The “Golden Thread”: Coercive Control and Risk Assessment for Domestic Violence

Abstract

Research on risk assessment for domestic violence has to date focused primarily on the predictive power of individual risk factors and the statistical validity of risk assessment tools in predicting future physical assault in sub-sets of cases dealt with by the police. This study uses data from risk assessment forms from a random sample of cases of domestic violence reported to the police. An innovative latent trait model is used to test whether a cluster of risk factors associated with coercive control is most representative of the type of abuse that comes to the attention of the police. Factors associated with a course of coercive and controlling conduct, including perpetrators’ threats, controlling behavior and sexual coercion, and victims’ isolation and fear, had highest item loadings and were thus the most representative of the overall construct. Sub-lethal physical violence—choking and use of weapons—was also consistent with a course of controlling conduct. Whether a physical injury was sustained during the current incident, however, was not associated consistently either with the typical pattern of abuse or with other context-specific risk factors such as separation from the perpetrator. Implications for police practice and the design of risk assessment tools are discussed. We conclude that coercive control is the “golden thread” running through risk identification and assessment for domestic violence and that risk assessment tools structured around coercive control can help police officers move beyond an “incident-by-incident” response and toward identifying the dangerous patterns of behavior that precede domestic homicide.

Keywords

domestic violence coercive control risk assessment police

Introduction

Extant studies relating to domestic violence risk assessment have focused overwhelmingly on the validity of risk assessment tools in predicting further physical assaults. This bias in the literature toward predictive validity exists despite a lack of consensus as to whether a risk-based approach to addressing domestic violence should focus on (one-time) predictions of the likelihood of revictimization or on managing and reducing the risk of further abuse (for a discussion see Bennett Cattaneo & Goodman, 2007). Proponents of a “risk management” approach argue prediction is problematic because the abuse in many intimate relationships is a process, as opposed to a series of discrete acts, and revictimization a virtual given in respect of a majority of cases that come to the attention of, for example, criminal justice agencies. In addition, being correct about the likelihood of further abuse does not necessarily equal being helpful to the victim (see Bennett Cattaneo & Goodman, 2007). The goal of risk assessment, it is suggested, should be to identify sources of risk and match them to interventions; the output of the process a safety plan, as opposed to a statistical measure.

Research on practitioners’ use of risk assessment tools has suggested assessors frequently use only a sub-set of the available information when classifying the level of risk (see Robinson & Howarth, 2012). This finding is problematic, as it raises the prospect of inconsistency in the absence of robust narrative to help guide practitioners’ judgments. Curiously absent, then, from much of the literature on the development of risk assessment instruments for domestic violence is the application of theories of abuse in intimate relationships. Although existing risk assessment tools draw on what Kropp (2008) suggests is a well-established set of empirically validated risk factors, there is little sense, even for tools that require assessors to apply “professional judgement,” of how these factors can or should be prioritized, particularly in combination (see also Robinson, 2010).

There is evidence that cases involving “coercive control” are more likely to result in serious harm, including domestic homicide, than cases that involve discrete acts of physical violence (see Campbell et al., 2003; Dobash & Dobash, 2015; Myhill, 2015; Stark, 2007). Stark (2013) defines coercive control as “a strategic course of oppressive conduct that is typically characterized by frequent, but low-level physical abuse and sexual coercion in combination with tactics to intimidate, degrade, isolate, and control victims” (p. 18). While acts of physical violence may be present, they are not required in all cases to instill the level of entrapment Stark (2013) contends should be the principal calculus of harm. For police practitioners, however, the narrative that has developed around risk and harm appears, perhaps not surprisingly, to have coalesced around a “violence model” (Stark, 2007, p. 11) of discrete acts of (injurious) physical violence, as opposed to patterns of abusive behavior that may or may not involve injurious violence.

Using data from risk identification interviews with victims in an English police force area, the study presented here tested whether physical assault with injury or coercive and controlling behavior was most indicative of cases of domestic violence that came to the attention of the police. The remainder of this article is structured as follows. First, we briefly outline the types of risk assessment for domestic violence and evaluate them against the backdrop of two competing theoretical approaches: coercive control and the “violence model.” Then, we test a set of hypotheses that derive from this theoretical discussion using a random sample of cases reported to a police force in England. We conclude with a discussion of the findings and their implications.

Types of Risk Assessment for Domestic Violence

Kropp (2008) recognizes three types of risk assessment for domestic violence: unstructured clinical judgment, actuarial methods that employ statistical algorithms to make “one-time” predictions, and structured professional judgment (SPJ) models that allow the assessor to apply specialist knowledge to overrule scores obtained by summating established risk factors in cases where they may simply “know otherwise” (Walklate & Mythen, 2011, p. 110).

Risk assessment tools for domestic violence specifically began to appear in the 1980s and 1990s. Starting with the Danger Assessment (DA; see Campbell, 1986), there are now a number of actuarial tools, such as the Ontario Domestic Assault Risk Assessment (ODARA; see Hilton et al., 2004), and SPJ tools, such as the Spousal Assault Risk Assessment (SARA; see Kropp, Hart, Webster, & Eaves, 1995). The majority of research studies on risk assessment for domestic violence have focused on the accuracy with which specific tools can predict future assaults. A meta-analyses by Messing and Thaller (2013) concluded broadly that existing risk assessment tools had moderate levels of predictive validity and that it was difficult to advocate one over another due to methodological variability in the validation studies and the different stated purposes of the tools themselves.

Risk of What?

Coercive control has become in recent years a prominent theory relating to violence and abuse in intimate relationships. Though notions of power and control have been reflected in feminist theories of intimate violence since the 1970s, the concept of coercive control as a distinct form of domestic violence has only recently influenced directly public policy in England and Wales. Evan Stark (2007, p. 198) presents a theory of coercive control as a course of conduct comprising multiple abusive behaviors and tactics, and distinguishes it from the low-level and bi-directional physical “fighting” and psychological aggression that others have termed “situational” violence (see Johnson, 2008).¹

Though some elements of coercive control are analogous to multidimensional measures of psychological and emotional abuse that have been tested (see, for example, Murphy & Hoover, 1999), coercive control goes beyond psychological aggression in several key respects. While both men and women in ostensibly healthy relationships may from time to time engage in verbal abuse or some degree of monitoring of their partner’s activity, coercive control involves the continuous “micro-regulation” of everyday life (see Stark, 2007). Coercive control often involves physical violence and sexual coercion to some degree, and is more likely than situational couple violence to involve severe forms of violence and have harmful consequences for victims (Myhill, 2015). Severe acts of violence may not always be present, however, at least initially. Previous studies have suggested psychologically abusive behaviors are predictive of future physical abuse (see, for example, Murphy & O’Leary, 1989), and, crucially, the everyday process of surveillance, threats, and low-level but routinized coercion is sometimes sufficient for an abuser to maintain control without recourse to injurious physical assaults. The impact on the victim however of this continuous form of abuse is cumulative and ultimately devastating.

Perhaps the crucial difference between coercive control and situational violence is that, whereas the latter is perpetrated to some degree by both men and women, and with many of the same motivations (see Johnson, 2008), coercive control is highly gendered (see Myhill, 2015). Though it may be possible for men to suffer coercive control at the hands of female partners (see, for example, Hines & Douglas, 2010), according to Stark (2007) much of the abusive behavior that characterizes coercive control operates through sexual inequality. So while situational fighting may arise from attempts to reconcile disputes that arise in intimate relationships, coercive control, notwithstanding victims’ defensive or retaliatory violence, is a unilateral attempt at denigration and denying a person their basic human rights, and it most often occurs through constructions and deconstructions of gender identities.

What is apparent then from extant research is that there are different forms of domestic violence that have different underlying dynamics and represent different levels of harm. The challenge of risk assessment for the practitioner is to identify accurately which form of domestic violence he or she is presented with, and recognize the degree of harm it is likely to represent.

A Critique of Risk Assessment Tools for Domestic Violence

The theory of coercive control, which recognizes the importance of multiple abusive tactics and behaviors beyond physical assault, raises questions for a literature on risk assessment that has focused thus far primarily on the statistical precision with which specific tools can predict future physical assaults. Most tools were developed following literature reviews to identify factors proven empirically to be associated with committing or experiencing domestic violence (see, for example, Kropp et al., 1995), which those in the field suggest are now widely accepted (Kropp, 2008; Robinson, 2010). Correlation does not necessarily equal causality, however, and “when factors become too numerous . . . we are in the hopeless position of arguing that everything matters” (Matza, 1964, as cited in Wikström, 2012, p. 53).

The developers of some risk assessment tools have, consequently, used statistical modeling techniques such as logistic regression to reduce the number of items and/or create scoring criteria to “upweight” specific risk factors (see Campbell, Webster, & Glass, 2009; Hilton et al., 2004). Such analysis has tended to prioritize factors associated with physical violence. Regression models have identified specific risk factors that are associated with (physical) victimization “controlling for” or independent of other factors in the model. This approach cannot, however, identify clusters of factors that might represent an underlying pattern of abusive behavior. Coercive control theory, rather than focusing on individual risk factors operating independently of one another, is based on the notion that the perpetrator employs a range of abusive tactics that, in conjunction, serve ultimately to entrap the victim.

An additional issue with extant research is that regression models require an objective measure of risk of future harm. The samples that have thus far been used to identify factors that predict revictimization have, however, been drawn from a sub-set of cases where the abuse has likely already escalated, such as cases known to the police where physical violence has occurred previously (Hilton et al., 2004), cases where victims are engaged with specialist support services (Robinson & Howarth, 2012), and cases of domestic homicide that are small in number relative to the volume of domestic violence in the population (Campbell et al., 2009). This issue would be less problematic if risk assessment for domestic violence was undertaken only in such “clinical” settings. Some tools have though been designed or adapted for use by frontline police officers, who respond to a much wider variety of incidents. These tools also follow the violence model to a large extent. Messing et al. (2014) reported findings from an evaluation of the “Lethality Assessment Program,” an element of which is a “lethality screen”—a shortened 11-question DA for use by frontline police officers. Victims are screened in as “high danger” if they respond positively to any of the first three questions, concerning use of weapons and threats to kill. They are also considered high danger if they respond positively to four of the subsequent eight questions, only two of which relate to coercive and controlling behavior. The ODARA, also designed to be administered by frontline police officers, follows broadly the violence model in prioritizing current and previous physical assaults (see Hilton et al., 2004).

This bias toward physical violence would not pose a problem for police risk assessment if the typical profile of abuse that comes to the attention of the police is characterized by injurious physical assaults. If, however, the typical profile of abuse is coercive control that may or may not involve regular physical assaults, then existing risk assessment tools may contribute to a proportion of high risk cases staying “under the radar” (Robinson, Pinchevsky, & Guthrie, 2016). Also of concern would be the issue of “false positives.” It is likely that some proportion of incidents attended by the police involve situational couple violence—disputes that become “violent enough or public enough that either the victim or bystanders call” (Johnson, 2008, p. 76). Such incidents likely present little or no ongoing risk, but may be prioritized by existing risk tools over cases that involve little injurious physical violence but high levels of entrapment.

The Present Study

The focus of the present study is to evaluate individual items on a risk assessment to see which represent most consistently domestic violence as experienced by victims whose situation comes to the attention of the police. As our goal is to examine the interaction of individual risk factors for domestic violence, as opposed to the independent association between individual items and an outcome variable, a latent variable or factor analysis approach is required. Assuming some relationship between established individual risk factors and domestic violence as a broader construct, a latent variable approach permits identifying not only whether a particular cluster of factors is most representative of domestic violence that comes to the attention of the police, but also which factors are associated most strongly with other risk factors across the construct—factors that may represent abusive behavior less consistently or, perhaps, only in a specific context.

The present study tests specifically the theory that risk factors indicative of coercive control will be most representative of a latent construct of domestic violence that comes to the attention of the police. While numerous behaviors and tactics comprise coercive control, certain elements are central to a course of conduct. Stark (2007) breaks the concept into four key sets of behaviors: violence, intimidation, isolation, and control. Degrees of confinement and isolation are essential elements in all scenarios that involve coercive control and “the restriction of the battered woman’s free movement is probably the most important technique” (Okun, 1986, p. 116). We expect therefore to find perpetrators’ jealous, controlling, and stalking behavior and victims’ sense of isolation to be key indicators. Similarly, we expect to find sexual coercion and intimidation through threats to be important. We also expect victims’ perceptions of the abuse they are suffering be key indicators (see Wheller & Wire, 2014). Victims’ sense that the abuse is getting worse is indicative of an escalating course of conduct, as is the presence of the generalized sense of fear that distinguishes victims of coercive control from those engaged in situational conflict.

We also expect acts of severe or sub-lethal physical violence, such as choking and use of weapons, which are not consistent with situational “fighting,” to be consistent with a coercive and controlling course of conduct. Sub-lethal (and ultimately lethal) violence is present in many cases of coercive control, especially if the victim has attempted to separate from the abuser, or indicated that they intend to. We expect, however, the simple presence of physical injury during the current incident not to be indicative of the typical pattern of abuse, due to extant research suggesting first that physical violence is not prominent in all cases of coercive control, and, second, that where present it frequently takes the form of low-level but repeated coercion (see Stark, 2013).

The present study advances the literature on risk assessment for domestic violence in three key respects. First, we bring to bear existing theories of abuse in intimate relationships to explore whether risk factors representative of a course of coercive and controlling conduct are most representative of abuse to which the police respond. Second, we use a random sample of incidents that came to the attention of the police, as opposed to a sub-set of cases containing only criminal offenses, domestic homicides, or higher risk cases drawn from support services. And third, we believe this study is the first to apply a latent trait model in an attempt to identify clusters of, as opposed to individual, risk factors that might be prioritized in relation to risk identification, assessment, and management. Our analysis tests the following specific hypotheses:

Hypothesis 1 (H1): Defining features of coercive control—isolation, threats, control, sexual coercion, fear—will be most indicative of domestic violence that comes to the attention of the police.

Hypothesis 2 (H2): Severe acts of physical violence at some point in the relationship, such as choking/strangulation and use of weapons, will be consistent with a pattern of coercive and controlling behavior.

Hypothesis 3 (H3): The simple presence or absence of physical injury at the current incident will be less consistently representative of a pattern of coercive control in a random sample of cases reported to the police.

Method

Sample and Sampling Process

In England and Wales, attending officers are required to undertake a 27 question risk identification interview with victims of domestic violence using the national “Domestic abuse, stalking and harassment risk identification, assessment and management model” (DASH). As well as a positive or negative response to a question, officers are expected to record explanatory and contextual information in freetext boxes on a DASH risk assessment form. Based on victims’ responses, officers are required to submit a DASH form, using their professional judgment to allocate a grade of “standard,” “medium,” or “high” risk. The DASH is therefore an SPJ tool, sharing most in common with the DA and SARA, and the risk factors represented by the questions are consistent with those established in the literature (see Robinson, 2010). It is a tool for identifying and managing immediate and future risk of harm; it is not a tool intended simply to predict the likelihood of future discrete acts of abuse.

The data used in the present study are a sample of completed DASH forms held on the Information Management System (IMS) of a medium-sized police force in the south of England. To generate a random probability sample of DASH forms, a list of all incidents “flagged” as “domestic-related”² during the police financial year 2011 to 2012 was extracted from the IMS and represented the sampling frame (n = 22,156). Although the best possible sampling frame available, it was not immune to human error. Police officers may, for example, have failed to classify an incident as “domestic-related,” resulting in some relevant incidents not being included the sampling frame.

The lead author selected a random (n = 600) sample of incidents from the sampling frame, stratified by calendar month to avoid potential seasonality bias. The lead author and a colleague accessed the DASH forms of the sampled cases and transcribed victims’ binary yes/no responses to the DASH questions to a data set. The researchers also interrogated the history of each case—examining case files and risk assessments from previous incidents—to identify whether any DASH forms in the sample had been completed for a “primary perpetrator” rather than the primary victim (for which the DASH interview is intended). A total of 15 such cases were excluded from the sample. The authors further screened the sample for repeat victims. In our sample, only six cases were identified where the police had been called on two occasions in the yearlong sampling period (other cases may have involved repeat calls outside the sampling period, or repeat victimization that had not been reported to the police). Incidents “clustered” in victims could lead to biased standard errors in the analysis; thus, only the most recent incident for each victim was kept in the sample, resulting in five incidents being excluded (in one repeat case, no DASH form was completed). Finally, a further 92 cases (15.7%) in which officers had submitted either no or a blank DASH could not be analyzed. The total number of cases suitable for analysis was n = 488.

Measures

The DASH risk assessment comprises 27 items derived from questions asked of the victim at the scene. Responses to the DASH questions included in the sample were coded binary as 1 = yes, risk factor present and 0 = no, risk factor not present. All but one item are included in the analysis. The excluded item (Question 3—“What are you afraid of?”) is an open-ended follow-up to Item 2 (“Are you very frightened?”). The 26 remaining items include risk factors representative both of coercive control and physical violence. It should be noted though that the DASH form was not designed to represent different “perspectives” on domestic violence. Freetext data on DASH forms shows coercive control can be evident from responses to questions ostensibly measuring physical violence (consistent with Stark’s theory), and from “circumstantial” factors such as conflict over child contact. The DASH items can be grouped, however, into those most representative of coercive and controlling behaviors, those most representative of physical violence, items that represent the victim’s subjective assessment of the abuse, and items capturing what might be termed circumstantial risk factors.

Coercive control

This includes seven binary items: isolation from family and friends, the perpetrator controlling everything the victim does and displaying excessive jealousy, threats to kill the victim, threats to kill the children/dependents, constant texting, phoning, stalking, or harassing, sexual humiliation or abuse, and the perpetrator threatening to commit suicide.

Physical aggression/violence model

This includes six binary items: physical injury in the current incident, past perpetrator attempts at choking, strangulation, suffocation, or drowning the victim, the perpetrator hurting the children; the perpetrator hurting someone else, use of objects/weapons to hurt the victim, and the perpetrator mistreating animals/family pets.

Victim’s subjective assessment of the abuse

This includes four binary items: whether the victim feels the abuse is happening more often, whether the victim feels the abuse is getting worse, the victim feeling very frightened, and the victim feeling depressed or having suicidal thoughts.

Circumstantial factors

This includes nine binary items: children/stepchildren living in the household, current pregnancy or recent birth, conflict over child contact, financial issues, attempts to separate in the past/current separation from perpetrator, whether the perpetrator has been in trouble with police, whether the perpetrator has breached bail or an injunction, whether the perpetrator has problems with alcohol/drug abuse or mental health, and whether there are other persons threatening the victim/who the victim is afraid of.

Statistical Analysis

A latent trait model was used to test the hypotheses. The measurement level of the observed DASH items is categorical, more specifically binary, which precludes standard factor analysis that requires the measurement level of the observed items to be continuous. A latent trait model provides an analogous analytical method for categorical variables. The resulting item loadings can be interpreted similar to factor loadings, the latent trait similar to a “factor” in factor analysis (Bartholomew, Knott, & Moustaki, 2011). The software program Mplus was used to fit the model, and results are presented as standardized item loadings. The significance of item loadings was assessed by examining p values and standard errors of the estimates. Goodness of fit was assessed using the fit statistics available within Mplus, specifically the comparative fit index (CFI; Bentler, 1990) and the Tucker–Lewis index (TLI; Bentler & Bonett, 1980). CFI and TLI values above 0.9 indicate satisfactory model fit.

Multidimensionality of the DASH was explored by fitting a two latent trait model in LatentGold. LatentGold provides Akaike information criterion (AIC) and Bayesian information criterion (BIC) fit statistics to assess whether goodness of fit is improved by adding a further latent trait. BIC suggested adding a second latent trait did not improve model fit (BIC = 11,461.3 [one-trait] vs. BIC = 11,502.4 [two-trait model]). However, the AIC statistic suggested a two-trait model had a somewhat better fit (AIC = 11,243.4 [one-trait] vs. AIC = 11,176.6 [two-trait model]). AIC tends to favor bigger models, sometimes overfitting the data, while BIC penalizes complexity more harshly, favoring simpler models (Kuha, 2004). Inspection of item loadings suggested that adding a second trait did not alter the interpretation of the results; no meaningful grouping of items emerged on the second trait, suggesting the DASH can be summarized adequately by a one-trait model.

Results

Table 1 shows the item wording together with basic frequencies, percentages, and the standardized item loading coefficients of a latent trait analysis with one latent trait. The CFI and TLI goodness-of-fit indicators suggest a satisfactory model fit (CFI = 0.921; TLI = 0.914).

Table 1.

Latent Trait Model.

Items	Yes		Item Loading
Items	Frequency	%	b	SE
Is the abuse getting worse?	144	29.5	0.842	0.031
Do you feel isolated from family and friends?	93	19.1	0.785	0.039
Does the perpetrator try to control everything you do and/or is he or she excessively jealous?	172	35.2	0.782	0.033
Are you very frightened?	212	43.4	0.772	0.038
Has the perpetrator ever threatened to kill you or someone else and did you believe it?	91	18.6	0.719	0.043
Is the abuse happening more often?	167	34.2	0.714	0.039
Has the perpetrator ever attempted to choke/strangle/suffocate/drown you?	98	20.1	0.700	0.044
Does he or she constantly text, call, contact, follow, stalk, or harass you?	120	24.6	0.645	0.047
Do you know if the perpetrator has hurt anyone else?	112	23.0	0.607	0.051
Does the perpetrator do or say things of a sexual nature that makes you feel bad or physically hurt?	48	9.8	0.580	0.060
Has the perpetrator ever breached bail/injunction/[. . .]?	65	13.3	0.527	0.065
Has the perpetrator ever threatened to hurt or kill the children or a dependent?	12	2.5	0.505	0.119
Has the perpetrator ever used weapons or objects to hurt you?	66	13.5	0.502	0.062
Are you feeling depressed or having suicidal thoughts?	119	24.4	0.495	0.057
Are there any children or stepchildren in the household?	85	17.4	0.485	0.071
Has the perpetrator ever hurt the children or dependents?	23	4.7	0.485	0.085
Has the perpetrator ever threatened or attempted suicide?	118	24.2	0.473	0.055
Do you know if the perpetrator has ever been in trouble with the police [. . .]?	314	64.3	0.444	0.057
Has the perpetrator had problems in the last year with drugs, alcohol, or mental health?	261	53.5	0.407	0.056
Has the perpetrator ever mistreated an animal or the family pet?	26	5.3	0.403	0.092
Is there any other person that has threatened you or that you are afraid of?	26	5.3	0.391	0.082
Is there conflict over child contact?	94	19.3	0.369	0.067
Are there any financial issues? For example, are you dependent on them for money?	152	31.1	0.349	0.059
Have you separated or tried to separate from your abuser within the past year?	258	52.9	0.343	0.057
Has the current incident resulted in injury?	124	25.4	0.242	0.064
Are you currently pregnant or have you recently had a baby?	92	18.9	0.125 ns	0.074

Note. Sample size: n = 488. ns = not significant at p < .10. All others significant at p < .001.

The most frequently mentioned and thus most prevalent risk factors in the sample were the victim stating that the perpetrator had been in trouble with the police previously (64.3%), and the perpetrator having alcohol, drug, or mental health problems (53.5%). These perpetrator characteristics were followed in prevalence by a range of characteristics describing the victim–perpetrator relationship: the victim having separated or previously attempted to separate from the perpetrator (52.9%), the victim being very frightened (43.4%), the perpetrator controlling everything and being excessively jealous (35.2%), and the abuse getting worse (34.2%).

Items with low prevalence in the sample concerned mainly violence and threats to children and others: the perpetrator threatening to hurt or kill the children/other dependent (2.5%), the perpetrator having hurt the children or other dependents (4.7%), the perpetrator having mistreated an animal or family pet (5.3%), and the presence of another person who has threatened the victim (5.3%). Sexually abusive behavior was also not especially prevalent (9.8%).

As we turn to the results of the latent trait analysis, it is important to note that these base rates don’t determine the item loadings; item loadings are in fact largely unrelated to the percentage of cases in which a particular risk factor is present. Instead, the model “up-weights,” on the latent trait, items where the response is consistent with the responses to the other items in the scale, and down-weights items where the response is not consistent with the responses to other items in scale. The item loadings thus enable an assessment of each item in terms of how informative or indicative it is as to the case’s overall location on the latent trait. In turn, the items with high item loadings define the latent trait and are suggestive of its interpretation.

The results showed support for H1. The items with the highest loadings were the victim perceiving the abuse getting worse (b = 0.84), the victim feeling isolated from family and friends (b = 0.79), the perpetrator controlling everything the victim does and being excessively jealous (b = 0.78), the victim feeling very frightened (b = 0.77), and the perpetrator making believable threats to kill (b = 0.72). The victim reporting they had suffered sexual abuse of some type was also consistent with coercive and controlling behavior (b = 0.58), as was the perpetrator ever having breached an injunction (b = 0.53). Items with the highest loadings give clues as to the interpretation of the underlying latent trait: perpetrators’ coercive and controlling behaviors formed the consistent response pattern to the DASH that resulted in high item loadings. Furthermore, the findings suggested police officers should heed victims’ instincts: the victim feeling unsafe—stating they are very frightened, and that the abuse is getting worse/happening more often—had high loadings. This finding is consistent with recent reviews of the literature (see Wheller & Wire, 2014).

The results also showed support for H2. Risk factors representative of severe acts of physical violence had medium to high item loadings: past attempts of the perpetrator to strangle, choke, suffocate, or drown the victim (b = 0.70), and the perpetrator using weapons/objects to physically hurt the victim (b = 0.50). Responses to items relating to physical violence alone did not, however, form in this sample a separate and consistent response pattern that was associated with the presence of other risk factors, and also did not form a second latent trait (a latent trait model with two latent traits did not significantly improve model fit, see above).

Finally, the results showed support for H3. The presence of physical injury during the current incident had one of the lowest item loadings (b = 0.24), suggesting it was not indicative consistently of a coercive and controlling situation, or the presence of other risk factors across the latent construct. This is an important finding as it suggests that the “objective” indicator that may be most apparent to police officers at the scene—physical injury—is less useful in discriminating a course of conduct than both non-physical coercive behavior and the “subjective” indicator of the victim feeling unsafe.

Discussion

Analysis of risk assessment data from a random sample of cases showed a number of risk factors indicative of coercive control were present consistently in cases of domestic violence attended by the police. Not only did perpetrators’ jealous, controlling, threatening, and sexually coercive behaviors, and victims’ sense of isolation, fear, and escalation, form a consistent cluster of factors at the heart of a latent construct of domestic abuse, these factors were associated most consistently with other factors across the construct, such as separation from the perpetrator and conflict over child contact. This analysis also showed that indicators of sub-lethal physical violence—such as choking and use of weapons—were consistent with the typical pattern of coercive and controlling abuse. Physical injury at the current incident, however, was at the periphery of the latent construct, suggesting it was associated much less consistently with the typical pattern of coercive and controlling abuse.

This study has implications for both the design and implementation of risk assessment tools, and the police response to domestic abuse more widely. Previous research has shown that police officers tend to prioritize indicators of physical violence when assessing risk, and in particular whether there has been injurious violence during the incident to which they have been deployed (Robinson et al., 2016; Robinson, Pinchevsky, & Guthrie, 2018). A major inspection of the police response to domestic violence in England and Wales also concluded that officers struggled frequently to identify patterns of abusive behavior in the absence of overt physical violence, and suggested that “officers need to see beyond the incident they are dealing with and look at the wider context of the situation they find” (Her Majesty’s Inspectorate of Constabulary, 2014, p. 55). This situation may in part be explained by physical assault being one of the traditional crime categories with which the police are familiar, but Stark (2013) suggests risk assessment tools following the violence model may exacerbate this process of “rendering the typical pattern of abuse invisible in plain sight by disaggregating the ongoing pattern into discrete episodes . . . rather than grasping them in their interrelated whole as victims and their children are experiencing them” (p. 20).

To date, it appears the focus in developing risk assessment tools has been on individual risk factors and their summation and/or weighting, as opposed to applying theories of abuse in intimate relationships to understand how combinations of factors may represent particular patterns of abusive behavior. Indeed, Kropp et al. (1995) suggested “the task of clinical prediction invites evaluators to isolate key variables that might accentuate or diminish the possibility of violence” (p. 2, emphasis added), and “critical items” may be “sufficient on their own” to conclude there is risk of harm (p. 20). There is no sense of how a group of related risk factors might comprise more than the sum of its parts, representing a dangerous pattern of behavior that should be “upweighted” when applying professional judgment to the level of risk. Where specific risk factors are upweighted for risk assessment, it tends to be those associated with the violence model. The revised version of the DA is scored; specific risk factors are upweighted, but not groups of factors. Specifically, factors associated with extreme physical violence (perpetrator’s access to a gun and weapon use; threats to kill) are upweighted, but the perpetrator exhibiting a more general pattern of controlling behavior is not.

While a focus on physical violence will no doubt identify many cases where there is a risk of current and future harm, domestic homicide reviews have identified numerous cases where the context preceding the homicide event was characterized not by frequent injurious assaults, or even sub-lethal violence, but by high levels of coercion and control (Dobash & Dobash, 2015; Monckton-Smith, Williams, & Mullane, 2014; Regan, Kelly, Morris, & Dibb, 2007). This finding has significant implications for practice in England and Wales, where frontline officers’ initial assessments of risk can have a direct bearing on whether a case is referred to specialist police teams and/or victim support services. Furthermore, an emphasis on physical assault, and especially injury during the current incident, may lead to cases being put forward for intervention where violence was in fact a “one off” or sporadic, and where there is no course of (escalating) conduct suggesting risk of serious harm.

Coercive control, by contrast, may be seen as the “golden thread” running through risk identification, assessment, and management. A focus on patterns of controlling behavior permits early identification prior to the onset of physical violence, or escalation to lethal or sub-lethal violence. It also gives an indication of the level of entrapment and consequent potential risk posed to a victim who attempts to escape their situation. Knowledge of the level and frequency of physical violence is clearly crucial to any assessment of risk. As important, however, is knowledge of the full range of tactics employed by an abuser, and the current level of entrapment of the victim. In the present study, it was controlling behavior rather than physical injury that was associated with other risk factors across the construct. Our findings support those of Campbell et al. (2003) in suggesting coercive control provides the context in which circumstantial risk factors such as separation from the perpetrator are likely to be triggered. This finding is important, as our study suggested such factors—including conflict over child contact, and the perpetrator’s mental health and substance abuse issues—are prevalent in a representative sample of incidents. Prioritizing such risk factors out of context again risks generating false positives.

Crucially, a focus on coercive control may help frontline officers move beyond taking an “incident-by-incident” approach to domestic violence. In-depth work with victim-survivors (see Kelly, 1999) has shown that many take a long time to name what they are experiencing as abuse, and that they frequently deploy coping strategies of minimizing the actions of the perpetrator while modifying their own lifestyle and behavior. Minimization by the victim, and manipulation of the situation by the perpetrator, can make it difficult for a first responding officer to see beyond the current incident and recognize the pattern of abusive behavior that underlies it. It will be of increasing importance for frontline officers in England and Wales to recognize such a course of conduct as in December 2015 a new criminal offense of “coercive and controlling behavior in an intimate or family relationship” was introduced. While officers have traditionally been restricted by working only with traditional crime categories such as assault and criminal damage, police in England and Wales now have the power to arrest and lay charges against a perpetrator for a course of (non-violent) abusive conduct.

Kelly (1999) proposed a model for police intervention in domestic violence that recommended officers encourage victims to recognize and acknowledge how they are limiting their own lives in response to the perpetrator’s patterns of behavior. A risk assessment tool that focuses on coercive control may help officers to achieve that goal and raises also the possibility of risk assessment as an intervention in and of itself; one that helps victims to recognize and name their experiences as abuse (see Robinson, 2010). It is likely, however, that, without such specific direction, police officers will continue to focus on risk factors with which they are most familiar and regard as most serious—specifically, those relating to (injurious) physical violence.

There are limitations to the present study, concerned primarily with the suitability of the DASH data for quantitative analysis. The accuracy and completeness of data from victim interviews depends both on the victim’s willingness to disclose and the officer’s ability to elicit and record the information. It is likely victims are less comfortable disclosing certain aspects of the abuse they are suffering, particularly to frontline police officers. Sexual coercion, for example, is a relatively common tactic of coercive control (Stark, 2007), yet was not prevalent in this sample. Perhaps most important is that we could not distinguish accurately whether specific items were not present or not disclosed. If largely blank DASH forms resulted from victims withholding information, as opposed to officers not recording the information, and such withholding was systematically linked to certain cases, bias may have been introduced to some degree.

As discussed, a key problem with this type of research more generally is the absence of a robust benchmark of risk that is broader than homicide/no homicide. A latent trait model is a useful first step in suggesting that a cluster of factors representing coercive control may be helpful in identifying risk in cases of domestic violence that come to the attention of the police. What is required beyond that is an objective measure of risk across a random sample of cases, alongside accurately measured risk factors. That would permit regression analysis with interaction effects to determine how specific factors (such as conflict over child contact, or the presence of stepchildren) and index measures (such as coercive control) operate in combination and in context.

Concluding Remarks

Flexibility is an advantage of a structured judgment model of risk assessment. Yet without a clear, evidence-based narrative to guide structured judgments, it runs the risk of practitioners making inconsistent decisions based on their own preferences and biases, akin to unstructured clinical judgments. The findings of this study suggest coercive control is the “golden thread” running through risk identification and assessment in cases of domestic violence. “Male sexual proprietariness and female attempts to escape male control” have long been recognized as underlying domestic homicide (Daly & Wilson, 1998, as cited in Campbell, Sharps, & Glass, 2000, p. 136). Identifying the “jealous surveillance” that characterizes coercive control would allow resources to be focused more sharply on early intervention (Regan et al., 2007, p. 6). A focus on coercive control does not, however, mean physical violence is unimportant. Physical violence is a key tactic of coercive control, and many victims of controlling abusers experience severe physical abuse. Yet others experience little physical violence, and many experience the type of low-level but repeated violence that fails to register on existing risk assessment tools (Stark, 2013). A focus on acts of physical violence to the exclusion of patterns of abusive behavior more widely will condemn many victims to years of abuse characterized not by injurious physical assaults but by entrapment and continuous coercive bullying. Coercive control, as well as being harmful in and of itself, also provides the context for assessing when other more circumstantial risk factors are more likely to be triggered, to devastating effect.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Notes

Author Biographies

Andy Myhill is an evidence and evaluation advisor at the UK College of Policing. He has published both government reports and academic journal papers on topics including survey methodology, public attitudes to the police, and domestic violence.

Katrin Hohl is a lecturer in criminology at City, University of London, UK. Her research centers on public trust in and legitimacy of the police and the wider criminal justice system, with a recent focus on criminal justice responses to sexual violence.

References

Bartholomew

Knott

Moustaki

(2011). Latent variable models and factor analysis: A unified approach (3rd ed.). Chichester, UK: John Wiley.

Bennett Cattaneo

Goodman

. (2007). New directions in risk assessment: An empowerment approach to risk management. In Kendall-Tackett

Giacomoni

(Eds.), Intimate partner violence (pp. 1-17). Kingston, NJ: Civic Research Institute.

Bentler

(1990). Comparative fit indexes in structural models. Psychological Bulletin, 107, 238-246.

Bentler

Bonett

(1980). Significance tests and goodness of fit in the analysis of covariance structures. Psychological Bulletin, 88, 588-606.

Campbell

(1986). Nursing assessment for risk of homicide with battered women. Advances in Nursing Science, 8(4), 36-51.

Campbell

Sharps

Glass

(2000). Risk assessment for intimate partner homicide. In Pinard

Pagani

(Eds.), Clinical assessment of dangerousness: Empirical contributions (pp. 136-157). New York, NY: Cambridge University Press.

Campbell

Webster

Glass

(2009). The danger assessment: Validation of a lethality risk assessment instrument for intimate partner femicide. Journal of Interpersonal Violence, 24, 653-674.

Campbell

Webster

Koziol-McLain

Block

Campbell

Curry

. . . Laughon

(2003). Risk factors for femicide in abusive relationships: Results from a multisite case control study. American Journal of Public Health, 93, 1089-1097.

Dobash

(2015). When men murder women. Oxford, UK: Oxford University Press.

10.

Her Majesty’s Inspectorate of Constabulary. (2014). Everyone’s business: Improving the police response to domestic abuse. Retrieved from https://www.justiceinspectorates.gov.uk/hmic/publications/improving-the-police-response-to-domestic-abuse/

11.

Hilton

Harris

Rice

Lang

Cormier

Lines

(2004). A brief actuarial assessment for the prediction of wife assault recidivism: The Ontario Domestic Assault Risk Assessment. Psychological Assessment, 16, 267-275.

12.

Hines

Douglas

(2010). Intimate terrorism by women towards men: Does it exist? Journal of Aggression, Conflict and Peace Research, 2(3), 36-56.

13.

Holtzworth-Munroe

Stuart

(1994). Typologies of male batterers: Three subtypes and the differences among them. Psychological Bulletin, 116, 476-497.

14.

Johnson

(2008). A typology of domestic violence: Intimate terrorism, violent resistance, and situational couple violence. Boston, MA: Northeastern University Press.

15.

Kelly

(1999). Domestic violence matters: An evaluation of a development project. London, England: Home Office.

16.

Kropp

(2008). Intimate partner violence risk assessment and management. Violence and Victims, 23, 202-220.

17.

Kropp

Hart

Webster

Eaves

(1995). Manual for the Spousal Assault Risk Assessment Guide (2nd ed.). Vancouver, Canada: The British Columbia Institute on Family Violence.

18.

Kuha

(2004). AIC and BIC: Comparisons of assumptions and performance. Sociological Methods & Research, 33, 188-229.

19.

Messing

Campbell

Wilson

Brown

Patchell

Shall

(2014). Police departments’ use of the Lethality Assessment Program: A quasi-experimental evaluation. National Criminal Justice Reference Service. Retrieved from https://www.ncjrs.gov/pdffiles1/nij/grants/247456.pdf

20.

Messing

Thaller

(2013). The average predictive validity of intimate partner violence risk assessment instruments. Journal of Interpersonal Violence, 28, 1537-1558.

21.

Monckton-Smith

Williams

Mullane

(2014). Domestic abuse, homicide and gender. Basingstoke, UK: Palgrave Macmillan.

22.

Murphy

Hoover

(1999). Measuring emotional abuse in dating relationships as a multifactorial construct. Violence and Victims, 14, 39-53.

23.

Murphy

O’Leary

(1989). Psychological aggression predicts physical aggression in early marriage. Journal of Consulting and Clinical Psychology, 57, 579-582.

24.

Myhill

(2015). Measuring coercive control: What can we learn from national population surveys? Violence Against Women, 21, 355-375.

25.

Okun

(1986). Woman abuse: Facts replacing myths. Albany: State University of New York Press.

26.

Regan

Kelly

Morris

Dibb

(2007). “If only we’d known”: An exploratory study of seven intimate partner homicides in Engleshire. London, England: London Metropolitan University, Child and Woman Abuse Studies Unit.

27.

Robinson

(2010). Risk and intimate partner violence. In Kemshall

Wilkinson

(Eds.), Good practice in risk assessment and risk management (pp. 119-138). London, England: Jessica Kingsley.

28.

Robinson

Howarth

(2012). Judging risk: Key determinants in British domestic violence cases. Journal of Interpersonal Violence, 27, 1489-1518.

29.

Robinson

Pinchevsky

Guthrie

(2016). Under the radar: Policing non-violent domestic abuse in the US and UK. International Journal of Comparative and Applied Criminal Justice, 40(3), 195-208.

30.

Robinson

Pinchevsky

Guthrie

(2018). A small constellation: Risk factors informing police perceptions of domestic abuse. Policing and Society, 28, 189-204. doi:10.1080/10439463.2016.1151881

31.

Stark

(2007). Coercive control: How men entrap women in personal life. Oxford, UK: Oxford University Press.

32.

Stark

(2013). The dangers of dangerousness assessment. Family & Intimate Partner Violence Quarterly, 6(2), 13-22.

33.

Walklate

Mythen

(2011). Beyond risk theory: Experiential knowledge and “knowing otherwise.” Criminology & Criminal Justice, 11, 99-113.

34.

Wheller

Wire

(2014). Domestic abuse risk factors and risk assessment: Summary of findings from a Rapid Evidence Assessment. London, England: College of Policing.

35.

Wikström

(2012). Does everything matter? Addressing the problem of causation and explanation in the study of crime. In McGloin

Sullivan

Kennedy

(Eds.), When crime appears (pp. 53-72). New York, NY: Routledge.