Participants
This study was conducted using a subsample of the Internet User Cohort for Unbiased Recognition of Gaming Disorder in Early Adolescence (iCURE) study, which is an ongoing prospective cohort study in Korea. The 3rd, 4th, and 7th grade students participated in the iCURE study. The protocol for the iCURE study was published previously 21. Three-hundred-and-ninety-nine children participated in the iCURE study from an eligible population of 1,061 students at six elementary schools in the cities of Seoul and Uijungbu. To examine possible mediating effects of parent-child attachment through self-esteem in the relationship between parental marital conflict and increases in IGD features in children, we included students who met the following criteria: anyone classified as a “non-case of high risk of IGD” in the initial assessment, anyone living with both parents, and current game use at baseline. A “non-case of high risk of IGD” was defined as a student with a total score of < 10 in the Internet Game Use-Elicited Symptom Screen (IGUESS) survey. A current game user was defined as having played games over the past year. Of 399 participants, 294 children met the inclusion criteria. To investigate the effect of parental marital conflict on increases in IGD features at 12 months, we excluded 26 participants who did not complete a 12-month follow-up assessment, leaving 268 children in the final analysis.
Informed consents were acquired from all participants and from one of their parents after explaining the principles of the present research, confidentiality, and the freedom of choice to participate. This study received approval from the Institutional Review Board of The Catholic University of Korea for data analysis (MC19ENSI0001). The iCURE data management board released de-identified iCURE data.
Measurements
Data collection
Data collection was conducted at the above-mentioned schools during school hours. In the baseline assessment, third-grade and fourth-grade students completed the questionnaires in a class setting; a research assistant read the questions with a standard script to help with comprehension and minimize time demands. In the 12-month follow-up assessments, all students completed the questionnaires on their own, using a web-based self-administration method, with a supervising research assistant available to answer questions. Average times spent playing internet games during weekdays and weekends (days and minutes) were obtained from the iCURE study baseline data. One of each child’s parents completed a demographic questionnaire at the participant’s home or at a private space at the school, according to the participant’s preference. The parental questionnaire was administered by trained interviewers at the baseline interview in which, the parents’ year of education, employment status, and socioeconomic status (SES) data were obtained.
Parental marital conflict
Parental marital conflict was measured using the Children’s Perception of Intraparental Conflict Scale (CPIC), which was developed by Grych, Seid and Fincham (1992) 22 at baseline. Although the full version of the CPIC consists of 49 questions, we only adopted the characters of conflict subscale, which contained 19 items23 and included four dimensions: conflict frequency (four questions), conflict intensity (five questions), (lack of) conflict resolution (seven questions), and stability of conflict (three questions) to measure children’s perception of their parents’ marital conflicts by conducting a confirmatory factor analysis. The response format was a five-point Likert scale ranging from 1 (never) to 5 (always). Total scores ranged from 19 to 95. Cronbach’s alpha is 0.90 for this subscale.
Parent-child attachment
Parent-child attachment was measured using the Inventory of Parent and Peer Attachment-Revised version (IPPA-R) 24 at baseline. The original questionnaire was developed to assess children’s perceptions of the positive and negative affective/cognitive dimensions of their relationship with their parents, and the questionnaire aimed to be a strong measure of psychological security. The IPPA-R is a 25-item questionnaire with responses on a five-point scale, ranging from 1 (“almost never or never true”) to 5 (“almost always or always true”), with higher scores indicating stronger attachment. Children reported on father-child attachment (25 questions) and mother-child attachment (25 questions) separately. The Cronbach’s alphas of the IPPA-R in the original version were 0.87 in mothers and 0.89 in fathers, and in the present study they were 0.93 and 0.93, respectively.
Self-esteem
Rosenberg’s self-esteem scale has 10 items, each rated on a five-point Likert scale, ranging from 1 (strongly disagree) to 5 (strongly agree) 25, with a total score range of 10 to 50. Rosenberg defined self-esteem as an individual’s sense of worthiness, which integrates self-respect and self-confidence. This measure detects feelings of self-acceptance, self-respect, and generally positive self-evaluation. Higher scores indicate greater feelings of self-acceptance, self-respect, and generally positive self-evaluation. Cronbach’s alpha was 0.85 in this study. Self-esteem was evaluated at baseline.
High risk of internet gaming disorder
Self-reported IGD features were assessed by the IGUESS at both baseline and 12-month follow-up. This instrument was created based on the nine IGD criteria established by Diagnostic and Statistical Manual of Mental Disorders (DSM–5). Students were instructed to respond reflecting their gaming behavior within the last 12 months, with each item rated on a four-point scale: 1 =strongly disagree, 2 =somewhat disagree, 3 =somewhat agree, 4 =strongly agree). A higher score indicates greater IGD severity 26. A Cronbach’s alpha of 0.80 was observed in this study.
Covariates
Possible confounding factors, including gender, age, family SES, and baseline IGUESS score, were obtained from the iCURE baseline data to control for these variables in the final model. The SES data were obtained from parents’ self-evaluations using a seven-point visual analog scale from very low (1) to extremely high (7). SES was reclassified into the lower level of SES (from 1 to 3) and the higher level of SES (from 4 to 7).
Statistical analysis
Descriptive and correlation analyses were performed using SAS 9.4 (SAS Institute Inc., Cary, NC, USA). Descriptive data were summarized with numbers and percentages for the categorical variables or mean ± SD and ranges for the continuous variables. Structural equation modeling (SEM) was conducted to examine the measurement and mediation models using the Analysis of Moment Structures, version 23.0. (IBM Inc., Chicago, IL, USA). In a confirmatory factor analysis, convergent and discriminant validity examine the extent to which measures of a latent variable share their variance and how they are different from others, respectively. The criterion of Fornell–Larcker (1981) is commonly used to assess the degree of shared variance between the latent variables of the model 27. According to this criterion, convergent validity is assessed by means of: (1) factor loadings (standardized regression weights), (2) reliability (Cronbach’s alpha), and (3) average variance extracted (AVE). The last entity measures the level of variance captured by a construct versus the variance due to measurement error. Values above 0.7 are considered very good and those between 0.7 and 0.5 is acceptable. Construct reliability (CR) measures whether a set of indicators representing a construct are consistent in their measurement, and it is customary to use the Cronbach’s alpha 28— values above 0.7 demonstrate that a scale is internally consistent — for this purpose. In addition, we resolved to estimate the CR, which is a less biased estimate of reliability with an acceptable value of 0.7 and above. According to the Fornell–Larcker testing system, discriminant validity can be assessed by comparing the amount of the variance capture by the AVE construct and the shared variance with other constructs (standard error). The levels of the AVE for each construct should be greater than the squared correlation involving the constructs.
A bootstrapping procedure was used to test and verify the paths and indirect effects for statistical significance with bias-corrected methods 29. A model fit was assessed using the multiple fit indices in terms of absolute fit, incremental fit, and parsimony fit indices. The absolute fit indices included a chi-square ratio over the degrees of freedom (x2/df), the goodness-of-fit index (GFI), and the root mean square error of approximation (RMSEA). The incremental fit indices were assessed using the Tucker–Lewis Index (TLI), the normed fit index (NFI), and the comparative fit index (CFI). The adjusted goodness-of-fit index (AGFI) was used for parsimony fit indices. SEM literature suggests that a model fit is good when x2/df ≤ 3; CFI ≥ 0.90, TLI ≥ 0.90, GFI ≥ 0.90, NFI ≥ 0.90, RFI ≥ 0.90, AGFI ≥ 0.90, and RMSEA ≤ 0.06 30. If the path was not statistically significant in the full hypothetical model, we deleted the insignificant path in the alternative model. We compared the model fit index of the two models and adopted the final model with paths that fit significantly better than others.