Associations of bullying victimization with problematic internet gaming and problematic social media use among adolescents: moderators and differences

Regarding gender, a variety of meta-analyses suggest that IGD is generally more prevalent in boys [15, 31, 32], while girls show a higher prevalence for PSMU [21, 33]. In line with these findings, male bullying victims seem to be more at risk for the development of PIG [34, 35], whereas for female victims, the relationship between social media use and bullying victimization was found to be stronger [14, 16]. One explanation for these gender-specific patterns could be differences in their coping styles: While boys were found to prefer avoidant coping (e.g., by escaping into a virtual gaming world), girls appeared to seek social support (e.g., in an online community) [36]. Despite these findings, some research points out that PIG is also common among girls, for instance, depending on the gaming genre, and that games are becoming increasingly attractive to females [37, 38]. Thus, PIG might still be a hidden problem in girls. In addition, some studies on PSMU suggest that females are often over-represented, which might influence the findings [21, 24].

Results on the role of age are even more controversial. For instance, in the age group of 12- to 18-year-olds, some studies reported a higher probability of PIG for younger participants [39], whereas other studies found higher probabilities for older age [40]. In contrast, a meta-analysis with studies on 9- to 21-year-olds reported no association between mean age and the prevalence of IGD [41]. For PSMU, the findings are also inconsistent. While one study found higher prevalence rates for 13- and 15-year-olds compared to 11-year-olds [42], another study reported the highest rate at 13 years followed by 11 years and lastly by 15 years [43]. In addition, findings on how bullying victimization interacts with age regarding internet use disorders in adolescents are inconsistent [16, 44].

In some countries there are different types of secondary schools to which students transition after primary school – among other factors – based on their academic achievements. Typically, these school types consist of high schools preparing students for university enrollment, and alternatively, vocational high schools. Thus, the former school type is usually regarded as higher educational background. Some studies suggest that vocational high school students might be more prone to problematic internet use, perhaps because of less awareness for its negative consequences [45]. However, while PIG seems to be associated to a lower education [46‐48], the role of the educational background in terms of PSMU [42, 49] and the interaction with bullying victimization regarding PIG and PSMU is still inconclusive [35].

One further potential predisposing variable for the development of problematic internet use suggested by the I-PACE model is psychopathology [30]. Indeed, research indicates that internalizing mental health problems such as symptoms of depression or anxiety, and externalizing problems like attention deficit hyperactivity disorder, are very common in internet-related addictions [50, 51] and that mental health problems are associated with PIG [52] and PSMU [53]. Furthermore, research suggests that mental health problems mediate the associations between bullying victimization and PIG [54] as well as PSMU [55]. However, to our knowledge, no study has investigated the influence of mental health problems on the relationship between victimization and PIG in comparison to PSMU.

Bullying was assessed using an adapted version of the Bullying Screening since it is a short instrument which still comprises different facets of bullying [61, 62]. In previous adolescent samples the victimization questions had a Cronbach’s alpha of 0.73, and predictive validity was demonstrated [62, 63]. In the current sample the Cronbach’s alpha was 0.67. The relatively low internal consistency observed is likely attributable to the scale’s primary function as a screening tool for different forms of bullying experiences. Although these forms share similarities, they represent distinct, yet related, constructs. More precisely, the Bullying Screening contains six questions: three on bullying victimization (direct, indirect, cyber) and three on perpetration (direct, indirect, cyber). Each question includes a clear definition and concrete examples of bullying to ensure validity. In this study, only the questions on victimization were analyzed (supplementary material, Table A1). A six-point scale was used to rate the frequency within the past six months. If participants indicated “often (2–3 times a month)” or more on at least one of the three questions on bullying victimization, they were classified as victims. This cut-off is widely used in research [64].

The German version of the Internet Gaming Disorder Scale was used to assess PIG [65]. In previous adolescent samples the reliability coefficients were between 0.56 and 0.83, and criterion validity was demonstrated [66, 67]. In the current sample the Cronbach’s alpha was 0.76, and Harman’s single factor test showed 35.09% of explained variance (see supplementary material). The scale was introduced by the sentence: “The following questions are about your use of computer games in the last year. By computer games, we mean all games played on a computer, tablet, game console or smartphone (offline and online).“ It includes nine questions with one question per DSM-5 criterion for IGD [19]: preoccupation (“Have there been periods when all you could think of was the moment that you could play a game?”), tolerance (“Have you felt unsatisfied because you wanted to play more?”), withdrawal (“Have you been feeling miserable when you were unable to play a game?”), persistence (“Were you unable to reduce your time playing games, after others had repeatedly told you to play less?”), escape from negative feelings (“Have you played games so that you would not have to think about annoying things?”), continuation despite problems (“Have you had arguments with others about the consequences of your gaming behavior?”), deception (“Have you hidden the time you spend on games from others?”), displacement (“Have you lost interest in hobbies or other activities because gaming is all you wanted to do?”) and conflicts (“Have you experienced serious conflicts with family, friends or a partner because of gaming?”). The dichotomous response format (“yes”=1, “no”=0) leads to a sum score ranging between 0 and 9. This sum score with a higher value indicating a greater severity of PIG in the last 12 months was used in the current study. In line with the DSM-5, participants with a sum score of five or more were classified as having IGD in Table 1 [19].

PSMU was assessed by the German version of the Social Media Disorder Scale [27]. In previous adolescent samples the reliability coefficients were between 0.57 and 0.81, and good convergent and criterion validity were demonstrated [23, 27]. In the current sample the Cronbach’s alpha was 0.80, and Harman’s single factor test showed 39.85% of explained variance (see supplementary material). The scale was introduced by the sentence: “The following questions relate to your use of social media such as WhatsApp, SnapChat, social networking sites such as Facebook, Twitter, Instagram, Google+, Pinterest and forums and blogs/weblogs.” Similar to the DSM-5 criteria for IGD [19], it examines the same nine symptoms with similar wording as the Internet Gaming Disorder Scale [65] but in relation to the use of social media (e.g., for preoccupation: “Have you regularly found that you can’t think of anything else but the moment that you will be able to use social media again?”). Analogously, the dichotomous response format (“yes”=1, “no”=0) leads to a sum score ranging between 0 and 9. This sum score with a higher value indicating a greater severity of PSMU in the last 12 months was used in the current study. A sum score of at least five was classified as indicative of Social Media Disorder (SMD) in Table 1.

Mental health problems were assessed by the Strengths and Difficulties Questionnaire (SDQ; [68]) which is a behavioral screening questionnaire and is not intended for diagnostic purposes. Its five scales consist of five questions each, resulting in overall 25 questions. The five scales are emotional symptoms (e.g., “I am often unhappy, depressed or tearful.”), conduct problems (e.g., “I fight a lot. I can make other people do what I want.”), hyperactivity/inattention (e.g., “I am restless, I cannot stay still for long.”), peer relationship problems (e.g., “I get along better with adults than with people my own age.”) and prosocial behavior (e.g., “I am helpful if someone is hurt, upset or feeling ill.”). A three-point scale was used to rate the degree of agreement (“not true”=0, “somewhat true”=1, “certainly true”=2) concerning the last six months. To calculate a total difficulties score for the main analyses, the first four scales were added, resulting in a range of 0–40. In previous adolescent samples the Cronbach’s alpha of the total difficulties score was = .80, and validity was demonstrated [69]. One of the questions on the peer relationship problems scale addresses bullying (“Other children or young people pick on me or bully me”), which might cause confounding. Therefore, this question was excluded, resulting in a maximum score of 38. To ensure the usual maximum range of 0–40, the total difficulties score was divided by 38 and then multiplied by 40. Neglecting one question in the total difficulties score did not affect the internal consistency in the current sample: Cronbach’s alpha including the bullying question was .80, Cronbach’s alpha excluding the bullying question was .79. Moreover, the factor structure remained coherent as principal axis factoring for the total difficulties score showed a change in loadings of maximum 6.50%.

For further analyses, the scores of the emotional symptoms and peer relationship problems scales were summed to obtain an internalizing score, and the scores of the conduct problems and hyperactivity/inattention scales were summed to an externalizing score. In previous adolescent samples the Cronbach’s alpha of the internalizing score was 0.66, the Cronbach’s alpha of the externalizing score was 0.76, and validity was demonstrated [70]. In the current sample the Cronbach’s alpha of the externalizing score was 0.70. When the bullying question was excluded, the maximum range of the internalizing score was 18 instead of 20. To ensure the usual maximum range, the internalizing score was divided by 18 and then multiplied by 20. This did not affect the internal consistency in the current sample: Cronbach’s alpha including the bullying question was 0.77, Cronbach’s alpha excluding the bullying question was 0.75. In addition, the factor structure remained coherent (principal axis factoring: maximum change in loadings = 3.29%).

Multilevel modelling was used to examine the associations of the independent variables bullying victimization (yes, no), gender (boy, girl), grade (grade 5, 6, 7, 8, 9; as a correlate of age), school type (A-level, B-level) and mental health problems (total difficulties score: 0–40) with PIG and PSMU. The model is graphically illustrated in Figure A1 in the supplementary material. We operationalized PIG and PSMU using their nine binary items. The associations were modelled as direct effects from the independent variables to the PIG and PSMU questions using a logit link function. Effects and thresholds were constrained to be equal across the nine questions but may differ between PIG and PSMU. This specification accounts for the bounded count nature of the data, similar to a binomial regression with nine trials, while avoiding violations of the Gaussian residual assumption. Simultaneously, it remains close to a continuous sum score in which all items are weighted equally, as prescribed in the manual for diagnosis. The constraints were validated using a principal component factor analysis showing one dominating factor on which all items loaded approximately equally (see Harman’s single factor test in the supplementary material).

The data had a three-level structure: participants nested within classes and classes nested within schools. The null model of a binomial regression with nine trials showed a small intraclass correlation coefficient (ICC) on the school-level (PIG: ICC = 0.01, 95% CI [0.00;0.01]; PSMU: ICC = 0.01, 95% CI [0.00;0.02]). The school-level predictor school type mostly explained the variance. A likelihood ratio tests indicated no significant improvement from adding random effects at the school level (PIG: χ²(1) = 0.26, p = .61; PSMU: χ²(1) = 3.33, p = .07). Given the minimal unexplained school-level variance, random effects at the school level were omitted. Instead, two random intercepts were specified at the class level (one for PIG and one for PSMU), with unconditioned covariance.

Data processing and descriptive statistics were processed using Stata 18. The model was estimated using the “TWOLEVEL” analysis in Mplus 8.10 with the robust maximum likelihood estimator (MLR). Results are reported as odds ratio (OR) referring to odds of fulfilling a PIG or PSMU criterion. A change of more than one criterion is interpreted as large effect and a change of less than 0.25 criterion is interpreted as small effect. This translates to a large (OR ≤ 0.64, OR ≥ 1.56), medium (0.64 < OR ≤ 0.90, 1.12 ≤ OR < 1.56) and small (0.90 < OR < 1.12) effect size. For continuous variables with standard deviation (SD), the thresholds were transformed by raising them to the power of 1/SD. Wald tests were used to assess potential differences in effects between PIG and PSMU and comparison to the null model. Statistical significance was set to α = 0.05. Estimates and covariance matrices from Mplus were imported into Stata and results were plotted using the marginal command, assuming a two-level binomial regression with nine trials.

We increased the granularity of bullying victimization and calculated the model by replacing the dichotomous predictor bullying by three dichotomous variables direct, indirect and cyberbullying. In the same manner, we split mental health problems into internalizing and externalizing problems.

Interactions of bullying victimization with gender, grade, school type and mental health problems were examined in separate models. In each model, an additional interaction term was added as an independent variable acting separately on the PIG and PSMU questions. The models were compared to models without interaction terms.