Measures of Violence within the United Kingdom Household Longitudinal Survey and the Crime Survey for England and Wales: An Empirical Assessment

Abstract

Criminology has been hampered by a lack of longitudinal data to examine the consequences of victimisation. However, recently, ‘Understanding Society’, the United Kingdom Household Panel Survey (UKHLS), began fielding a small battery of questions relating to violence experience. Here, we examined the strengths and weaknesses of these UKHLS measures with similar indices from the Crime Survey for England and Wales (CSEW), a widely used and regarded but cross-sectional survey. We empirically assessed the extent to which the UKHLS variables are comparable with those in the CSEW to determine the viability of the UKHLS for the longitudinal study of (fear of) violence and its consequences. Overall, we regarded the UKHLS to provide an important resource for future panel research on the consequences of victimisation. We found the indicators measuring physical assault to be similar in both sets of data, but also noted differences in prevalence and/or different distributions by socioeconomic group for the indices relating to being threatened and of feeling unsafe. Nonetheless, we maintain their utility for researchers in this field, allowing researchers to uncover new inequalities in violence exposure.

1. Introduction

Government expenditure on crime and violence prevention justifiably requires a robust evidence base not only to ensure accurate measures of violence occurrence and its effects, but also, crucially, to monitor the costs of violence and the effectiveness of interventions that have sought to decrease it. This demands analyses of high-quality data; however, there has been little scholarship in the UK to date into violence and its effects using large-scale representative panel data. To facilitate such analyses, we examined the strengths and weaknesses of two highly regarded and widely used datasets, one of which is primarily used by criminologists and is cross-sectional in nature; the other, which is a panel, has only recently begun to field items on violence experience of all its respondents. Our paper contributes to the field in its empirical assessment of the extent to which a series of indicators measuring violence experience and fear of violence in Understanding Society, the United Kingdom Household Panel Survey (UKHLS), are comparable with equivalent and long-standing measures from the annual Crime Survey for England and Wales (CSEW, previously known as the British Crime Survey).

While the UKHLS was not primarily designed to measure and support the study of violence and its effects, and is consequentially rarely used by criminologists, it has recently extended a battery of questions originally intended for its ethnic minority booster sub-sample which measure discrimination and harassment to all adult interviewees. These new variables, which index components of violence experience, have the potential to support significant developments in criminological research given the breadth of variables the UKHLS collects alongside its panel data structure. The CSEW, on the other hand, is already widely used in both policy and academic circles amongst those interested in crime prevalence/incidence as well as trends (Allen and Harding 2021; Heeks et al. 2018; Office for National Statistics 2023; Oliver et al. 2019); yet, its annual cross-sectional nature does not allow for assessments of causal dynamics or of the possible long-term effects of criminal actions on its victims.

Below, we compared and contrasted the UKHLS with the CSEW to uncover overlapping operationalisation and measurement of these data, with the aim to assess the quality and usability of the UKHLS for future victimisation research. Our analyses will therefore allow researchers to determine the utility of the UKHLS as a much-needed source of longitudinal data for criminological research, which, crucially, allows assessments of the causal ordering of events that the CSEW currently does not support. Though there have been attempts to introduce a panel component to the CSEW, these appear to be under-review given concerns that self-completion modes would risk data quality and may also result in a structural break in the time series of measures of crime (Hamlyn et al. 2018). We therefore investigated the UKHLS as one of the few sources of panel data with indicators of use for criminological research. While the UKHLS offers a limited number of indicators of violence, it does offer a broader range of covariates than is true of the CSEW, allowing for assessments of possible cross-over effects of violence on different life domains, including assessments of its effects on labour market dynamics (e.g., Gash and Blom 2023), socioeconomic positioning, family circumstances, and health. Here, we hope to open up the field of criminology through an assessment of the quality of the UKHLS violence victimisation measures, allowing fellow researchers to determine the utility of the UKHLS for their own research agendas.

Panel Data Structures in the Social Sciences

Nationally representative panel data conducted via the ‘gold-standard’ face-to-face interview allows social scientists to conduct sophisticated analyses, which, in principle, allow for an assessment of causal ordering. Without this longitudinal component, we cannot disentangle correlation from causation, nor can we examine the long-term effects of significant life events, such as exposure to violence on personal and familial trajectories. These data, however, require considerable financial investment and expertise to set up and maintain. Consequentially, there are not many countries which provide panel data of such scope. One of the first nationally representative panel data was set up in the United States of America in 1968, the Panel Study of Income Dynamics, followed by Germany in 1984, with the UK following suit in 1991, with the British Household Panel Survey, which has since been combined into the larger UKHLS (see Turek et al. (2021) for an overview of these data as well as efforts to offer tailored, harmonised panel data for analyses).

Yet, the substantive focus of most panel data tends to include few indicators of interest to criminologists. The field of criminology has, nonetheless, benefited from some panel data, though the majority of these are US based. For example, the National Crime Victimisation Survey (NCVS) (Thompson and Tapp 2022), the National Longitudinal Surveys of Youth (NLSY) (Bares et al. 2021), and the National Longitudinal Study of Adolescent to Adult Health (Add Health) (Landeis et al. 2021) have allowed researchers to demonstrate the contribution longitudinal data can make to the study of violence and its consequences (e.g., Rezey 2020; Sabia et al. 2013). Yet, even these data are limited in their scope: the NLSY and Add Health are not nationally representative of the entire US population, and while the NCVS is nationally representative, it has a three-year rotating panel design, and so suffers from a comparatively short timeframe compared to other panel studies. Similarly, in the UK, other than the UKHLS, there is currently no longitudinal data on violence victimisation which is nationally representative. This is a concern, as criminology is therefore limited by its poor access to data that allow for assessments of the possible long-term effects of violence and fear of violence on the populace. The UKHLS remains, therefore, the only data currently available which allows researchers to examine the possible long-term effects and consequences of violence in the UK. Moreover, while the variables we examined here have been successfully used by others to examine racial discrimination, including associations between harassment and mental health (e.g., Nandi et al. 2020; Nandi and Luthra 2021; Wallace et al. 2016), we are not aware of anyone to date who has used these variables to explicitly measure violence prevalence or fear of violence for the general population.

2. Literature on Measuring Violence and Violent Crime

2.1. Measurement, Conceptualisation, and Item Design

There is little agreement on how best to conceptualise violence, and therefore how best to define and measure it. This lack of consistency is further magnified by the divergent institutional orientations towards it. Healthcare workers respond to the physical and psychological effects of violent actions, and the sector recognises violence as a public health issue (Krug et al. 2002). The judiciary and police force address violence in so far as it is criminal. Amongst NGOs, in the third sector, different organisations tend to be targeted to specific expressions of interpersonal violence, e.g., women-only domestic abuse services or services for teenage victims of violent crime. Indeed, apart from the CSEW, which is widely referenced within the criminal justice system as well as various third sector groups, many sectors tend to use their own administrative data targeted to their own purposes with consequent silos in orientation to the conceptualisation and measurement of violence.

This lack of consistency extends to governance structures. Looking at the United Nations, the World Health Organisation, and the Council of Europe’s definitions of violence, Blom et al. (2023) sought to develop a new conceptual framework of violence to allow for a consistent measurement of violence between these groups. They noted a need for a broadened conceptualisation of violence and advised the inclusion of non-criminal acts in their definition, which can have significantly negative effects on victims; for example, coercive control has only been a crime in the UK since 2015 (McMahon and McGorrery 2020) even though it has large mental health consequences (Lohmann et al. 2023). The authors further advised better integration of non-physical acts in the conceptualisation of violence, alongside a need to establish agreed thresholds for non-physical acts and harms. Finally, they underscored the utility of repeat measures and of longitudinal data to better understand and cater for repeat victimisation and its possible long-term effects. Skafida et al. (2023) similarly offered a series of recommendations for an improved measurement and conceptualisation of intimate partner violence (IPV). They systematically assessed the item design of IPV within seven different UK-based questionnaires, including the CSEW analysed here. They noted an overemphasis on measures of physical abuse, which coincides with Blom et al.’s (2023) call for a broadened conceptualisation to include non-physical harms. Skafida et al. (2023) also noted the need for better longitudinal data, as well as the related need to examine patterns of abuse, rather than the current focus on incidents, alongside a need to determine the long-term impacts of abuse on victims. Nandi and Luthra (2021), in their investigation into ethnic minorities’ exposure to racial and ethnic harassment, argued that the UKHLS offers a more comprehensive measure of racist harassment than those which rely on police reports, with the UKHLS catering for a multifaceted assessment of psychological and physical fears of violence and harassment under multiple different contexts. They also noted that the UKHLS reveals higher rates of racist harassment than the CSEW allows for as a result of the questionnaire routing. In combination, these studies suggest the need for a broader conceptualisation of violence than is frequently adopted in criminological research. They also suggested the need to include non-criminal acts in item design and an improved assessment of the psychological effects of violence.

2.2. Representation and Mode of Data Collection

In addition to the lack of agreement on violence conceptualisation, there are methodological concerns on how best to collect indicators of violence. Estimates derived from representative survey data are generally regarded to be the best means of accessing precise indicators of social phenomena, including violence prevalence/incidence and its effects. Statisticians and survey methodologists ensure an optimal overlap between estimates derived from survey data and those from the ‘true’ population through a minimisation of survey error. A comprehensive review of different sources of error, based on earlier work (e.g., Andersen et al. 1979; Deming 1944), is provided in Groves et al.’s (2004) total survey error (TSE) framework. Broadly, survey error is grouped into two areas: failures of representation (poorly representing the intended sample through poor sample design or non-random missingness) and failures of measurement (such as poor item design or mismatch between question type and mode of data collection). Therefore, survey design can have significant implications for data quality. For example, Hamlyn et al. (2018) recently examined the implications of changing the mode of data collection of the CSEW from computer-assisted personal interviews (CAPIs), with skilled interviewers collecting the majority of data for the CSEW via face-to-face interviews in people’s homes, to data primarily collected via online self-completion. The authors concluded that it would not be possible to change the mode of data collection without serious adverse effects on the time series of measures of crime. In their report, the authors noted the often-forgotten role of skilled interviewers for the maintenance of data quality, whose activity minimises survey errors which can more easily arise in self-completion modes. CSEW interviewers were found to work outside the questionnaire script, offering frequent clarifications to respondent queries regarding correct interpretation of the questions posed. Interviewers were also given the flexibility to fill in respondents’ answers to questions which may have already been asked or divulged at earlier points during the interview. This was done to avoid irritating the interviewees with unnecessary repetition, which can have adverse effects on item responses and can also lead to satisficing. Face-to-face interviewers therefore played a key role in improving accuracy in reporting in what is known to be a long and complex questionnaire.

While there are mixed findings regarding the correct mode of data collection for sensitive research questions, the assumption in the field is that self-completion is preferable to CAPIs for modules where face-to-face interviews may be embarrassing or awkward for interviewees. For example, disclosure of domestic violence was found to be higher via the self-completion module of the CSEW than via indicators collected from its face-to-face component (Walby et al. 2014). Yet, in a recent empirical analysis of variance in the measurement of sexual victimisation in four different surveys in the UK, Brunton-Smith et al. (2022) noted that the type of survey, rather than its mode (here, they examined face-to-face as opposed to self-completion), appeared to be key to eliciting different prevalence rates. They found that surveys targeted to the measurement of violence against women and/or sexual behaviours and health reported a higher prevalence than traditional crime surveys. Similarly, research by Walby and Myhill (2001) found a tendency for the under-reporting of crime in dedicated crime surveys, especially in contexts where victims of crime may be unaware and/or unwilling to recognise the criminal status of harms they have experienced. In what follows, we also aimed to contribute to this body of work which reviews the empirical utility of different sources of data for the systematic measurement of violence and its effects.

3. The Data—Comparing the UKHLS and the CSEW

Below, we reviewed both datasets along with variations in the questionnaire wording. Both the UKHLS and CSEW have large sample sizes, are representative of national populations (with the CSEW only for England and Wales), and have a broad and international user base. The CSEW is currently managed by the Office for National Statistics (ONS), though previously it was conducted for the Home Office, with its focus on prevalence and trends in personal and household crime, as well as public perceptions of crime and judicial systems, key for judicial policy and monitoring purposes.

3.1. The UKHLS

We used wave 11 of the UKHLS, which was fielded between 2019 and 2021. The UKHLS is a nationally representative panel survey of households in the UK. The sample consists of the general population sample, which was clustered and stratified for Great Britain, though a simpler random sample was drawn for Northern Ireland. The data were supplemented with two targeted booster samples for ethnic minority and immigrant groups to increase the sample size for smaller populations of exceptional interest. All adults aged 16+ in sampled households have been interviewed annually since 2009, allowing for analyses of change on both the individual and household level. These data collected a range of objective and subjective indicators on a broad range of subjects covering health and wellbeing, working-life, education, as well as detailed indicators on labour and non-labour income. The first wave of the panel sampled approximately 40,000 households, collecting data via computer-assisted personal interviews (CAPIs) in respondents’ homes. The sample size was maintained via outreach actions with respondents, including between wave contacts and postal letters sent to respondents’ homes notifying them of the eminent start of data collection. While the survey was largely collected via CAPIs, there were also some self-completion components (including modules on general health, neighbourhood social capital, and partner relationship quality).

The violence variables, discussed in detail below, were fielded in waves 1, 3, 5, 7, 9, and 11. The questions were originally only posed to ethnic minority booster respondents as well as a small comparator reference group from the general population sample, with the questions originally operationalised to measure harassment and discrimination incidents including violence (Wallace et al. 2016). In wave 11, which we used here, the variables relating to violence and fear of violence were, for the first time, collected from all adult respondents, increasing the sample significantly.

3.2. The CSEW

We used the CSEW that was fielded in 2019/2020. The CSEW seeks to determine respondent attitudes, perceptions, and experiences of crime in the 12 months prior to interview from a representative sample of households in England and Wales. This survey started in 1982 and was collected biennially until 2001 when it became an annual survey (Kantar 2020), allowing for trend analyses. It was collected on behalf of the UK’s Home Office before it was transferred to the UK ONS in 2012 (Kantar 2020). Like the UKHLS, the CSEW is collected via the gold standard face-to-face interviews by expert interviewers, with data inputted via the CAPIs by both the interviewer and via self-completion for sections of the questionnaire where disclosure is likely to be enhanced through self-completion (such as respondent exposure to domestic violence, sexual victimisation, stalking, and substance use) (Kantar 2020). The CSEW has been recognised as an important supplement and counter balance to police data on prevalence, given its broader scope and the fact that not all crimes are reported to the police, which may lead to differing trends in violence (e.g., shown for the US by Lauritsen et al. 2016) and spatial variation (Buil-Gil et al. 2021).

3.3. Questionnaire Wording—The UKHLS

There is a good conceptual overlap for three of the four violence indicators in the UKHLS with CSEW indicators, and below we detailed their face validity through a comparative assessment of questionnaire wording. Full wording of each UKHLS indicator alongside the wording of equivalent variables within the CSEW is presented in Table S1 of the Supplementary Materials (and the full UKHLS questionnaire sequence of this module is provided in Section S2 of the Supplementary Materials). The first indicator, relating to fear of violence, measures whether respondents feel ‘unsafe.’ The UKHLS asks respondents if ‘in the last 12 months, you [have] felt unsafe in any of these places?’, with respondents shown a list of eleven locations/places in both the public and private sphere, including in one’s workplace or university. Respondents can answer positively to any of the listed locations. UKHLS respondents who respond positively are then asked to offer a ‘reason’ for the violence, ‘did you feel unsafe for any of these reasons? If so, which ones?’, with a showcard presented to respondents with nine sociodemographic characteristics, including respondent sex, age, ethnicity, sexual orientation, and health or disability (please refer to Section S2 of the Supplementary Materials for the full list). Next, respondents are asked if they have avoided going to or being in any of the places listed on the card, for which we found no CSEW equivalent, and which was therefore excluded from further analyses.

The UKHLS collects an additional two variables on violence prevalence/experience. The first combines non-criminal behaviours with criminal behaviours and asks ‘have you been insulted, called names, threatened or shouted at’, with threatening someone a criminal act, whereas the other behaviours, whilst potentially abusive, are not necessarily illegal.1 The final question in this battery has the greatest conceptual cross-over with the CSEW. It asks: ‘In the last 12 months, have you been physically attacked in any of the places listed’. The questionnaire routing is the same for each of the four variables, with the number of locations and the ‘reasons’ behind the incident broached each time. Here, it should be noted that given the length of the sequence, there may be some satisficing in responses and indeed learning on the sequence of the follow-up questions, with greater drop-off and non-response as the more problematic, and also criminal, forms of violence are collected.

3.4. Questionnaire Wording—The CSEW

As will be demonstrated below, the measures of violence operationalised in the CSEW are more clearly targeted to the measurement of crime than is the case with the UKHLS (see also Table S1 of the Supplementary Materials). The closest comparator in the CSEW to the UKHLS’s question on feeling unsafe asks respondents if they ‘feel unsafe walking in their neighbourhood (15 min’ walk from here) after dark’ (see Table S1 of the Supplementary Materials for full question wording). This battery is also asked to a subsample of respondents to avoid overburdening the sample. Responses are collected using a Likert scale, ranging from 1 (very safe) to 4 (very unsafe). For our prevalence estimates, we combined the categories ‘very unsafe’ with ‘a bit unsafe’ as that appeared to offer the best conceptual overlap to the wording of the equivalent variable in the UKHLS.

The CSEW’s collection of indicators on respondent exposure to threats and physical crimes are obtained after a series of screener questions. These questions seek to determine whether respondents experienced various crimes in the last twelve months, with respondents’ description of their experiences then collated in a series of ‘victim forms.’ The battery sequencing starts with questions relating to property crime, then moves to physical violence, followed by threats, sexual violence, or abuse, and ends with questions relating to violence in the household. If respondents report being victims of any of these crimes, a victim form, with a maximum of six per respondent, is filled in for each criminal offence. In the victim form, respondents provide (using CAPIs) their own detailed accounts of each incident alongside other key questions relating to each incident. After the interview, trained coders assess whether what has been reported represents a crime and, if so, which one. Coding hierarchies ensure that when multiple offences happen at the same incident the most serious offence is coded (but see also Pullerits and Phoenix (2023) for the implications for measures of violence). Offenses outside the scope of the survey are excluded, such as if they occurred more than 12 months prior to the interview or outside England and Wales.

The CSEW indicator we used to examine the prevalence of threats was obtained from the screening process described above, with the full wording available in Table S1 of the Supplementary Materials. Respondents who confirmed that they had been threatened then went on to provide descriptions of the ‘threat incident’, which is then subsequently categorised into offence categories post-interview. Inevitably, there is an imperfect overlap between the indictors measuring threats derived from the screener question, where the respondents themselves report their view that they have been threatened, and the criminal offence category. Of those who reported that they had been threatened in the screener question, about 27.3 percent did not have their incidents subsequently classified as a threat offence by the professional coders; however, of this 27 percent, about half of these were coded as the more serious crime of physical violence due to priority coding (see also Section S3 of the Supplementary Materials). For both indicators of being threatened, derived from the screener questions and the post-interview categorisation of reported incidents into threat offences, we can expect the CSEW measures to differ significantly from those in the UKHLS. The UKHLS indicator is based on a question which is worded quite differently; respondents are not prompted with multiple examples of various forms of threatened action and, perhaps, most crucially, the UKHLS indicator includes ‘being called names, and insults and being shouted at’ in their indicator, which are not, per se, covered by the CSEW. Thus, we can expect the CSEW to record a lower prevalence of ‘threats’ than the UKHLS indicator, as the UKHLS indicator includes a broader range of threatening behaviours.

Lastly, the screener question for physical violence asks “…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?” (see also Table S1 of the Supplementary Materials). For the violent crime code, we included the offences serious wounding, other wounding, common assault, attempted assault, serious wounding with sexual motive, and other wounding with sexual motive, robbery, and attempted robbery. Again, there was an imperfect overlap between the screener question of physical violence and the recording of a violent offence, but here this was mostly due to respondents misclassifying violent offenses as the lesser ‘threat offence’, with 25% percent of offences coded as physical violence coming from a re-classification of incidents from the threat screener question into (mainly) common assault or attempted assault (see also Section S3 of the Supplementary Materials). This was carried out by professional coders based on the narrative provided by the respondent. Thus, the CSEW physical violence screener question (which is similar to the UKHLS question on physical violence) disproportionally picks up more severe instances of physical violence.

To conclude, both datasets are statistically representative surveys sampled using multistratified sampling techniques. Both are collected using CAPIs in people’s homes by skilled interviewers, which should improve item response and data quality. The CSEW differs notably in its provision of a self-completion component regarding sexual assault and domestic abuse to maximise disclosure on sensitive subjects. However, we did not use variables collected via self-completion in our analyses here. While differentials in estimates due to representation should be minimised in our comparison, with both datasets adopting similar sampling strategies alongside similar modes of data collection, we can expect differentials in measurement as the wording and the routing of the questionnaires differ in important ways, especially relating to the measurements of ‘feeling unsafe’, and also of ‘being threatened’. Conversely, as the wording relating to ‘being physically attacked’ is similar (comparing the UKHLS and the screener question for physical violence in the CSEW), we can expect both datasets to offer similar estimates of ‘being attacked.’ Nonetheless, it is possible that the CSEW screener question and the UKHLS question on being attacked overlook some ‘less severe’ physical violence instances, as indicated by the non-perfect overlap between the offence categorisation and the screener question. Potentially, there are also risks that the CSEW under reports less significant violent and non-criminal behaviours, with respondents perhaps sensing that their experiences do not warrant collection in a survey on crime (e.g., Brunton-Smith et al. 2022). There is also the possibility that the UKHLS’s questionnaire routing, which prompts multiple different contexts where unwanted behaviours may have taken place, may illicit a higher prevalence.

Overall, an assessment of the questionnaire wordings suggests that while there is some good conceptual overlap, and from a comparative perspective the face validity is relatively sound, we expect differences in the prevalence of feeling unsafe and of being threatened, though the prevalence of being attacked should be consistent in both sets of data. Here, it will also be interesting to examine the variance in the predictors of these variables, which may offer insights into the different purchase of different wordings and survey types with different socioeconomic and demographic groups.

3.5. Control Variables and Analytic Strategy

We empirically tested the comparability of the three UKHLS and CSEW indicators, measuring feeling unsafe, threating behaviours, and physical violence through analyses of their prevalence (in Table 1) alongside their association with key sociodemographic and socioeconomic variables (Figure 1 and Figure 2). Our key controls were sex/gender, which is measured as a binary variable and reflects whether people are female or male. We categorised respondents’ age according those who are aged between 16–24, 25–34, 35–44, 45–54, and 55 years and older. For ethnicity, we used the ONS categorisation of White (British), Asian (British), Black (British), mixed or multiple ethnicities, and other. Relationship status at the time of the interview was categorised as married or cohabiting, single or widowed, and separated or divorced. We included a binary control for those who have any limiting physical disability alongside the number of dependents in the home, distinguishing between those who have none, one, or two or more. We controlled for socioeconomic differences using a series of variables, including annual gross household income. Here, we were limited to the CSEW’s questionnaire wording, which collects household income categorically. We therefore coded the UKHLS continuous measure into the six categories the CSEW applies: less than or up to £10,400, £10,400–£20.800, £20.800–£31.200, £31.200–£41.600, £41,600–£52,000, and over £52,000. Housing tenure was measured as homeowner, renter, or other; lastly, we included a binary control for benefit receipt.

We further assessed the construct validity of the violence indices through analyses of their associations with self-assessed general health. In the UKHLS, respondents are asked “in general, would you say your health is: excellent, very good, good, fair, or poor”. CSEW respondents are asked “How is your health in general? Would you say it is: very good, good, fair, bad, and very bad”. To harmonise the variables, we recoded them into three categories: ‘good’ (including excellent, very good, and good), ‘fair’, and ‘poor/bad’ (including poor, bad, and very bad). This categorisation led to quite similar estimates for the two datasets, with 6.7 and 5.8 percent reporting poor health, and 17.6 and 16.7 percent reporting fair health in the CSEW and UKHLS, respectively.

The data for the UKHLS and CSEW were weighted to account for the complex survey designs and to correctly reflect the sampled population. As the scope of the CSEW only includes England and Wales, while the UKHLS also includes Scotland and Northern Ireland, we excluded observations from Scotland and Northern Ireland for the purposes of this comparison. After listwise deletion of missing values, the CSEW was composed of 27,922 respondents, and the UKHLS had 24,142 respondents.

We compare prevalence of the violence and feeling unsafe in the datasets in Table 1. We employed logistic regression to assess the associations between key sociodemographic indicators and the violence indices and also to examine how these associations may differ between the UKHLS and the CSEW. For the CSEW estimates, we focused on the offence categorisation variant, as that is the most widely used by academic and policy research. We used the average marginal effects displayed in Figure 1 (based on estimates in Section S4 of the Supplementary Materials; Table S4). Figure 2 underscores the statistically significant differences in average marginal effects in Figure 1 between the two datasets based on pooled analyses, where every variable is interacted with a dataset indicator. Finally, associations between the violence indices with self-rated general health were estimated using ordinal logistic regression to test for comparative construct validity (presented in Table 2).

4. Finding

4.1. Prevalence of Fear of Violence and of Violence Exposure

Table 1 presents weighted estimates of the UKHLS violence prevalence indicators with the CSEW equivalent indices by sex. Despite notable variances in the questionnaire wording, we found a similar prevalence of feeling unsafe in both sets of data, with 27% of UKHLS respondents and 22% of CSEW respondents reporting feeling unsafe. The estimate using the CSEW data includes both those who claim to feel both very and a bit unsafe. Interestingly, while both sets of data found that women were more likely to report feeling unsafe (at about 30–34%), the proportions of men who express this fear was substantially higher for the UKHLS (21%) than for the CSEW (13%). This may be due to variances in the questionnaire wording, with men less likely to report feeling unsafe in their neighbourhoods and at night-time (which is how the item is worded in the CSEW) and more likely to report feeling unsafe when the range of locations is broader (as is the case in the UKHLS).

Table 1. A comparative analysis of the prevalence of violence in the UKHLS and the CSEW via sex/gender.

	The UKHLS (N = 23,883)			The CSEW (N = 27,922/6490 for Unsafe)
	Women	Men	All	Women	Men	All
Fear of violence; feeling unsafe
Unsafe	33.51	20.82	27.48	30.44	12.91	21.66
Violence exposure; was insulted/threatened (0–1)
Threatened/insulted	14.81	13.36	14.12
Threat offence				2.99	2.43	2.71
Threat screener questions				4.04	3.27	3.66
Violence exposure; was attacked (0–1)
Attacked	2.27	2.32	2.29
Violent offence				1.66	2.51	2.08
Attacked screener questions				1.22	1.78	1.50

Source: UKHLS wave 11 and CSEW 2019/2020.

Table 1. A comparative analysis of the prevalence of violence in the UKHLS and the CSEW via sex/gender.

	The UKHLS (N = 23,883)			The CSEW (N = 27,922/6490 for Unsafe)
	Women	Men	All	Women	Men	All
Fear of violence; feeling unsafe
Unsafe	33.51	20.82	27.48	30.44	12.91	21.66
Violence exposure; was insulted/threatened (0–1)
Threatened/insulted	14.81	13.36	14.12
Threat offence				2.99	2.43	2.71
Threat screener questions				4.04	3.27	3.66
Violence exposure; was attacked (0–1)
Attacked	2.27	2.32	2.29
Violent offence				1.66	2.51	2.08
Attacked screener questions				1.22	1.78	1.50

Source: UKHLS wave 11 and CSEW 2019/2020.

We had anticipated a poor overlap in the measurement of being threatened/insulted, given large differences in the wording in both sets of data, and indeed found the prevalence of being threatened to be considerably higher using the UKHLS’s multi-component variable, which captures respondent experiences of being threatened, called names, insulted, and/or shouted at. Using the UKHLS, we found that 14% of the population experienced threats, insults, and/or were shouted at, while the CSEW comparator suggested that only 3–4% of the population have been threatened. The CSEW estimate holds for two different operationalisations of the variable using both the screener question and the offence classification. Once again, we found that in both sets of data women reported a higher prevalence of being insulted and/or threatened, though the differential was much smaller with this variable.

Finally, as hypothesised, we found consistency in the estimand for being physically attacked, which has the highest amount of conceptual overlap in its wording. Using both the UKHLS and the CSEW, we found that about 2% of the population have been the victim of a physical attack in the past year, with the UKHLS having a slightly higher prevalence than both the screener and offence classification variables of the CSEW. In both datasets, there was again consistency in tendency, with men reporting slighter higher rates in both datasets, though this difference was somewhat larger in the CSEW. Note that the higher prevalence of the CSEW violent offence indicator compared to the attacked screener indicator was due to the screener question being more likely to record severe physical violence at the cost of recording attempted and common assault.

4.2. Inequalities in Violence Exposure

Next, we analysed patterns of inequality in violence experience and determined whether these differ for different groups in both sets of data. Figure 1 presents predicted probabilities from a series of models to allow for a visual assessment of the variance in the predicted probabilities of the risk of violence using average marginal effects (AMEs) as well as variance in the key predictors of violence in both sets of data. This figure was based on a series of models shown in Table S4 of the Supplementary Materials which, alongside the five covariates displayed, also control for employment status, household tenure, benefit entitlement, the number of children in the household, and respondent disability in an attempt to avoid confounders.

Sex/gender: Figure 1 shows the predicted probabilities from a model with multiple controls, which again confirms that women report a higher prevalence of feeling unsafe in both sets of data. It also establishes women to be slightly more likely to be threatened and/or insulted or shouted at than men and finds men to be more likely to be physically attacked. While these tendencies were confirmed with both sets of data, there are also important differences between them. Specifically, men’s differential risk of being attacked is not statistically significant in the UKHLS, whereas it is in the CSEW, where women are statistically less likely to report having been attacked than men.

Age: Age is another known predictor of violence exposure, with younger people often found to be disproportionately at risk (Cooper and Obolenskaya 2022; Thompson and Tapp 2022). We found a very strong age-related gradient to feeling unsafe and being threatened/insulted using the UKHLS data, which was not replicated with the CSEW (Figure 1). However, the age gradient to physical violence was similar in both datasets.

Ethnicity: There are known differences in violence exposure by ethnic group (Cooper and Obolenskaya 2022). While we confirm some of these differences in Figure 1, the figure also reveals different tendencies by ethnic group in each set of data. The CSEW indicates that Asian (British) respondents feel less safe than White (British) respondents, while to our surprise, the UKHLS indicates the opposite—that White respondents feel less safe than Asian respondents. Similarly, UKHLS Asian (British) respondents reported fewer instances of being threatened/insulted than their White (British) counterparts, whereas there were no ethnic differences on this index for the CSEW. Lastly, Figure 1 shows that the CSEW data finds lower risks of physical violence for those of Asian (British) and Black (British) ethnic heritage compared to those from a White (British) background, whereas we find no differences by ethnic group on the risk of being attacked using the UKHLS. Note that researchers have found ethnic minorities to underreport crime (Davis and Henderson 2003). They have also been found to develop avoidance strategies if exposed to or fearful of racist harassment, and to exhibit fewer behaviours associated with becoming a victim (e.g., lower alcohol consumption, see Hurcombe et al. 2010), which is thought to reduce victimisation recorded in surveys. Nonetheless, it is unclear why they may do so in one survey rather than another, which we presented here. Nevertheless, other than the notable differential in tendency for Asian British groups to report less exposure to feeling unsafe and to being threatened in the UKHLS compared to the CSEW, both surveys do show similar patterns of ethnic inequalities in victimisation.

Relationship status: Both sets of data confirm that those who are married or cohabitating have reduced risks of feeling unsafe and of being threatened or attacked compared with those who are separated/divorced (Figure 1). The CSEW also indicated that those who are single/widowed are at a higher risk of physical violence than those who are married or cohabiting.

Household income: Though our model controls for a series of indicators of socioeconomic status (SES), with labour force status, benefit receipt, and housing tenure also included in our model, we restricted the discussion to the findings relating to household income here, with SES in general understood to be a key predictor of violence exposure and victimisation. We found a clear differential in the reporting of fear of violence by household income in both sets of data. In the UKHLS, we found that as household income increases the risk of feeling unsafe also rises, while the opposite effect was found with the CSEW. A significant portion of this differential may be due to questionnaire wording, with the CSEW clearly linking feeling unsafe to one’s own neighbourhood, while it appears that the UKHLS’s broader conceptualisation of the variable allows it to be relevant for a different tranche of respondents. Yet, we also noted strong household income differentials in tendency between both sets of data for those who have been threatened/insulted. Using the CSEW, there was a significant association between receiving threats and household income, with higher income groups receiving more threats. For the UKHLS, with its broader classification, there was no clear relationship between household income and risk of being threatened/insulted. Importantly, we found no differences between the data in the risk of physical violence by household income.

Figure 1. Predicted probabilities of violence and feeling unsafe for key sociodemographic groups. Based on average marginal effects and models in Table S4 of the Supplementary Materials.

Figure 1. Predicted probabilities of violence and feeling unsafe for key sociodemographic groups. Based on average marginal effects and models in Table S4 of the Supplementary Materials.

4.3. Differences in Prevalence Estimates by Subgroup—The UKHLS versus the CSEW

While Figure 1 sought to determine the variance in the risk of violence for different sociodemographic groups within each set of data, Figure 2 extends that analysis by presenting tests of significant differences in the prevalence between the UKHLS and the CSEW per subgroup. Broadly, with the exception of the index on ‘being threatened’, the data appear to offer similar measures of prevalence for different socioeconomic groups, with many differentials in prevalence statistically insignificant and/or substantively small. Nonetheless, there were noteworthy differentials which we reviewed below.

We found that the UKHLS recorded a slightly higher prevalence of feeling unsafe for many of the sociodemographic groups examined. The UKHLS’s reference to numerous locations where one may have felt unsafe, compared to the CSEW, which references one’s neighbourhood and night-times, appears to have allowed more people to report having felt unsafe. This was found to be true irrespective of one’s sex/gender, age (with the exception of those aged 55+), relationship status, and for those in wealthier households. The fact that the disparity between the data was most notable on income again suggests that while wealthier households may be able to afford to live in neighbourhoods with lower rates of crime and/or anti-social behaviours, and so may avoid feeling unsafe where they live, the UKHLS’s conceptualisation of being ‘unsafe’ in multiple locations allows for wealthier respondents to report feeling unsafe in other contexts. Differentials in prevalence for ethnic minority groups revealed that the UKHLS recorded a significantly lower prevalence of feeling unsafe for Asian (British) respondents compared to the CSEW, and a higher prevalence for White (British) respondents, whereas there were no other significant differences for other ethnic groups.

Figure 2. Differences between the UKHLS and the CSEW in predicted probabilities of violence and feeling unsafe for key sociodemographic groups. Based on the average marginal effects of models, where we interacted all variables with the type of dataset. Estimates are significantly different between the UKHLS and the CSEW if the confidence interval does not cross 0.

Figure 2. Differences between the UKHLS and the CSEW in predicted probabilities of violence and feeling unsafe for key sociodemographic groups. Based on the average marginal effects of models, where we interacted all variables with the type of dataset. Estimates are significantly different between the UKHLS and the CSEW if the confidence interval does not cross 0.

As anticipated, whilst reviewing the differential face validity of ‘being threatened’ in each data, we found that the UKHLS records statistically significantly higher levels of being threatened/insulted than the prevalence of the more specific ‘threat offences’ recorded in the CSEW. We found this to be the case for all groups.

Lastly, we revealed significances in the differential prevalence of being attacked in the UKHLS compared to the CSEW. Figure 2 shows that the UKHLS records more violence against women than is true of the CSEW, whereas there are no differences in the prevalence for men. We also found that the prevalence of being attacked was higher for (British) Asians in the UKHLS and noted no other significant differences for other ethnic groups. The UKHLS showed a marginally larger prevalence of being attacked among married/cohabiting people compared to the CSEW, and we again found no difference between the CSEW and the UKHLS in the predicted probabilities for each household income group, apart from a slight difference among those making from £20,800 to £31,200 per year.

4.4. Differential Effects of Violence on Health

To test differentials in the construct validity of the UKHLS and CSEW indicators, we examined their associations with self-rated general health, given the known associations between violence and health (Sundaram et al. 2004). Here, we could expect differential effects of the measure of feeling unsafe given the UKHLS’s broader conceptualisation. We could also expect differences for our indicators which measure being threatened, as the UKHLS index includes non-criminal behaviours and so may have weaker associations with poor health. In fact, what we found was that all three variables in each set of data are strongly correlated with poor health (Table 2). Whilst we might have expected the CSEW variables relating to fear of violence and to being threatened/insulted to have a stronger association with poor health, given the data’s greater emphasis on criminal actions, we did not find this to be the case, with no statistically significant differences found between the UKHLS and the CSEW in the associations between each violence indicator and self-rated general health. This suggests that the UKHLS’s indexes, while less aligned to legal definitions of criminal behaviour, remain strongly correlated with negative outcomes for respondents. It also suggests that restrictions to legally defined harms may be limiting the scope of analyses into other forms of detrimental behaviour.

Table 2. Comparing predictors of self-rated general health (0 = good health, 1 = fair health, and 2 = poor health) for the UKHLS and CSEW datasets using ordinal logit models. Odds ratios are presented, and SEs are in parentheses.

	Model 1a	Model 1b	Model 2a	Model 2b	Model 3a	Model 3b
	UKHLS	CSEW	UKHLS	CSEW	UKHLS	CSEW
Unsafe	1.774 ***	2.122 ***
	(0.081)	(0.183)
Threatened			1.683 ***	1.692 ***
			(0.100)	(0.180)
Attacked					2.048 ***	1.701 ***
					(0.253)	(0.229)
Cut 1	22.339 ***	39.151 ***	19.044 ***	33.935 ***	17.189 ***	33.977 ***
	(2.808)	(9.786)	(2.388)	(4.163)	(2.127)	(4.182)
Cut 2	122.202 ***	252.523 ***	103.110 ***	215.274 ***	92.709 ***	215.395 ***
	(15.926)	(64.410)	(13.383)	(27.003)	(11.892)	(27.102)
Observations	23,883	6490	23,883	27,922	23,883	27,922

*** p < 0.001, ** p < 0.01, and * p < 0.05. No associations between attacked, threatened, unsafe, and self-rated general health are statistically significantly different between the UKHLS and CSEW. Partial model shown, with full model controlling for sex/gender, age, ethnicity, relationship status, number of children, employment status, benefits, household income, and housing tenure. These models do not control for disability given the risks of over-specification given its inevitable correlation with the dependent variable.

Table 2. Comparing predictors of self-rated general health (0 = good health, 1 = fair health, and 2 = poor health) for the UKHLS and CSEW datasets using ordinal logit models. Odds ratios are presented, and SEs are in parentheses.

	Model 1a	Model 1b	Model 2a	Model 2b	Model 3a	Model 3b
	UKHLS	CSEW	UKHLS	CSEW	UKHLS	CSEW
Unsafe	1.774 ***	2.122 ***
	(0.081)	(0.183)
Threatened			1.683 ***	1.692 ***
			(0.100)	(0.180)
Attacked					2.048 ***	1.701 ***
					(0.253)	(0.229)
Cut 1	22.339 ***	39.151 ***	19.044 ***	33.935 ***	17.189 ***	33.977 ***
	(2.808)	(9.786)	(2.388)	(4.163)	(2.127)	(4.182)
Cut 2	122.202 ***	252.523 ***	103.110 ***	215.274 ***	92.709 ***	215.395 ***
	(15.926)	(64.410)	(13.383)	(27.003)	(11.892)	(27.102)
Observations	23,883	6490	23,883	27,922	23,883	27,922

*** p < 0.001, ** p < 0.01, and * p < 0.05. No associations between attacked, threatened, unsafe, and self-rated general health are statistically significantly different between the UKHLS and CSEW. Partial model shown, with full model controlling for sex/gender, age, ethnicity, relationship status, number of children, employment status, benefits, household income, and housing tenure. These models do not control for disability given the risks of over-specification given its inevitable correlation with the dependent variable.

5. Conclusions and Discussion

We started this paper by highlighting the need for longitudinal data on victimisation for the discipline of criminology and noted the absence of representative panel data with a criminological focus for the UK. As an interim solution to this knowledge gap, we investigated whether the UKHLS may provide an important resource on violence victimisation, within a longitudinal framework, given its recent decision to field a series of variables on violence alongside its pre-existing indicators. By comparing the prevalence of violence experience and its associations with demographic, socioeconomic, and self-rated health indicators between the UKHLS and the more frequently used CSEW, this paper provides an empirical basis for assessing the validity and use of the UKHLS measures on violence victimisation for criminological research. We specifically focused on the indices ‘feeling unsafe’, ‘being threatened’, and ‘being physically attacked’.

We found notable differences in the conceptualisation and wording of the index on ‘feeling unsafe’. While the UKHLS asked about feeling unsafe in a wide variety of locations, the CSEW asked about feeling unsafe in one’s own neighbourhood and at night-time. This notable difference in wording and conceptualisation affected both aggregate prevalence and differential prevalence via socioeconomic and demographic groups. While the overall prevalence of feeling unsafe was similar for women (at around 30 percent) in both sets of data, it was different for men (20.8 in the UKHLS and 12.9 percent in the CSEW). Moreover, those who reported feeling unsafe varied in other respects in each set of data, with young people and those living in more wealthy households more likely to report feeling unsafe using the UKHLS indicator (see also Figure 1). We regard the UKHLS’s index on feeling unsafe to be of particular interest, therefore, to others in the field who have researched the topic using the standard conceptualisation of feeling unsafe in one’s own neighbourhood (e.g., Farrall et al. 2021), with the UKHLS’s extended conceptualisation allowing for an assessment of different settings, including commercial and public spaces. We found far fewer people were victims of a threat offence in the CSEW than were threatened or insulted based on UKHLS data. Additionally, we found different patterns in victimisation by socioeconomic group, with the UKHLS indicator showing a very high prevalence amongst younger age groups. Yet, in other aspects, we found similar tendencies, and most crucially, tests of construct validity, which sought to determine associations with self-rated general health, found both indices to be similarly correlated with poor health. These findings suggest that the two variables may be measuring a similar latent construct, and so we concluded that the inclusion of (possibly) non-criminal abusive harms, namely being called names, shouted at, and insulted, may be important for future researchers who are interested in investigating the negative consequences of inter-personal abuse.

Lastly, we found a similar prevalence, about 2 percent, of ‘being physically attacked’ in each set of data. Additionally, the associations between key demographic factors and being attacked were similar. Nonetheless, we did find some differences, with a higher prevalence of physical abuse in the UKHLS for women being one of the most noteworthy differentials. This finding is consistent with others’ research who found the CSEW to underestimate violence against women (Cooper and Obolenskaya 2021; Pullerits and Phoenix 2023). We also found statistically significant differences in the estimands of being attacked by ethnic group. Asian British respondents had a greater prevalence of being attacked using the UKHLS than was true of the CSEW. Again, it is unclear why this might be the case, though others have found that ethnic minorities tend to disclose less violence than White people in the CSEW (Cooper and Obolenskaya 2022), and this has also been shown to be true in the USA for Asians/Native Hawaiians/Pacific Islanders (Thompson and Tapp 2022). Why the UKHLS appears to be ‘better’ at capturing more violent incidents from British Asians is still unclear. Though we found the UKHLS index on physical violence to be similar to the violence screener question of the CSEW, we did note inconsistencies in the measurements of physical violence within the CSEW. The CSEW provides researchers with two different measures of physical violence. The first is the violence screener question; the second is the CSEW offence classification, an indicator generated by professional coders of interviewee accounts of possible criminal incidents. Our analyses suggest that interviewees were less likely to divulge an experience of physical assault in contexts where interviewers prompt them by listing a series of examples of criminal assault. Combined with the slightly stronger association between violence and health using the UKHLS data (see Table 2), this suggests that the UKHLS may be picking up more serious forms of physical violence. Nonetheless, we regard the variable measuring being attacked to capture a similar construct in both sets of data and see no reason for these variables not to be considered equivalent for future studies on violent victimisation and its associations.

It is worth noting that the CSEW ‘violent offence’ measure has been criticised for its failure to fully capture both domestic (physical) violence and sexual violence (Cooper and Obolenskaya 2021). We believe that there is a similar risk of this omission in the UKHLS given its wording and questionnaire sequencing, with few specific prompts for domestic violence, and no clear prompting of sexual violence. Moreover, other types of violent crime, including coercive control and stalking, are unlikely to be covered by the existing measurements in the UKHLS, whereas the CSEW has a dedicated self-completion section to ensure accurate measures of these crimes (although see also Hester et al. 2023 for a critique). Thus, while we have found the UKHLS to perform similarly to the CSEW in many respects, for the three variables tested, it remains limited by its restricted range of variables. We therefore recommend expanding the UKHLS self-completion questionnaire to include questions on domestic violence and abuse. We recommend this whilst simultaneously noting that caution would be required in the collection of these data, as the UKHLS, unlike the CSEW, interviews all adult household members, and there is a risk therefore that an abuser and his/her victim would be alerted to the line of questioning if both were posed the same battery of questions. This would pose a risk to victims of domestic violence and abuse, and so a randomised allocation of one household member to such a module is also advised.

Overall, this study found a considerable overlap alongside face and construct validity in the measures of violence in both the UKHLS and the CSEW. We hope that this will encourage future criminologists to exploit the longitudinal nature of the UKHLS, which can be usefully examined alongside its wide variety of socioeconomic, family, health, and lifestyle indicators. We also found that the UKHLS’s wording of its indicators regarding experiences of ‘being insulted or threatened’ and of ‘feeling unsafe’ to be consistent with calls to broaden measurement and therefore understandings of violence beyond criminal definitions.