David Buil-Gil and Angelo Moretti
Hot-spot policing strategies involve more focused police attention on small geographical areas where crime (and disorder) is more prevalent. During the last two decades, multiple evaluations have shown that police patrols targeting micro-locations with relatively high crime rates within a territory are effective in reducing offending, mediating what has been called a “micro-deterrence” process (e.g. Braga and Bond, 2008; Braga et al, 1999; Braga et al, 2012; Braga et al, 2014; Sherman and Weisburd, 1995; Sherman et al, 2014; Telep et al, 2014). The efficacy of hot-spot policing for targeting such areas depends, therefore, on available data geocoded at low-geographical level. Mapping crime is the first step for a smart policing intervention. Police-detected crimes are easily mapped at low-geographical level, as are calls for police service. However, an appropriate hot-spot policing approach cannot depend solely on known or officially-detected crimes, but also on non-detected offences, citizens’ perceptions of disorders, fear of crime and attitudes towards the police, among others. The latter variables are usually based on surveys with small samples that cannot be easily mapped at low-area level. The following sections are going to detail a new generation of statistical techniques that are helpful for obtaining small area estimates of survey variables; estimates that can play a vital role on a multidimensional-based hot-spot policing.
Some of the first approaches to hot-spot policing focused mainly – and sometimes only – on detected crimes as the unique data source for targeting small area patrols and evaluating those interventions (e.g. Sherman and Weisburd, 1995). However, more recent studies have established the need for including other variables in the databases and maps used by police officers to decide where to target major efforts and how to assess their impacts (e.g. Braga et al, 2014; Ratcliffe et al, 2015; Weisburd et al, 2011). Hot-spot policing interventions, whilst having crime reduction as their main objective, also tackle citizens’ perceptions of deviance and aim to decrease feelings of unsafety. Trust and police legitimacy also need to be analysed when assessing hot-spot patrolling (Weisburd and Telep, 2014). Officially-detected crimes collect three to five times fewer offences than victim surveys (Coleman and Moynihan, 1996) and might be influenced by criminal policy and police propensity to record offences (Aebi, 2010). It is therefore necessary to assess survey-detected victimisation as well as other variables related to subjective perceptions of disorder and trust in the police. As noted, the main source of data for obtaining information about these necessary variables beyond detected-crimes is victimisation surveys.
Victimisation surveys are sample surveys designed to obtain data on personal experiences with crime and deviance. The Crime Survey for England and Wales (CSEW), formerly the British Crime Survey, is the main source of data for analysing these variables in the territories included in its sample. However, the CSEW sample is mainly designed for being representative of the entire English and Welsh territory. With a few exceptions, the survey yields a minimum sample of 650 households for each of the 42 Police Force Areas (PFA), but the sample is not representative of small geographical areas within each PFA. In other words, the small number of households sampled in each of the 181,408 output areas included in the survey – zero in numerous cases – makes it impossible to support reliable direct estimates that can be used for police-targeting purposes. Output areas are small geographical areas designed by the Office for National Statistics to have similar population sizes (125 households on average). Whilst some police forces and local authorities run their own surveys, they are very few and their samples are not large enough to run direct estimates of adequate precision either. We must therefore rely on indirect estimates. In order to obtain indirect reliable estimates of the variables of interest at low-geographical level, small area estimation techniques are an option with great potential for the future (e.g. Taylor, 2013; van den Brakel et al, 2013; Whitworth, 2012).
Direct estimation methods use only area-specific sample data and sampling weights to produce small area estimates. This means that when the area sample size is small these estimates may suffer from low precision and are therefore unreliable for patrol targeting purposes. As noted, crime surveys usually sample few (or even zero) people from each micro-location/area of interest (e.g. output areas). Small area estimation aims to develop statistical techniques which produce efficient and precise estimates at small area level (also when the sample is not large enough to support direct estimates). In order to calculate estimates of adequate precision for small areas, indirect estimation methods make use of auxiliary information from other sources of data (e.g. Census) and produce linking models that borrow strength from related areas or time periods. Thus, this type of statistical inference is ‘model-based’.
According to Rao and Molina (2015), small area models can be classified into two types: area-level and unit-level models:
- Unit-level models, which relate individual or unit values of a study to unit-specific covariates, can only be used when data is available at individual or unit-level. Due to confidentiality criteria, data from victimisation surveys is rarely released at individual level, thus making unit-level small area estimation impossible.
- Area-level models relate small area direct estimates to area-specific auxiliary information available from other sources of data. Auxiliary data from the Census and other administrative records can be used in area-level small area estimation to produce reliable estimates at small area level of crime surveys’ variables.
Weisburd and Neyroud (2013) argue that a wider involvement of science is necessary in the arena of evidence-based policing. Officially recorded crime data mapped at small area level appears insufficient for targeting hot-spot patrols, as they only collect a small proportion of crime and can be biased. Victimisation surveys are today a useful tool for obtaining further information on crime and deviance, but their small samples cannot be used to produce reliable direct small area estimates. Model-based small area estimation techniques are today a potential option for obtaining further reliable information at low-area level for a more complex, efficient and evidence-based hot-spot policing.
David Buil Gil is a PhD student at University of Manchester researching small area estimation techniques for victimisation surveys data. Email: firstname.lastname@example.org. Twitter: @DavidBuil
Angelo Moretti is a PhD student at University of Manchester researching multivariate small area estimation methods for poverty and wellbeing indicators. His research interests cover also multivariate statistical techniques for data dimensionality reduction. Email: email@example.com