The responses of 3,148 CSE students, from 12 to 18 years old (M=14.01, SD=1.39), are analysed. Of the sample, 48.1% are girls. Sex is distributed homogeneously across the different ages of the sample, χ2(6)=11.695, p=.069, Cramer's V=.061. Some 95.1% of the evaluated adolescents have their own mobile phones, 84.1% surf the internet in their free time for non-school tasks, 93.6% use instant messaging programs (Messenger, WhatsApp, or others), and 77.9% participate in social networks (Tuenti, Facebook, or others) in their free time.

The participants evaluated attended 19 schools, selected using stratified random sampling from all publicly funded schools in which CSE is taught in Asturias. To select the sample, the population of schools was divided according to their ownership (public or state-subsidised private) and, in each stratum, a number of schools proportional to the population was selected. As a result, 11 public and 8 state-subsidised private schools were selected. The selected schools are in predominantly urban areas where residents are in the middle socioeconomic class. All groups were evaluated in each selected school. A total of 28.2% of the participants are in their 1st year, 25.3% are in their 2nd year, 24.4% are in their 3rd year, and 22.1% are in their 4th year of CSE.

Cyber-Aggression Questionnaire for Adolescents (CYBA). The CYBA questionnaire designed and tested in this study is a self-report composed of 19 items, each of which describes an aggression conducted through a mobile phone or the internet (see Appendix). The evaluated participant should indicate the frequency that he/she has conducted the different actions in the last three months, using a Likert-type scale with four options (1=never, 2=rarely, 3=often, and 4=always).

Aggression at school. A self-report scale that was developed ad hoc was used; the scale was based on a selection and adaptation of six items from the Aggression Questionnaire by Álvarez-García, Menéndez, González-Castro, and Rodríguez (2012). This study uses a scale composed of the following items related to the frequency with which the participant says he/she conducts aggressions in the physical school environment: “No he dejado participar en mi grupo a algún compañero, durante alguna actividad de recreo o de educación física” [“I have not let a classmate participate in my group during the break or physical education”], “No he dejado participar en mi grupo a algún compañero en alguna actividad de clase” [“I have not let a classmate participate in my group in a class activity”], “Me he reído y burlado de algún compañero/alguna compañera” [“I have laughed at or teased a classmate”], “He hablado mal de algún compañero/alguna compañera a sus espaldas” [“I have talked badly about a classmate behind his/her back”], “He insultado a la cara a algún compañero/alguna compañera” [“I have insulted a classmate to his/her face”] and “He pegado a algún alumno/alguna alumna del centro, dentro o a la salida del recinto escolar” [“I have hit a student of the school, within or outside the school grounds”]. The response is a Likert-type list with four options (1=never, 2=rarely, 3=often, 4=always). The internal consistency of the scores obtained with this scale in this study sample is high (Raykov's rho=.917).

Impulsivity. The self-report scale is developed ad hoc from some of the criteria for impulsivity proposed by the DSM-5 for the diagnosis of ADHD (American Psychiatric Association, 2013). The scale used is composed of five items, referring to the extent that the respondent considers it true that he/she conducts certain behaviours of an impulsive nature: “En clase o en juegos, a menudo me cuesta esperar turno, por lo que me cuelo o interrumpo” [“In class or when playing games, it is often difficult for me to wait my turn, so I jump in or interrupt”], “A menudo contesto antes de que se haya completado la pregunta” [“I often answer before the question has finished”], “A menudo digo lo que me viene a la cabeza, sin pensar primero sus consecuencias o si es oportuno para la conversación” [“I often say what comes to mind without thinking first of the consequences or whether it is appropriate for the conversation”], “A menudo hago cosas sin pensar en las consecuencias” [“I often do things without thinking of the consequences”] and “Habitualmente me resulta difícil esperar turno, por lo que me adelanto a hablar cuando no me corresponde o interrumpo a quien está hablando” [“Usually, I find it difficult to wait my turn, so I jump in to speak when it is not my turn or I interrupt the person talking”]. The response is a Likert-type list with four options (1=totally false, 2=somewhat false, 3=somewhat true, 4=totally true). The internal consistency of the scores obtained with the scale in this study sample is high (Raykov's rho=.865).

Empathy. The self-report scale was developed ad hoc and composed of five items referring to the extent that the respondent considers it to be true that he/she is capable of identifying with others and sharing their feelings: “Siento las desgracias de los demás” [“I feel the misfortunes of others”], “Si se burlan de un compañero, me siento mal pensando en lo mal que lo está pasando” [“If a classmate is teased, I feel bad thinking about what is happening to him/her”], “Soy paciente con las personas que hacen las cosas peor que yo” [“I am patient with people who do things worse than I do”], “Cuando veo que un/a amigo/a está triste, yo también me entristezco” [“When I see that a friend is sad, I also become sad”] and “Me alegro cuando le pasa algo bueno a un conocido” [“I am happy when something good happens to someone I know”]. The response is a Likert-type list with four options (1=totally false, 2=somewhat false, 3=somewhat true, 4=totally true). The internal consistency of the scores obtained with the scale in this study sample is adequate (Raykov's rho=.755).

First, the CYBA questionnaire was designed. The wording of the items came from a theoretical model of six factors or types of cyber-aggression (Table 1), based on the model of four factors (written-verbal, visual, impersonation, and exclusion) proposed by Nocentini et al. (2010). The authors produced the written items. Once the questionnaire was designed and the schools selected, permission was requested from the schools’ respective leadership teams to implement the questionnaire. Each leadership team was informed of the objectives and procedures of the study, its voluntary and anonymous nature, and the confidential treatment of the results. Once the school agreed to participate, informed consent was requested from the parents or guardians of the students, given their status as minors. The questionnaire was applied in all the schools during the second and third quarter of the 2014-2015 school year. Before answering the questionnaire, students were also informed of the anonymous, confidential, and voluntary nature of their participation. In general, students had 20minutes to answer the questionnaires, although the timing was flexible depending on the age and characteristics of the respondents. The research team administered the test to all groups from each school during school hours.

First, in the data matrix, the presence of values that were missing or outside the scale was analysed. All items were answered in an acceptable way by more than 90% of the respondents, so none were removed from the final analysis of the test on that basis. Of the 3,233 students who responded to the CYBA, 85 responded in an acceptable way to less than 90% of the items, so they were removed from the data matrix. Once they were removed, the number of missing values in the CYBA matrix was reduced to 122 (0.2% of the total).

Second, the statistical program EQS 6.2 (Bentler, 2014) was used to analyse the normality of the data and the dimensionality of the test. Normality was assessed by analysing the multivariate kurtosis. Dimensionality was examined using confirmatory factor analysis. The multivariate kurtosis pointed to the non-normality of the obtained data (Mardia coefficient=2571.16). Consequently, Arbitrary Generalized Least Squares (AGLS) was used as estimation method. Given the ordinal nature of the scale, the analyses were completed based on the matrix of polychoric correlations (Lila, Oliver, Catalá-Miñana, Galiana, & Gracia, 2014). Missing values were treated by computing the matrix of covariances using the pairwise method because AGLS requires large sample sizes and to avoid removing more subjects and be able to use all available data.

To determine the degree of fit of the tested models, the chi-square (χ2)/degrees of freedom (df) ratio, the Comparative Fit Index (CFI), the Bentler-Bonett Non-normed Fit Index (NNFI), the Standardized Root Mean Square Residual (SRMR), the Root Mean Square Error of Approximation (RMSEA), and the Akaike Information Criterion (AIC) were utilized. Usually, the fit is considered good when CFI ≥ .95, NNFI ≥ .95, SRMR ≤.08, and RMSEA ≤ .06 (Hu & Bentler, 1999), and χ2/df<3 (Ruiz, Pardo, & San Martín, 2010). The AIC allows models to be compared, with those having lower values being preferable.

The discriminant validity of the model with best fit was studied, analysing the correlation between its factors. Very high correlations (r ≥ .85) warn of possible collinearity or redundancy between factors and, therefore, poor discriminant validity (Brown, 2015). Given that a very high correlation was indeed found between some factors, post hoc re-specifications in the model were completed.

The indices of fit and the correlations between factors of the two re-specified models were analysed. Once the model with the most appropriate discriminant validity and fit to the data was identified from those tested, the factor weights and standard errors of each item for that model were found. Usually, factor weights greater than .30 are considered acceptable (Izquierdo, Olea, & Abad, 2014).

Subsequently, the reliability of the scores obtained in the test was analysed. The reliability of each subscale was analysed in terms of internal consistency, finding its respective Raykov's rho from the matrix of polychoric correlations. This coefficient is considered as an alternative to the weaknesses of the Cronbach's alpha coefficient, highlighted by authors such as Sijtsma (2009). In general, the Raykov's rho coefficient is considered acceptable if it has a value above .70, high if it is greater than .80, and very high if it is greater than .90. Additionally, Squared Multiple Correlations of the scores for each item were found, which was an indicator of the proportion of variance on the item explained by the latent variable and, therefore, of the item's reliability to measure that variable (Bollen, 1989).

Finally, the statistical software SPSS 21 (IBM Corp., 2012) was used to analyse the criterion validity. For this, the Spearman correlation coefficient was calculated between the score on each of the CYBA's three factors and three external criteria about which previous evidence demonstrates their association with cyber-aggression: aggression at school, impulsivity, and empathy.

Factorial validity. At first, the indices of fit of the initial theoretical model, composed of significantly interrelated six factors or types of cyber-aggression (6FM), were obtained and compared with those from two additional plausible models from a theoretical perspective (Table 1). On the one hand, they were compared with a unifactorial model (1FM), whereas all the items from the test could measure only one construct (cyber-aggression). On the other hand, they were compared with a hierarchical model of six first-order factors–the same as 6FM–and one second-order factor–cyber-aggression–(6FM2), whereas each item could be explained by a first-order factor and each first-order factor by the common construct cyber-aggression. As presented in Table 2, the 6FM presents good indices of fit to the obtained data, the best from among the three models initially tested.

Discriminant validity. All the 6FM factors have positive and statistically significant correlations among themselves (Table 3). Some of these correlations are very high (r>.85): the factor Visual Cyber-Aggression Teasing/Happy Slapping is very highly correlated with Visual-Sexual Cyber-Aggression, Impersonation, and Online Exclusion; and the factor Online Exclusion is very highly correlated with Verbal-Written Cyber-Aggression and Verbal-Oral Cyber-Aggression, in addition to its relationship with Visual Cyber-Aggression Teasing/Happy Slapping referenced above.

Due to the high correlation between some factors, two modified models of the initial 6FM were designed and tested: a model with five factors (5FM) and another with three (3FM). In both, items referring to Visual Cyber-Aggression Teasing/Happy Slapping were removed from the model and are now considered additional indicators included in the questionnaire but outside the model. Thus, the 5FM is composed of the same factors as the 6FM, with the exception of the factor Visual Cyber-Aggression Teasing/Happy Slapping. Furthermore, in the 3FM, the items initially included in the factors Verbal-Written Cyber-Aggression, Verbal-Oral Cyber-Aggression, and Online Exclusion now constitute only one factor, called Verbal Cyber-Aggression and Exclusion (Figure 1). As indicated in Table 4, both models present good indices of fit. The 5FM presents the best fit of both. Nonetheless, the correlation between the factor Online Exclusion and the two factors referring to Verbal Cyber-Aggression (Oral and Written), remain very high (r>.85). In the case of the 3FM, no correlations exist between factors greater than .85 (Figure 1). That is why this model (3FM) is considered the one that best represents the obtained data among those tested.

Figure 1.

Factorial Structure of the CYBA.

Note. I=impersonation, V-S=visual-sexual cyber-aggression, V-E=verbal cyber-aggression and exclusion.

(0.26MB).

Factor weights. In the 3FM, the factor weights of each item in its factor are high (Figure 1), with standardised regression coefficients equal to or greater than .70 on all items except 13 (λ=.69).

Reliability. The reliability of scores in each subscale, evaluated in terms of internal consistency, could be considered high, and for each item the reliability scores could be considered moderate or high (Table 5). The reliability coefficient Raykov's rho of each factor is greater than .80. The proportion of variance of the item explained by the latent variable (R2) is between 47% and 82%.

Criterion validity. The scores of the three factors of the CYBA have a statistically significant correlation with the three external criteria analysed. The correlation is positive with the frequency of aggression conducted in the physical school environment and impulsive behaviours, reported by the evaluated adolescents themselves; the correlation is negative with empathy (Table 6).