Abstract

Purpose: To assess the quality of police recorded crime and survey estimates of crime in a unified framework that does not treat either as a ‘gold standard’. Understanding the quality of crime estimates has implications for all empirical studies assessing the geographic distribution of crime. Methods: Police records and survey data are combined using a Multi-Trait Multi Trait Model treating both sources as potentially unreliable. Data are from England and Wales between 2011 and 2019 distinguishing between multiple crime types (burglary, damage, vehicle, and violence offences) and different area levels (Police Force Areas and Community Safety Partnerships). Results: We find an important trade-off between systematic bias in means – with police records capturing on average fewer crimes – and consistent variance estimates – with surveys showing higher levels of random error. Reliability ratios are higher for vehicle crime and burglary than for violence crimes and criminal damage, and there is evidence that reliabilities are getting worse over time. Conclusions: Our results suggest that analyses using police recorded crime data may be less at risk from measurement error than previously thought, with potential biases resulting from systematic undercounting substantial easier to adjust for than random error.

Keywords: criminal, statistics, reliability, validity, regions, UK

Introduction

Crime measurements commonly stem from one of two data sources: police records or crime victimisation surveys. Police records have the benefit of being highly granular (spatially, temporally, and in terms of offence types) but are generally accepted to be flawed since they fail to capture crimes that are not reported, and because of the different recording practices followed across police officers and forces. Victimisation surveys provide less widespread coverage of the range of crimes and their precise locations, but their focus on victims’ experiences and the use of representative samples allow them to overcome key limitations of police recorded crimes, including recording information about crimes unknown to the police (Aebi and Linde, 2014; Lohr, 2019). As a result, survey data is often taken as the most valid approach to estimate national levels of crime at one point in time or across medium and long terms, and are also often used to assess the extent of measurement error in police data (Gibson and Kim, 2008; Tarling and Morris, 2010). However, because victimisation surveys are generally carried out annually and their sample design is optimised to produce country or regional level estimates, most research into the causes and consequences of crime is still explored using police data. This includes those studies focusing on comparatively short time-intervals (e.g., the immediate effect of a lockdown), more granular spatial resolutions (e.g., comparison of crime prevention initiatives at the city or neighbourhood level) and offence types (e.g., knife crime or homicides) (Ariel and Bland, 2019; Bland and Ariel, 2020).

The reliance on police data means that a good deal of crime research is suspected to be heavily affected by measurement error, and therefore bias. However, recent studies have shown that this does not need to lead to substantial bias. If the form and prevalence of the measurement error mechanisms affecting police data can be estimated, we can accurately anticipate the direction and extent of the subsequent bias in multivariate analyses relying on this data (Pina-Sánchez et al., 2022; 2023). Yet, as much as it is widely acknowledged that police data is flawed, we still do not have accurate estimates of its measurement properties. This represents an important and surprising gap in the crime data literature.

The most common approach used to explore the presence of measurement error in police data involves making direct comparisons with estimates from victimisation surveys (e.g., Gibson and Kim, 2008; Pina-Sánchez et al., 2022; Tarling and Morris, 2010). Typically, this involves treating victimisation surveys as a ‘gold standard’, which implies that any discrepancies between crime estimates derived from survey or police sources are evidence of measurement error in the latter. Such an approach can help us assess the extent to which certain crimes are under-reported (i.e., systematic errors), but it is bound to overestimate the amount of noise (i.e., random errors) in police data, and in so doing, underestimate its reliability. This is because although one could take crime estimates derived from surveys to be unbiased, they are still estimates, and, as such, uncertain. Importantly, the amount of uncertainty in survey-based crime estimates is likely to vary across areas and over time (Rosenbaum and Lavrakas, 1995). Alternative approaches have relied on manual reviews of documents and recordings of incidents (Her Majesty Inspectorate of Constabulary, 2014; Klinger and Bridges, 1997). Reliability estimates based on this methodology are as internally valid as they can possibly be. However, this is an expensive approach, and therefore estimates are limited to relatively small samples of police forces, crime types and time points, which affects their external validity.

In this study we take a different approach, explicitly accounting for the uncertainties in both police recorded crime data and survey data using a Multi-Trait Multi-Method (MTMM) model. As a latent variable estimation method, MTMM models allow us to take crime rates derived from each data source as imperfect measures of the true extent of crime, while, crucially, estimating how well each of the measures captures the true (but unobserved) crime rate. That is, MTMM directly provides us with estimates of the reliability ratio of both police recorded and survey-based estimated crime rates. This allows us to assess their quality profile more accurately, eschewing the need to give one of the data sources primacy as a ‘gold standard’. Specifically, we use police data from the Home Office and survey data from the Crime Survey for England and Wales (CSEW) to estimate annual crime rates across Police Forces (N=43) and Community Safety Partnership (CSP) areas (N=312) from 2011 to 2019.

The paper proceeds as follows. We first discuss the collection of official crime records as well as the use of surveys as a way to capture the same information. We then present the data used and formally introduce the MTMM model. Finally, we present the results and discuss their implications for crime research and crime prevention practices.

The measurement of crime

Measuring the extent of crime has a long history, with the first maps of recorded crime across France being produced as early as 1829 as part of the collection of so-called ‘moral statistics’. Police records of crime have since become one of the longest running sources of data about the population, with data collated since the late 1800s charting a dramatic rise in criminal activity from the middle of the 20^th Century and an equally dramatic drop since the mid-1990s (Berg and Lauritsen, 2016; Farrell et al., 2014).

Yet, in spite of the long running nature of recorded crime data and the consistent approach to counting crimes, researchers have criticized the lack of reliability of official crime statistics for almost as long as the data has been collected (de Candolle, 1830[1987]; Sellin, 1931). Critics have argued that police records are affected by several external factors which could themselves vary in space and time: for instance, victims may be unaware of crime or choose not to report it, and the police may fail to identify or arrest the offender or choose not to record incidents for a variety of reasons. The very definition of criminal activities is also historically contingent and subject to change, whilst the policing priorities of the day and size of the police force can present the illusion of a spike in crime whilst the underlying rate of crime remains constant. And despite the existence of a comprehensive set of counting rules and protocols for the police to follow (Home Office, 2011, 2013), once brought to their attention the decision to record an incident as a crime is down to the personal discretion of the officer(s) involved (Boivin and Cordeau, 2011; Burrows et al., 2000) as well as the prevailing culture within the specific force (Warner, 1997). As such, police recoded crime data must be viewed as a record of the extent of crime which is necessarily affected by policing activities and their interactions with the public.

In part reflecting these problems with police recorded crime data, the 1960s saw the advent of a new approach to measuring crime using survey data. Pioneered in the US as part of the President’s Commission on Law Enforcement and the Administration of Justice (Ennis, 1967), sample surveys measure crime by asking victims directly to report on their experiences. By adopting a probability-based sampling strategy coupled with a consistent measurement tool, they are unaffected by the sorts of variations in reporting and recording practice that have been so prevalent in recorded crime data. Crime surveys are now a mainstay of counts of crime in a range of countries throughout the world.

Whilst police-recorded crime statistics and victim surveys remain the main sources of crime data, we are witnessing an increasing number of data sources about offending and victimisation. These include calls for police services, probation statistics, incidents recorded by health emergency services, social media data, gunshot detection technology and self-report crime surveys (Hibdon et al., 2021; Koziarski et al., 2022; Piza et al., 2023; Sutherland et al., 2021; Williams et al., 2017). However, these continue to be deployed rather sporadically and there is no centralized system for collating this information at a national scale.

Researchers and practitioners have subsequently compared estimates from crime surveys against recorded police crime data, confirming the presence of a large number of offences that go unreported and unrecorded each year. Known colloquially as the ‘dark figure of crime’ (Biderman & Reiss, 1967; Skogan, 1977), studies have since demonstrated that a victims’ willingness to report an incident to the police is contingent on characteristics of the victim and offence (Hart & Rennison, 2003; Tarling & Morris, 2010), with reporting rates differing systematically according to the victims’ sex, age, ethnicity and income, as well as their relationship to the offender (Baumer, 2002; Hart & Rennison, 2003; Xie and Baumer, 2019). Studies have also identified differences in reporting rates by crime types, with theft of motor vehicle and burglary typically being those with the highest reporting rates, and petty crimes such as theft and shoplifting being less likely to be reported to the police (Hart & Rennison, 2003; Tarling & Morris, 2010).

However, these comparisons are generally simplistic, with victim survey data frequently treated as an error-free ‘gold standard’ that police records are compared against. This is despite the fact that survey data are themselves subject to a range of widely known measurement errors, with estimates of victimisation likely affected by interviewer effects, non-response bias, memory failures, question wording, and social desirability bias (Brunton-Smith and Allen, 2009; Schneider and Sumi, 1981). Crime surveys also tend to have limited sample sizes at the level of small geographic areas (Rosenbaum and Lavrakas, 1995) and are subject to sampling errors, meaning they may be less suited to identifying differences in the levels of crime between areas. Studies have also tended to assume that ‘more is better’, with the higher crime counts typically observed in victim surveys taken as evidence of systematic bias in police recorded crime data (Gibson and Kim, 2008; Pina-Sánchez et al., 2022). And when discussing discontinuities between data sources over time, studies have (implicitly) assumed a consistent survey design in addition to the lack of confounding between changes in reporting practices and changes in levels of crime over time (Office for National Statistics, 2023).

More robust approaches to comparison that combine multiple data sources to quantify measurement error are now becoming available to researchers. In particular, MTMM models have enabled researchers to go beyond the assumption of a ‘gold standard’ and consider measurement error in multiple data sources concurrently. For example, Oberski et al. (Oberski et al. 2017) showed how the MTMM approach can be used to estimate measurement error in administrative records and survey data in the context of employment statistics. This approach was recently used by Cernat et al. (2022) in the context of crime data, although their focus was on comparisons between different survey-based estimates. They found that the distribution of survey-based offense location estimates, as opposed to victim residence estimates, is highly similar to police-recorded crime statistics, and there is little trade off in terms of the reliability and validity of offense location and victim residence measures.

Data

In this study we combine police recorded crime data with the CSEW for the time period 2011-2019, comparing the number of crimes estimated by each data source separately for each Police Force Area (PFA) and Community Safety Partnership (CSP). There are a total of 43 PFAs in England and Wales, differing substantially in size and internal composition (with the smallest, the City of London, covering just 2.6 km² and the largest, Dyfed-Powys, spanning nearly 11,000 km²). Each PFA contains an average of 7 CSPs, with a total of 312 in England and Wales. Both geographies structure police activity in some meaningful way. PFA represent the main structure for policing, with each PFA operating as a distinct policing unit with responsibility for its own budget and for setting its own policing priorities. Each PFA is also responsible for collecting and cleaning their own recorded crime data, though all police forces are required to follow common counting rules established by the Home Office (2011, 2013). CSP, by contrast, are a police administrative boundary that has primarily been established to facilitate intelligence sharing across a range of local services that feed into policing activity within each PFA, as well as contributing to local anti-social behaviour strategies.

Data from police recorded crimes and CSEW were harmonised prior to analysis, with four distinct crime categories identified that could be consistently measured by both data sources – violent crime, burglary, vehicle crime (including theft and damage to vehicles) and criminal damage (to households). Counts of crime within each category were then converted into rates per 1,000 residents to ensure comparability across areas. Data from the City of London PFA were omitted from the analysis because of the very low resident population and outlier crime distribution.

Police recorded crime data

Crime counts were derived directly from police recorded crime data for each PFA/CSP from Home Office published data tables (https://www.gov.uk/government/statistics/police-recorded-crime-open-data-tables). Annual counts of crime in each crime category were first combined into the four main crime types (see Supplementary Materials for the precise offence mapping), with the annual count of crime for each crime type and area converted into an equivalent crime rate using PFA/CSP population and household totals from the 2011 Census (including the appropriate upward adjustment for years after 2011).

Crime Survey data

We produce crime counts from CSEW data (Office for National Statistics, 2021a) following the same approach used to produce national statistics (Office for National Statistics, 2021b). Respondents to the survey are first asked whether they had experienced a range of different types of incidents in the last twelve months, with these initial screening questions used as the basis for a series of more detailed follow up survey questions. From these ‘victim forms’ we identify the same four crime types and count the total number of incidents experienced by each individual, as well as the number of these incidents subsequently reported to the police. These crime totals are then capped at a maximum of five to minimise year-to-year and area-to-area fluctuations1 and used to estimate the weighted incidence rate per area (annually by PFA and biannually by CSP). This is calculated as the (weighted) mean number of incidents experienced by each sampled individual in each area of residence, with the incidence rate then re-scaled to rates per 1,000 people/households using the 2011 Census counts. Incidence rates are calculated annually for each PFA and biannually for each CSP to increase the area sample sizes at the smaller spatial scale.

Method

To assess the consistency of crime survey and police recorded crime data we adopt a MTMM approach. This combines the three separate sources (‘Methods’) of crime estimates – police recorded crimes, crime survey estimates of all experienced crime, and crime survey estimates of all reported crimes – and four separate crime types (‘Traits’) – violent crimes, household burglaries, vehicle crimes and household criminal damage incidents. The MTMM approach treats each crime estimate from the same source of crime data as an indicator of a latent ‘Method’ variable (police recorded, CSEW experienced, CSEW reported), whilst simultaneously treating all estimates of the same crime type from across the data sources – an indicator of a latent ‘Trait’ variable (violence, burglary, vehicle, damage) (Figure 1). As a result, we have three latent method variables (same method/different crime type) and four latent trait variables (same crime type/different method), with each latent variable itself measured by three indicators. This allows us to identify the unique variance associated with each method, and separate the unique variance associated with each crime type. Any residual variation (i.e., the proportion of variation that is not explained by the latent method or latent trait) is then treated as random measurement error.

The MTMM model can be formally written as:

Where y is the observed variable measured for each method m and trait t. The observed variables are explained by the trait latent variables ($T_{t}$), method latent variables ($M_{m}$) and indicator specific residual ($\epsilon_{mt}$). The traits capture the variance of interest that is consistently measured across the different methods. We refer to this as the corrected crime estimate or the trait. The method latent variable captures the variance that is specific to each measurement approach. This is the variance that is common across traits within each method. This is the complement of validity as it does not capture the concept of interest but is consistent across topics. Finally, the item specific residual represents the random error of the measures and is the complement of reliability. In the context of this model, we define validity as the proportion of the total observed variation that is due to the latent trait variable ($T_{t}$). We define lack of validity, or method effects, as the proportion of the total observation that is due to the method latent variable ($M_{m}$). This is defined as lack of validity as it is a consistent source of variation that is present over multiple measures that do not capture the substantive topic of interest. Finally, we define lack of reliability as the proportion of variation that is due to the residual ($\epsilon_{mt}$). We can estimate these statistics for each question indicator ($y_{mt}$) and also aggregate them by trait (t), method (m) or both.

Loadings are fixed to one for identification purposes (Cernat and Oberski, 2019; 2021). This is used as a way to define the latent variables as the common variation for all items measuring either the same trait (t) or method (m). This does not imply that the validity or reliability are the same across the indicators as these depend on the standardized coefficients which vary due to different amount of item variation. We also include correlated errors between observed variables measuring the same topic and coming from the CSEW to account for the survey specific shared variance. We estimate the model separately for PFA and CSP at each time point.

Final models use the logged crime rate to address the skewed distribution of crimes across areas and adjust for missingness (areas with a CSEW estimated crime rate of 0 due to sparse data) using a full information maximum likelihood approach that assumes the data are Missing At Random given our measurement model. Models are estimated using Mplus Version 8.

Figure 1. Visual representation of the MTMM model estimated. Squares represent observed variables and circles represent latent variables. Loadings are fixed to 1 for identification and correlated errors are included between observed variables measuring the same topic from the survey to account for the shared variance. Residual variances are not presented to facilitate reading.

Results

Figure 2 plots annual crime rates across all PFAs in England and Wales for each crime category using the three measurement approaches. A similar trend is evident across all four crime types, with the survey data tending to identify a higher number of crimes than recorded crime data (a fact confirmed when considering the intercepts from our MTMM). There is also a general reduction in experienced crimes (the dotted line) and (to a lesser extent) reported crimes (the dashed line) over time. By contrast, there has been a general increase in police recording (the black line) that results in the gap between the three estimated crime lines generally closing. This is most notable for violent crimes. So much so that recording rates actually mirror (or even exceed) reported rates at the latter part of the data.

Figure 2. Crime trends using police recorded crime and CSEW data (2011/12-2018/19) PFA data. Annual crime counts are calculated by summing the total number of crimes per PFA (n=43)

Examination of the bivariate correlations for the (logged) crime counts present a consistent picture, with strong positive correlations and modest evidence of stronger correlations between the estimates from the same methods than between the survey and police (Supplementary materials, figure S1). There is, however, some indication that the correlations have got weaker over time. A similar picture is evident, albeit with more modest correlations, when focus is moved to biannual estimates at CSP level (Supplementary materials, figure S2).

The full results from our MTMM models are reported in tables S1-S2 (Supplementary materials), where the generally higher intercepts from the two survey estimated crime rates confirm the systematic undercount commonly associated with police recorded crime data. We also see the modest convergence between the datasets over time, with the logged (conditional) recording rates of vehicle and violent crime becoming higher than the equivalent CSEW estimates. Turning our attention to the reliability ratios (Figures 3-5) a complex picture emerges. Looking first at overall crime rates (Figure 3), it is evident that crime can be measured with a generally high degree of accuracy, with at least 75% of the crime variation allocated to ‘true’ trait variance across all years. There is, however, evidence of a modest reduction in data quality over time, with both the random and method-specific variance increasing. The picture is similar when looking across PFA (top panel) and CSP level (bottom panel), albeit with somewhat reduced data quality over all when looking at the more spatially granular CSP estimates. This is not altogether surprising, with the CSP data presenting a somewhat noisier picture of the true crime landscape and method-specific variation also more evident.

Police Force Area
Community Safety Partnership

Figure 3. Measurement error over time (all crimes) across Police Force Areas (n=42) and Community Safety Partnerships (n=312)

A similar picture emerges when the four different crime types are considered separately (Figure 4), with evidence that data quality has gradually worsened between 2011 and 2019 for all crime types and a reduced quality profile when considering crimes measured across CSPs (bottom panel). In addition, we note substantially more random error evident when considering criminal damage and violent crime (~30%) when compared to burglary and vehicle crimes (~13%). This may reflect the fact that violence and criminal damage offence categories are somewhat broader in their scope, and suggests that there is considerably more evidence of individual differences in meaning and interpretation when considering these offences. As a result, they are generally measured with less quality. By contrast, method-specific errors remain generally similar across all crime types.

Police Force Area
Community Safety Partnership

Figure 4. Measurement error across crime types at Police Force Area level (n=42) and Community Safety Partnerships (n=312)

The results so far present a somewhat reassuring picture, suggesting that crime can be measured with a reasonable degree of accuracy, at both the PFA and CSP level, particularly for those crime categories like burglary and vehicle crime that are subject to the least amounts of discretion and conceptual ambiguity. However, of most interest to us are the relative differences in quality between police recorded crime data and survey estimates of crime (Figure 5). Here, a more complex picture emerges. Whilst it is clear that the same pattern is evident overall – more error amongst violent crime and criminal damage, and a generally worsening picture over time and at the smaller spatial scale – this is not uniform across the data sources. In particular, there is evidence of considerably less random error when considering police recorded crime (4.5%) and no clear indication that the magnitude of this error has increased over time. Interestingly, whilst police recorded crime data does not appear to be subject to much random error, we do observe a comparatively large contribution of method effects (around 6%), meaning that the reliability estimates for police recorded crime data are around 90%. Method effects are particularly noticeable for police recorded crime when the focus shifts to CSP.

By contrast, both survey estimates have substantial random errors (24% and 35% for experienced and reported incidents at PFA level), with reported incidents of violence and criminal damage particularly prone to random error across all years. As a result, around half of the variation in these PFA estimates is attributable to random error. The picture for CSEW is similar when considering CSP estimates, although here we do not see such a clear difference between experienced and reported crimes. Unlike police recorded crime data, method effects are far less prominent when survey estimates are considered.

Police Force Area
Community Safety Partnership

Figure 5. Comparing the magnitude of measurement error between crime survey and police recorded crime data across Police Force Areas (n=42) and Community Safety Partnerships (n=312)

Conclusion

In this paper we show how MTMM models can be used to investigate data quality in crime measures. Here we treated three data sources – police recorded crimes, experienced CSEW crimes and reported CSEW crimes – as different ‘methods’ that measure the same concepts of interest (‘traits’): burglary, criminal damage to a dwelling, vehicle crime and violence. Our results identify a complex pattern in data quality. At its core there appears to be a trade-off between the systematic bias in the means – with police records capturing on average fewer crimes – and consistent variance estimates – with surveys showing higher levels of random error. Additionally, while the police records are more reliable, they show higher levels of method effects. This implies that while recording practices are relatively consistent, they are systematically different from the survey estimates.

Overall, then, when the research goal is not to retrieve true estimates of crime, but rather explore crime variability across areas, then police data is not as bad as commonly thought. In fact, we find that the reliability of police data is high, with only around 10% of the variation between areas identified as error (including both random and method-specific variation). There is also no evidence that this reliability ratio has changed over time. This means that this data is well equipped to explore variability of crime across time. On the other hand, CSEW data exhibits considerably more random errors and a seemingly worsening picture over time, even when this is used at the PFA level.

This finding should not be understated. Recent work from Pina-Sánchez et al. (2022) demonstrated that it is straightforward to adjust for the potential biases resulting from the systematic undercounting inherent in crime data, but pointed to a more uncertain way forward when random errors are considered and a need to incorporate more complex sensitivity analyses in this context. The fact that we find these random errors are modest in nature suggests they can be reasonably ignored in most analyses unless there are reasons to expect the levels of undercounting are associated with other focal variables of interest.

Another important take away is that, whilst the measurement quality of these four types of crimes is overall good, it does vary significantly by topic and aggregation level. The average amount of variance that is attributed to the trait over all the questions, years and methods is 76%. This implies that around a quarter of the observed variance in crime measures does not capture the concept of interest. The measurement quality is better when burglary and vehicle crimes are considered (83% and 87% variance is trait, respectively) than when measuring damage and violence (63% and 71%, respectively). Also, using data aggregated at CSP level shows worse quality (62% trait variance) than when using PFA, confirming that crime data quality worsens at smaller spatial levels (Buil-Gil et al., 2022). Trait variance is likely to be even larger in crime records aggregated at the highly detailed levels of analysis typically considered in place-based criminology and policing, such as neighbourhoods, micro places and street segments (Braga and Weisburd, 2010). Further work is thus needed to study the measurement qualities of estimates of crime obtained from police and survey data for small geographic areas.

Like all studies, this work is not without its limitations. In particular, whilst the MTMM model has enabled us to relax the assumption that survey data represent an error-free ‘gold standard’, this is replaced by other identifying assumptions. One of the most important ones being that the different ‘methods’ capture the same phenomenon. Nevertheless, differences observed across crime data sources will not be solely due to measurement error. How crime is conceptualised is also central, with different sources of crime data measuring distinct, but related, phenomena. For instance, police statistics record crimes that happen in an area, while estimates of crime obtained from surveys show crimes in places where victims live, thus systematically underestimating crime in places with a low residential population and a large ‘ambient’ population (Cernat et al., 2022). The crimes captured by the two sources may also be different, with many of the incidents identified in a victim survey unlikely to be reported to the police (perhaps because they are not deemed serious enough, or because of public distrust or fear of reprisals) and many of those incidents reported to the police not ultimately recorded (if they are less serious or details are ambiguous).

Keeping these limitations in mind, this paper has shown how the MTMM model can be used to better understand the measurement quality of crime data without having to afford one source of data the status of a ‘gold standard’ for comparison. Using this approach, we have demonstrated that whilst recorded crime data may systematically undercount the true extent of crime, it exhibits a high level of data quality and outperforms survey data when considering variations across PFA and CSP areas. This is important, with existing work demonstrating that systematic errors are considerably more straightforward to anticipate and correct without the need for recourse to complex measurement error approaches. Consequently, whilst survey data remains the most important resource to provide evidence of the dark figure of crime and the nature of crime underreporting, recorded crime data can and should be given primacy in empirical studies that aim to examine the causes and consequences of crime across and within geographic areas.

Acknowledgements

Funding: This work was supported by the Economic and Social Research Council [grant number: ES/T015667/1].

References

Aebi, M.F., & Linde, A. (2014). National Victimization Surveys. In G. Bruinsma and D. Weisburd (Eds.), Encyclopedia of Criminology and Criminal Justice (pp. 3228-3242). New York: Springer.

Ariel, B., & Bland, M. (2019). Is crime rising or falling? A comparison of police-recorded crime and victimization surveys. In M. Deflem & D. M. D. Silva (Eds)., Methods of criminology and criminal justice research (pp. 7-31). Emerald Publishing Limited.

Baumer, E. P. (2002). Neighborhood disadvantage and police notification by victims of violence. Criminology, 40(3), 579-616.

Berg, M. T., & Lauritsen, J. L. (2016). Telling a Similar Story Twice? NCVS/UCR Convergence in Serious Violent Crime Rates in Rural, Suburban, and Urban Places (1973–2010). Journal of Quantitative Criminology, 32, 61-87.

Biderman, A.D., and Reiss, A.J. (1967). On Exploring the “Dark Figure” of Crime. The ANNALS of the American Academy of Political and Social Science, 374(1), 1-15.

Bland, M. P., & Ariel, B. (2020). Targeting Domestic Abuse with Police Data. Springer.

Boivin, R., & Cordeau, G. (2011). Measuring the impact of police discretion on official crime statistics: A research note. Police Quarterly, 14(2), 186-203.

Braga, A. A., & Weisburd, D. L. (2010). Editors’ introduction: Empirical evidence on the relevance of place in criminology. Journal of Quantitative Criminology, 26, 1-6.

Brunton-Smith, I., and Allen, J. (2009). Measuring crime, in M. Bulmer, J. Gibbs and L. Hyman (eds) Social Measurement through Social Surveys: an Applied Approach. Ashgate.

Buil-Gil, D., Moretti, A. & Langton, S.H. (2022). The accuracy of crime statistics: assessing the impact of police data bias on geographic crime analysis. Journal of Experimental Criminology, 18, 515-541.

Cernat, A., & Oberski, D. L. (2019). Extending the within‐persons experimental design: The multitrait‐multierror MTME approach. In P. Lavrakas, M. Traugott, C. Kennedy, A. Holbrook, E. de Leeuw, & B. West (Eds.), Experimental methods in survey research (1st ed., pp. 481–500). Wiley. https://doi.org/10.1002/9781119083771.ch24

Cernat, A., Buil-Gil, D., Brunton-Smith, I., Pina-Sánchez, J., & Murrià-Sangenís, M. (2022). Estimating Crime in Place: Moving Beyond Residence Location. Crime & Delinquency, 68(11), 2061–2091.

Cernat, A., & Oberski, D. L. (2021). Estimating stochastic survey response errors using the multitrait-multierror model. Journal of the Royal Statistical Society Series A, 185, 134–155.

de Candolle, A. (1830[1987]). Considérations sur la statistique des délits. Déviance et Société, 11(4), 352–355.

Ennis, P. H. (1967). Criminal Victimization in the United States. A Report of a National Survey. Washington DC: US Government Printing Office.

Farrell, G., Tilley, N., & Tseloni, A. (2014). Why the crime drop? Crime and Justice, 22, 421-490.

Gibson, J., & Kim, B. (2008). The effect of reporting errors on the cross-country relationship between inequality and crime. Journal of Development Economics, 87(2), 247–254.

Hart, T. C., & Rennison, C. (2003). Reporting crime to the police, 1992-2000. Special Report. Bureau of Justice Statistics.

Her Majesty Inspectorate of Constabulary (2014). Crime-recording: making the victim count—The final report of an inspection of crime data integrity in police forces in England and Wales. Justice Inspectorates. https://www.justiceinspectorates.gov.uk/hmicfrs/wp-content/uploads/crime-recording-makingthe-victim-count.pdf

Hibdon, J., Telep, C. W., & Huff, J. (2021). Going beyond the blue: The utility of emergency medical services data in understanding violent crime. Criminal Justice Review, 46(2), 190–211.

Home Office (2011). National standard for incident recording counting rules. Retrieved from: https://www.gov.uk/government/publications/the-national-standard-for-incident-recording-nsir-counting-rules

Home Office (2013). Counting rules for recorded crime. Retrieved from: https://www.gov.uk/government/publications/counting-rules-for-recorded-crime#full-publication-update-history

Klinger, D. A., & Bridges, G. S. (1997). Measurement error in calls-for-service as an indicator of crime. Criminology, 35(4), 705-726.

Koziarski, J., Ferguson, L., & Huey, L. (2022). Shedding light on the dark figure of police mental health calls for service. Policing: A Journal of Policy and Practice, 16(4), 696-706.

Lohr, S. H. (2019). Measuring crime: Behind the statistics. Boca Raton: CRC Press.

Oberski, D. L., Kirchner, A., Eckman, S., & Kreuter, F. (2017). Evaluating the Quality of Survey and Administrative Data with Generalized Multitrait-Multimethod Models. Journal of the American Statistical Association, 112(520), 1477–1489. https://doi.org/10.1080/01621459.2017.1302338

Office for National Statistics (2021a). Crime Survey for England and Wales, 2011-2020: Secure Access, Low-Level Geographic Data. [data collection]. 7th Edition. UK Data Service. SN: 7311, http://doi.org/10.5255/UKDA-SN-7311-8

Office for National Statistics (2021b). User guide to crime statistics for England and Wales: March 2020. Retrieved from: https://www.ons.gov.uk/peoplepopulationandcommunity/crimeandjustice/methodologies/userguidetocrimestatisticsforenglandandwales#comparison-of-the-csew-and-police-recorded-crime

Office for National Statistics (2023). Crime in England and Wales: year ending December 2022. Retrieved from: https://www.ons.gov.uk/peoplepopulationandcommunity/crimeandjustice/bulletins/crimeinenglandandwales/yearendingdecember2022

Pina-Sánchez, J., Buil-Gil, D., Brunton-Smith, I., & Cernat, A. (2022). The Impact of Measurement Error in Regression Models Using Police Recorded Crime Rates. Journal of Quantitative Criminology, 1-28.

Pina-Sánchez, J., Brunton-Smith, I., Buil-Gil, D., & Cernat, A. (in press). Exploring the Impact of Measurement Error in Police Recorded Crime Rates through Sensitivity Analysis. Crime Science.

Piza, E. L., Hatten, D. N., Carter, J. G., Baughman, J. H., & Mohler, G. O. (2023). Gunshot Detection Technology Time Savings and Spatial Precision: An Exploratory Analysis in Kansas City. Policing, 17, paac097.

Rosenbaum, D. P., & Lavrakas, P. J. (1995). Self-reports about place: The application of survey and interview methods to the study of small areas. In J. E. Eck & D. Weisburd (Eds.), Crime and place. Crime prevention studies (pp. 285–314). Criminal Justice Press.

Schneider, A. L., & Sumi, D. (1981). Patterns of forgetting and telescoping: An analysis of LEAA survey victimization data. Criminology, 19(3), 400-410.

Sellin, T. (1931). The basis of a crime index. Journal of Criminal Law & Criminology, 22(3), 335–356.

Skogan, W. G. (1977). Dimensions of the dark figure of unreported crime. Crime and Delinquency, 23(1), 41–50.

Sutherland, A., Strang, L., Stepanek, M., Giacomantonio, A., Boyle, A., & Strang, H. (2021). Tracking violent crime with ambulance data: How much crime goes uncounted? Cambridge Journal of Evidence-Based Policing, 5(1), 20–39.

Tarling R, Morris K (2010) Reporting crime to the police. The British Journal of Criminology, 50(3), 474–490.

Walby, S., Towers, J., and Francis, B. (2016). Is violent crime increasing or decreasing? A new methodology to measure repeat attacks making visible the significance of gender and domestic relations. British Journal of Criminology, 56(6), 1203-1234.

Warner, B. D. (1997). Community characteristics and the recording of crime: Police recording of citizens’ complaints of burglary and assault. Justice Quarterly, 14(4), 631-650.

Williams, M., Burnap, P., & Sloan, L. (2017). Crime sensing with big data: The affordances and limitations of using open-source communications to estimate crime patterns. The British Journal of Criminology, 57(2), 320–340.

Supplementary materials

Police Recorded Crime offence codes

Figure S1. Annual correlation matrices of crime types and methods at Police Force Area level- 2011/12 - 2018/19 (n = 42)

2011/12	2012/13	2013/14
2014/15	2015/16	2016/17
2017/18	2018/19

Figure S2. Biannual correlation matrices of crime types (and methods) at Community Safety Partnership level - 2011/13 - 2017/19 (n=312)

2011/13	2013/15	2015/17
2017/19

Table S1. Full annual MTMM results (unstandardized) at Police Force Area level

Table S2. Full annual MTMM results (unstandardized) at Community Safety Partnership level