Abstract

Victimization is concentrated among few individuals, commonly referred to as polyvictims. Yet, there exists a lack of consensus regarding the operationalization of polyvictimization. This study investigates the impact of using different measures of polyvictimization on the identification of polyvictims and outcomes of regression models. Operationalizations used in this research are meta-analyzed using the 2019/20 Crime Survey for England and Wales. Results show that while calculating the top 10% of respondents with more victimizations as polyvictimized and applying logistic regression may lead to false positives, the preferred approaches to analyze polyvictimization are either selecting respondents who suffered multiple victimizations and applying bivariate probit regression, or calculating numbers of victimizations and applying negative binomial regression. Mental health conditions are the strongest correlate of polyvictimization.

Introduction

Previous research shows that crime is concentrated among a few victims (O et al., 2017), offenders (Martinez et al., 2017) and places (Lee et al., 2017). The focus of this paper is on the former - individuals who experience crime repeatedly. This phenomenon was first described as “repeat victimization” in the late 1970s (Sparks et al., 1977; Nelson, 1980), but many different terms have been used since then, including “chronic victimization”, “multiple victimization”, “polyvictimization” and “revictimization” (Farrell and Pease, 2014). Throughout this paper, we will use the terms “repeat victimization” and “polyvictimization”, which refer to victims who experience the same crime type twice or more, and those who experience at least two crime types and/or incidents, respectively.

Crime victimization is a relatively rare event. Using data from the Crime Survey for England and Wales (CSEW), Tseloni and Pease (2005) found that while 80% of households did not experience any property crime over a 1-year period, 34% of household victims were repeatedly victimized. The proportion of repeat victims varies across the studies, and it often depends on the number of variables included to measure polyvictimization and how victimization variables are operationalized (e.g., binary or count). In relation to the number of variables included, for example, while some studies only investigate repeat personal (Tseloni and Pease, 2003, 2004) or property victimization (Osborn et al., 1996), others analyze repeat victimization of multiple crime types (Tseloni and Pease, 2005). Online victimization is rarely included in existing studies. Importantly, inconsistencies in how polyvictimization is operationalized, measured, and analyzed can lead to significant variations in research outcomes and the theoretical frameworks built upon them. This article investigates if the measurement of polyvictimization affects the identification of polyvictims and the results of regression models.

We first construct several polyvictimization measurements based on vehicle crime, burglary, personal theft, violence, and digital crime using various sub-methods used in previous studies (Radtke et al., 2024). Second, we check the level of consistency between the computed polyvictimization variables to evaluate if they classify polyvictims similarly or otherwise. Third, we investigate predictors of polyvictimization applying a variety of statistical models. Finally, we conduct a meta-analysis of all applied operationalizations of polyvictimization and statistical regression models. We use the CSEW as our data and focus on studies that used this dataset to investigate repeat victimization.

Measurement of polyvictimization: Review of existing studies

Preventing repeat victimization reduces overall crime rates when tactics are properly developed and applied (Pease, 1998; Grove et al., 2012). In the early 1990s, the UK Home Office was set up to roll out a repeat victimization prevention strategy to police forces, followed by the introduction of repeat victimization as a performance indicator (Laycock, 2001). However, as Pease et al. (2018) note, “[d]espite the substantial research and practice base, the first two decades of the new millennium have seen a progressive decline in attention to repeat victimization in policy formulation, although not in research” (p. 258). Tseloni et al. (2005) argue that the reason why repeat victimization has largely been overlooked in crime prevention efforts is related to the technical and inconsistent terminology often used in academic research.

Several related terms have been used in the study of repeat victimization. Furthermore, Tseloni et al. (2005) discuss some of the methodological aspects that have historically influenced the measurement of repeat victimization, including the analysis of prevalence rates (proportion of victims in the population), incidence rates (number of crimes) or crime concentration (number of incidents over the number of victims, or the ratio of incidence over prevalence). Crime incidence is almost-always greater than prevalence, with some victims experiencing crimes repeatedly and a few victims concentrating very large volumes of crimes.

In order to systematically review how polyvictimization has historically been constructed and analyzed, we have carried out a rapid evidence assessment of the literature exploring polyvictimization from the CSEW. The search was conducted using Web of Science[1] and returned 55 outputs; or 35 after excluding duplicates. Twenty of these studies analyzed the CSEW and 10 of them had a form of polyvictimization as their outcome variable.

The terms prevalence, incidence, and concentration are used to explore crime variation descriptively or to identify proportions of individuals, households or areas that suffer repeat victimization (Trickett et al., 1992; Ellingworth et al., 1995). Previous studies construct polyvictimization variables differently. Studies that analyze the CSEW tend to construct polyvictimization as counts of one type of crime victimization, such as burglary (Osborn and Tseloni, 1998; Tseloni et al., 2004; Hunter et al., 2021) property crime (Tseloni, 2006) or personal crime (Tseloni and Pease, 2015). These studies tend to use negative binomial regression, which is a suitable approach for over-dispersed count variables (Hope and Trickett, 2008). Some note that analyzing crime counts rather than binary variables (i.e., victim/non-victim) predicts the entire distribution of victimization incidents and accounts for crime concentrations (Osborn and Tseloni, 1998; Tseloni and Pease, 2003).

Hope et al. (2001) constructed victimization as binary variables and used bivariate probit regression. Osborn et al. (1996) used both multinomial regression and bivariate probit modeling. Ignatans and Pease (2016) calculated the top 10% respondents with more victimizations and used contingency tables but not regression analysis. Finally, Hope and Norris (2013) and Tseloni and Pease (2005) used Latent Class Analysis (LCA) and Lorenz curve approaches, respectively, to identify polyvictims.

We also reviewed previous studies that analyze databases other than the CSEW. These studies tend to construct polyvictimization differently. For example, Segura et al. (2018) first coded individual victimization variables as binary (0=non-victim; 1=victim) and then summed these variables up to create an overarching numeric polyvictimization variable. Then, using this newly created polyvictimization variable, polyvictims are classified as those who suffer two or more crime types (Segura et al., 2018) or at least one more crime than the victim population average (also called “one-above the mean” approach; Finkelhor et al., 2005). The polyvictimization variable is usually a categorical ordinal variable with categories of “non-victim”, “(single) victim” and “polyvictim”. Importantly, previous studies mostly used binary variables with categories of “non-victim” and “victim” (Sampson and Wooldredge, 1987).

To our knowledge, no previous study has brought these methods together in a single paper to investigate (1) if different approaches lead to significantly different populations of polyvictims, and (2) if identified predictors of polyvictimization differ across operationalizations and statistical models. Importantly, no previous study has conducted a meta-analysis of predictors of polyvictimization considering the various existing operationalizations of polyvictimization. Segura et al. (2018) explored the measurement of polyvictimization of juveniles using the Juvenile Victimization Survey; and Radtke et al. (2024) undertook a systematic review of conceptualization and operationalizations of youth polyvictimization. They, however, neither included methods such as negative binomial regression or bivariate probit models nor investigated the predictors of polyvictimization.

Theories of (poly)victimization and profile of polyvictims

This section explains the theories of crime victimization and summarizes the predictors of (repeat) victimisation based on the existing literature. The three main theories used to explain victimization are lifestyle-routine activity theory (Hindelang et al., 1978), routine activities theory (Cohen and Felson, 1979) and social disorganization theory (Sampson and Groves, 1989).

Although lifestyle and routine activity theories are slightly different, both argue that the demographic and socioeconomic characteristics of individuals and households, and their routine activities and lifestyles, influence their chance of exposure to crime through enabling offenders to interact with targets in the absence of guardians (Turanovic, 2023). Social disorganization theory argues that low economic status, ethnic heterogeneity, residential mobility, and family disruption lead to community-level social disorganization, which then increases offending and victimization (Sampson and Groves, 1989). Predictor variables used in the present study were chosen following these theories.

As noted above, we carried out a rapid evidence assessment to identify the studies that analyzed the CSEW to investigate the predictors of polyvictimization. Only 10 studies had a form of polyvictimization as their outcome variable and investigated its predictors. Existing studies usually consider individual-level (lifestyle-routine activities) and the macro-level (social disorganization) theories of victimization (Sampson and Wooldredge, 1987).  For example, Tseloni and Pease (2015) investigate (repeat) personal victimization drawing on the above theories and find that the number of visits to the pub and club increase personal victimization. Positive area correlates of repeat victimization are poverty and population density (Tseloni and Pease, 2015). Sampson and Wooldredge (1987) explored victimization – but not repeat victimization – and found that those who frequently go out at night or leave their homes empty experience personal crime victimization the most. Family disruption (i.e., lone parent households), unemployment, and housing density are all associated with burglary victimization as indicators of community context (Sampson and Wooldredge, 1987). 

Osborn et al. (1996) also investigated both micro- and macro-level predictors of property crime victimization without explicitly referring to routine activity or social disorganization theories. They found that multiple victims do not differ significantly from single victims. At the micro level, living at the same address for 2 to 5 years increases property crime victimization risk. Conversely, having an Afro-Caribbean ethnic background and being older are factors that reduce the risk of becoming a victim of property crime. At the macro-level, single parents not in self-contained accommodations, the percentage of young people aged 16-24, and percentage of households without car increase the risk of property crime victimization.

It is important to note that previous research using the CSEW has tended to focus on either property crime or personal crime victimization and included similar predictors in their analysis, which were found to be robust indicators of victimization risks (Trickett et al., 1995; Osborn et al., 1996). This is the reason why instead of reporting a detailed list of predictors of repeat victimization, here we explain how victimization theories guide us to choose our predictor variables.

Methodology

Data

The distributions of four crime types (vehicle crime, burglary, personal theft, violence, and digital crime) are taken from the 2019/20 CSEW to construct polyvictimization variables. We effectively employ data from 32,410 respondents. The CSEW, formerly known British Crime Survey, employs a multistage stratified sample, which is representative of the adult (16 years or older) population living in private accommodation in England and Wales. The sampling frame of the 2019/20 CSEW is Postcode Address File. The sample does not consider students living in accommodations, people residing in care homes, permanent residents in hotels, hospitals, prisons, and schools, people living in trailers, and homeless.

We used the non-victim form. Respondents (randomly selected among the adults living in the household) answered questions about household property crime on behalf of the whole household, and personal crime regarding their own individual experiences (Hope et al., 2001). Household crimes include vehicle crimes and burglaries, while personal crimes include theft, violence, and digital crime.

Outcome variables: Polymeasures of polyvictimization

The strength of this study is constructing several polyvictimization outcome variables out of various victimization types by bringing together the methods used in previous studies. Newly constructed polyvictimization measures are then used as dependent variables in different statistical models which investigate the predictors of polyvictimization.

We have two main approaches for constructing polyvictimization dependent variables. The first one (henceforth, “Method 1”) is treating individual victimization types (i.e., vehicle crime, burglary, personal theft, violence, digital crime) as binary variables (0=non-victim; 1=victim) (see Segura et al., 2018). The second approach (henceforth, “Method 2”) is treating victimization measures as count variables (e.g., range=0-12) (see Hope et al., 2001, Table 1 and Supplementary Materials Table 1). Using both approaches, we then construct overarching polyvictimization variables by summing up these binary and count variables, separately. In turn, the “number crime types” variable is constructed using the binary variables and it ranges from 0 to 5 (i.e., Method 1), while the “number crime types and times” variable is constructed using the count variables and ranges from 0 to 90 (i.e., Method 2). Having done so, we use different sub-methods to identify polyvictims.

            The first sub-method identifies polyvictims as those who experience “2 or more crime types” (e.g., burglary and personal theft) following Method 1, or “2 or more crime types and/or times” (e.g., repeat victims of the same or different crimes) following Method 2. Given that we use binary variables in Method 1, and count variables in Method 2, our new indicators of polyvictimization effectively measure respondents who report “2 or more crime types” and “2 or more types and/or times of victimization”. Our new variables classify “polyvictims” (2 or more types/times) against “others” (non-victims and one-off victims), and we use this measure as a dependent variable in binary logistic regressions (henceforth, “2+/One above mean* (binary logistic)”[2] and “2 or more** (binary logistic)”). We then disaggregate the “others” group as “non-victims” and “one-off victims” in order to estimate multinomial logistic regressions (see “2+/One above mean* (multinomial logistic)” and “2 or more** (multinomial logistic)” in Table 1 and Supplementary Materials Table 1).

The second sub-method is the “one-above-the-mean number of victimizations” approach. With this method, if we utilize the “number crime types” variable, “polyvictims” are those who experience “2 or more crime types” (effectively the same individuals as with the previous method, “2+/One above mean* (binary logistic)”); and “3 or more crime types/times” if we use the “number crime types and times” variable (henceforth, “One above mean** (binary logistic)”).  Similarly, we recoded the “number crime types” and “number crime types and times” variables as binary variables to be used in binary logistic regression, and nominal variables with three categories to be used in multinomial logistic regression (henceforth, “2+/One above mean* (multinomial logistic)” and “One above mean** (multinomial logistic)”).

The third sub-method classifies polyvictims as top “10% of the sample who experienced the highest number of victimizations” (Finkelhor et al., 2009; Radford et al., 2013). With this method, if we use the “number crime types” variable, “polyvictims” are those who experience “3 or more crime types” (henceforth, “Top 10%* (binary logistic)”), and “4 or more crime types/times” if we use the “number crime types and times” variable (henceforth, “Top 10%** (binary logistic)”). Similarly, we recoded the “number crime types” and “number crime types and times” variables as binary variables to be used in binary logistic regression, and nominal variables with three categories to be used in multinomial logistic regression (henceforth, “Top 10%* (multinomial logistic)” and “Top 10%** (multinomial logistic)”).

The fourth sub-method in our study classifies polyvictims following LCA. In order to estimate LCA models, we used both individual binary variables and count variables recoded as categorical variables, with categories 0, 1, 2, 3, 4, and 5 or more, respectively (Latent Class Analysis Variables in Table 1 and Supplementary Materials Table 1). When we used binary variables, the best model identified two latent classes: “non-polyvictims” and “polyvictims”. Similarly, when we used nominal variables with six categories, the best model identified two latent classes: “non-polyvictims” and “polyvictims” (Hope and Norris, 2013).

The fifth sub-method classifies polyvictims following Lorenz curve analysis. Following Tseloni and Pease (2005), we used the “number crime types and times” measure as a count variable. We identified polyvictims as the top 10% of the victim population and constructed two outcome variables, a binary variable and a nominal variable with three categories. We then estimated binary logistic and multinomial logistic regression models, (henceforth, “Lorenz curve top 10%** (binary logistic)” and “Lorenz curve top 10%** (multinomial logistic)”, respectively).

Binary logistic regression, multinomial logistic regression, LCA, Lorenz curve analysis and other statistical methods used (e.g., negative binomial, bivariate probit regression analyses) are explained in detail in the “Data analysis” section. Table 1 presents descriptive statistics of the polyvictimization variables constructed.

Table 1. Descriptive statistics of dependent variables

Predictor variables

Inclusion of predictor variables in the study was driven by evidence from the literature. Socio-demographic characteristics were found to be correlates of victimization in previous research (Trickett et al., 1995). Therefore age, sex, ethnicity, marital status, and education level were included. As a proxy for accessibility and desirability (Miethe and Meier, 1990), we included socioeconomic status, number of cars, household tenure, and accommodation types. To account for physical guardianship component of lifestyle/routine activities theory (the ability of people to prevent crime) number of adults was included.  Length of residence at the same address, tenure, and household composition and accommodation type proxy social guardianship (Meier and Miethe, 1993). Time spent away from home, and frequency of going to pubs and clubs were included as proxy for routine activities and exposure to crime (Tseloni and Pease, 2015). We also included reported mental health status because previous research found a correlation between polyvictimization and trauma (Finkelhor et al., 2007). Area type (i.e., inner city, urban, or rural) was included to reflect proximity to crime (Miethe and Meier, 1990). Table 2 presents descriptive statistics of predictor variables. All predictor variables are coded as dummy variables and hence treated as categorical variables.

Table 2. Descriptive statistics of predictor variables

Data analysis: Regression modelling

Before estimating regression models, we tested the inter-reliability of the newly constructed categorical polyvictimization variables using Cohen’s κ (McHugh, 2012).

We then conducted binary logistic regression if the polyvictimization outcome variable has two categories (i.e., “polyvictims” versus “others”), and multinomial logistic if it has three categories (e.g., “non-victim”, “single victim” and “polyvictim”). Following Osborn et al. (1996), after fitting multinomial logistic regressions, we also estimated bivariate probit regression models using the “2+/One above mean* (multinomial logistic)” and “2 or more** (multinomial logistic)” variables.

While binary logistic regression is used to predict dichotomous outcome variables, multinomial logistic regression is used when there are three or more possible outcome categories (Britt and Weisburd, 2009). As opposed to binary logistic regression, multinomial regression considers the changes in the odds of multiple outcomes, which then leads to multiple comparisons. In the current study, we first chose “non-victims” as the reference category and compared them with “one-off victims” and “polyvictims”. Subsequently, we changed the reference category to “one-off victims”. We do not report the results from the latter due to space constraints.

Bivariate probit regression is used to estimate several correlated binary outcomes jointly. In the current study, our first binary variable measured “victims” versus “non-victims”. For the second binary variable, the logic was that only when the victim “hurdle” is crossed (i.e., the first binary variable) we have information about the process of repeat versus single victimization, which becomes the second “hurdle” (i.e., second binary variable; Osborn et al., 1996). Therefore, the process of becoming a polyvictim which is effectively estimated in our bivariate probit regressions is as follows: first non-victim, then a victim, and finally a polyvictim. Due to space constraints, in this article we only present the analysis of polyvictims versus non-polyvictims.

LCA is used to explore similarity of responses (e.g., victimization) among respondents. LCA captures the clustering of individuals through analysis of manifest variables that are related to an underlying categorical construct (Muthén, 2002), which is polyvictimization in our case (see Supplementary Materials Table 1). In our analyses, the best models identified two latent classes (“non-polyvictims” and “polyvictims”), therefore we used binary logistic regressions after conducting LCA.

Lorenz curve was originally developed as a graphical representation of the distribution of wealth. We apply it to victimization (Tseloni and Pease, 2005). Lorenz curves plot the cumulative percentage of victimization against the cumulative percentage of the population. We identified the top 10% of the victim group as polyvictims and constructed two outcome variables: a binary and a nominal variable with three categories. We used binary logistic regression and multinomial logistic regression after conducting Lorenz curve analysis.

Furthermore, using the “number crime types” and “number crime types and times” variables, we then estimated negative binomial regression models, which are appropriate for over-dispersed count outcome variables (Cameron and Trivedi, 1986).

Data analysis: Meta-analysis

Finally, we compared the eighteen regression models calculated using the multiples measures of polyvictimization. Initially, we present regression estimates in tables and graphs, facilitating a visual examination of variations in statistically significant predictors of polyvictimization. Subsequently, we employ a meta-analytic linear fixed-effects model to compute weighted averages of the “true” effects across all eighteen regressions. Linear fixed-effects models, also referred to as equal-effects models, operate under the assumption of homogeneous “true” underlying effects across models (Fleiss, 1993; Rice et al., 2018). This assumption proves particularly fitting in our study, given our comparison of models derived from the same dataset to predict the same underlying construct of polyvictimization. By employing meta-analytic fixed-effects models, we calculate inverse-variance weighted averages of regression estimates for each predictor, thereby identifying predictors of polyvictimization that remain statistically significant across regression models. We interpret our meta-analytic estimates as reflective of the “true” effect of each predictor on the underlying multi-measured outcome of polyvictimization.

Beyond presenting meta-analytic odds ratios and their corresponding confidence intervals, we explore the homogeneity in regression estimates across models through several metrics:

1. Q statistic, indicating substantial variability between effect sizes across model;

2. I², which quantifies the proportion of total variation across models that is due to heterogeneity rather than chance;

3. Higgins I², a modification of the I² test for meta-analysis, which estimates the proportion of total variation in effect estimates due to genuine differences between the models rather than random error; and

4. τ², estimating the amount of true heterogeneity beyond observed variation.

Furthermore, utilizing our meta-analytic estimates as the “true” effect for each predictor across all models, we compute the rates of false positives (FPR), false negatives (FNR), true positives (TPR), and true negatives (TNR) for each regression model. TPR and TNR reflect a positive performance of a model, while FNR indicates a low capacity to identify significant predictors, and FPR entails the risk of mistakenly identifying significant predictors. This analysis not only elucidates the “true” predictors of polyvictimization across models but also informs future research regarding preferred approaches to mitigate the risk of false positives and negatives. We accomplish this by directly comparing statistically significant predictors in each model against the meta-analytic estimates using the following formulae:

• FPR = FP/(FP + TN),

• FNR = FN/(FN + TP),

• TPR = TP/(TP + FN), and

• TNR = TN/(TN + FP);

where FP represents a “false positive” estimate, TN a “true negative”, TP a “true positive”, and FN a “false negative” estimate.

Data can be accessed via the UK Data Service (Office for National Statistics, 2021) and analytic codes are available on Github (anonymized for review: https://anonymous.4open.science/r/CSEW_polyvictimization/).

Results

The Results section is structured as follows: First, we present the results from the inter-reliability analysis in order to check if there is agreement between the computed categorical polyvictimization variables. Second, we summarize and present the results from the regression models. Finally, we present the meta-analysis findings.

Inter-reliability results

Cohen's κ is a statistical measure used to evaluate the level of agreement between two or more raters that classify items into mutually exclusive categories. In this paper, we used two overarching methods (i.e., binary variables versus count variables) and several sub-methods (e.g., 2 or more types of crime) to construct polyvictimization variables. We test the level of consistency in the classification of polyvictims according to these different approaches. The κ value can range from -1 to 1, where 1 indicates perfect agreement, 0 suggests no agreement better than random chance, and -1 indicates perfect disagreement (McHugh, 2012).

In Table 3, the κ for “2+/One above mean* (binary logistic)”-“2 or more** (binary logistic)” (0.73) suggests a strong level of consistency in the classification of polyvictims. “2+/One above mean* (multinomial logistic)”-“2 or more** (multinomial logistic)” (0.91) suggests that the measurements are almost always in agreement. “2+/One above mean* (binary logistic)”-“One above mean** (binary logistic)” (0.56) represents moderate level of agreement. While there is a fair level of consistency, it is not as strong as in the first two pairs, indicating some variability in across these two approaches. “2+/One above mean* (multinomial logistic)”-“One above mean** (multinomial logistic)” (0.9) indicates almost perfect agreement, showing a very high level of consistency in the classifications. The “Top 10%* (binary logistic)”-“Top 10%** (binary logistic)” (0.45) κ value suggests a moderate agreement. It is on the lower side compared to other pairs, indicating less consistency in the identification of polyvictims. “Top 10%* (multinomial logistic)”-“Top 10%** (multinomial logistic)” (0.95) indicates almost perfect agreement, the highest among the listed pairs. “LCA* (binary logistic)”- “LCA** (binary logistic)” (0.88) shows also very strong agreement, indicating that these variables are highly consistent with each other.

In summary, most pairs show a moderate to high level of consistency in classifications across variables. A couple of pairs show moderate agreement, suggesting some variability but still a reasonable level of consistency. These results suggest overall strong reliability in the way variables are classified, meaning that for polyvictimization identification purposes it carries minimal weight whether binary variables or count variables are used. Another way to test the level of agreement between classifications could be across the sub-methods, but due to space constraints we do not report results here.

Table 3. Inter-reliability results

Regression model results

Table 4 presents the set of statistically significant predictors of polyvictimization across all regression models; the table displays the direction of the statistically significant association (i.e., positive or negative) (see Supplementary Materials Tables 2-20 for full model results). First, we report those predictors of polyvictimization that are statistically significant across all models. Second, we briefly discuss those predictors which hold statistically significant associations in certain regression models but not others. Figure 1 visually displays all regression estimates for each predictor; thus enabling exploring variation in effect sizes.

First, we report those predictors with statistically significant effects across all models. Car ownership is a statistically significant predictor of polyvictimization, and it suggests that households with fewer than 3 cars are less likely to suffer polyvictimization compared to those who have 3 or more cars. The effect sizes are similar across models, as shown in Figure 1. Second, those with a mental health condition are more likely to become a polyvictim than those without. Effect sizes, nonetheless, vary extensively across models, with “Top 10%* (multinomial logistic)” showing an effect size almost 3 times larger than “2+/One above mean* (bivariate probit)” and “2 or more** (bivariate probit)” (see Figure 1). Finally, those who have less education than A levels are less likely to become a polyvictim compared to those with A levels. 

Second, the predictors which hold statistically significant associations in certain regression models but not others are as follows. Females aged 16-29 are more likely to become a polyvictim compared to males 16-29 in models “Top 10%** (binary logistic)” and “Top 10%** (multinomial logistic)”. Females aged 60 or above are less likely to become a polyvictim compared to males 16-19 in “Number crime types and times** (negative binomial)” only. Persons from mixed/multiple ethnic background are more likely to experience polyvictimization compared to those from White ethnic background in five models, and those with “other” ethnicity in four models. Single and separated/divorced individuals are more likely to experience polyvictimization compared to married individuals in eleven and sixteen models out of eighteen, respectively. Individuals in a routine/manual occupation or unemployed are less likely to become polyvictims compared to individuals in a higher managerial occupation in ten and three models, respectively.

Households with three or more adults are more likely to become polyvictims compared to households with two adults in “Number crime types and times** (negative binomial)” only. Households that reside at the same address for less than two years and two to five years are more likely to become polyvictims compared to households that have resided at the same address for more than 10 years in five and four models, respectively. Social and private renters are more likely to experience polyvictimization compared to owners in seventeen and ten models, respectively. Terraced houses and flats are more likely to be polyvictimized than detached houses in ten and seventeen models, respectively. Households with no children, with children, and with a reference person aged 60 or above are less likely to experience polyvictimization than lone parent households in sixteen, sixteen, and fourteen models, respectively.

Those who do not leave their houses are more likely to become polyvictims compared to those who leave their houses 7+ hours on a weekday in models “2+/One above mean* (binary logistic)” and “2 or more** (multinomial logistic)”. Those who leave their houses one to three hours or five to seven hours on a weekday are less likely to become polyvictims compared to those who leave their houses more than 7 hours in models “Lorenz curve top 10%** (multinomial logistic)”, “Lorenz curve top 10%** (multinomial logistic)” and “Number crime types and times** (negative binomial)”. Those who go to pubs fewer than 9 times during the last month are less likely to become polyvictims in at least twelve models. Those who go to night clubs one to three times in the evening during the last month are more likely to become polyvictims compared to those who go night clubs more than four times in “Number crime types and times** (negative binomial)” only.

Finally, those living in urban and inner-city areas are more likely to become polyvictims compared to those who reside in rural areas in thirteen and twelve models, respectively. Among all the predictors, results indicate that individuals with mental health condition are the group with the highest polyvictimization risk.

Table 4: Summary of the model results

Model 1: Number crime types* (negative binomial), Model 2: 2+/One above mean* (binary logistic), Model 3: 2+/One above mean* (multinomial logistic), Model 4: Top 10%* (binary logistic), Model 5: Top 10%* (multinomial logistic), Model 6: LCA* (binary logistic), Model 7: 2+/One above mean* (bivariate probit), Model 8: Number crime types and times** (negative binomial), Model 9: 2 or more**(binary logistic), Model 10: 2 or more** (multinomial logistic), Model 11: One above mean** (binary logistic), Model 12: One above mean** (multinomial logistic), Model 13: Top 10%** (binary logistic), Model 14: Top 10%** (multinomial logistic), Model 15: Lorenz curve top 10%** (binary logistic), Model 16: Lorenz curve top 10%** (multinomial logistic), Model 17: LCA** (binary logistic), Model 18: 2 or more** (bivariate probit)

Figure 1. Correlation coefficient by dependent variable (Odds ratio)

Meta analysis results

We have identified a set of predictors which hold statistically significant associations across all models, as well as other predictors which only show significant associations in some models but not others. While we can almost certainly assume that those predictors with significant effects across all models (i.e., car ownership, mental health, and education level) are indeed important determinants of polyvictimization; one should not directly disregard those predictors which do not always hold significant associations with polyvictimization. It is possible, for instance, that certain measurements of polyvictimization, or regression modelling approaches, fail to identify relevant predictors due to factors such as measurement error and validity of measurement, range of variation in outcome measures, uncontrolled interaction effects, or the statistical power of the models.

In all cases, we found that there was substantial variability between the effect sizes across the models (i.e., a significant Q statistic), providing evidence that various regression models effectively lead to significantly varying estimates. Moreover, I² metrics indicate that effect sizes for the categories car ownership, number of visits to pubs, mental health status, unqualified (no education), urban area, inner city, no children, and separated or divorced exceed 50%, suggesting that effect sizes are particularly heterogeneous - and hence susceptible to variation due to the measure being used - for these predictors. This is further reflected in the particularly large Higgins’ I² metrics estimated for these variables.

In Figure 2 we display the estimates of our meta-analytic fixed-effects model, alongside their 95% confidence intervals (used to identify significant predictors across models). In other words, Figure 2 displays the weighted averages of the “true” effects across all eighteen regressions. The meta-analysis results show that car ownership, mental health status and education level are indeed important predictors of polyvictimization. For instance, mental health status displays the largest positive effect size on polyvictimization (i.e., those with mental health conditions are between 1.69 and 1.79 times more likely to be polyvictims than those without). Having no car (as opposed to three or more cars) and no qualification (as opposed to A level of above) also show large negative impacts on polyvictimization. Moreover, we also identify other significant predictors which were not previously observed across all models.

First, females over 60 are only between 72 to 87% as likely as males under 30 to experience polyvictimization. Second, while we find no effect of Black or Asian ethnicity on polyvictimization, respondents with a mixed ethnicity background and those belonging to other ethnic groups are around 1.31 and 1.25 times more likely to be polyvictimized than Whites. Single and separated respondents are around 16% and 24% more likely to be polyvictims than married individuals, respectively. Unemployed and those with manual occupations are only around 86% and 82% as likely be polyvictimized as those in higher managerial positions.

Attending pubs fewer than 9 times during the last month is negatively and strongly associated with polyvictimization; while spending less than 7 hours aways from home on a weekday is also negatively associated with polyvictimization but with a much weaker effect. Households with no children, with children, or with a reference person aged over 60 are only around 65%, 72% and 60% as likely as lone parent households to experience polyvictimization. Finally, the fewer years residing at the same address, the larger chances of polyvictimization; social and private renters are around 24% and 15% more likely to be polyvictims than homeowners; terraced houses and flats have 1.15 and 1.25 times higher chances to be polyvictimized than detached ones; and there are higher chances of experiencing polyvictimization in urban and inner-city areas compared to rural zones.

Figure 2. Meta-analysis results (Odds ratios and confidence intervals)

Beyond exploring the “true” predictors of polyvictimization across models, the meta-analysis further informs about the measurement approaches which mitigate or exacerbate the risk of false positive and negatives (see Figure 3).

The only model with a FPR higher than zero is “Top 10%* (binary logistic)”, indicating that this approach is the only one in our study which has misidentified a variable as having a significant effect on polyvictimization. The risk of false positives should be reduced to zero whenever possible. Similarly, “Top 10%* (binary logistic)” is the only approach with a TNR below 100%, and the approach with the largest FNR, indicating the overall low accuracy of this approach compared with the others. Importantly, “Top 10%* (multinomial logistic)” also performed poorly both in its FNR and TPR.

On the other hand, the approaches with the largest TPR and smallest FNR are “2+/One above mean* (bivariate probit)”, “2 or more** (bivariate probit)”, and “Number crime types and times** (negative binomial)”. However, it is important to note that such metrics may be influenced by the fact that fewer independent variables are included in the final bivariate probit regression models than in the rest of the models. This is not the case for negative binomial regressions. “Number crime types* (negative binomial)” and “LCA** (binary logistic)” also performed comparatively well, with a FNR smaller than other approaches, and a larger TPR.

Figure 3. Accuracy tests calculated for each measurement/regression model

Discussion and conclusions

This paper had three main aims: (1) reviewing various ways of measuring polyvictimization, (2) investigating if the identification of polyvictims is dependent upon operationalization decisions, and (3) conducting a meta-analysis to identify correlates of polyvictimization based on the 2019/20 CSEW. Previous research has applied heterogeneous approaches to conceptualizing, operationalizing, and analyzing ‘polyvictimization’. There is an overall lack of consistency in the methodological approaches used to study the phenomenon, which can have important implications for research as well as crime prevention efforts.

Our inter-rater reliability analysis of the various methods applied in previous research revealed that the use of binary and count variables for constructing polyvictimization measures has a negligible impact on the identification of polyvictims. This is likely driven by the fact that the number of participants who experience more than 2 types of crime tends to be relatively small, and in turn the use of binary or count measures has little impact on how polyvictims are identified. Researchers and practitioners solely interested in identifying polyvictims can choose the simpler of these methods without compromising the accuracy of the outcome. However, we encourage researchers to explore other aspects of polyvictimization measurement such as intra-familial crime victimization, which was not considered in our study. While the operationalization of polyvictimization may have negligible effects on the identification of polyvictims, it is also key to ensure methodological decisions do not have significant impacts on regression models exploring the causes and consequencies of polyvictimization.

We explored the impact of using various operationalizations on regression estimates exploring predictors of polyvictimization. Operationalization choices do have a strong impact on regression estimates, both regarding the identification of statistically significant predictors and their effect sizes. Upon completing the meta-analysis of results obtained from using various operationalizations and regression approaches, our study revealed that one of the best approaches to explore the predictors of polyvictimization is the use of negative binomial modelling with a dependent variable that is computed using count variables of victimization. Therefore, where possible future research should use count victimizations to compute a count polyvictimization variable and estimate negative binomial models. Alternatively, researchers might want to use bivariate probit modelling as these models performed relatively well compared to other approaches. If researchers would like to compute a binary dependent variable, our results indicate that the preferred approach would to be to first apply LCA with count victimization variables to compute the polyvictimization variable and then estimate binary logistic regression models. Conversely, researchers should refrain from using the “Top 10%* (binary logistic)” and “Top 10%* (multinomial logistic)” approaches to reduce the risk of false positives and false negatives.

Our meta-analytical approach also enables exploring which predictors of polyvictimization hold a stronger effect size to explain individual differences in polyvictimization. Across all measures, the strongest correlate of polyvictimization is self-reported mental health condition. While we note that most previous studies did not consider the effect of mental health, Chan (2017) and Tanksley et al. (2020) also identified the relevance of mental health in explaining polyvictimization. Criminal Justice System practitioners, including the police, are encouraged to develop interventions to target polyvictims considering the substantial amount of time spent responding to calls from individuals with mental-ill health (Langton et al., 2021). We advocate for increased funding for research on polyvictimization and mental health, as well as for services that support victims (The Lancet, 2024). We recommend the development of policies that recognize the intersection between victimization and mental health and emphasize the importance of early detection of signs of polyvictimization in mental health patients.

Furthermore, our meta-analytic findings generally support routine activities and social disorganization theories of victimization, in line with previous literature (Tseloni, 2000, 2006; Tseloni and Pease, 2004). For example, in line with social disorganization explanations of polyvictimization, mixed/multiple ethnic and other ethnic individuals are more likely to experience polyvictimization compared to Whites. Single or separated/divorced individuals are more likely to experience polyvictimization compared to married respondents (Tseloni and Pease, 2004). These results align with the expectations of social disorganization theory. Furthermore, households living at the same address less than 10 years are more likely to experience polyvictimization, which supports social disorganization theory and highlights the importance of social cohesion. Lone parent households are more likely to experience polyvictimization; such a finding also supports social disorganization theory as lone parent households proxy economic instability, which might exacerbate the effects of social disorganization. Finally, those who reside in urban and inner-city areas experience polyvictimization more often, which is in line with social disorganization theory.

Our meta-analysis also partly supports the expectation of the routine activity models of victimization. Individuals who are unemployed or who ae in routine/manual occupations are less likely to experience polyvictimization compared to individuals in higher managerial occupations. While these results do not necessarily align with previous international studies (Block et al., 1985), Tseloni (2006) argued that they are supportive of routine activities models of victimization, since individuals in higher managerial occupations might have lifestyles that expose them to different risks and make them targets for theft (Tseloni, 2006). Households with no car, one car, or two cars are less likely to become polyvictims compared to households with three or more cars. This finding is in line with previous research (Tseloni and Pease, 2004; Tseloni, 2006) and aligns with the routine activity theory, given that households with three or more cars might symbolize wealth, attracting offenders and increasing visibility and accessibility. Social and private renters are more likely to experience polyvictimization compared to homeowners, which is in line with both social disorganization and routine activities theories given that renters experience high rates of social mobility thereby hindering the development of social ties (Wickes et al., 2019). Terraced houses and flats are also more likely to become polyvictims, which supports both social disorganization and routine activities theories. These types of accommodations tend to be concentrated in densely populated areas with lower levels of social cohesion (Haynes et al., 2007). In addition, the close-knit structure and share access points of these type of households might provide opportunities to offenders. Those who leave their houses often, and those who frequently visit pubs, are more likely to experience polyvictimization, which supports lifestyle-routine activity theories.

Finally, individuals with a qualification below A level are less likely to become polyvictims. Tseloni (2000) argues that this finding should be treated with suspicion given that educated people might be able to identify the criminogenic nature of the incidents and complete complex surveys, such as the CSEW, more accurately.

The study is not without limitations. Our review focused on adult victimization, but other studies have investigated polyvictimization of children analyzing other datasets (Fisher et al., 2015; XXXX et al., 2023). The CSEW dataset categorizes gender in a binary manner, which impedes our ability to assess the experiences of nonbinary individuals (Messinger et al., 2022). Finally, we only analyzed data from the screener questionnaire of the 2019/20 CSEW and did not investigate the victim forms in which victims provide in-depth information about the nature of victimization.

References

Block, R., Felson, M., & Block, C.R. (1985). Crime Victimization Rates for Incumbents of 246 Occupations. Sociology and Social Research, 69(3), 442-451.

Blom, N., & Gash, V. (2023). Measures of violence within the United Kingdom Household Longitudinal Survey and the Crime Survey for England and Wales: an empirical assessment. Social Sciences, 12(12), 649.

Britt, C.L., & Weisburd, D. (2010). Logistic regression models for categorical outcome variables. In A.R. Piquero & D. Weisburd (Eds.), Handbook of quantitative criminology (pp. 649-682). London: Springer.

Cameron, A.C., & Trivedi, P.K. (1986). Econometric models based on count data. Comparisons and applications of some estimators and tests. Journal of Applied Econometrics, 1(1), 29-53.

Chan, K.L. (2017). Family polyvictimization and elevated levels of addiction and psychopathology among parents in a Chinese household sample. Journal of Interpersonal Violence, 32(16), 2433-2452.

Cohen, L.E., & Felson, M. (1979). Social change and crime rates trends: A routine activity approach. American Sociological Review, 44, 588–608.

Ellingworth, D., Farrell, G., & Pease, K. (1995). A victim is a victim is a victim-chronic victimization in four sweeps of the British Crime Survey. British Journal of Criminology, 35(3), 360-365.

Finkelhor, D., Ormrod, R., Turner, H., & Hamby, S.L. (2005). Measuring poly-victimization using the juvenile victimization questionnaire. Child Abuse & Neglect, 29(11), 1297–1312.

Finkelhor, D., Ormrod, R.K., & Turner, H.A. (2007). Polyvictimization and trauma in a national longitudinal cohort. Development and Psychopathology, 19(1), 149-166.

Finkelhor, D., Ormrod, R.K., & Turner, H.A. (2009). Lifetime assessment of poly-victimization in a national sample of children and youth. Child Abuse & Neglect, 33(7), 403-411.

Fisher, H.L., Caspi, A., Moffitt, T.E., Wertz, J., Gray, R., Newbury, J., Ambler, A., Zavos, H., Danese, A., Mill, J., Odgers, C.L, Pariante, C., Wong, C.C.Y., & Arseneault, L. (2015). Measuring adolescents’ exposure to victimization: The Environmental Risk (E-Risk) Longitudinal Twin Study. Development & Psychopathology, 27(4pt2), 1399–1416.

Fleiss, J.L. (1993). The statistical basis of meta-analysis. Statistical Methods in Medical Research, 2(2), 117-195.

Grove, L.E., Farrell, G., Farrington, D.P., & Johnson, S.D. (2012). Preventing repeat victimization: A systematic review. Stockholm: Brottsförebyggande rådet.

Haynes, R., Daras, K., Reading, R., & Jones, A. (2007). Modifiable neighbourhood units, zone design and residents’ perceptions. Health & Place, 13(4), 812-825.

Hindelang, M.J, Gottfredson, M.R., & Garofalo, J. (1978). Victims of Personal Crime: An Empirical Foundation for a Theory of Personal Victimisation. Cambridge, MA: Ballinger Publishing Company.

Hope, T., & Norris, P.A. (2013). Heterogeneity in the frequency distribution of crime victimization. Journal of Quantitative Criminology, 29, 543-578.

Hope, T., Bryan, J., Trickett, A., & Osborn, D.R. (2001). The phenomena of multiple victimization. The relationship between personal and property crime risk. British Journal of Criminology, 41(4), 595-617.

Hope, T., & Trickett, A. (2008). The Distribution of Crime Victimisation in the Population. International Review of Victimology, 15(1), 37-58.

Hunter, J., Ward, B., Tseloni, A., & Pease, K. (2021). Where should police forces target their residential burglary reduction efforts? Using official victimisation data to predict burglary incidences at the neighbourhood level. Crime Science, 10(1), 11.

Ignatans, D., & Pease, K. (2016). On whom does the burden of crime fall now? Changes over time in counts and concentration. International Review of Victimology, 22(1), 55-63.

Langton, S., Bannister, J., Ellison, M., Haleem, M.S., & Krzemieniewska-Nandwani, K. (2021). Policing and Mental ill-health: Using Big Data to Assess the Scale and Severity of, and the Frontline Resources Committed to, mental ill-health-related calls-for-service. Policing: A Journal of Policy and Practice, 15(3), 1963-1976.

Laycock, G. (2001). Hypothesis-based research: The repeat victimization story. Criminal Justice, 1(1), 59-82.

Lee, Y., Eck, J.E., O, S., & Martinez, N.N. (2017). How concentrated is crime at places? A systematic review from 1970 to 2015. Crime Science, 6, 1-16.

Martinez, N.N., Lee, Y., Eck, J.E., & O, S. (2017). Ravenous wolves revisited: a systematic review of offending concentration. Crime Science, 6(1), 1-16.

Mayhew, P., Aye Maung, N., & Mirrlees-Black, C. (1993). The 1992 British Crime Survey, Home Office Study No 132. London: Home Office

McHugh, M.L. (2012). Interrater reliability: the kappa statistic. Biochemia Medica, 22(3), 276-282.

Meier, R.F., & Miethe, T. (1993). Understanding theories of criminal victimisation. In M. Tonry (Ed.), Crime and Justice: a review of research (pp. 459-499). Chicago: University of Chicago Press.

Messinger, A.M., Guadalupe-Diaz, X.L., & Kurdyla, V. (2022). Transgender polyvictimization in the US Transgender Survey. Journal of Interpersonal Violence, 37(19-20), NP18810-NP18836.

Miethe, T.D., & Meier, R.F. (1990). Opportunity, choice, and criminal victimization: A test of a theoretical model. Journal of Research in Crime and Delinquency, 27(3), 243-266.

Muthén, B.O. (2002). Beyond SEM: General latent variable modeling. Behaviormetrika, 29, 81-117.

Nelson, J.F. (1980). Multiple victimization in American cities: A statistical analysis of rare events. American Journal of Sociology, 85(4), 870-891.

Office for National Statistics (2021). Crime Survey for England and Wales, 2019-2020. [data collection]. UK Data Service. SN: 8812, DOI: http://doi.org/10.5255/UKDA-SN-8812-1

O, SooHyun, Martinez, N.N., Lee, Y., & Eck, J.E. (2017). How concentrated is crime among victims? A systematic review from 1977 to 2014. Crime Science, 6(1), 1-16.

Osborn, D.R., Ellingworth, D., Hope, T., & Trickett, A. (1996). Are repeatedly victimized households different? Journal of Quantitative Criminology, 12, 223-245.

Osborn, D.R., & Tseloni, A. (1998). The distribution of household property crimes. Journal of Quantitative Criminology, 14, 307-330.

Pease, K. (1998). Repeat victimization: Taking stock. London: Home Office

Farrell, G. & Pease, K. (2014). Repeat victimisation. In G. Bruinsma & D. Weisburd (Ed.), Encyclopaedia of Criminology and Criminal Justice (p. 4371). New York: Springer.

Pease, K., Ignatans, D., & Batty, L. (2018). Whatever happened to repeat victimisation? Crime Prevention and Community Safety, 20, 256-267.

Radford, L., Corral, S., Bradley, C., & Fisher, H.L. (2013). The prevalence and impact of child maltreatment and other types of victimization in the UK: Findings from a population survey of caregivers, children and young people and young adults. Child Abuse & Neglect, 37(10), 801-813.

Radtke, S.R., Wretman, C.J., Fraga Rizo, C., Franchino-Olsen, H., Williams, D.Y., Chen, W.T., & Macy, R.J. (2024). A Systematic Review of Conceptualizations and Operationalizations of Youth Polyvictimization. Trauma, Violence, & Abuse, 0(0).

Rice, K., Higgins, J.P.T., & Lumley, T. (2018). A Re-Evaluation of Fixed Effect(s) Meta-Analysis. Journal of the Royal Statistical Society Series A: Statistics in Society, 181(1), 205–227.

Sampson, R.J., & Wooldredge, J.D. (1987). Linking the micro-and macro-level dimensions of lifestyle-routine activity and opportunity models of predatory victimization. Journal of Quantitative Criminology, 3, 371-393.

Sampson, R.J., & Groves, B.W. (1989). Community structure and crime: Testing social disorganisation theory. American Journal of Sociology, 94, 774–802.

Segura, A., Pereda, N., & Guilera, G. (2018). Poly-victimization from different methodological approaches using the juvenile victimization questionnaire: are we identifying the same victims? Journal of Trauma and Dissociation, 19(3), 289-306.

Sparks, R.F., Genn, H.G., & Dodd, D.J. (1977). Surveying victims: A study of the measurement of criminal victimization, perceptions of crime, and attitudes to criminal justice. London: Wiley.

Tanksley, P.T., Barnes, J.C., Boutwell, B.B., Arseneault, L., Caspi, A., Danese, A., Fisher, H., & Moffitt, T.E. (2020). Identifying psychological pathways to polyvictimization: evidence from a longitudinal cohort study of twins from the UK. Journal of Experimental Criminology, 16, 431-461.

The Lancet (2024). Severe mental illness in the UK: a crisis of compassion. The Lancet, 403(10427), 587.

Trickett, A., Osborn, D.R., Seymour, J., & Pease, K. (1992). What is different about high crime areas? The British Journal of Criminology, 32(1), 81-89.

Trickett, A., Osborn, D. R., & Ellingworth, D. (1995). Property crime victimisation: the roles of individual and area influences. International Review of Victimology, 3(4), 273-295.

Tseloni, A. (1995). The modelling of threat incidence: Evidence from the British Crime Survey. In R.E. Dobash, R.P Dobash & L. Noaks (Eds.), Crime and Gender (pp. 269-294). Cardiff: University of Wales Press.

Tseloni, A. (2000). Personal criminal victimization in the United States: Fixed and random effects of individual and household characteristics. Journal of Quantitative Criminology, 16, 415-442.

Tseloni, A. (2006). Multilevel modelling of the number of property crimes: Household and area effects. Journal of the Royal Statistical Society Series A: Statistics in Society, 169(2), 205-233.

Tseloni, A., & Pease, K. (2003). Repeat personal victimization. ‘Boosts’ or ‘Flags’? British Journal of Criminology, 43(1), 196-212.

Tseloni, A., & Pease, K. (2004). Repeat personal victimization: Random effects, event dependence and unexplained heterogeneity. British Journal of Criminology, 44(6), 931-945.

Tseloni, A., Wittebrood, K., Farrell, G., & Pease, K. (2004). Burglary victimization in England and Wales, the United States and the Netherlands: A cross-national comparative test of routine activities and lifestyle theories. British Journal of Criminology, 44(1), 66-91.

Tseloni, A., & Pease, K. (2005). Population inequality: the case of repeat crime victimization. International Review of Victimology, 12(1), 75-90.

Tseloni, A., & Pease, K. (2015). Area and individual differences in personal crime victimization incidence: The role of individual, lifestyle/routine activities and contextual predictors. International Review of Victimology, 21(1), 3-29.

Tseloni, A., Knusstsson, J., & Laycock, G. (2005). Repeat Victimization: Introduction. International Review of Victimology, 12(1), 47-49.

XXXX et al., (2023).

Turanovic, J.J. (2023). Victimization and Its Consequences over the Life Course. Crime and Justice, 52(1), 265-342.

Wickes, R., Zahnow, R., Corcoran, J., & Hipp, J.R. (2019). Neighbourhood social conduits and resident social cohesion. Urban Studies, 56(1), 226-248.

Acknowledgements

We thank Dr Christopher Long, who is a Principal Academic in Healthcare Statistics at Bournemouth University, for the support we have received to fit the Bivariate Probit models and the comments on the analyses conducted in the paper.

Competing interests

The authors declare that they have no competing interests.

Funding

No outside funding was used to support this work.

Availability of data and materials

Data can be accessed via the UK Data Service (Office for National Statistics, 2021) and analytic codes are available on Github (anoymized for review: https://anonymous.4open.science/r/CSEW_polyvictimization/).

[1] ((Repeat victimization) OR (Chronic victimization) OR (Frequent victimization) OR (High-frequency victimization) OR (Multiple victimization) OR (Near repeats) OR (Near-repeat victimization) OR (Polyvictimization) OR (Prior victimization) OR (Recidivist victimization) OR (Recurrent victimization) OR (Reoccurrence of victimization) OR (Repetitive victimization) OR (Revictimization)) AND ((British Crime Survey) OR (Crime Survey for England and Wales))

[2]  Henceforth, * indicates Method 1, and ** Method 2. The statistical techniques in brackets indicate the statistical model used to predict the dependent variable. The original variable names used in the analyses can be found in Appendix Table 1.