Victims and suspects of modern slavery: Identifying subgroups using latent class analysis

We explored heterogeneity (subgroups) amongst victims and suspects of modern slavery offences using a cross-sectional extract of police data from a large metropolitan area in the UK with details of 540 victims and 380 suspects (reported to the police between April 2015 and June 2018). Latent class analysis was used to identify subgroups of victims and suspects based on the manifest demographic (age, sex, and place of birth) and exploitation type variables. Amongst suspects, classes distinguished between ‘Male sex traffickers’ (82%) and ‘Labour and domestic traffickers’ (18%). Amongst victims, four classes were identified as: ‘Sexually and domestically exploited women’ (30%) and ‘Sexually exploited girls’ (35%), ‘Men exploited in licit and illicit markets’ (26%), and ‘Criminally exploited boys’ (9%). These findings reveal trafficking as structured by gender, generation, and migration, but caution against defining the problem as one of exclusively male offenders exploiting women and children victims, even while this is commonplace.


Background
Attention to human trafficking and modern slavery 1 has increased since the turn of the century. The Protocol to Prevent, Suppress, and Punish Trafficking in Persons, Especially Women and Children embedded in the UN Convention against Transnational Organized Crime is the basis for the many approaches to trafficking and modern slavery globally. The Protocol framed human trafficking as a problem of transnational organized crime and border security, with law enforcement as the primary approach to tackling human trafficking (Jordan, 2002). Many non-governmental organizations (NGOs) argued that a human rights-based framing would be more effective, particularly for victim care (Jordan, 2002;Doezema, 2010). However, international policy development ran ahead of research and binary stereotypes proliferated, embedding a model that counterpoises 'innocent victims'-often women and children-with 'evil traffickers'-assumed to be foreign national men engaged in organized crime (Berman, 2004;Gadd and Broad, 2018).
This crime control framing was reflected at national levels where trafficking was defined as a problem best addressed through border controls and policing. This included the UK where modern slavery became a policy priority in the context of reducing immigration (Broad and Turnbull, 2019). Sexual exploitation became a primary focus of early UK policing efforts (Broad, 2015), though the diversity of exploitative markets was soon recognized, with cannabis cultivation, forced criminality, 'county lines', sham, and forced marriage identified as priorities alongside sexual and labour exploitation as well as domestic servitude (Home Office, 2017). Inspectorate reports have highlighted failings in the police response to modern slavery, but with limited recognition of the economic and legal factors that contribute to exploitation (HMICFRS, 2017;NAO, 2017). Arguably, intelligence-based action around particular topics raised expectations in terms of law enforcement capability to tackle exploitation that reached beyond police capacity to intervene in global markets (Bjelland, 2018;Home Office, 2019). Awarenessraising tends to focus on sites of exploitation, for example, nail bars or car washes, or specific nationalities, without recognizing the factors that lead migrants to participate in low paid work (Scholenhardt et al., 2012). Recommendations to improve policing have highlighted the importance of multi-agency working and collaboration with NGOs to provide victim support (HMICFRS, 2017;National Audit Office, 2017), but the need to reach across international borders and understand how people become indebted to traffickers is rarely anticipated in domestic policy (Gadd and Broad, 2018).
Most academic and policy attention has quite rightly focussed on victims and their needs (Surtees, 2008;Home Office, 2014;Choi-Fitzpatrick, 2016). However, there has been a lack of research on perpetrators, delimiting the scope for a more holistic, evidence-based approach within criminal justice. The political use of victim/perpetrator binaries has made it difficult to develop responses that acknowledge that some victims are also exploiters and some perpetrators have been exploited/victimized. This is most evident among women convicted of trafficking offences (Broad, 2015), and has become increasingly apparent in 'county lines' cases, where a complex division of labour entraps young people who become reliant on the profits of illicit drug-dealing and are frequently criminalized (Coomber and Moyle, 2018). The introduction of legislative measures to support people who commit criminal offences whilst being victimized through modern slavery is said to create a loophole which traffickers exploit, rendering those the law intended to protect at greater risk of exploitation. To date, the UK government has provided 'little clarity' for practitioners to resolve this conundrum (IASC, 2019, p. 14). This matters because it is not always clear who has exploited whom. Ethnographic research has revealed that exploitation across international borders often occurs because of two failed migration journeys (Mai, 2010) where parties trade poorly paid or illicit services against the backdrop of legal precarity and complex vulnerabilities. Home Office research with people convicted of facilitating 'organised immigration crime' has revealed that such offenders often regard themselves as providing assistance, while making a little extra cash to supplement income from legitimate work (Webb and Burrows, 2009).
The United Nations Office on Drugs and Crime (UNODC) has reported data on perpetrators' and victims' age, gender, and nationality since 2006. The 2018 report (UNODC, 2018) found that although men predominate in trafficking, women continue to represent a significant proportion of offenders-35%. Females are more vulnerable to victimization than males, but the latter are by no means immune. Nearly half (49%) of victims are adult females. One-fifth (21%) of victims are adult males, while 23% of victims are girls and 7% are boys. Trafficking for sexual exploitation remains the most prevalent form of exploitation globally (59%), followed by forced labour (34%). This global overview is useful in providing a context for understanding exploitation at local levels. Local studies have found that pathways into trafficking for offenders coalesce around pre-existing links between people who need support to travel or work, and often have little to do with transnational organized crime groups (Keo et al., 2014;Denton, 2016;Gotch, 2016;Viuhko, 2018;Cockbain, 2018). Cockbain et al. (2019) and Bjelland and Dahl (2017) highlight the potential for police data to enhance understanding of modern slavery, despite its limitations, particularly in relation to informing evidence-based approaches to policing and anti-slavery policy and understanding the particular processes that define relationships between perpetrators and victims.

Aim
This study examined the characteristics of victims, suspects, and the crimes in which they are involved. It sought to explore the degree of heterogeneity among offenders and victims in cases classified by the police as potential 'modern slavery', by identifying subgroups using a person-centred approach. Doing so can reveal distinct subgroups of victims and perpetrators that require different forms of intervention, given the diversity of modern slavery and the importance of tailoring prevention and support work (Cockbain et al., 2019). We refer to 'suspects' throughout this article, as at the time of data collection, it had not been determined that these individuals were guilty of an offence. Latent class analysis (LCA) was employed to discern the subgroups within the dataset. LCA is growing in popularity in Criminology and has been used to identify subgroups of perpetrators of domestic abuse (Carbone-Ló pez et al., 2006;Klostermann et al., 2009;Ansara and Hindin, 2010) and drinking patterns associated with violence (Lightowlers, 2017). LCA has been used to profile risk factors associated with human trafficking victimization (Reid et al., 2019) and factors associated with recovery from victimization Ottisova et al., 2018;Iglesias-Rios et al., 2020). Hargreaves-Cormany and Patterson (2016) used LCA to study perpetrators of child sex trafficking. They discerned relationships between the criminal history severity, the severity of violence, and charismatic offender behaviour.
LCA categorizes individuals, not by testing theoretical assumptions, but by exploring patterns in indicators empirically to examine the underlying structure among cases (i.e. latent taxonomic structure) (Uebersax, 2009). LCA offers a personcentred methodological approach to identifying typologies/subgroups of victims and offenders of modern slavery without preceding theoretical assumptions of how these groups should be defined (Aebi et al., 2015). This study identifies classifications of victims and suspects using police data. The results of this analysis offer insights into the distinct subtypes of victims and suspects involved in modern slavery, the potential impact of anti-slavery initiatives, and the opportunity to reflect on the limitations of official recorded crime data. With attention to statistical and theoretical interpretation, we identify four distinct subgroups of victims and two of suspects.

Data
Details of 540 modern slavery offences were obtained from a large metropolitan area in the UK. The original dataset contained 1,502 enquiries pertaining to modern slavery reports to the police from April 2015 to June 2018. Of these enquiries, 540 were recorded as crimes entailing 540 instances of victimization. Suspects were identified for 380 such incidents. As with many crimes, not all investigations resulted in the identification of a suspect. Some victims and suspects featured more than once in the dataset. Some victims were subject to multiple forms of exploitation and some suspects committed more than one offence. The following analyses focus exclusively on those 540 separate incidents recorded as crimes given the greater certainty that something akin to modern slavery had occurred. Analysis was conducted on victims in the 540 instances of victimization and analysis of suspects on the 380 offences for which an individual was identified (288 individuals). Further discussion of the 92 repeat offenders is included below. Within the data, details of the following demographic characteristics of victims and offenders were included: adult/minor (<18 or !18 years); sex; place of birth and exploitation type. The data contained textual entries on the modus operandi (MO) recorded by police officers to describe the activities. We have made use of example MOs in the findings section to provide qualitative context to the latent classes.

Data quality and missing data
The police data were cleaned and thoroughly anonymized. Several data quality concerns arose during the data cleaning process which are common in administrative police data often due to operational demands (Bjelland and Dahl, 2017;Cockbain et al., 2019). There were issues with missing data, some of which we were able to partially resolve. Crime details are most often recorded by front-line officers or referred from external agencies and relayed to specialist modern slavery teams. Where information is missing from the initial record there is limited capacity to backfill these details, for example, with information that transpires during prosecutions. Consequently, there remained large numbers of 'unknown' entries for some of the variables of interest. These incomplete data have been included in this analysis for the sake of completeness and to highlight the need to improve data quality. Where we could meaningfully backfill gaps (e.g. the type of exploitation and sex of victim/suspect were backfilled with information from the MO), this was done. However, these derived variables only represent the information available in the MO which varied depending on the quality of the information input by police officers.
Some victims are subject to multiple types of exploitation (Home Office, 2017). In the current analyses, only the primary exploitation type is considered. The dataset did include some information on secondary exploitation but numbers were too small to perform meaningful analysis. Similarly, some victims of trafficking are subject to other types of crime, for example, rape, assault, and abduction (Home Office, 2017). This information was not recorded in the dataset but in some instances, there may have been charges and convictions for other related offences, such as fraud, people smuggling, or rape.

Methods
As well as descriptive analyses, LCA was employed to identify distinct subgroups of victims and suspects of modern slavery offences. 'Latent class analysis (LCA) provides a framework to identify latent classes by observed manifest variables' (Zhang et al., 2018, p. 1). LCA is a technique which probabilistically assigns cases to distinct subgroups in a heterogeneous population based on different response patterns to manifest categorical variables. Here it is used in an exploratory manner, with the objective of identifying best fitting or most parsimonious models relating to victims and suspects of modern slavery offences. Necessarily constrained by data quality and completeness, as outlined above, these were derived from demographic and exploitation type variables. For victims these included: exploitation type, sex, whether the victim was an adult or minor, and place of birth. For suspects, where data were less complete, these included: exploitation type, sex, and place of birth. A search of the database to determine whether any of the perpetrators and victims were the same person (i.e. whether any of the perpetrators had been subject to prior victimization) was negative-although there are many factors which may have contributed to this such as their victimization being unreported.
Model specification was conducted incrementally with due attention to statistical criteria, parsimony and theoretical interpretability (Collins and Lanza, 2010). One to four and one to five class models for suspects and victims were run, respectively, for comparison. Whilst in statistical terms, preferred models are those that minimize values of the Bayesian information criterion (BIC) and/or Akaike information criterion (AIC), the parsimony and substantive interpretation of solutions must be considered. To this end, feedback was academically informed as well as obtained by discussing the results with a range of policy and practice stakeholders advising on the project, including the UK Independent Anti-Slavery Commissioner's Office and Home Office as well as the modern slavery team at the police service providing the data.
All models were fitted using the open source R package Polytomous Variable Latent Class Analysis (PoLCA), which allows for the estimation of latent class models with polytomous outcome variables Lewis, 2011, 2013;Zhang et al., 2018). Models were estimated using expectation-maximization and Newton-Raphson algorithms to find maximum likelihood estimates of the parameters. PoLCA employs listwise deletion; a case is excluded from analysis if any single value is missing Lewis, 2011, 2013). To ensure global, rather than a local, minimum was obtained, models were run 10 times with different start values and a maximum of 1,000 iterations (the maximum number of iterations through which the estimation algorithm will cycle).

Victims
There were 43% female, 30% male victims (unknown in 27% of cases (n ¼ 393)). These sex differences are less pronounced than those reported globally (UNODC, 2018). However, in comparison to the UK National Referral Mechanism (NRM) (the mechanism through which victims are recognized as such in the UK) women are over-represented among victims in this police area. NRM referrals comprised 39% females and just <60% male in 2018 (National Crime Agency, 2019). Differences may stem from women being less likely to report victimization, as research has indicated that women, particularly migrant women, are less likely to report for fear of reprisals or mistrust of authorities (Walby et al., 2016). Victims' ages were not always recorded (22%), but where known ranged from 6 to 73 (mean ¼ 24) and comprised 61% adults and 39% minors (n ¼ 422). The national picture is similar, with 55% of NRM referrals recorded as adults and 45% minors (NCA, 2019).
Over half of details for victim nationality were missing (n ¼ 294; 54%). Of that recorded, the most common nationality was UK (n ¼ 60; 11%) followed by Hungarian (n ¼ 17; 3%); Vietnamese (n ¼ 15; 3%) and the Czech Republic and Romanian (n ¼ 10; 2% for both). This is similar to the national picture with the top five countries of origin for NRM referrals being UK; Albania; Vietnam; China; and Romania with differences potentially reflecting the local population and patterns of migration there (Turner and Kelly, 2008). We coded this country information into four main categories-UK (24%), EU (17%), Other (22%), and Unknown (37%).
Half of crimes recorded related to sexual exploitation and one quarter to labour exploitation. One in six cases (13%) involved criminal exploitation, defined as 'acquisitive crimes such as forced begging, forced theft. . . as well as financial exploitation' (NCA, undated a, p. 2 ), around half of which related to cannabis cultivation. Domestic servitude consisted 9% of the cases, while 2% of crimes were recorded as facilitating travel for the purposes of exploitation. This differs from national trends in the NRM where referrals for labour Victims and suspects of modern slavery Article Policing exploitation (including criminal exploitation) were most common. Even when the labour and criminal categories were aggregated in the current data, they still comprised a smaller percentage than sexual exploitation, likely reflecting the geographical composition of the study area and positive local relationships between the police service and sex worker support organizations.
Latent class models were run incrementally up to five classes, the four-class solution was selected due to theoretical interpretability of the classes. Whilst the BIC is lowest for the two-class solution (Table 1) BIC underestimates the number of classes when samples are small (Nylund et al., 2007). It is thus meaningful to opt for either the three or four-class solution as indicated by the sample size adjusted BIC (aBIC). The aBIC is lower for the four-class solution (compared to two and five) and only marginally higher than the three-class solution. As the four-class solution aids substantive interpretation it is favoured in this instance. Collins and Lanza (2010, p. 88) note that 'because of varying penalties associated with each criterion, they often do not identify the same model as optimal. Information criteria are likely to be more useful in ruling out models and narrowing down the set of plausible options than in pointing unambiguously to a single best model'. The substantive interpretability of classes is key in model selection and class sizes may be considered in doing so and 'subject-matter knowledge should also be considered when considering the number of classes' (Zhang et al., 2018. p. 3). As such 'a combination of fit indices (rather than sole reliance on one index), coupled with a consideration of theory and interpretability is recommended to determine the optimal latent class model' (Zhang et al., 2018, p. 3). Having examined response probabilities of the two, three, and four class solutions, the additional classes in the four-class solution builds on the two-class solution by discerning predominantly between male and female adults and minors. This is of substantive importance as it points to the profile of exploitation these groups experience and opportunities for bespoke intervention which are both age and sex specific. The entropy value (0.78) represented a clear class assignment in the four-class model. Moreover, in statistical terms classes containing <5% of the sample are more likely to be deemed spurious, often resulting from extracting too many classes (Hipp and Bauer, 2006) and so the five-class model was rejected.
Response probabilities for the four-class model are visualized in Fig. 1. This highlights distinctions between the four classes. Two classes were characterized by high levels of sexual exploitation (Fig. 1, Classes 1 and 3) comprised nearly two-thirds of victims (30 and 35%, respectively). These two classes differed mainly by age. Class 1 comprised mainly adult females who may also have been victims of domestic servitude (response probabilities 0.89 and 0.24, respectively)-albeit to a lesser extent than sexual exploitation (response probability 0.61)-whereas Class 3 comprised mainly female children (response probabilities 0.86 and 0.8) almost exclusively sexually exploited (response probability 0.92). We will refer to these classes as 'Sexually and domestically exploited women' and 'Sexually exploited girls', respectively. The following examples selected from the MOs illustrate these two classes: 'Sexually and domestically exploited women': The victim was brought to the UK by her husband. She was told that she would be cleaning his house and his relatives' houses in the day and then 'going with' different men at night. She was beaten on a daily basis by her husband and was locked in her bedroom when not meeting relatives or sex buyers.
'Sexually exploited girls': The victim-a girl-travelled from her home country to the UK via several different countries and entered the UK in a container on a lorry. She was housed in different locations over six months. She was visited by numerous men who sexually abused her.
The third class (Class 2) was characterized by high levels of labour exploitation and to a lesser extent cannabis cultivation (response probabilities 0.75 and 0.20). It comprised just over a quarter of cases (26%). These victims were mostly adult males (response probabilities 0.83 and 0.95). We will refer to this class as 'Men exploited in licit and illicit markets'.
The following example selected from the MOs illustrates this class: 'Men exploited in licit and illicit markets': The adult male victim was brought to the UK to wash cars for an agreed amount of money. Once in the UK the victim was made to work longer hours, six days per week for a Victims and suspects of modern slavery Article Policing rate less than that promised. The victim was treated poorly at the workplace and was not able to leave due to his legal status. He was provided with accommodation as part of the job but was then locked out of the property and became homeless.
The fourth class (Class 4) were mainly male children (response probabilities 0.79 and 0.83) associated with the offence classification of 'Criminal -Other' (response probability 0.68). We will refer to this class as 'Criminally exploited boys'. This reflects those victims exploited for the purposes of criminality, for example, shoplifting and activity which has been termed 'county lines' (Coomber and Moyle, 2018). This is a 'term used when drug gangs from big cities expand their operations to smaller towns [. . .] exploiting children and vulnerable people to sell drugs' (NCA, undated b). This problem was highlighted towards the end of the 2010s and has received significant media attention, with police commenting that criminal exploitation of children is 'almost back to Victorian times' (Guardian, 2020). This represented a small proportion of the sample (9%) but may indicate a growing class as police attention is turned to this priority area in line with media and policy concern (Serious and Organized Crime Strategy, Home Office, 2018).
The following example selected from the MOs illustrates this class: 'Criminally exploited boys': The victim was a 15-year old boy who was reported as missing from home and suspected of being criminally exploited through county lines. The victim was subsequently arrested and found in possession of class A drugs, cash and two mobile phones.
With the available data, it was not possible to discern well between classes based on victims' place of birth given the large number of 'unknown' data entries (missing observations). Although not conventional, we felt it was important to include such entries in our analyses, so as not to artificially emphasize findings from those who do have valid entries on this field and to highlight the need to improve data collection by the police-and other reporting agencies-on the characteristics of victims beyond their age and sex, as this can help improve insights into the profile of victims coming to the attention of authorities.

Suspects
Of the 380 suspects, 75% were male and 21% female (4% missing). The relatively high number of female suspects reflects patterns elsewhere for modern slavery and human trafficking crimes (UNODC, 2018) and has been attributed to intimate relationships with male co-defendants, roles performed in offending activity, and prior experiences of modern slavery/human trafficking victimization (Broad, 2015). The age range of suspects was between 11 and 75 years of age (mean ¼ 33). Most suspects were adults (94% were >18 years). The suspects were on average 7 years older than victims and older across all categories of exploitation. This varied from suspects 5 years older than victims for 'Criminal -other' to 11 years older for 'Sexual exploitation' and 'Domestic servitude'.
There were many missing entries regarding suspect nationality (n ¼ 163; 43%). This is probably because suspects were either not apprehended and/or their nationalities were unknown to the victim when reporting the crime. Of the nationalities recorded, the most common country of origin was the UK (n ¼ 77; 20%) followed by Romania, Czech Republic (n ¼ 23; 4%), Pakistan (n ¼ 20; 4%), Hungary (n ¼ 19; 3%), and Vietnam (n ¼ 13; 2%).
There were 92 repeat suspects (24% of suspects). These were of a similar age profile to suspects more generally (mean 32, range 17-61 years). There was a higher proportion of males among the repeat suspects (82%, n ¼ 75) than amongst the suspects more generally. Repeat suspects were most commonly engaged in sexual exploitation (61%, n ¼ 56), followed by labour exploitation (24%, n ¼ 22) and cannabis cultivation (10%, n ¼ 9), with only a few involved in 'Criminalother', domestic servitude or facilitating travel for the purposes of exploitation.
Whilst there were 540 crimes included in this dataset, only 40 were known (at the time of data extraction) to have culminated in charges specifically for modern slavery related offences. In 45% of all crimes, the outcome of the case was unknown as no suspect had been identified for these offences at the point of data extraction. In 19% of crimes, a suspect was identified but there was no further action taken due to insufficient evidence or because the victim did not support a prosecution. In 16% of crimes, no suspect was identified (In 18% of these, this was because the victim refused to disclose the identification of the suspect (n ¼ 14)). In 9% of crimes, a suspect had been charged with an offence (85% of which were modern slavery offences (n ¼ 40)). In 3% of crimes, a suspect had been identified and was outstanding.
Latent class models were run incrementally up to four classes, with the two-class solution providing the best fit based on the log-likelihood values, Akaike and Bayesian Information Criteria (AIC and BIC) (where a lower value indicates a better fit) (see Table 2) as well as parsimony (number of parameters to be estimated) and theoretical interpretability of the classes. When assessing the fit of the two-class model, it is noteworthy that adding a third class did not offer tangible benefits in interpretation, as two classes were broadly similar (and the AIC only lowered slightly for three classes). Adding a third class would have only differentiated between whether suspects came from the UK or 'Other' for Class 1-namely amongst 'Male sex traffickers'. As this is the least reliable variable in the dataset, the two-class model was selected due to its parsimony and interpretability. The entropy value (0.61) represented reasonable class assignment in the two-class model, and was likely not as clear as we would have liked due to many missing/ unknown data observations.
Response probabilities for the two-class model are visualized in Fig. 2 which highlights the distinctions between the two classes; one class comprising the majority of suspects (82%) who were male and perpetrated mostly sexual exploitation (response probabilities 0.87 and 0.68, respectively). These were less discernible by their country of origin. We will refer to this class as 'Male sex traffickers'. The second class comprised less than one in five cases (18%) containing males and females perpetrating labour exploitation and domestic servitude (response probabilities 0.40 and 0.27, respectively). The country of origin for this group was mostly unknown (response probability 0.71). We will refer to this class as 'Labour and domestic traffickers'. When contrasting these two classes, the former class ('Male sex traffickers') was more likely to be engaged exclusively in sexual exploitation, whereas the latter ('Labour and domestic traffickers') was Victims and suspects of modern slavery Article Policing more likely to be versatile in the exploitation type-that is, domestic servitude, labour exploitation as well as sexual exploitation-and was more likely to be perpetrated by both men and women.

Discussion
We have analysed heterogeneity amongst suspects and victims of modern slavery offences in a large UK metropolitan region by deriving subtypes based on their offence and demographic characteristics. Although the profile of modern slavery cases varies across police areas, the region discussed here features a comparatively larger number of cases (in relation to case numbers reported to the NRM (NCA, 2019)) and has featured in independent reviews of modern slavery cases as an effective and high-performing police area. The cases include most exploitation types and so the findings can be useful to other police areas in considering their anti-slavery activities. Whilst these cases only represent those known to the police, and are thus likely only a small proportion of exploitative activity, the detailed information from police data on both victims and suspects of these relatively rare forms of offending allow for an investigation of the local manifestation of what has become a national policy priority. A person-centred approach using LCA revealed four distinct subtypes of modern slavery victims and two suspect subtypes. Victim subtypes: Sexually and domestically exploited women; Sexually exploited girls;

Men exploited in licit and illicit markets; and
Criminally exploited boys Suspect subtypes:

Male sex traffickers and
Labour and domestic traffickers (male and female).
The sexual exploitation of women and girls by men represented the most common circumstances in this police area, a pattern repeated at a global level (UNODC, 2018). Although this is not the case across all police areas as labour exploitation is most commonly referred to the NRM, all police areas report varying levels of sexual and labour exploitation and can therefore identify with these research findings (NCA, 2019). The relationship between the police and sex worker NGOs was reportedly very positive in the region to which our dataset pertained. Police areas reporting lower levels of sexual exploitation may benefit from developing more inclusive multi-agency models of working, particularly with local sex worker NGOs. Positive relationships between the police and sex worker NGOs have been recognized elsewhere as facilitative of crime reporting of the exploitation of sex workers (Penfold et al., 2004) and thus may account for the high levels of reporting in this study. Conversely, the co-occurrence of sexual exploitation and domestic servitude in adult women represents a particular problem for policing. It can be difficult for service providers to reach victims confined to a single premise where they are being domestically and sexually exploited. The cogendered offending illustrated in suspect Class 2 in relation to sexual exploitation and domestic servitude is representative of reported cases of couples involved in domestic servitude cases. 2 Targeted campaigning in relation to domestic servitude in this region along with multi-agency models of working that involve groups representing local populations may have assisted in creating a positive environment for reporting domestic servitude activity. In areas where domestic servitude remains harder to reach, these approaches can support progress to identify such activities.
Labour exploitation of adult men was the next most common category of exploitation in the region studied. The increased reporting of labour exploitation since the advent of legislation in this area has seen a shift from the initial prominence of sexual exploitation. Adult males can face particular barriers in reporting exploitation. Some potential victims of labour exploitation assume that they are not being exploited because they are earning more than they would in their countries of origin. Others are aware of their exploitation but choose not to report because they are fearful of deportation as 'illegal workers'. There is thus a need to understand more about how victims of labour exploitation in the UK define themselves, what they understand about their rights under labour and criminal laws, and how they can be supported through the criminal justice process. Research pertaining to victims and perpetrators of labour exploitation remains 'limited and fragmented' in the UK Cockbain and Brayley-Morris 2018) and the UK points based immigration system announced in early 2020 may further impact on the marginalization of people who need to work on the periphery of the legitimate labour market (Broad and Gaddd, 2020).
Prevalence probabilities identified 'Criminally exploited boys' as an emerging victim group in the police region studied (and is a problem identified across the UK (Guardian, 2020)). Section 45 of the Modern Slavery Act 2015 provides that those involved in criminal activity have a defence for activity that they were compelled to do. This legal clause had the potential to move the UK policy agenda regarding young people and vulnerable adults involved in exploitative criminality away from criminalization and towards supportive safeguarding approaches. However, despite improvements in the multi-agency response to those with complex needs, there remains a need to better identify and raise awareness of criminal exploitation and for statutory and non-statutory agencies to work with parents and the wider community to protect children and vulnerable adults (HMICFRS, 2018). The difference between the suspect and victim age groups in criminal exploitation evidences the challenges young people face when they become involved in gangs: some doing so because they are vulnerable; others becoming vulnerable because of debt or drug dependence. Such young people can, because of the effects of cumulative disadvantage and previous criminalization, struggle to find alternatives to criminal lifestyles and are often exploited by other young people whom they regard as peers, or by family members upon whom they are reliant (HMICFRS, 2018). Policing approaches that work strategically in collaboration with relevant services, such as youth justice, education, and substance use NGOs could most effectively identify and support young people on the edge of these activities. The very young age of some perpetrators also underlines the importance of working with young people in order to identify their needs and to divert them from the criminal justice system where possible.
The LCA did not identify a specific group of female traffickers (within reasonable statistical bounds, as explained above). However, this does not mean that there were no female traffickers, indeed, they represented 21% of the suspect group. Women's involvement in modern slavery may be too diverse to categorize, as they are sometimes accomplices or former victims (Broad, 2015). In accordance with the international picture (UNODC, 2018), the high proportion of female offenders-relative to women's involvement in almost all other crime types-raises questions about the role investigative and prosecutorial processes play in shaping the profile of traffickers known to the criminal justice system. More research relating to women's and men's roles in modern slavery activity and about sentencing for these offences would help to increase understanding in this area.

Limitations
These data only represent the beginning of the criminal justice process. The pipeline for modern slavery from investigation to prosecution is complex and resource intensive. Prosecutions are largely dependent on victim evidence which can be difficult to secure. Unfortunately, the final outcome of the prosecution was not recorded in this dataset due to limited information exchange between the police and Crown Prosecution Service and a lack of resources to manually follow up on and input sentencing outcomes, so it was not possible to know how many of these cases resulted in a conviction for modern slavery, or other offences. However, the numbers of convictions nationally were 251 (2016), 444 (2017), and 377 (2018) which remain relatively low considering the number of crimes reported in this one police area (Home Office, 2019).
This study is an analysis of cases known to the police. Thus, this sample may represent a subsample of high-risk/public-facing offenders and particularly vulnerable victims which impedes generalizability to wider populations. In addition, there were considerable data quality issues, as described above. There are also inherent limitations associated with LCA; especially the subjective nature of selecting the appropriate model (Liu et al., 2019, p. 189). In naming classes, it is also important to note the probabilistic nature of LCA; not all individuals within a latent class share identical characteristics (Lanza et al., 2007).

Conclusion
Our findings illustrate that trafficking, as conceived through police reports, is structured by sex, generation, and migration. Our study suggests it is important not to define trafficking as a problem of exclusively male offenders exploiting women and children victims, even though this is commonplace. Recognizing heterogeneity amongst offenders and victims is key to understanding local patterns of modern slavery and devising policing approaches that are responsive to its specific manifestations. It is also critical for recognizing the unintended consequences of legislation designed to safeguard vulnerable and young people, for example, when clauses meant to protect those forced into exploitation lead to them being targeted by exploiters. Anti-slavery policing is complex and as demonstrated through these analyses, must be flexible and diverse in order to identify the spectrum of exploitative activities involved in these offences. These findings suggest the need for police intervention that is supported by multi-agency models to support victims, particularly young people who may be perceived as offenders and women exploited in private homes. The proportion of repeat offenders also suggests the need to ensure monitoring and information sharing in relation to these individuals as they seek more opportunities to exploit others. However, the breadth and complexity of these activities also underlines the need for adequate resourcing for police and other relevant services in order to carry out anti-slavery practice effectively.
We hope these findings will inspire further research to understand the trajectories of traffickers and their victims within the criminal justice process as well as sensitivity to the specifics of different offence and victimization patterns. Police data are necessarily a representation of a particular policy problem, and whilst this can pose problems, it nonetheless provides a useful starting point for understanding local manifestations of modern slavery from which wider lessons can be learned.