victimization research. In this article the authors applied a new conceptualization of the way of calculating the risk of victimization, by arithmetically linking the prevalence, incidence and concentration (called vulnerability) of victimization in the formula:

2) Interpretive flexibility of the modelling of distribution

16Exploring the explanation of repeat victimization further, Ken Pease and Graham Farrell (1993) put forward a dual hypothesis to account for the mechanism underlying victimization: the repeat victimization of a person or a home can be seen either as a flag signalling a relatively stable risk of victimization, or as an indication that the incidents of victimization boost the probability of the occurrence of subsequent incidents. Translated into statistical terms, the hypothesis of the flag is called risk heterogeneity, and that of the boost is said to be state-dependent.

18In the late 1970s the first modelling studies on victimization survey data showed that the negative binomial model was satisfactory for representing the distribution of victimization in the population. The explanation was given in terms of certain social groups’ propensity to be victimized repeatedly, the idea of contagion being considered as irrelevant (Sparks preferred more sociological explanations for repetition, and maintained that epidemiological reasoning encountered its limits here). In the mid-1990s, however, the two explanations seemed relevant and the QCG set out to untangle the boost from the flag. How could the repetition mechanism be examined specifically, from a statistical point of view? Through what process does an incident of victimization repeat itself or cease? And when the victimization continues, what frequency does it attain over time? What are the specific (individual and local) risk factors of the first hurdle and of the second one? What is the correlation between the first and second hurdles?

19The QCG wanted to integrate into the same empirical examination what had until then been excluded: the statistical description of the shift from the state of non-victim to that of victim (first hurdle), then from victim to repeated victim (second hurdle). To produce this type of model, the econometrics literature proposed a solution consisting in elaborating a system of simultaneous equations, a multivariate statistical modelling technique that was well known in the econometrics of education. In 1996, inspired by this standardized statistical solution, the QCG published the article Are repeatedly victimized households different (Osborn et al., 1996), which was considered as a significant methodological step forwards in research on the modelling of victimization. The Group was however unable to identify a factor that was more significant than the others to explain repetition. As one of the members commented: we inferred that we were not measuring any predictors that might distinguish repeat victims from victims in general, and that was different from those common to all victims and distinguished them from non-victims (Hope, Trickett, 2004a, 389).

II - Research around the boost hypothesis and the development of predictive policing

21In the late 1990s, in-depth analysis of boost mechanisms became a major research topic in situational crime prevention. How could detailed insight into the boost phenomenon be gained that could be used to fight and even predict crime? From a situational crime prevention perspective, explaining a crime amounted to relating it to the behaviours of delinquents in the situation where they act, and to the target’s vulnerability. Research on repeat victimization henceforth included situational crime prevention strategies into a complex and evolving space-time. It was from this perspective that Richard Townsley, Shane Johnson and Kate Bowers, in close collaboration with Ken Pease, embarked on a series of studies in the early 2000s, at the newly-created Jill Dando Institute of Crime Science (JDICS), England’s first academic research and development laboratory entirely devoted to situational crime prevention.

1) Modelling contagion

23The JDICS team of researchers wondered how to represent the spread of victimization in statistical terms, and how to identify the more or less repetitive spatio-temporal configurations on which prevention strategies could be built. To answer these questions they applied spatial analysis statistical tools – a standard reflex when one wishes to identify spatio-temporal concentrations of events. In particular, they were interested in Knox’s spatio-temporal statistical test, well-known for the analysis of contagion phenomena. Developed in the 1960s in an epidemiological research framework, Knox’s test uses a Poisson test to determine whether the time-space distribution of events differs significantly from an independent, random distribution. This method was used to test the simultaneous existence of groups of victims in time and space – a statistical refinement that enables the spatio-temporal interdependence of crimes to be represented in the form of clusters. By inputting police statistics on the time and place of crimes, the researchers were able to use this test to model the way in which burglaries were repeated and spread by contagion.

24Based on the Knox test, the researchers embarked on a series of studies on the communication of the risk of victimization. Townsley published one of the first articles on contagion, based on Knox tests, and proposed the notion of infectious burglaries (Townsley, 2000, 2003). In his PhD thesis he had already shown important elements of spatio-temporal analysis, such as the fact that the most unstable burglary hotspots corresponded largely to the addresses of repeat victims, whereas the most stable hotspots were linked more to aspects of the immediate social and physical environment. In his article on infectious burglaries, Townsley pursued his analysis of neighbourhood similarities, but showed this time that a first burglary increases the risk of another one following on a close target, provided that the new target has similar social and physical characteristics to the first.

25Using the same Knox test, Kate Bowers and Shane Johnson showed that a first burglary increased the risk of a burglary in a 400-meter radius, in an estimated interval of one to two months (Johnson, Bowers, 2004b; Bowers, Johnson, 2005). They also showed that although the poorest areas are those with the most repeat victimization, clusters of burglaries in space and time are found more in the wealthiest areas. Hence, a burgled house next to a first burglary is far likelier to be burgled than a house situated a few meters away, especially during the first week after the first burglary. Moreover, in the same street, burglaries are repeated on houses with the same architectural characteristics. In other words, a house with a different architectural layout in a street with identical houses can reduce the risk of being burgled.

2) Developing a predictive machine

26Owing to the Knox test and data geo-referenced on the scale of a street and dated, the scientific discourse on repeat victimization gradually came to contain more and more details on the spatio-temporal boost process. But it was not enough for the researchers at JDICS to provide ever-more detailed elements of analysis on the spatio-temporal dynamics of repeat victimization. The JDICS research policy was also to develop innovative technology to fight crime. How could the spatio-temporal analysis be used to devise prevention strategies? To answer this question, Pease joined the JDICS researchers in the early 2000s, to develop an operational tool for crime prediction: Prospective Crime Mapping, known as PROMAP. It was based on the modelling of spatio-temporal changes in repeat victimization in a given area (Bowers, Johnson, 2004a). Their aim was to differentiate their work from prediction as it had been practised until then with calculation methods and hotspot representations (Dupont, Ratcliffe, 2000; Weisburd, 2009). With the latter approach, simple representations of spatial concentrations of past burglaries were considered as good predictors of future crime. The objective of PROMAP was to map not the past as a good means of revealing the future (the classical practice of hotspot policing), but rather future crimes.

III - Tim Hope and the theory of reflexive securitization 

1) The start of the controversy

31The main traces of this critique are found in the article published in 2004 with Trickett, but the main results were formalized from 1995 (Hope, Walklate, ibid.). Hope and Trickett (2004a) considered that too little analysis had been granted to the fact that far more non-victims were measured than victims (victimization is a rare phenomenon), and that the surveys showing the highest frequencies of victimization were the rarest. According to them, if one wanted to account for the over-dispersion of distribution properly, one had not only to take into account the concentration of victimization on a small part of the population (the right-hand tail of the distribution curve), but also to seek to explain the extreme concentration of non-victimization on a very large portion of the population (left-hand tail of the distribution curve). They then showed that most victims (people who had suffered a first instance of victimization) did not remain victims in the long term; on the contrary, they usually returned to the status of non-victims after that first time. Hope and Trickett thereby concluded that the distribution of victimization in the population could be explained by a differential and decreasing tendency towards immunity from crime risk.

32To demonstrate this immunity hypothesis, Hope and Trickett reasoned by simulation, based on equations of the bivariate probit model. They imagined three levels of probability of experiencing a first instance of victimization (p = 1 or 0.4 or 0.005). They used the bivariate probit model to simulate the probability of a subsequent victimization event, in relation to the three levels of risk, and deduced the probability of a subsequent non-victimization. For the most vulnerable people, the risk of experiencing a first instance of victimization increased the risk of a subsequent one (a 22% risk), but not to the extent of having less chance of not being a victim. For people with an average risk of suffering a first instance of victimization (40% chance), the risk of experiencing a second one remained the same, and with the extreme hypothesis of 100 chances out of 100, the risk of a subsequent victimization declined (0.60%). Thus, Hope and Tricket were able to show that the risk of a second repeat victimization depended on the level of risk a priori, with a general tendency, over time, towards non-victimization.

33Based on these results, Hope argued that the hypothesis of an increase in exposure to risk over time (boost) and related theories should be dropped, and that other interpretive frameworks of distribution of victimization in the population should rather be adopted. For his new framework of interpretation, Hope drew on Giddens’ theory of structuring, which was very popular in the 1990s. I will not to examine this well-known theory here. Let us simply bear in mind that Giddens’ structure is a set of rules and resources organized repetitively, updated and coordinated in the form of traces in actors’ memories (Rojot, 2010) – in this case, in those of the victims. Most important is the focus on "structures" rather than on epidemiological factors. This is how, according to Hope, the notion of structure can be applied to conceptualize repeat victimization:

This [structuration theory] provides one way of understanding the dynamism between, for example, the structural location of women (one way of understanding women’s powerlessness, a defining characteristic of being a victim), and women’s negotiation of their structural location (one way of understanding the term survivor). It is this kind of theoretical starting point, which neither treats individual characteristics as given nor individual events as products of such givens, which permits the development of a critical edge to the work on repeat victimization (Hope, Walklate, 1995).

34To continue his research on repeat victimization, Hope explored the ways in which victims showed their competencies and oriented their behaviours based on their knowledge on how victimization took place. From this point of view, it was less risk that interested Hope, than "security" in the broad sense of the term, close to the concept of ontological security of the actor, defined by Giddens. How do victims explain their own vulnerability? What makes them review this explanation? How to they adapt? What are the conditions of this adaptation?

2) Explaining immunity: the clubbization of security

35By raising these questions, Hope undertook research that radically turned the tide and paved the way for prevention policies which differed substantially from situational crime prevention. There was thus a shift from the issue of distribution of victimization to that of the distribution of reflexive securitization. Surveys on victimization provided abundant useful data on the consumption of security products and individuals’ engagement in their own quest for security. Based on these data, Hope identified three forms of preventive actions that people engaged in to make their private sphere safer: neighbourhood watch (all the activities related to residential security, such as marking property and keeping a watch in the neighbourhood), security technologies (e.g. alarms) and fortress security (everything pertaining to physical security related to the home). The analysis of the data from surveys on victimization shows that the tendency to engage in these preventive actions is related to the perception of the risk of victimization, and that the people who are most readily take security measures are those in the most privileged classes.

The theory of reflexive securitisation would seem a plausible way of linking burglary victimisation and private security trends. The residential clubbing and consequent intensification of private security, may affect mid-range communities the most: very low crime communities have an excess of community over risk, and thus an excess of immunity; their investment in private security may be primarily symbolic. In contrast, high crime neighbourhoods have an excess of victimisation, which negatively affects trust and social capital formation […] Some part of the decline in burglary may have been introduced asymmetrically into the trend by the socio-spatial distancing between victims and offenders brought about by the changing ecology of tenure in the UK over the period (Hope, 1999, 2001). Thus, reflexive securitization may explain why the bulk of burglary reduction has come about through a reduction in prevalence, since a large number of erstwhile victims may have been removed from risk, and thus would be eligible no longer for selection as repeat or multiple victims (Hope, 2007).

37By conceptualizing security as a mutualized good in his explanation of trends in crime against goods, Hope was better able to account for the role of non-governmental actors (citizens) and the relations they maintain with one another and other institutions. He was thus able no longer to focus on government action only, particularly that of the police, when evaluating major burglary reduction experiments. Tim Hope drew up a model of crime prevention based on the notion of the "good". As the diagram below shows, it enables us to see crime from a global and structural perspective:

39The controversy stemmed largely from the broader disagreement between the immunity model and the boost model, which related to two different conceptions of what produces security: the assumption of the immunity model is that protection factors actually say more about security than about the risk itself (Hope, Trickett, 2004a). According to Hope, crime prevention becomes a matter of detecting and predicting the disintegration of these factors of protection through an ever finer understanding of the unequal conditions of access to security.

IV - Turning the interpretation around

40Peace’s prevention (predictive policing) strategy and Hope’s conceptual innovation (reflexive securitization) were both based on flexible interpretations – respectively, the boost model and the immunity model – of statistical studies on the distribution of victimization in the population. They therefore had to carry out more fundamental research on this distribution. The advocates of situational crime prevention still had to solve the riddle of two explanations of repeat victimization that had never been untangled: boost or flag? As for the group represented by Hope, a comprehensive model still had to be found, that took into account the inflation of zero victimization, in order to account for the role of immunity in the distribution of victimization.

41To build these two models, the researchers relied primarily on so-called "frequentist" statistical tools. They applied logistic regression econometric methods, with a view to isolating the effects of certain variables. From the end of the 2000s this type of reasoning seemed to have reached its limits, and the two groups of researchers turned to other statistical techniques to further their analysis. The development of predictive tools and the use of increasingly sophisticated computational techniques (particularly the Monte Carlo simulation used in Knox tests) gradually led the JDICS researchers to investigate the social simulation techniques of Nigel Gilbert (1994) that had been taken up in all the social sciences from the early 1990s. At the same time, the untimely death of Alan Trickett prompted Tim Hope to join another sociologist/statistician with whom he sought to apply the tools of exploratory data analysis, based more on the analysis of co-occurrences of victims’ properties than on the idea of pure effects of the variables. I now explore each of these two changes of method individually, and show that they are two ways of turning the interpretation around.

1) Social simulation of boost and flag

42Social simulation as defined by Gilbert interested Pease and his colleagues as it afforded the possibility of representing artificial social situations from which the researchers could observe spatio-temporal dynamics. Of the many simulation methods available, multi-agent simulation enabled them to push their analysis furthest. It afforded a high degree of flexibility in representing burglars, their behaviours and their interactions in an environment.

43Multi-agent simulation enabled criminologists to represent a delinquent’s behaviour by means of a computerized program. In an article published in 2011 (Pitcher, Johnson, 2011), Johnson and a young mathematician defined a multi-agent model offering a computer representation of burglars in the form of agents moving about in an environment consisting of targets (homes) whose attractiveness is both stable (depending on the situation) and changing (depending on the burglars’ action). The burglars can see the attractiveness of a target and surrounding targets, and react according to that attractiveness. The rules defining the burglar’s movement are translated into an algorithm that the authors represent as follows:

44Once the mathematical model had been parameterized and the algorithm implemented, the researchers could execute the model by varying its dynamic and static aspect and thus simulate different configurations: heterogeneity of the risk, dependence on the event without heterogeneity of the risk, and heterogeneity of the risk with dependence on the event. After many experiments, they argued in favour of a boost hypothesis that differed from that of the 1990s. Whereas in the classical multivariate models there was always the possibility of having left out hidden variables that might explain repetition in terms of risk heterogeneity, with their simulation the researchers showed that, on the contrary, repeat victimizations could not be explained by the configuration of risk heterogeneity alone:

In the introduction to this article, we discussed the possibility that observed space–time patterns of crime may be explained by a statistical artefact that occurs when results are aggregated for populations with very different risks (the flag account). However, the results of a series of simulations suggest that such models (as specified here) were insufficient and did not generate the types of pattern that are observed in real-world data [...] In contrast to the flag hypothesis, our results suggest that the boost account may offer a plausible explanation for why crime clusters in space and time (Pitcher, Johnson, 2011).

45The logic of statistical inference was reversed. It was a matter no longer of measuring the gap between a statistical model and the empirical data collected from victims in large surveys on the population at large, but rather of evaluating the explanatory significance of hypotheses on the mechanisms generating repetition. Whereas in statistical modelling the theoretical models of the heterogeneity of risk and of event-dependence were centred on the intensity of relations between the variables, in social simulation studies these same models were centred on the plausibility of the mechanisms underlying repeat victimization. This new approach enabled Johnson and his colleague to show the interdependency between risk heterogeneity and event-dependency in the explanation of repeat mechanisms. On this basis they inferred the importance of articulating conventional measures to reduce occasions, and the prediction of repetition. The authors’ interpretation of this "fundamental" result still supported situational crime prevention.

2) Immunity as a latent variable: the mobilization of data analysis

46As we saw in the second part of this article, Hope’s main intention was – even more so than Pease and his colleagues – to pursue his statistical work. This consisted in deducing, from samples of surveys on victimization, the "distribution of the probability" of victimization in the population of England and Wales. But we have seen that this estimation was not a research goal per se. It was an essential and strategic step in pushing forward a model that might justify a particular representation of the right way of going about prevention. In 2004 Hope and Tricket proposed the following model:

• A general trend towards non-victimization, independently of subsequent risks;

• A tendency for non-victims to maintain themselves in a state of non-victimization over time;

• A tendency for victims who are repeatedly victimized to retain a high frequency of victimization over time;

• A tendency for victims with a low level of victimization to revert to the general tendency of non-victimization.

In 2010 Hope sought a statistical technique that would enable him to group together and to take into account in more detail these different components of the repeat phenomenon, without worrying about their observability, as the BCS data available to him were not longitudinal (he could not directly account for the stochasticity of the process).

49Hope sought to model the distribution of each of these classes, based on the calculation of the probability of a respondent belonging to a class, according to the frequency of victimization. Respondents who had never been victims had a higher probability of being non-victims and a low probability of being intermediate victims. Only respondents with a high frequency of victimization belonged to the class of chronic victims. Moreover, the distribution of the probability of being a victim was heterogeneous for the two extreme classes (non-victims and chronic victims), whereas it was homogeneous for the intermediate class. Hope therefore wondered whether intermediate victims (which constituted a relatively large set) and chronic victims (which consisted of few victims but a very high level of victimization) might not pose etiological questions peculiar to them. LCA did not enable him to explore this further, but by raising this question Hope showed all the nuances that could still be brought to the modelling of the distribution of victimization. LCA enabled him to criticize the limited category of “repeat victimization” by showing forms of distribution according to class. This supported the idea of a heterogeneous mixture of the distribution of victimization in the population.

50The main contribution of LCA was that it enabled Hope to account for a general model of distribution. He showed that non-victimization had a strong influence on the general pattern of distribution. The class of non-victims accounted for 80% of the population but only 20% of victimization. The remaining 20% of the population (intermediate and chronic victims) accounted for 80% of the cases of victimization.

In sum, while the broad 80:20 inequality of the distribution persists, we can now see how that inequality is constructed out of a heterogeneous mixture of sub-distributions that lie between the two poles of maximum immunity and maximum exposure (Hope, forthcoming).

51Hope inferred that victimization could be predicted only on the basis of combined modeling of immunity and exposure to victimization risk. Immunity and exposure are two mutually interacting dimensions. But what was his intention when he showed that the general model of distribution of victimization could be conceptualized as propensities for immunity from and exposure to crime victimization, respectively? Hope explained his reasoning as follows:

In view of the nature of the distribution, non-victimization would appear to be a better predictor of non-victimization than would victimization be of victimization; in other words, long-run safety may be a more certain and reliable outcome for the general population than is the prospect of short-term, periodic risk. Of course, this is no consolation for the minority who, at any one time, suffer excessive victimization. It may be even more galling for them to know that a much larger proportion of their fellow citizens might be luxuriating in more than their “fair share” of safety while they suffer more than their fair share of harm (Hope, forthcoming).

52While research on the prediction of victimization focuses only on a part of the distribution model (repetition of victimization in the population), Hope showed the importance of understanding the causes of the zero. He therefore challenged all the underpinnings of predictive policing. According to him, crime is a problem that can be understood only by looking at how immunity and chronicity are established in the population. This helps to explain Hope’s complex proposition that has been ignored by all the advocates of situational crime prevention. Basing a public policy exclusively on the risk exposure model means disregarding the fact that the immunity of some is linked to the exclusion of others from security. It also means the failure to take timely and in-depth action to abate crime, and it prevents the protection of victims from being approached from a perspective of solidarity. Furthermore, predictive policing solutions can have counter-intuitive effects such as reinforcing logics of immunization that worsen the exclusion of some, or disrupt the virtuous dynamics that lead to immunity. This complex idea could also be explored by means of social simulation tools, taking into account the interaction of these two components of the distribution of victimization in the population.

To conclude: the public problem of confinement of a scientific controversy

54Seeking to sort the wheat from the chaff in this controversy will not help to further the debate. One can hope, rather, that this comparison has shed new light on the status of knowledge in public crime prevention policies. This exploration of the dynamics of the production of scientific knowledge, and particularly the plurality of possible and contradictory perspectives on prevention, points to the need to focus on the public implications of scientific controversies that remain confined to the scientific sphere whereas they actually concern the entire political sphere.

55As I have shown in this article, Hope has been battling for nearly twenty years to get prevention specialists of to reach agreement on what they disagree on. But the dynamics of the controversy between the two groups has ever taken the form of confrontation. Ken Pease and his colleagues are fully aware of Tim Hope’s criticism, yet they are largely unconcerned because they consider it too distant from their main objective: to develop innovative crime reduction technologies. With the use of computational methods (social simulation and the development of a predictive machine), situational crime prevention is entering a knowledge production regime that is increasingly integrated into the action and commercialization of predictive analysis software. For its promoters, this is the most important criterion of success. The Promap and Predpol software, two policing innovations that are a direct spin-off of research on boost simulations, are examples of this success of situational crime prevention. Yet the analysis of Tim Hope’s critique enables us to show the extent of the ignorance that has gone with the development of knowledge on the prediction of repeat victimization and the associated prediction algorithms. It also enables us to see the political dimension of the statistical modelling of victimization. Whereas Hope’s hypotheses cannot be dissociated from a conception of public policy on security that is based on solidarity, those of Pease and his colleagues rest on a minimal and short-term conception. By refusing to consider Tim Hope’s results, the advocates of situational crime prevention are limiting the expansion of their disciplinary field and simultaneously protecting the speciality that they have only just created.

56The attitude that seems most urgent today is to ensure that what the scientists do not know is also known by those who have to implement prevention policies. In a context where the latter are increasingly dependent on prediction algorithms built in research laboratories, criminological controversies are becoming public issues and should not be the preserve of a scientific elite. How can public security policies be defined, not by knowledge produced by experts, but on the basis of what researchers do not know? That, I think, would be a reasonable way of attaining the scientification (Erikson, Shearing, 1985) of prevention and security policies that seem to have been at work over the past three decades.