Participants

A total of 3,059 adolescents (51.5% boys) 12 to 18 years of age (M = 14.01, SD = 1.39) from Asturias (Spain) were assessed. In total, 95% had their own mobile phone, 93.5% used instant messaging programs in their free time, and 77.9% participated in social networking sites in their free time. The sample was selected through stratified random and cluster sampling from the total number of students in Compulsory Secondary Education (CSE) in Asturias. To that end, the population of CSE schools supported by public funds (97.6% of the total) in Asturias was divided into seven geographical areas. In each area, a number of schools proportional to the total number of schools in the area in the population was randomly selected. As a result, 19 CSE schools were eventually selected. The questionnaires were given to all of the groups in years 1 to 4 of CSE in each center selected. The selected centers were found in both rural and urban areas and were attended by students of heterogeneous socioeconomic levels.

In the 19 schools selected, there were 3,697 students in Compulsory Secondary Education at the time of the study. Of those, 3,233 (87.5%) completed the questionnaires. The questionnaires were not given to the remaining 464 students (12.5% of the total) due either to them being absent on that day or to not having obtained the consent of their parents or legal guardians. Of the 3,233 students who did complete the questionnaires, 174 (5.4%) were eliminated from analysis due to errors or omissions in a significant proportion of their responses in one or more of the questionnaires. The final sample described above was therefore made up of 3,059 students.

Measures

Demographic data and management of communication technologies. An ad hoc questionnaire was used to collect information regarding age, sex, and the grade year of respondents as well as information referring to the availability and frequency of use of communication technologies. Age and school grade were evaluated through two open-ended questions. The other variables were evaluated by means of dichotomous Yes/No response items (“I own a mobile phone”, “In my free time, I participate in social networks —Tuenti, Facebook, or other”—, and “In my free time, I use instant messaging programs —WhatsApp, Messenger, or other”—), or polytomous items (“In general, from Monday to Friday, how many hours per day do you spend using the Internet for non-school tasks?”, and “In general, during the weekend, how many hours per day do you spend using the Internet for non-school tasks?”) with five options (none, less than an hour, between one and two hours, between two and three hours, and more than three hours).

Behavioral control. To measure parental control of activities as perceived by adolescents, the “behavioral control” factor from the Dimensions of Parenting Style Questionnaire by ^{Álvarez-García, García, Barreiro-Collazo, Dobarro, and Antúnez (2016)} was used. It consists of four items (“My parents try to know where I am going when I leave home”, “If I return home late, my parents ask me why and who I was with”, “My parents set limits on the time that I should return home”, and “My parents ask me how I spend money”), in which respondents are asked to assess the extent to which each statement is true (from 0, completely false, to 3, completely true). The total score for each respondent in this factor corresponds to the sum of the scores on each item (minimum 0, maximum 12). High scores indicate high levels of behavioral control. The internal consistency of this scale with this study sample is adequate (α = .75).

Rules for Internet use. Parents establishing limits on the Internet use, as perceived by adolescents, was evaluated by using the “rules for Internet use” factor from the Parental Control on Internet Usage Questionnaire by ^{Álvarez-García, García, Cueli, and Núñez (2017)}. It consists of three items (“My parents or legal guardian set some rules regarding what I can or cannot do on the Internet”, “My parents or legal guardian limit the content I can access on the Internet through computer filters”, and “My parents or legal guardian limit the hours I use the Internet, either by word or by configuring the computer”, with four response options (from 0, completely false, to 3, completely true). The total score for each respondent in this factor corresponds to the sum of the scores on each item (minimum 0, maximum 9). High scores indicate greater establishment of standards by families. The internal consistency of this scale with the sample for this study is adequate (α = .71).

Supervision of Internet use. The opinions of adolescents regarding the extent to which their parents supervise how they use the Internet were evaluated with the “monitoring Internet use” factor from the Parental Control on Internet Use Questionnaire by ^{Álvarez-García, García et al. (2017)}. It consists of four items (“When I access the Internet in my free time, my parents monitor and take a look at the screen”, “My parents know the passwords to access my email accounts, social networking sites, and messaging programs”, “My parents monitor my profiles on social networks, e.g., Facebook, Tuenti, Twitter, Instagram”, and “My parents know my contact lists”), with four response options (from 0, completely false, to 3, completely true). The total score for each respondent in this factor corresponds to the sum of the scores on each item (minimum 0, maximum 12). High scores indicate high levels of supervision. The internal consistency of this scale with the sample for this study is high (α = .80).

Affection and communication. The perceptions of adolescents regarding their parents’ support, emotional closeness, and willingness to talk were assessed by the “affection and communication” factor from the Dimensions of Parenting Style Questionnaire by ^{Álvarez-García, García et al. (2016)}. It consists of four items (“When I talk to my parents, they show interest and pay attention”, “My parents encourage me to tell them my problems and concerns”, “If I have a problem, I can count on my parents’ help”, and “My parents show interest in me when I am sad and angry”), with four response options (from 0, completely false, to 3, completely true). The total score for each respondent in this factor corresponds to the sum of the scores on each item (minimum 0, maximum 12). High scores indicate high levels of affection and communication. The internal consistency of this scale in this study sample is high (α = .87).

Impulsivity. The degree of the respondents’ impulsivity was assessed using a self-reported scale previously used by the research team (^{Álvarez-García, Barreiro-Collazo, Núñez, & Dobarro, 2016}). The scale was created using part of the impulsivity criteria proposed by the DSM-5 for the diagnosis of Attention Deficit and Hyperactivity Disorder (American Psychiatric Association, 2013). It consists of five items: “In class or when playing games, I often find it difficult to wait, so I jump in or interrupt others”, “I often blurt out an answer before a question has been completed”, “I often say what comes to mind without thinking of the consequences first or whether it is appropriate for the conversation”, “I often do things without thinking of the consequences”, and “I often find waiting difficult, so I do not wait for my turn to speak or I interrupt the person who is speaking”. The response is a Likert-type scale with four options (from 0, completely false, to 3, completely true). The total score for each respondent in this factor corresponds to the sum of the scores on each item (minimum 0, maximum 15). High scores indicate high levels of impulsivity. The internal consistency of the scores obtained with the scale in this study sample is adequate (α = .75).

Empathy. The degree of empathy in the adolescents evaluated was measured by a self-reported scale previously used by the research team (^{Álvarez-García, Barreiro-Collazo et al., 2016}). It is composed of five items that refer to the extent to which a respondent believes that he/she is capable of identifying with others and sharing his/her feelings: “I feel the misfortunes of others”, “If a classmate is teased, I feel bad thinking about what is happening to him/her”, “I am patient with people who do things worse than I do”, “When I see that a friend is sad, I also become sad”, and “I am happy when something good happens to someone I know.” The response is a Likert-type scale with four options (from 0, completely false, to 3, completely true). The total score for each respondent in this factor corresponds to the sum of the scores on each item (minimum 0, maximum 15). High scores indicate high levels of empathy. The internal consistency of the scores obtained with the scale in this study sample is adequate (α = .70).

Aggression at school. To assess the extent to which the respondents have engaged in aggressive behavior on school premises over the past three months, a self-reported scale previously designed and used by the research team was used (^{Álvarez-García, García et al., 2016}). It is composed of six items: “I have not let a classmate participate in my group during the break or physical education”, “I have not let a classmate participate in my group in a class activity”, “I have laughed at or teased a classmate”, “I have talked badly about a classmate behind his/her back”, “I have insulted a classmate to his/her face”, and “I have hit a student from the school, on or off school grounds”. The response is a Likert-type list scale four options (from 0, never to 3, always). The total score for each respondent in this factor corresponds to the sum of the scores on each item (minimum 0, maximum 18). High scores indicate high levels of aggression. The internal consistency of this scale with this study sample is adequate (α = .75).

Antisocial behavior. To assess the extent to which respondents recognize engaging in different types of antisocial behavior, a self-reported scale previously designed and used by the research team was used (^{Álvarez-García, García et al., 2016}). It consists of six items: “I consciously soiled, damaged, or destroyed public furniture, e.g., a wall, a trashcan, a lamppost, seats on the bus”, “I stole something from a shop, school, or a private home”, “I trespassed on private property”, “I have hit or fought with a stranger to the point of harming him/her”, “I used illegal drugs”, and “I have gotten drunk.” The response requested from respondents is dichotomous (1 = true, 0 = false), indicating whether they have performed these activities at least once in the last year. The total score for each respondent in this factor corresponds to the sum of the scores on each item (minimum 0, maximum 6). High scores indicate high levels of antisocial behavior. The internal consistency of the scale in this sample is adequate (KR20 = .73).

Antisocial friendships. To assess the extent to which respondents relate to antisocial friendships, a scale previously used by the research team was used (^{Álvarez-García, García et al., 2016}). It comprises four items, in which the respondents indicate whether the situation described has occurred during the past year: “One or some of my best friends have soiled, damaged, or destroyed public furniture, e.g., a wall, a trashcan, a lamppost, seats on the bus”, “One or some of my best friends have stolen something from a shop, school, or a private home”, “One or some of my best friends have had a real physical fight with another young person”, and “One or some of my best friends have consumed illegal drugs”. The response requested from the respondent is dichotomous (1 = true, 0 = false). The total score for each respondent in this factor corresponds to the sum of the scores on each item (minimum 0, maximum 4). High scores indicate high levels of antisocial friendships. The internal consistency of the scale in this sample is adequate (KR20 = .72).

Cyber-victimization. The frequency with which the respondents have been victims of aggression via mobile phone or the Internet during the last three months was assessed using the Cyber Victimization Questionnaire for Adolescents (CYVIC; ^{Álvarez-García, Núñez, Barreiro-Collazo, & García, 2017}). It consists of 19 items that refer to five types of cyber-victimization: verbal cyber-victimization (e.g., “I have been made fun of using offensive or insulting comments on social networks”); online exclusion (e.g., “A group of people has agreed to ignore me on social networks”); visual-sexual cyber-victimization (e.g., “I have been pressured to do things that I did not want to, regardless of whether I agreed to do them or not, under the threat of having intimate conversations or images of me disseminated”); visual cyber-victimization – teasing/happy slapping (e.g., “I have been forced to do something humiliating, and this has been recorded and then disseminated to make fun of me”); and impersonation (e.g., “I have been impersonated on the Internet, and comments have been posted in my name, as though coming from me”). A Likert-type response format is used (from 0 = never to 3 = always). In this study, the total score on cyber-victimization for each respondent was obtained by adding the scores from the 19 items (minimum 0, theoretical maximum 57). High scores indicate high levels of cyber-victimization. The internal consistency of the scale in this study sample is adequate (α = .79).

Cyber-aggression. The frequency with which the respondents admit having committed aggressions via mobile phone or the Internet during the last three months was evaluated through the Cyber-Aggression Questionnaire for Adolescents (CYBA; ^{Álvarez-García, Barreiro-Collazo et al., 2016}). This self-reported scale has the same indicators and response format as the CYVIC, but this time referring to aggression. The total score for each respondent regarding cyber-aggression responds to the sum of the scores on the 19 items (minimum 0, theoretical maximum 57). High scores indicate high levels of cyber-agression. The internal consistency of the scale in the sample of this study is high (α = .82).

Procedure

Permission to administer the questionnaires was requested from the administration in each center selected. Each teaching center acquired family consent for the participation of the students in the study because they were underage. The questionnaires were completed by the students at the educational center during school hours. At the time of the application of the questionnaires, participants were informed of the voluntary and anonymous nature of the test as well as the confidential treatment of the data obtained.

Data Analysis

First, the two polytomous variables used (“In general, from Monday to Friday, how many hours per day do you spend using the Internet for non-school tasks?” and “In general, during the weekend, how many hours per day do you spend using the Internet for non-school tasks?”) were recoded as dichotomous variables (three hours or fewer/more than three hours).

Then, the sample was divided into three subgroups, depending on the degree of cyber-aggression exerted. The “no cyber-aggression” group is composed of those who answered never to all of the statements of the CYBA questionnaire. The “severe cyber-aggression” group is composed of those who obtained a score equal to or greater than the 99^th percentile in the CYBA (direct score ≥ 14.40). To determine this cut-off point, the data on severe cyber-aggression in adolescence – which is approximately 1% – offered by recent national studies published in Spain were taken as a reference (^{Calmaestra et al., 2016}; ^{Díaz-Aguado, 2010}). The “occasional cyber-aggression” group comprises those who acknowledged having engaged in aggressive behavior at least once but whose total score in the CYBA is below the 99^th percentile.

Once the database was closed, analyses were conducted to address the study objectives. To that end, first, a preliminary analysis and, then, a multinomial logistic regression analysis were conducted. The preliminary analysis consisted of an initial descriptive analysis of the sample set and potential predictors analyzed, in terms of means and standard deviations or frequencies and percentages, depending on the categorical or continuous nature of the variable analyzed. Second, the means or percentages of non-cyber-aggressors, occasional cyber-aggressors, and severe cyber-aggressors were compared, focusing on the potential predictors analyzed. To that end, one-way analyses of variance (ANOVA) or chi-square tests were conducted, depending on the nature of each predictor variable; the effect size was calculated using the eta squared or Cramer’s V, respectively. Next, the simple correlation between cyber-aggression and each of the potential predictors analyzed was studied, in addition to the degree of association between each of the predictors. To that end, the Pearson’s correlation coefficient, the point-biserial correlation coefficient, or the phi coefficient was used, depending on the nature of the variables involved in the association analyzed.

Subsequently, multinomial logistic regression analyses were conducted to identify the predictive capacity of the variables analyzed regarding being an occasional or a severe aggressor, in both cases taking the non-cyber-aggressor group as reference. For both occasional and severe cyber-aggression, first, crude odds ratios were calculated using univariate regression analysis. Subsequently, multivariate regression analyses were conducted, aiming to provide adjusted estimators that approximate as closely as possible the real relationship between potential predictors and being a cyber-aggressor. To that end, each group of predictive variables was progressively included, in blocks. As a result, there were six predictive models. Crude and adjusted odds ratio were compared, in addition to those of each fitted model and those of the next. When the inclusion of a new indicator variable in the model significantly affected the odds ratio of the rest, the degree of association between them was analyzed to determine the reason. In this manner, we were able to identify not only the variables with higher predictive capacity but also potentially confounding variables, indirect relationships, and even spurious relationships between the potential predictors analyzed and cyber-aggression. Finally, to analyze the validity of the six models studied and, therefore, the explanatory contribution of each block of variables introduced in each step, the Nagelkerke’s pseudo R² and the percentage of correct classifications were calculated.

Preliminary Analysis

Description of the sample on the basis of potential predictive variables.Table 1 provides an overview of the sample with respect to the potential predictors analyzed. There are statistically significant differences, depending on the degree of cyber-aggression exercised in all the potential predictors analyzed, except for sex. The greater the severity, the greater the age, Internet use, impulsiveness, antisocial behavior, and cyber-victimization, and the lesser the parental control and empathy.

Table 1 Descriptive Analysis of the Sample and Comparison between Non-Cyber-aggressors, Occasional Cyber-aggressors, and Severe Cyber-aggressors in the Analyzed Variables 

Note. N-CBA = non-cyber-aggressors; O-CBA = occasional cyber-aggressors; S-CBA = severe cyber-aggressors; ES = effect size. ¹Mean (standard deviation); ²Frequency of affirmative responses (percentage of affirmative responses); ³F-test; ⁴Pearson’s; ⁵Eta-squared; ⁶Cramer’s V.

Table 2 Measures of Association among the Variables Analyzed in the Study 

Note. AGE = age; SEX = sex (boy); SNS = participated in social networks; UIM = use of instant messaging programs; IMF+3 = Internet use from monday to friday for non-school tasks more than 3 hours a day; ISS+3 = Internet use on weekends for non-school tasks more than 3 hours a day; BEC = behavioral control; RIU = rules for Internet use; MIU = monitoring Internet use; AAC = affection and communication; IMP = impulsivity; EMP = empathy; SAG = school aggression; ABE = antisocial behavior; AFR = antisocial friendships; CBV = cyber-victimization; CBA = cyber-aggression; r = pearson correlation coefficient; r_pb = point-biserial correlation coefficient; ϕ = phi coefficient.***p ≤ .001.

Association between the variables analyzed. As shown in Table 2, the correlation between cyber-aggression and each of the potential predictors analyzed is statistically significant in all cases. It is positive in most of them, except for empathy and the four variables relating to parental control, in which case cyber-aggression correlates negatively. The magnitude of these correlations is small, except for impulsivity, cyber-victimization, and the three variables relating to antisocial behavior, with which cyber-aggression correlates moderately.

The relationship between potential predictors is statistically significant in 112 out of 120 pairs analyzed (Table 2). The magnitude of the association between indicators is small, except in 18 of the pairs, in which it is moderate.

Multinomial Logistic Regression

Predictors of occasional cyber-aggression. The unadjusted (univariate) regression analysis suggests that 15 of the16 predictive variables analyzed (all except sex) are statistically significantly related to becoming an occasional cyber-aggressor. These 15 variables would constitute risk factors, except for empathy and the variables concerning parental control. These two would, in turn, act as protective factors (Table 3). However, the multivariate regression analysis, which included all variables analyzed (Model 6), shows that only seven of them had a statistically significant relationship with being an occasional cyber-aggressor. These seven variables include participating in social networks, using instant messaging programs, degree of impulsivity, frequency with which they attack other students in the educational center, antisocial behavior outside the center, antisocial friendships, and degree of cyber-victimization endured. All of these are risk factors. Teenagers who participate in social networks and those who use instant messaging programs are 1.6 and 1.7 times more likely, respectively, to be occasional cyber-aggressors than those who do not use them. The probability of being an occasional cyber-aggressor is 1.1 times higher in those who score one point higher in impulsivity, antisocial behavior, or antisocial friendships and 1.4 times higher in those who score one point higher in exercising school aggression or in suffering cyber-victimization.

Table 3 Results of the Multinomial Logistic Regression Analysis Regarding the Probability of Being an Occasional Cyber-aggressor 

Note. Reference category: Non-cyber-aggressor. AGE = age; SEX = sex (boy); SNS = participated in social networks; UIM = use of instant messaging programs; IMF+3 = Internet use from monday to friday for non-school tasks more than 3 hours a day; ISS+3 = Internet use on weekends for non-school tasks more than 3 hours a day; BEC = behavioral control; RIU = rules for Internet use; MIU = monitoring Internet use; AAC = affection and communication; IMP = impulsivity; EMP = empathy; SAG = school aggression; ABE = antisocial behavior; AFR = antisocial friendships; CBV = cyber-victimization. *p ≤ .05, **p ≤ .01, ***p ≤ .001.

Table 3 shows how the inclusion of each new block of variables affects the estimators of the previous model. Changes in the magnitude of the estimators after the inclusion of each new block of variables are consistent with the association found between predictors, which is provided in Table 2.

Predictors of severe cyber-aggression. The univariate regression analysis suggests that 15 of the 16 predictive variables analyzed (all except the use of instant messaging programs) are statistically significantly related to being a severe cyber-aggressor. The 15 variables would constitute risk factors, with the exception of empathy and the variables relating to parental control, which would constitute protective factors (Table 4). However, in the multivariate regression analysis that included all variables analyzed (Model 6), only four of them showed statistically significant correlations with becoming a severe cyber-aggressor. These four variables are degree of impulsivity, frequency with which attacks on other students are committed in the educational center, and degree of cyber-victimization suffered, as risk factors, and degree of empathy, as a protective factor. The probability of being a severe cyber-aggressor is 1.2, 1.6, or 1.9 times higher in those who score one point higher in impulsivity, cyber-victimization endured, or school aggression exerted, respectively, and 1.3 times lower in those who score one point higher in empathy.

Table 4 shows how the inclusion of each new block of variables affects the estimators of the previous model. Changes in the magnitude of estimators after the inclusion of each new block of variables are consistent with the association found between predictors, shown in Table 2.

Table 4 Results of the Multinomial Logistic Regression Analysis for the Probability of Being a Severe Cyber-aggressor 

Note. Reference category: Non-cyber-aggressor. AGE = age; SEX = sex (boy); SNS = participated in social networks; UIM = use of instant messaging programs; IMF+3 = Internet use from monday to friday for non-school tasks more than 3 hours a day; ISS+3 = Internet use on weekends for non-school tasks more than 3 hours a day; BEC = behavioral control; RIU = rules for Internet use; MIU = monitoring Internet use; AAC = affection and communication; IMP = impulsivity; EMP = empathy; SAG = school aggression; ABE = antisocial behavior; AFR = antisocial friendships; CBV = cyber-victimization.**p ≤ .05, **p ≤ .01, ***p ≤ .001.

Analysis of model validity. The variables included in the final model (Model 6) together explain 45% of the variance in cyber-aggression scores (Table 5). The percentage of cases that this final model is able to correctly predict is 76.6%. In both cases, these values are the best obtained in all models analyzed. The largest increase in R² and in the percentage of correct classifications occurs when variables related to antisocial behavior (Model 5) are included. The smallest increase occurs with the inclusion of variables related to parental control (Model 3).

Table 5 Validity Indicators of the Six Models Analyzed