Participants

Participants were 600 students aged between 14 and 18 years (M = 15.54; SD = 1.22) from five public high schools in Seville and Córdoba, with medium and medium-low economic, social and cultural status. Of these initial participants, only those who had been in a romantic relationship over the past year and had an Internet-enabled mobile telephone during that time were selected for our study. The final sample comprised 307 adolescents aged between 14 and 18 years (M = 15.57; SD = 1.26). Of these, 46.9% were boys (n = 144), 51.8% girls (n = 159), 1% non-binary (n = 3), and 0.3% preferred not to say (n = 1). In terms of education level, 31.6% (n = 97) were in the third year of secondary school (ESO), 45.3% (n = 139) were in the fourth year of ESO, 8.1% (n = 25) were in the first year of the Spanish Baccalaureate (SB) and 11.4% (n=11) in the second year of the SB. The 74.1% (n = 223) defined themselves as heterosexual, 20.6% (n = 62) as bisexual, 3% (n = 9) as homosexual, 1.7% (n = 5) said they did not know and 0.7% (n = 2) as asexual.

Instruments

The instrument was developed in 4 phases: a) Literature review; b) Focus groups with young adults; c) Review by experts; and d) Creation of the final instrument.

a) Literature review

We reviewed the extant literature on adolescent cyber dating violence with the aim of identifying the theoretical dimensions, the most representative behaviours of the construct and the principal limitations of previous studies (^{Kim & Ferraresso, 2022}; ^{Martínez-Soto & Ibabe, 2022}; ^{Rodríguez-deArriba et al., 2021}). We also reviewed measures of face-to-face dating violence to adapt certain behaviours to the online context (^{Fernández-Fuertes et al., 2006}; ^{Foshee, 1996}). This analysis returned 34 behavioural items that were grouped into three theoretical dimensions: verbal/emotional, control and sexual.

b) Focus groups with young adults

During the second phase, we organised 6 focus groups with psychology, education, and agricultural engineering undergraduates. Participants were 14 boys and 30 girls aged between 18 and 23 years. Participants were given the list of 34 items and were asked to state their opinion in terms of representativeness, severity and perceived types. Participants identified public forms of violence and those that included multimedia content as being the most serious, whereas those entailing control and surveillance were identified as the most frequent. The classification of the items resulted in the differentiation between sexual and non-sexual forms, with those referring to cyber control being the ones identified most clearly. Participants recommended eliminating some items, arguing that they were unrealistic or obsolete in the online environment (e.g., ‘Threatening to damage or break something belonging to one's partner') and recommended using the verb ‘pressuring' instead of other more ambiguous terms such as ‘asking'.

C) Review by experts

Based on the opinions expressed by the young adults, the 34 original items were transformed into 28 (Table 1). Some items were divided into two, depending on the audience or witnesses of the aggression (e.g., public insults vs. Private insults), some were eliminated altogether, and others were grouped into a single item, such as items describing different means of threatening one's partner. The resulting items were reviewed by expert psychologists, who made improvement suggestions and adapted some linguistic expressions. The items were then written using neutral wording so they could be used as double items for both aggression and victimisation. We also ensured that all three behavioural dimensions were represented and that each dimension included public and private behaviours.

The 28 items corresponded to three dimensions:

d) Creation of the final scale

The final Cyber Dating Violence Instrument for Teens (CyDAV-T) comprised 28 double items (for both aggression and victimisation) rated on a 5-point Likert-type scale ranging from 0 = ‘This has never happened to me' to 4 = ‘This almost always happens to me', which measures the frequency with which respondents have experienced the behaviour described over the past year.

Dating status and the use of technology were controlled for through the following questions: ‘Do you have or have you had a romantic partner over the past year?', (response options being 0 = ‘No' and 1 = ‘Yes'); and ‘Did you have an Internet-enabled mobile telephone and profiles of social media (Instagram, Facebook, Twitter, WhatsApp, etc.) during that time?', (response options: 1 = ‘I had neither of these two things', 2 = ‘I had an Internet-enabled mobile telephone but no social media profiles'; 3 = I had an Internet-enabled mobile telephone and at least one social media profile').

To analyse the instrument's evidence of criterion and convergent validity, we included the online moral disengagement scale (^{Paciello et al., 2020}) and the Non-Consensual Sharing (NCS) scale, that measures the non-consensual re-sending of other people's intimate sexual content (^{Walker et al., 2021}). Moral disengagement is closely linked to the explanation of violent online phenomena (^{Lo Cricchio et al., 2021}) and NCS is considered a specific type of sexual cyber aggression (^{Walker & Sleath, 2017}). The mean for the scales was calculated.

Procedure

An accessibility-based sampling technique was used for this study. To recruit participants for the focus groups, the members of the research team posted notices in the first-year classrooms of the faculties of Psychology, Education and Engineering informing about the study aims and the conditions for participation. The decision was made to interview first-year university undergraduates because they were the closest in age to the adolescent population, and because gaining informed consent was simpler with these participants, since they were all legal age.

To recruit adolescent participants, an e-mail describing the aims of the study and the participation conditions was sent to a list of public schools known by the research team. Researchers then met with representatives of any interested schools. Once an agreement had been reached, their participation was approved by the School Board and informed consent was requested from families. Those students who agreed to participate completed the questionnaire during class time and in the presence of teachers. The questionnaire was available in both pen and paper and online format, with each school choosing one of the two options in accordance with its COVID-19 regulations. Participants were informed by the research team of the aims of the study and the voluntary nature of their participation and were reassured that any information provided would be treated confidentially. The study was approved by the ZZ's Ethics Committee.

Data Analysis

Given the high degree of non-normality of the item distribution, we dichotomised the answers following the recommendations of previous studies (^{Menesini et al., 2011}): 0 (no involvement) and 1 (involvement). Descriptive item and scale analyses were performed with SPSS26.

First, items with a frequency of involvement less than 3% in the aggression and victimisation scales were eliminated due to their low representativeness in the sample. Second, confirmatory factor analyses (CFA) for aggression and victimisation were conducted on a theoretical three-factor starting model (^{Byrne, 2013}) using MPlus8 (^{Muthén & Muthén, 1998-2017}). Delta parameterisation and the weighted least squared mean variance (WLSMV) estimator were used. The recommended fit indexes for structural equation models (^{Ferrando et al., 2022}; ^{Kline, 2015}) were reported: χ ² of the model, CFI (Comparative Fit Index) to explore the comparative fit of the proposed solution with respect to the null independence model, and RMSEA (Root-Mean-Square Error of Approximation) to indicate the relative fit of the model with respect to its complexity. The optimal values for each are CFI >.95 and RMSEA <.06 (^{Hu & Bentler, 1998}; ^{Kline, 2015}).

Next, a two-parameter IRT model was developed with the IRTPRO application to examine the items quality and representativeness. This analysis provides the discrimination (a) and difficulty (b) parameters of each of the items in the latent factor. The item discrimination parameter indicates how well an item may differentiate people with different levels of ability, with the most discriminating items being those with the highest scores (^{De Ayala, 2013}). The item difficulty parameter represents the location of the item along the latent trait scale of cyber dating violence, where the participant has a 50% chance of giving a positive response on the dichotomous scale (^{Hambleton et al., 1991}). Higher scores indicate greater difficulty or severity (i.e., a higher risk for cyber dating violence). To develop two symmetrical scales of aggression and victimisation with the most substantive items, the items were analysed according to their item curve characteristic (ICC) in both scales and their parameter characteristics. The fit of the refined version of the scales was then tested using CFA.

Descriptive analysis

The descriptive analyses are presented in Table 2. For aggression, involvement ranged from 2% (items 4 and 9) to 40.7% (items 6 and 7) for verbal/emotional forms, from 5% (item 12) to 35.5% (item 19) for control and from 0.7% (item 27) to 8.1% (item 20) for sexual violence. General prevalence rates were 56.3% for verbal/emotional violence, 54% for control and 15% for sexual violence.

For victimisation, involvement ranged from 3% (item 9) to 45.7% (item 7) for verbal/emotional forms, from 7.6% (item 12) to 33% (item 19) for control and from 1% (item 26) to 14.3% (item 24) for sexual violence. General prevalence rates were 64% for verbal/emotional violence, 54.8% for control and 29.2% for sexual violence victimisation (Tables 2 and 3).

Given the aims of the present study, items 23, 26 and 27 were eliminated from the aggression and victimisation scales due to their low representativeness in the study sample (less than 3% on both scales, see Table 2). The total number of remaining items was 25.

Confirmatory factor analysis

Cyber aggression model. The three-factor model with 25 items was found to have a good fit (χ ² (272) = 299.81, CFI = .98, RMSEA = 0.02, CI 0.00-0.03). The standardised factor loadings oscillated between .66 (item 17) and .92 (item 25). The correlations of between dimensions were .85 (verbal/emotional-control), .77 (verbal/emotional-sexual), and .59 (control-sexual). The internal consistency of the scales was acceptable; McDonald's ω: .73 (verbal/emotional), .80 (control) and .75 (sexual).

Cyber victimisation model. The three-factor model with 25 items was found to have a good fit (χ ² (272) = 376.07, CFI = .97, RMSEA = 0.04, CI 0.03-0.04). The standardised factor loadings oscillated between .66 (item 17) and .96 (item 16). The correlations between dimensions were .79 (verbal/emotional-control), .84 (verbal/emotional-sexual), and .64 (control-sexual). The internal consistency of the scales was .78 (verbal/emotional), .85 (control) and .74 (sexual).

Item Response Theory Model

Table 2 presents the discrimination and severity indexes of the items included in the analysis. All items discriminated correctly in their scale, with indexes of over 1, whereas the severity indexes varied between .15 (item 7 victimisation) and 3.80 (item 9, aggression). The ICC analysis (see Figures 1-3) revealed that the probability of answering positively each item was very low in some cases, for example, in relation to items 3 and 5 (less than .1 on the aggression scale). We therefore decided to eliminate those items with a low probability on both scales (aggression and victimisation) and, to maintain symmetry across the two scales, when an item was deleted on one scale, we decided to delete the parallel item from the other scale.

Verbal/emotional cyber dating violence. As shown in Figure 1, items 3 and 9 offered very little information for the latent aggression factor, whereas the results for item 5 reflected a low curve in both factors (aggression and victimisation). These three items had the highest severity indexes and were also those which discriminated least in the aggression factor. The decision was therefore made to eliminate them from both scales. The six remaining items reflected public and private behaviours of diverse difficulty levels, with good discrimination. Items 2 and 8, which reflected public aggression, had the highest severity values, whereas items 6 and 7 were less severe. Item 4 reflected private aggression and presented a high representative value (Table 2).

Control cyber dating violence. The ICC analysis revealed that items 14, 17 and 19 (Figure 2) provided very little information for both latent factors, particularly victimisation, and were also the least discriminatory in both factors. We recommended to eliminate them, leaving the dimension to be represented by seven private items, all with good discrimination indexes and diverse severity levels (Table 2).

Sexual cyber dating violence. The sexual aggression and victimisation items had a high level of discrimination and were also the items with the highest severity indexes of the entire scale (Table 2). No item was eliminated as a result of its ICC (Figure 3), leaving this factor represented by six items that encompassed diverse behaviours, ranging from sending or pressuring to send and share intimate content (items 22 and 24), to spreading private information publicly (item 21) and even using alcohol to obtain sexual favours from one's partner (item 28). These last two items were the ones with the highest severity rates.

Figure 1. Curves of the verbal/emotional cyber dating aggression and victimisation items in accordance with their information (right) and probability (left) 

Figure 2. Curves of the control cyber dating aggression and victimisation items in accordance with their information (right) and probability (left) 

Figure 3. Curves of the sexual cyber dating aggression and victimisation items in accordance with their information (right) and probability (left) 

A CFA was conducted for the 19-item aggression and victimisation scales. The fit of the models was good for both aggression (χ ² (149) = 163.81, CFI = .99, RMSEA = 0.02, CI 0.00-0.03) and victimisation (χ ² (149) = 227.71, CFI = .98, RMSEA = 0.04, CI 0.03-0.05). Figure 4 shown the correlation between dimensions, prevalence rates, and the items factor loadings. The correlations between the latent factors were reduced slightly in relation to the initial version, indicating greater discrimination in each of the two factors.

Figure 4. Confirmatory Factor Analysis Flowchart of the CyDAV-T refined version for aggression (left) and victimisation (right) 

Descriptive analyses of the refined version of the CyDAV-T

Verbal/emotional aggression was the most prevalent dimension (54.3%), followed by control (36.8%) and then sexual aggression (15%). The same trend was found for victimisation: verbal/emotional (63.4%), control (41.9%) and sexual victimisation (29.2%).

Gender analyses revealed that girls exercised more control but were more often victims of sexual violence. Boys had higher rates of sexual aggression. No differences were found for the verbal/emotional aggression scale or for control-related victimisation (Table 3).

Evidence of convergent and criterion validity

The means comparison analyses revealed that those involved in all three forms of cyber dating aggression and victimisation had higher means for online moral disengagement and NCS. Significant differences were found in all cases, except that of control aggression and sexual victimisation, for which values were close to statistical significance (Table 4).