When is violence not a crime? Factors associated with victims’ labelling of violence as a crime

Many people do not regard violence against them as a crime, but the factors that influence this response are unknown. Understanding how the ‘crimeworthiness’ of violence is interpreted allows an insight into how victims make sense of their experience, how communities influence attitudes towards victimisation and the reporting of crime to the police. A pooled cross-sectional sample of respondents to the Crime Survey for England and Wales was used to identify factors associated with the decision to label or discount a violent incident as a crime. Individual and neighbourhood-level effects were estimated using multilevel modelling. Harm, the perceived unjustness of the incident and victim-offender relational distance predicted labelling, whereas frequency of victimisation and victim initiation of the incident predicted discounting. Neighbourhood and neighbourhood crime had little effect on victims’ interpretations of the ‘crimeworthiness’ of violence. When victims interpret violence against them, they appear to do so unencumbered by social norms, but are influenced by the impact of the violence, the ‘prototypicality’ of the incident as a crime and their previous experience of violence.


Introduction
Criminal incidents are not objective phenomena: what is a crime to one person may not seem like a crime to another. That individuals vary in what they interpret as a crime is a central tenet of the sociology of law, but the reasons for this variability are not understood. In the few instances when the decision to 'label' something as a crime has been discussed, the focus has largely been on constructing and resisting 'victimhood' identities (Weiss, 2011) or exploring whether some currently legal behaviour should be criminalised (Nielsen, 2000). It appears that no research has been undertaken to identify the individual and contextual factors that constitute a crime in the eyes of the victim. Gaining an understanding of crime labelling is important for understanding how victims make sense of their experiences, for understanding victim behaviour such as reporting to the police and for providing insight into the socially constructed nature of crime. Focusing on violence, this paper seeks to identify factors that predict when an incident will be labelled as a crime -and, correspondingly, to predict when it will not. In essence, it seeks to identify the principal components of violent crime in the eyes of for those who have experienced violence.
The relevance of 'crime labelling' Labelling an incident as a crime is the first step in the use of law (Black, 1979). Before a victim of violence reports the incident to the police, they must decide that it is a crime, i.e. 'worthy of police attention' (Black, 1979: 20). This simple cognitive step can provide insight into the use of formal and informal methods of social control, to improving police ascertainment of crime, to understanding the 'dark figure' of crime, to increasing access to victim services, to understanding the democratic processes that influence the use of and changes in criminal law and to understanding the constructed nature of crime. As labelling exists only in the mind of the victim, it is free of implications for that victim, whereas reporting is not. As reporting crime is a behaviour, bringing with it consequences and uncertainty as well as motivational and logistical barriers, the factors that predict reporting cannot be generalised to be the factors that predict labelling.
It is important to draw a distinction between the use of the term 'crime labelling' in this context and the alternative use of the term 'labelling' in criminology, as used by Lemert (1951), Becker (1963) and others. Both approaches are concerned with the role of labels in the interpretation of the world. Where they differ is their chronological approach to the construct. The more common use of the term (as used by Lemert and Becker) is primarily interested in the consequences of attaching a label to a phenomenon. Importantly, the process of studying this labelling begins after the label has been attached. This paper, however, is interested in the factors that precede and influence the attachment of a label -crime -to a very specific phenomenon, violence. The difference between the two approaches, therefore, is that the one employed in this study treats labelling as the phenomenon to be observed, whereas the other is concerned with labelling as an active agent or mechanism that operates upon the world. In this study, respondents did not actively identify themselves as victims of crime or, without prompting, label violence against them as a crime: they responded to a question about their experience of violence and made a decision about whether they regarded the violence as a crime.

Theoretical background
Many studies have addressed victims' behavioural responses to violent crime, including avoidance (Baumgartner, 1988;Kilpatrick and Acierno, 2003), violent retaliation (Anderson, 1999;Felson and Steadman, 1983) and reporting -or not reporting -to the police (Brennan, 2011;Skogan, 1984), but these studies have generally implied that victims actually regard the violence as a criminal incident. As the concept of 'crime labelling' is novel, it lacks a theoretical basis from which to launch an empirical exploration. As the use of law is a consequence of labelling crime, it is possible that models of reporting crime can be translated into models of labelling crime. In one of the few theoretical models that can be applied to crime labelling, Ruback et al. (1984; see Figure 1) proposed that victims' decisions about crime are informed by two interacting and simultaneous factors: the impact of the incident on them and the extent to which the incident fits their model of what constitutes a crime, hereafter 'crime prototypicality'. This model is the starting point for this paper's exploration of the 'crime labelling' phenomenon.
Informed by Janoff-Bulman and Frieze (1983), Ruback et al. (1984) proposed that the impact of the crime is felt through several mechanisms: the harm or potential harm caused by the incident, the perceived injustice of the incident and the victim's perceived future vulnerability. The preference here for the term 'impact' over the more common 'harm' is important as it emphasises the extent to which an incident can affect an individual both immediately and prospectively. While their initial model focused on the incident as either legal or illegal, Ruback and colleagues' later work showed that the prototypicality of an incident is informed by more than this simple dichotomy. Using a series of vignette studies, they demonstrated that offender characteristics, victim characteristics, precipitatory factors and incident characteristics all contributed to hypothetical decisions about crime (Greenberg and Ruback, 1992;Ruback, 1994;Ruback et al., 1999Ruback et al., , 2004Ruback and Thompson, 2001). Later work found that a victim's construction of their 'victimhood' (Dunn, 2008) and their own prototypicality as a victim (Loseke, 2007;Zaykowski, 2015) can also affect the labelling process. Therefore, 'crime prototypicality' reflects not just the respondent's objective view of the legality of the event, but also their subjective understanding of how worthy such an incident is of the use of law (Black, 1979) and the 'affront' to their personhood that the incident represents.

Predictors of crime prototypicality and impact
Although the above model is a potentially valuable framework for understanding response to crime, it does not identify the specific factors that might influence this response. The most suitable model from which specific predictions can be made is Black's theory of law (1976). This theory seeks to explain variations in how individuals use law, such as why victims report or do not report victimisation to the police. Although not all parts of Black's theory have been shown to have predictive validity (Gottfredson and Hindelang, 1979; discussed further below), the combination of the models proposed by Black and Ruback et al. (1984) are a valuable starting point from which to explore what constitutes crime for those who experience violence.

Context Legality Harm Unjustness
Perceived future vulnerability Crime labelling Figure 1. Model of crime labelling adapted from Ruback et al. (1984). Black (1976) suggested that the use of law was influenced by relative social distance (stratification), relational distance (morphology), level of education (culture), victim group size (organisation) and desire to exercise social control. Importantly, the 'pure sociological' theory did not accommodate the seriousness of the incident, the effect of contextual factors such as incident location or internal factors such as victim attribution of blame. When the theory was tested by Gottfredson and Hindelang (1979) using victimisation survey data, the results, which offered little support for Black's theory, indicated that the failure to account for incident seriousness was a significant limitation. In the years since Black proposed his theory, incident seriousness, measured in a variety of ways, has been recognised as one of the few consistent predictors of the use of law. However, the theory should not be discounted entirely. The proposal that the likelihood of victims' use of law is proportional to the relational distance between victim and offender has received considerable empirical support as many studies have reported an unwillingness among victims to report the violence of their partners, families and friends (Felson and Messner, 1999;Gottfredson and Hindelang, 1979). A rational explanation of the observed phenomenon is that victims do not want to deal with the consequences of reporting, but -as has been demonstrated in the domestic abuse literature (Weiss, 2011) -it is also possible that this type of violence simply does not fit their prototype of a crime.
It is likely that the respondent's narrative of the violent event will have some impact on their interpretation of the incident (Gergen and Gergen, 1983). Therefore, their explanation of the cause of the incident will provide some insight into their understanding of the 'justness' of the violence. In the aftermath of violence, victims may attribute the cause of the incident to a range of different factors, such as bad luck, their own recklessness or an offender's violent disposition. Therefore, informed by Black (1976), victims are most likely to feel aggrieved by the incident when the violence was 'unbalanced' (i.e. they were victimised by more than one person) and when they were singled out as a target. These components combined constitute the 'unfairness' of the incident.
A mechanism of crime labelling and discounting Many victims of crime do not regard offences against them as a crime (Hough and Mayhew, 1985). In the context of violence, this is an important point. Being pushed, hit, kicked or assaulted in some other way to the extent that it is memorable is likely to reach the legal threshold for assault. Therefore, the gap between an assault being regarded as a crime (labelled) and not a crime (discounted) is filled with subjectivity. As suggested by Ruback et al. (1984), for crime discounters, the incident does not fit their prototype of 'a crime', it has caused them little harm or offence and it has done little to change their view of their own vulnerability or their view of the world. The event is therefore of little seriousness and can be disregarded. In these cases, it is proposed that the mechanism through which this has occurred is one of 'normalisation'. 1

Normalisation of victimisation
Violence between two adolescents is often regarded as less 'crimeworthy' than violence involving an adult (Finkelhor and Ormrod, 2001). Although this may be unsurprising, it is important to ask why this is the case. One potential explanation for this phenomenon is that of normativity. If fighting is regarded as a normal part of adolescent life and learning to deal with conflict is part of the maturing process, labelling the violence as a crime becomes inefficient. However, there are limits to this discounting approach; some violence is too serious, harmful, frequent or unusual to ignore, although the criteria for this decision are not defined. If such tacit rules exist for judging the 'crime prototypicality' of adolescent victimisation, it is likely that such a set of rules also exists for adult victims.
As with all learned concepts, an individual's prototype of crime is informed by their experiences and social norms. It is reasonable to expect that these prototypes can result in differing interpretations of what constitutes a crime: as Young has noted, 'what seems trivial to some would be serious to others ' (1988: 173). The effect of most stimuli dulls with repeated experience. Lottery winners eventually return to their prior levels of happiness (Brickman et al., 1978) and the quality of life of some (but not all) recently disabled people returns to its pre-injury levels (Diener et al., 2006). Similarly, repeated exposure to violence is associated with a reduced perceived significance of future violence (Garbarino, 1995). This phenomenon is known by several names including desensitisation, pathologic adaptation or normalisation and may explain differences in the labelling of some assaults. By this rationale -assuming similar levels of harm -over time, repeat victims will start to rate assault as less serious and less of a crime.
Even among those for whom experiencing violence is not normal, there is little doubt that the context of an assaultive action affects how it is interpreted (Black, 1979;Young, 1988). Codebased explanations of criminal etiquette, enforced at the social group level and experience at the individual level, may provide a useful insight here into understanding victim attitudes and behaviours. Just as a willingness to use violence in certain contexts can increase an individual's social standing in violent subcultures, perhaps stoicism in the face of assault can do the same (Anderson, 1999). Å kerström (2002) has detailed how care workers do not interpret the regular violence they suffer at work as crimes and there is a substantial body of literature on nurses discounting attitudes towards work-related assaults (e.g. Levin et al., 1998;Smith-Pittman and McKoy, 1999). Whether this discounting is out of loyalty to patients, sympathy, pragmatism or an implicit code of conduct remains to be seen. Nevertheless, that this reticence appears to be commonplace suggests that there is a normative attitude towards violence in certain contexts.
This reluctance to label violence as a crime is mirrored in other contexts. Anderson (1999) reported an avoidance of law among deprived communities despite high levels of crime. In these neighbourhoods, social norms ensured that a tendency towards showing 'heart' trumped a tendency towards authority. Ellickson's (1991) study of conflict resolution in a rural community demonstrated that efficient 'informal norms of neighbourliness' (p.viii) can trump legal norms when all members of a group cooperate. Unfortunately, the single-community samples upon which Ellickson and Anderson based their research do not permit generalisation to a wider population. However, it is feasible to test whether the discounting of violence by victims is normalised in a consistent manner across neighbourhoods and related to the level of crime in those neighbourhoods.
It is not clear whether the discounting of legality in these communities was the product of bottom-up interpersonal cooperation or that of top-down normative social pressure. If it was the former, then it provides support for Black's theory of relational distance. If it was the latter, it suggests some form of informal social control within a cluster or community. Both of these positions can be tested empirically.
Using models of crime reporting, this study seeks to identify the factors that predict crime labelling. Previous research has suggested that, for victims, crime is a combination of the impact of the incident on the victim (harm, unjustness and feelings of vulnerability) and the extent to which the incident matches their prototype of a crime (legality and context) and that the decision to label these incidents as a crime is informed by personal experience and social norms. Based on the evidence presented above, which has explored decision making about crime at individual and social levels and considered precipitatory, event-level and consequential factors, the following hypotheses are proposed.

H1:
The likelihood of crime labelling is positively related to the amount of incident harm H2: The likelihood of crime labelling is positively related to the unjustness of the incident H3: The likelihood of crime labelling is positively related to victim feelings of vulnerability H4: The likelihood of crime labelling is proportional to the relational distance between victim and offender(s)

H5:
The likelihood of crime labelling is negatively related to experience of violence H6: The likelihood of crime labelling is clustered within neighbourhoods H7: The likelihood of crime labelling is negatively related to the level of crime and deprivation in an area

Sample
The initial sample was all respondents to the Crime Survey for England and Wales (CSEW) between 2008/09 and 2010/11. The CSEW is a nationally representative survey of English and Welsh respondents' experience of crime in the twelve months preceding completion of the survey. Details of the CSEW sampling procedures are available in Bolling et al. (2007), Nicholas et al. (2007) and Fitzpatrick and Grant (2011).
Inclusion criteria were that survey respondents were aged 16 years or over and living in a household, as opposed, for example, to being homeless or living in a communal establishment, such as a hospital or prison and respondents answered yes to the question: 'has anyone, including people you know well, deliberately hit you with their fists or with a weapon of any sort or kicked you or used force or violence in any other way?'. CSEW data sets for the years 2008/09 to 2010/11 (Home Office Research, Development and Statistics Directorate and BMRB Social Research, 2010 a, b; 2012 a-d) were extracted electronically from the UK Data Service (http://www.ukdataservice.ac.uk/). Lower-level geographical data were obtained under a special license agreement with the UK Data Archive. In this data set, lowerlevel superoutput area (LSOA) codes represented neighbourhoods (Office for National Statistics, nd). The LSOA is the best proxy for neighbourhood captured in currently available administrative datasets, with an average population of 1500 residents.
Three sweeps of the CSEW were pooled to maximise the sample of relatively rare events (violent victimisation). CSEW methodology remained consistent during this period, allowing comparison between yearly data sets. To ensure consistency in the ranking of neighbourhood crime and disorder, only data from England were included. 2 These data were combined to form a working data set of 139,515 respondents describing 3,325 violent victimisations against 2,734 respondents. Following convention, only the respondent's most recent violent victimisation was included in the sample, resulting in a sample of 2,734 nested within 2,465 neighbourhoods.

Measures
Crime labelling. The outcome variable was crime labelling. Participants were asked to rate the extent to which they regarded the violent incident as a crime or not a crime. The variable was derived from a survey item wherein respondents indicated whether they regarded the incident as 'Just something that happens', 'Wrong, but not a crime' and 'A crime'. From this, a binary variable 'A crime' (1)/ 'Not a crime' (0) was created.
Impact. As suggested by Ruback et al. (1984), victim perceptions of the seriousness of a crime are likely to be informed by the incident's harmfulness (injury, weapon use and emotional harm), its perceived injustice (unjustness of the incident, number of offenders, victim initiation of the incident and attribution of reason for the incident) and the victim's subsequent feelings of vulnerability (victim feelings of vulnerability).
Harm. Incident harm was explored from three perspectives: physical harm, emotional harm and potential for harm, represented by injury, emotional harm and weapon present, respectively.
Injury. Whether the incident resulted in an injury to the respondent was coded as a binary variable in response to the question: 'Were you bruised, scratched, cut or injured in any way?'. Affirmative responses were coded as 1.
Emotional harm. Tied to the physical seriousness of an incident is the emotional harm of the incident. Respondents were asked to indicate the extent of any emotional response to the incident. A four-point ordinal scale was created reflecting, in increasing order to harm: 'No emotional reaction' (1); 'Just a little' (2); 'Quite a lot' (3); 'Very much' (4).
Weapon present. The use of a weapon has been cited as one of the key factors that influence the perceived and actual seriousness of an assault (Brennan, 2011;Brennan et al., 2010). Although the mechanism through which this occurs is not firmly understood, it has been proposed that this is due to the potential threat to life that weapon use presents (Weaver and Clum, 1995). A binary variable was created in response to the question, 'Did the person have a weapon?'. Affirmative responses were coded as 1.
Unjustness of the incident. The unjustness of the incident was measured by the number of assailants, whether the victim initiated the violence and whether the victim attributed the reason for the incident to some characteristic of themselves. These variables were selected as they reflect unfairness and discrimination in the characteristics of the incident.
Number of offenders. Respondents indicated how many people were involved in assaulting them. This was categorised as 'one' or 'more than one'.
Victim initiation of the incident. Respondents were asked to indicate whether they 'used force first'. This was a binary, Yes/No response.
Attribution of reason for the incident. In response to the question, 'Can you tell me why you think this incident happened?', twelve options were available, to which respondents could indicate more than one reason. To reflect respondent vulnerability, a binary variable was created and coded as 'Yes' if the respondent reported that the incident happened because of respondent negligence, the respondent was specifically targeted, the respondent was a geographically easy target or the incident was a hate crime. All other explanations for the offence including not knowing why the incident happened were coded as 'No'.
Perceived future vulnerability. Respondents were asked to rate their fear of walking alone in their neighbourhood at night. This measure is widely used as a proxy for fear of victimisation (Gray et al., 2008). Victims can respond along a four-point ordinal scale ranging from 'Very safe' to 'Very unsafe'. Although this is a valid measure of an individual's feelings of vulnerability, its cross-sectional nature limits its use as a measure of change in these feelings in response to victimisation. In addition, if a victim feels vulnerable in response to domestic violence, it is unlikely to be reflected in this variable.
Crime prototypicality. Crime prototypicality was explored through the use of two variables, relational distance (victim-offender relationship) and experience of violence (number of violent victimisations reported).

Relational distance.
Respondents were asked if the violence against them was committed by a person or persons they knew (Yes/No). If the respondent knew any of their assailants, they were asked to categorise their relationship with these individuals. Prior to the analyses, these relationships were grouped as (1) family member/intimate partner, (2) friend/acquaintance or neighbour or (3) other, resulting in an indicator of relational distance in increasing order of relational distance.

Number of violent victimisations reported.
Respondents were asked to recall how many times they had been victims of violence in the past year. This count variable was recoded as a three-part ordinal variable with categorised as 'one', 'two' or 'three or more assaults'.
Neighbourhood clustering effects. Given that likelihood of reporting crime to the police has been shown to be correlated within US neighbourhoods (Baumer, 2002), it is possible that the area in which a person lives may influence their attitudes towards the crimeworthiness of violent victimisation. Such a trend could reflect general attitudes towards the use of formal and informal social control as noted in a range of communities (e.g. Anderson, 1999;Ellickson, 1991). Therefore, the neighbourhood in which a respondent lived was used to cluster respondents and national measures of deprivation and crime within these neighbourhoods were used to determine the effect of exposure to these factors on crime labelling.
Neighbourhoods. Respondents were clustered at the Lower Layer Super-output Area (LSOA), which corresponds to between 400 and 1,200 households (Minimum population 1000 -Maximum population 3000). All LSOAs in England (n ¼ 32,842) were eligible for inclusion in the study.
Neighbourhood deprivation and crime. Neighbourhood characteristics were also included in the model to test the hypothesis that individuals living in higher crime areas would be less likely to label violence as a crime. LSOAs were ranked into deciles by the Office for National Statistics based on a composite of number of reported violent crimes, burglaries, thefts and cases of criminal damage per 1,000 at risk population (Communities and Local Government, 2011). This variable was reversed so that a higher decile reflected greater levels of neighbourhood crime and disorder. Then, neighbourhoods in deciles nine and ten were recategorised as high crime areas.

Statistical methods
Multilevel logistic regression analyses were undertaken to identify predictors of a binary variable, 'crime labelling'. A hierarchical structure was employed because it was theorised that attitudes towards crime may be correlated within neighbourhoods. The use of a multilevel modelling strategy counteracts the effect of clustering on the standard errors of the regression estimates. Failure to account for this clustering increases the risk of Type I errors. Furthermore, it allows the estimation of how much of the variability in crime labelling can be attributed to neighbourhoods and, because different mechanisms may be active in different neighbourhoods, allows the relationship between the independent and dependent variables to vary between neighbourhoods.
Four models were run. Firstly, an intercepts-only model (Model I) was run to estimate the individual and neighbourhood effects without any predictor variables. A single-level logit regression with standard errors adjusted for neighbourhood clustering (Model II). Model III was a multilevel model with individual-level predictor variables constrained as fixed effects. In Model IV, a neighbourhood-level variable, level of neighbourhood crime and disorder, was included as a random effect.
The analyses produced odds ratios and 95% confidence intervals (95% CI) of each variable. In order to describe the effect of neighbourhood on crime labelling, two indicators of interneighbourhood variance were calculated. The variance partition coefficient (VPC) -analogous to the Intraclass Correlation Coefficient -is a percentage estimate of the extent of similarity within neighbourhoods (Merlo et al., 2005) and the Median Odds Ratio (MOR) is an indicator of the median change in likelihood of crime labelling when comparing two randomly selected neighbourhoods.
As is typical of many multilevel models of nationally-representative social surveys, withinneighbourhood sampling was sparse (Bell et al., 2008). As multi-level models aggregate variance at the case and cluster levels (in this case, individuals and neighbourhoods, respectively), it is important to consider the effect that the low levels of clustering might have on variance calculations. Researchers in this context are often forced to decide between the use of smaller cluster units (e.g. LSOAs), which better represent neighbourhoods (Fone et al., 2007) or larger cluster units (e.g. Middle Layer Super Output Areas), which have less sparseness, but are worse proxies for neighbourhood. In this case, the benefit of using MSOAs was negligible, so the more representative indicator of neighbourhood, LSOA, was preferred.
An early rule of thumb has recommended approximately 20 cases per cluster (Hox, 1995) to avoid an inflated risk of Type I errors. However, more recent simulation studies with more sparsely-distributed samples have found low levels of statistical bias, with low number of clusters being more problematic than high sparseness (Bell et al., 2008). Furthermore, the proportion of 'singletons' (i.e. level 2 units with a solitary level 1 unit) in the simulation studies did not affect the estimation of fixed or random effects when the number of clusters exceeded 500. Nevertheless, the sparseness of the data modelled in the current study exceeds those simulated by colleagues (2008, 2010). To be confident that sparseness did not adversely bias the model coefficients and reported variances, alternative models with lower sparseness but smaller sample sizes were also created by limiting the sample to neighbourhoods with at least two respondents. These alternative models were used for confirmatory purposes only and the models reported below were based upon the full sample.
Missing values. Missing values can adversely affect the generalisability of study findings, particularly if data are missing for a systematic reason, e.g. a refusal by respondents to answer intrusive questions. Through casewise deletion, the inclusion of predictor variables resulted in a 4.8% loss of cases. Logistic regression failed to find any predictor of missing cases. Therefore, it was determined that these cases were unlikely to affect the validity of the results.

Results
Of the 2,729 respondents (99.8% of the eligible sample) who described violence against them, 64.8% labelled the incident as a crime. Descriptive statistics for the variables are featured in Table 1, including a measure of association (chi-square test) between the variable and the outcome ('crime labelling'). The results of the four multilevel regression models are described in Table 2. Model I: An 'intercepts only' model was run to determine whether likelihood of labelling violence as a crime was clustered within neighbourhoods. This 'null model' reported a level 2 variance of 0.025 (Standard error 0.11) and a VPC of 0.008, which indicated that neighbourhood explained less than 1% of the variance in the outcome. 3 The median odds ratio for this model was 1.16.
Model II: In the presence of small group size within clusters, the results of the logistic regression analysis are presented to permit examination of the consistency between single-and multilevel models. As the estimates are almost identical across models, the findings of the multilevel model are reported in detail. Model III: A fixed-effects model, controlling for neighbourhood-level effects was run to determine the relationship between the predictor and outcome variables while adjusting standard errors for the effect of clustering within neighbourhoods. In the model, the 'harm' variables, weapon use, injury and emotional damage predicted crime labelling, although the relationship between the outcome and level of emotional harm was non-linear (Hypothesis 1). Use of a weapon approximately increased the likelihood of the incident being labelled a crime by a factor of two. Injury increased the likelihood of the incident being labelled a crime by 71%. There was no difference in likelihood of labelling violence as a crime if the respondent felt that they suffered 'just a little' emotional harm compared to those victims who did not report an emotional reaction to the incident, but respondents who indicated that they were 'quite a lot' or 'very much' affected by the incident were three times and almost seven times more likely to label the violence as a crime, respectively. Labelling was more likely among respondents who reported multiple assailants or attributed the reason for the incident to some characteristic of themselves (Hypothesis 2). Relational distance was also a statistically significant predictor of crime labelling although the relationship was non-linear (Hypothesis 4). Respondents were just as likely to discount violence against them by friends, neighbours and work colleagues as they were to discount violence by family members or intimate partners. However, they were twice as likely to label violence by strangers and lesser-known acquaintances compared to violence by family members or intimate partners. Reporting more than one other violent victimisation was associated with a 20% lower likelihood of labelling violence as a crime (Hypothesis 5). Fear of walking alone at night was not related to labelling (Hypothesis 3). Finally, those living in high-crime neighbourhoods were more likely to label violence as a crime compared to those living in low-crime neighbourhoods (Hypothesis 7).
The introduction of the level 1 predictor variables led to a small decrease in VPC for level-2 effects and the MOR 4 remained constant, indicating that the addition of these variables decreased the ability of the multilevel model to explain between-neighbourhood variation. The fact that these were individual-level variables, suggests that the between-neighbourhood variation in the propensity for labelling violence as a crime is low (Hypothesis 6).
In order to further test Hypothesis 7, neighbourhood crime rates were introduced as a random coefficient in Model IV. The inclusion of neighbourhood crime rates as a random coefficient would allow the model to detect differences in the relationship between the outcome and the predictor variables depending on the level of crime in the areas. However, the inclusion of this variable as a random coefficient did not improve the predictive ability of the model -as evidenced by the reduction in VPC and MOR statistics -and individual-level estimates remained largely constant across models. Therefore, it was concluded that neighbourhood level of crime did not affect the relationship between the outcome and predictor variables.

Accounting for sparseness in clustering
There was an average of 1.1 respondents per neighbourhood. This represents considerable 'sparseness', meaning that in a large number of neighbourhoods, the group mean would be identical to the individual score, thus resulting in a neighbourhood-level variance of 0. Although Bell et al. (2008) have suggested that large numbers of 'singletons' (clusters containing a single case) are unlikely to present a problem for the validity of model estimates, their position is based on simulation studies with less sparseness and fewer clusters. Therefore, an alternative sample was generated that only included cases with more than one respondent per neighbourhood. This alternative sample consisted of 502 cases nested within 234 neighbourhoods (Average 2.1, Min 2, Max 4). Models I-IV were re-run using this alternative sample. Considering the differences between the models alongside the changes in the Model II estimates, it appears that any differences between the estimates using the full and restricted samples were the result of sampling variation rather than bias created through sparseness.

Discussion
This study provides the first exploration of the circumstances under which victims regard violence against them as a crime. Given its exploratory nature, the study took a broad range of theoreticallyinformed individual, contextual and ecological factors into consideration. This study identifies a number of factors that contribute to a victim's decision to label a violent incident against them as a crime and it has identified a number of characteristics of violence that make victims less likely to regard the incident as a crime. The findings of this research have demonstrated that the model provided by Ruback et al. (1984) is a valuable framework for understanding victim attitudinal responses to violence and that Black's (1976) theory of 'morphology' is relevant even when a decision about crime has no real-world consequences. In addition, the study found that neighbourhood effects on crime labelling are minimal, which can be tentatively interpreted to indicate that the social norms that have been shown to affect reporting behaviour (Baumer, 2002) are not reflected in individuals' personally-held beliefs about the criminal nature of violence.

Predictors of the labelling and discounting of violence
As hypothesised, impact played a fundamental role in the decision to label violence as a crime. The findings provide support for the model proposed by Ruback et al. (1984), which extends the concept of harm beyond psychological or physical injury to incorporate the victim's future vulnerability and the 'unjustness' of the incident in the construct of impact. Victims attributing the incident to some characteristic of themselves and the unjustness of being assaulted by more than one person, and the strong relationship between discounting and respondent initiation of the violence, demonstrate that impact should constitute the future as well as past impact of crime. The lack of a statistically significant relationship between fear of crime and crime labelling should not preclude feelings of vulnerability from future analysis of victim responses: the cross-sectional nature of the data set prevented change in fear of crime from being measured while the measurement of fear of crime is notoriously difficult.
Independent of the impact of the incident, it appears that the decision to discount violence was greatly influenced by the relationship between the victim and the offender(s), although this relationship was non-linear, with violence by relatives and acquaintances equally likely to be discounted. Victims were twice as likely to regard violence by a non-acquaintance as a crime compared to violence by a family member or partner. This strengthens the support for Black's (1976) 'morphology' theory that use of law is proportional to relational distance between victim and offender. While on one hand providing support for a component of Black's theory of law, the findings of this study demonstrate the limitations of Black's 'pure sociology' approach to this phenomenon (1993: 159). In this approach, there is no place for individual decision making or psychological factors. However, taken alongside the weak effect of neighbourhood, it is reasonable to surmise that the observed effect of 'morphology' is a bottom-up effect, with individuals making decisions about labelling based on rational decisions about interpersonal cooperation rather being influenced by downward social pressure to avoid formal methods of social control.
In support of the 'normalisation' hypothesis at the individual level, number of reported assaults was associated with discounting violence. At the neighbourhood level, contrary to hypothesis 6, there was only weak evidence of clustering in response to victimisation. Furthermore, neighbourhood crime and disorder -hypothesised to predict crime discounting through the mechanism of promoting anti-authoritarian attitudes -had little impact on discounting.
These findings indicate that decisions about crime are informed by personal and contextspecific factors rather than social or structural processes. Importantly, this finding about attitudes towards violence lies contrary to findings about crime reporting that indicate an effect of neighbourhood on this behaviour (Baumer, 2002). Taken together, it suggests that anti-authoritarian attitudes in areas that tend towards informal use of social control are enforced at the neighbourhood level rather than being engrained in individual attitudes. Therefore, normalisation, at least in terms of responding to crime, is a mechanism that operates at the individual level rather than the social.
As a methodological and cautionary note, the diagnostic tests (see the Appendix) demonstrate that the small group sizes employed in this study did not bias the estimates. However, it is unlikely that all 80% of respondents who were the sole representatives of victims of violence in their neighbourhood were archetypal of those populations. Therefore, the findings relating to neighbourhood effects should be interpreted with caution. As the sample of CSEW respondents increases over time it will become possible to use larger average group sizes wherein groups are more representative of victims of violence in their neighbourhoods. Although it will have its own methodological issues relating to changes in neighbourhood characteristics over long periods of time, the hypotheses of this study could be retested when larger group sizes become available.

Making sense of crime discounting
The observation that 35% of the violent incidents included in the study were not regarded as crimes demonstrates that some violence is acceptable to some people. Therefore, it is important to ask of crime discounting -what is in it for the victims? This can be examined from several different perspectives. By discounting the criminal nature of crime against them, the victim simultaneously expresses resilience, seizes control over the incident and prevents state involvement in their victimisation.
By discounting violence, victims exercise their ability to ignore the law while also contributing to the phenomenology of crime and legality. Robinson (2000) has asserted that 'the extent of criminal law's moral authority determines the extent of its ability to shape community norms and to influence people's conduct through normative forces' (p.1840). Therefore, by refusing to acknowledge violence as a crime, the victim undermines the credibility of law as a moral code. In their study of 'everyday law', Ewick and Silbey noted that 'every time a person interprets some event in terms of legal concepts or terminology -whether to applaud or to criticize, whether to appropriate or to resist -legality is produced (1998: 45)'. Therefore, the discounting of violence by victims is an example of individuals exercising and constructing the law that exists in their everyday lives.
The normalisation process described above is just one approach to understanding how victims manage their experience of violence. It is quite likely that, rather than being normatively-informed, crime labelling and discounting may have instrumental purposes. While discounting can be seen as an attempt to exercise self-presentation and autonomy, it could also be viewed as a post hoc attempt to protect one's self. An alternative explanation of the discounting phenomenon is that the respondent did indeed regard the incident as a crime, but employed some cognitive technique to neutralise it. Reasons for this could include a desire to avoid a 'victim' identity or to avoid facing up to an unjust, unpredictable world (Agnew, 1985). Taylor et al. (1983) have suggested that a victim, in an effort to protect the ego from the unpleasant condition of victimhood, may re-evaluate and minimise the seriousness or significance of the victimisation. Adapting Sykes and Matza's (1957) analysis of delinquents' rationalising of their criminal behaviour, Agnew (1985) collated evidence that victims can activate a range of neutralising techniques in order to discount and reconceptualise the effect and meaning of their victimisation. Such techniques have been observed by researchers studying domestic abuse (Mills, 1985) and sexual violence (Frazier and Burnett, 1994;Weiss, 2011). However, they do not appear to have been investigated in terms of more common interpersonal violence, nor does there appear to have been an investigation into the influence of setting or context on these outcomes. Unfortunately, it was not possible to explore this neutralisation phenomenon further using the existing survey questions and it remains a potentially valuable avenue of exploration. It should be remembered that the processes of normalisation and neutralisation are not mutually exclusive and, so, how these two processes interact could be a valuable area of research on victim responses to violence.

Study limitations
This study has a number of limitations that must be considered when interpreting the results. Firstly, the components of the incident, including harm, morphology and neighbourhood clustering were measured and analysed through proxy measures, which limits their validity. For example, the variable that referred to victim perceptions of vulnerability may reflect generalised social insecurity rather than individual feelings of fear, which might explain why it was not a statistically significant predictor of the outcome. The weak validity of variables measuring victim perceptions of a phenomenon is a constant limitation of victimological research that seeks to test theoretical explanations of victim behaviour. Future studies could explore these findings qualitatively, generating greater insight into the motivations and explanations that victims give for their labelling and discounting behaviour. Encouragingly, the quantitatively-informed findings of this study are consistent with more in-depth, qualitative work with small samples of victims, thus triangulating and strengthening the validity of the findings. Secondly, the context in which the data were collected may have some relevance. Respondents will have been aware that they were responding to questions as part of a 'crime survey'. This may have primed them to be more crime-focused in their responses, leading some to re-evaluate their beliefs about the incident in a more authoritarian way. Thirdly, the term 'victim' is not clearly defined. In this context, the 'victim' is simply someone who experienced violence and was asked to complete a survey: whether they had some culpability for its occurrence was not considered. The findings indicate that respondent initiation of the incident was a statistically significant predictor of discounting, suggesting that the antecedents of the incident play an important role in the construction of crime. Although Ellickson's (1991) and Anderson's (1999) studies focused on tactics for conflict resolution, the current study focuses on conflict discounting and does not seek to explore how or whether the respondents acted in response to the violence. In that sense, the present analysis is somewhat unsophisticated as it removes much of the precipitatory and residual context of the incident. As Black (1983) has noted, much violence is an attempt to regain or exercise control and so, to be fully understood, violent incidents should be examined in the context of their antecedents and sequelae. Importantly, the present study does not aim to study the way in which the incident was resolved. The resolution of the incident, if the conflict was resolved at all in the eyes of the respondent, may have influenced their later labelling of the incident. The survey, which only collects data on official resolution of the incident -which will inevitably covary with labelling -does not permit the exploration of the effect of informal conflict resolution on crime labelling. In the language of Pigliaru (1970), the focus is on the omerta, but not the vendetta.

Conclusion
The identification of factors that predict the discounting of violence as criminal demonstrates that victims, independent of the harm caused, consider a number of contextual factors before they decide that something is criminal. Hulsman (1986), Young (1988) and Hough and Mayhew (1985) -each approaching the phenomenon of victimhood from different perspectives -have all noted inconsistencies in response to victimisation. However, this is the first time that such predictors have been identified while controlling for the level of harm associated with the incident.
It is inevitable that not all violence is a crime and to argue that all violence should be regarded as such would be naïve. However, this study has demonstrated the potential for many victims to discount serious offences against them and serves to emphasise that many victims are unaware of their legal position or are unwilling or unable to seek help. A corollary of this is that there is an even greater number of unidentified victims of crime than previously estimated. Whether the mechanism for this is normalisation, neutralisation or otherwise is not as important as the need to ensure that victims are aware of their right to protection from violence and are aware of the services available to ensure their wellbeing. The statistical analyses used have allowed the identification of risk factors while controlling for the direct harm suffered. Therefore, the study shows that, regardless of harm inflicted, victims still do not label violence as a crime when it is caused by someone they know well. This suggests a problem that is even more deeply-engrained than the behavioural issue of not reporting this violence to the police as it compounds a behavioural problem with an attitudinal one. Therefore, organisations who work with victims should not just focus on changing victim behaviour, such as facilitating reporting, but they should also focus on changing the mindset of victims and the public. This paper has not sought to postulate about the legitimacy or rightness of labelling or discounting crime; it has merely been an attempt to understand the factors that influence how legality is constructed in the minds of victims and how it is produced between actors. The impact of the incident on the victim is the most important factor in crime labelling and is composed of at least the harm experienced and the unjustness of the incident. It remains to be seen whether the effect of the incident on individual feelings of vulnerability contributes to this construct. In addition, the extent to which something is regarded as a crime is depleted by repeated exposure and familiarity with the perpetrator. Finally, it is evident that victims interpret the 'crimeworthiness' of an incident at an individual level that is largely independent of social pressures or the normalising effects of exposure to crime at a neighbourhood level.
This study has shown that people respond to victimisation in systematically biased ways. Illuminating the contextual factors that interact with harm and impact to construct 'crime' in the mind of a victim can help explain implicit rules about the use of social control and the acceptability of harm among the general population. Furthermore, at a more practical level, identifying variations in the construction of crime can provide insight into such diverse criminological issues as police ascertainment of crime, predictors of repeat victimisation, and the democratic processes that influence changes in criminal law and allow us to recognise a population of victims beyond those who do not report to the police. Further research, particularly into the effect of specific 'lawless' situations, such as bar-rooms or war zones, may shed more light on the contextual aspects of victims' complex responses to violence.