Cumulative lifetime violence severity scale: development and initial testing among men

Experiences of violence can be traumatic events that evoke an adaptive acute stress response [10] that protects vital body functions through a process of allostasis [11]. However, recurrent intense stress such as that from violence experiences can produce allostatic overload that leads to dysregulation of the body’s natural stress response system, causing significant and persistent, biophysical (e.g., neuroendocrine, metabolic, immunologic) changes to the body and brain that have been implicated in the development and progression of many chronic diseases [12, 13]. An essential mechanism for precipitating changes in health and well-being in response to exposure to violence and other trauma is a distress reaction; that is, how distressing the experience is perceived to be [14]. Therefore, a key dimension of violence severity in health research is the degree of distress experienced from exposure to violence. Some evidence shows that how men conceptualize violence and how distressed they are consequent to being targets and/or perpetrators of violence differs from women [15]. Violence exposure, particularly IPV, for men often is measured using tools that were originally developed for women [16] and ignores gender, the lens through which violence is perceived, evaluated, expressed, and experienced [17, 18]. Thus, other critical underpinnings of the CLVS scale are gender and distress.

Researchers in violence and health often rely on a “more is worse” concept; that is, the dose-response relationship between health outcomes and the frequency of abuse exposure and/or the number of types of abuse, a practice that implies homogeneity among experiences of violence [1]. The concept of violence severity has been operationalized across studies differently, including by higher frequencies, more serious injuries, potential to cause harm, presence of a weapon, or accounts of the target or perpetrator only [19]. Some scales have pre-classified specific acts as more or less severe based on independent assessment of potential to cause harm, or egregiousness, by persons who may never have had similar experiences [20‐22]. When measures of violence severity do not consider the perspective of the target or perpetrator, the classification of severity is arbitrary because it is not made relative to an individual’s context [6]. Severity is a complex construct with potential to reflect a continuum in problem magnitude through multiple dimensions including frequency and multiplicity of behaviors, subjective appraisal of the problem seriousness, polyvictimization, and intrusiveness of behavior [5].

Together, these conceptual underpinnings informed the dimensions of CLVS included in our scale development. We also drew on our 20-year history in studying violence and health for women and men, using qualitative and quantitative methods [23‐25]. The scale development process included multiple steps, each of which was approved by our institution’s research ethics board. The conceptualization evolved and was revised at each step resulting in the final scale evaluated here. This work was conducted as part of the Men’s Violence, Gender and Health Study (MVGHS), conducted in Eastern Canada with 685 men ages 19 to 65.

Revisions based on feedback from experts resulted in a pilot measure of 49 items divided into 4 subscales: childhood target (13 items); adult target (16 items); childhood perpetrator (8 items); adult perpetrator (12 items). We pretested the measure with a convenience sample of 53 English-speaking men, ages 19 to 65 years, who resided in New Brunswick (NB) and agreed to participate in a pilot study of gender, violence, and health. Acceptable levels of internal consistency (Cronbach’s alpha = > .70) were obtained for all frequency and distress scales except childhood perpetration frequency (α = .63). Concurrent validity was supported by statistically significant positive correlations among the violence subscales and measures of depression and pain, suggesting men reporting more violence, whether as a target or perpetrator, tended to report higher levels of depression and pain. Lack of variability in men’s responses raised questions about the utility of some items and resulted in modification and/or deletion. The greatest variability in responses and highest scores were observed for the childhood violence target subscale, perhaps suggesting men were more willing to report childhood target experiences.

Through examination and discussion, we recognized that in an effort to reduce burden for respondents, we had included some items that tapped into multiple types of violence (i.e., physical, psychological, sexual) in more than one context. We reworked items so that each was conceptually distinct and represented only one cell in the CLVS dimension table (see Table 1). Specifically, we refined each item to measure one type of violence, in childhood or adulthood, as target or perpetrator, in one particular context (e.g., family, peer, workplace) to be scored in terms of frequency and distress. By averaging the frequency and distress scores for each item, we created a score for perceived violence severity; that is, how harsh, serious, forceful, unpleasant, and/or demanding that violence experience was felt to be by the respondent. This takes into account that distress initiates negative biophysical responses and permits the capture of the heterogeneity of men’s experiences that is often overlooked by using frequency only. For example, a man who scored frequency as rarely (2), and distress as very often (5) would have a severity score of 3.5, whereas another man who scored frequency as rarely (2), and distress as rarely (2 would have a severity score of 2. The subsequent revised CLVS scale had 64 items.

To explore stability of the measure, we recruited 31 additional NB men to complete CLVS-64 scale twice, 2 weeks apart. The test-retest reliability of the perceived CLVS-64 scale was r_s = .88, suggesting stability in the measure. Variability in responses remained poor for some items measuring sexual violence and for others measuring perpetration, with scores tending to be skewed toward the low end of the scale. Conceptually, these items were vital for capturing all dimensions of cumulative lifetime violence. We concluded that the poor distribution might be a result of the small sample, and opted to retain these items for the main study.

The study survey included the CLVS-64 as well as self-report questions on demographics, health, and gender. For each CLVS-64 item, we asked men to score on a 5-point Likert-like scale how often (never to very often) and how distressing (not at all to very). Because there is no existing measure of cumulative lifetime violence severity reflecting a gold standard for criterion validity, we used 4 global measures, each using a 10-point rating scale, for the frequency and distress of lifetime experiences of violence as perpetrator and target to determine convergent validity for the scale [26]. We summed and averaged these global measures for a Global Lifetime Violence Severity (GLVS) score of 0 to 10. We assessed concurrent validity with measures of depression, anxiety, and posttraumatic stress disorder (PTSD) because the relationship between each of these mental health problems and men’s exposure to violence such as IPV [27], workplace bullying [23] and child maltreatment [28] is well-established.

The Center for Epidemiological Studies Depression Scale Revised (CESD-R) is a 20-item 4-point Likert-type scale (rarely to most of the time) to assess depressive symptom frequency in past 2 weeks [29]. Summative scores range from 0 to 60 with higher scores indicating more depressive symptoms. Van Dam and Earleywine reported reliability and validity among diverse gender, age and community samples [30]. In this study, α = .95 (N = 685) [30]. The Generalized Anxiety Disorder Scale (GAD-7) is a 7-item, 4-point Likert-type scale (not at all to nearly every day) to measure severity of anxiety symptoms over the previous 2 weeks following DSM-IV criteria [31]. Summative scores range from 0 to 21, with higher scores indicating greater severity of symptoms. Reliability and construct validity have been established in the general population [32]. In this study, α = .94 (N = 685). The Posttraumatic Stress Disorder Checklist, Civilian Version (PCL-C) is a 17-item, 5-point Likert-type scale (not at all to extremely) to measuring how much the respondent has been bothered by PTSD symptoms based on DSM-IV criteria over the past month [33]. Summative scores range from 17 to 85 with higher scores reflective of greater symptomology [34]. Conybeare et al. reported good internal consistency, test-retest reliability, and convergent and discriminant validity in a non-clinical sample [35]. In the current study, α = .95 (N = 685).

We recruited and collected data in two steps. First, we recruited men from the province of NB using posters and online classified advertisements. Using phone or email, interested men contacted the research coordinator who forwarded a letter of information and an online link for eligibility and consent. After receiving their consent, the research coordinator directed participants to the online survey. Of 825 men who were eligible and gave informed consent, 611 (74%) completed the survey and received an honorarium of 20 Canadian dollars to acknowledge their time.

Our missing data were minimal (i.e., ≤ 5%). Missing values were replaced by case mean substitution if case missing values were fewer than 30% in validated health scales and 20% in survey-specific violence scale items [26]. The sample size with complete data on violence items was 590, large enough to develop a stable factor analysis solution [36]. Corrected item-scale correlations are an indicator of how well the item measures the construct; experts differ on the cut score of corrected item-scale correlations for deletion of items with some indicating .20 and others .30 [26]. For the CLVS scale, only 11 items had scores less than .30, and 5 of these, all focusing on sexual violence, had scores less than .20. Conceptually, retention of sexual violence items was important for this scale; therefore, deletion of these items was not feasible. The mean inter-correlation among the 64 items was low at .195 with a range of .06 to .768. Although a number of correlations were less than .30 suggesting some items might have limited congruence with the scale, there was no evidence of over-redundancy as no items were correlated above .80 [26].

Examination of individual violence items revealed poor variability with distribution skewed toward the floor of the scale, particularly for some items focusing on sexual violence and/or on violence perpetration. Combining the upper scores of often and very often for frequency and quite a bit and very for how distressing resulted in improvement in distribution; however, sexual violence items and some perpetration items still failed to use the highest possible score and were skewed toward the lowest scale score. Normal distribution of variables is not necessary if PCA is used to summarize the relationships among a large set of observed variables but can enhance the solution [36]. However, Field noted that if the intent is to use the analysis beyond the sample collected, roughly normally distributed data is more important [37]. Since this analysis is the first step toward developing a measure of CLVS to be widely used with men, better distribution of data was judged to be important. Therefore, we recruited another 100 participants to complete the online survey. We modified the inclusion criteria to include participants living in the provinces of Nova Scotia and Prince Edward Island and required self-identification of having experienced violence in their lifetime. We reasoned that a broader catchment area would increase recruitment and that those who self-identified lifetime violence experience might be more likely to have responses toward the higher end of the scale. After data cleaning, the total sample size was 685. For a descriptive profile of the participants in the study, see Table 2. Of the 685 men, only 15 (2.2%) reported no experiences of violence in their lifetimes; 567 (82.8%) reported experiences as target and as perpetrator, 100 (14.6%) as target only, and 3 (0.4%) as perpetrator only.

In order to minimize skewness in the violence items, we recoded violence frequency and distress scores to a 4-point scale in which 4 represented the upper two scores of the original scale and calculated severity scores (range 1 to 4) for each item by averaging the frequency plus distress score. With respect to normality in item distribution in the sample of 685, 58 of 64 items used the full range of severity scores from 1 to 4, with 5 items having a maximum score of 3.5, and 1 having a maximum score of 3. Although Kline suggested that a skewness score greater than 3 is an indicator for deletion of items, we retained the 14 items that exceeded 3 because most were about sexual violence and considered conceptually important [38]. Of the 64 items, 7 had corrected item-scale correlations less than .3 and all measured sexual violence. One, perpetration of sexual violence at work, had a corrected item-scale correlation of less than .2 and was deleted from the scale because it was unlikely to correlate with the other items [26, 36]. Among the remaining 63 items, the mean inter-item correlation was .218 (range − .023 to .754) and most were significant (<.05). Significance suggested reliability in relationships among pairs of items, an indicator of the matrix being factorable [36]. As well, a matrix should include several sizable (>.30) inter-item correlations, although larger sample sizes may produce smaller correlations. In our sample, 57 items had correlations above .30 with at least 4 and as many as 26 other items suggesting that the items overall are suitable for factor analysis. Thus, data were deemed suitable for PCA.

Using an oblique rotation to begin, we found correlations among components to be, with one exception, less than .30 indicating the components were largely uncorrelated; therefore, we judged a PCA with orthogonal rotation to be most appropriate [39]. With 63 items, we ran the PCA with orthogonal varimax rotation, specifying generation of a Scree plot, extraction of components based on eigenvalue > 1, display of coefficients sorted by size, and suppression of coefficients < .32 because these cannot not be interpreted in PCA [36]. This initial analysis provided further support for factorability of the data set. In the anti-image correlation matrix, the negatives of pairwise partial correlations coefficients adjusted for the effects of all other items were very small, indicating factorability. Additionally, the Kaiser-Meyer-Olkin (KMO) test, a ratio of the sum of squared correlations among items to the sum of squared correlations plus the sum of squared partial correlations, was .91; the closer this is to 1, the better for factor analysis [36, 39]. Additionally Measures of Sampling Adequacy (MSA), which is the KMO for individual pairs of items, were all above .810, greater than the .5 required for sampling adequacy [37]. The output from the initial Varimax rotation showed 14 components, all with eigenvalues > 1, that accounted for 62.85% of the variance. The Scree test showed a break in trajectory at 11 and 13 components. Thurstone as cited in Tabachnick and Fidell [36] suggested simple structure in the solutions; that is, each component should have several variables with high loadings (i.e., above .50) but each variable should have a primary or high loading on only one component. Following rotation, items without primary loadings on a single factor can be eliminated. Cross loadings (loadings on more than one component) that are separated by less than .20 from the primary loading are also problematic as these indicate a complex variable that contributes to more than one construct. We aimed for a simple solution by deleting items one at a time, first on the basis of lacking a primary loading greater than .50, and then on the basis of cross loadings separated by less than .2 from the primary loading. We ran a new PCA after each item was deleted. We deleted items starting from the bottom of the rotated factor matrix, reasoning that the components at the bottom accounted for less variance, consisted of fewer items, and likely were less stable.

Using this approach, we ran many PCAs with the 63 variables using varimax and quartimax rotations and examined solutions with differing numbers of components. Varimax rotations provided simple solutions that were quite consistent across 9, 10, and 11 factor components, although we judged the 11component, 44-item solution to be the best match for our conceptualization of CLV. To confirm this solution, we also randomly selected 400 cases from the sample of 685 and ran the PCA first on the 400 and then on the remaining 285. The 400-case PCA yielded a 10 component, 35-item solution and the 285-case PCA yielded a 9 component, 33 item solution. Both were remarkably similar to the 11 component solution with components that accounted for the majority of the variance similar in all three solutions. The consistency across solutions even with the smaller samples provided support for the original 11 component solution.