We used data from the Greek arm of the 2015 European School Survey Project on Alcohol and Other Drugs (ESPAD, www.espad.org), conducted by the Athens University Mental Health, Neurosciences, & Precision Medicine Research Institute (ESPAD Group, 2016). The target population was 16-year-old high school students (Grade 10). A nationwide stratified, clustered probability sample was drawn from the Ministry of Education’s list of schools, with the school class as the primary sampling unit. Stratification took into account geographical region and school type (private/public, comprehensive/vocational).

Of the 1,558 schools in the sampling frame, 175 were contacted and 167 (95.4%) participated. Self-completion, pencil-and-paper questionnaires were group-administered by trained research assistants to students in the sampled classes during school hours. Absences on the day of administration accounted for 7.0% of enrolled students, and 2.1% of those present either refused participation or lacked parental permission. Of the 3,507 questionnaires collected, 305 (8.7%) were excluded according to ESPAD data-cleaning rules, leaving 3,202 cases for analysis.

Statistical analysis was conducted using the Complex Samples module in IBM SPSS, Version 20.0 to take into account the stratified, clustered sampling design and case weights. The Pearson product–moment correlation coefficient was used to assess correlations between the numbers of days per week involving different Internet activities. Pearson’s chi-square test and Student’s t-test (via the GLM procedure) were used to compare categorical and scale variables, respectively, between genders. Listwise deletion of missing data was adopted for all analyses.

Latent class analysis (LCA) models were fitted to identify patterns of use across the six online activities investigated, using the R package poLCA (Linzer & Lewis, 2011). Models with one to five classes were fitted to determine the optimal number of classes, based on conceptual considerations and several statistical criteria, including Akaike’s Information Criterion (AIC), the Bayesian Information Criterion (BIC), sample size–adjusted BIC (a-BIC), G² fit statistic, and entropy (Linzer & Lewis, 2011). Models were fitted both with and without gender as a covariate. In both cases, the optimal number of classes identified was four. Since both the literature and our preliminary analyses indicated that types of Internet use are gender-sensitive, the four-class model including gender as a covariate was retained for further analysis. Posterior probabilities from this model were used to classify students into latent classes, applying the maximum probability assignment rule, which assigns individuals to the class with the highest posterior probability of membership (Bray et al., 2015).

The association between latent classes and CIUS status (compulsive/non-compulsive Internet use) was tested using a logistic regression model, with CIUS status as the dependent variable and gender and latent classes as explanatory variables. The associations between latent classes and family variables were subsequently tested in a multinomial logit regression model, with latent classes as the dependent variable and again controlling for gender. Class membership in this second model was calculated from the posterior probabilities of an inclusive LCA with the optimal number of classes, including gender and all family variables as covariates. Inclusive LCA was conducted to eliminate bias (Bray et al., 2015). This approach was not adopted for the relation between CIUS status and latent classes (the first model) due to the high correlation between them. Similarly to logistic regression, multinomial logit regression fits models for the ratios of the probabilities of belonging to each of the categories of the dependent variable, compared with the probability of belonging to the reference category. However, these relative probabilities are not odds, as in logistic regression, because the two probabilities involved are not complementary. Consequently, we refer to the exponentiated regression coefficients — which indicate the multiplicative effects of each explanatory variable on these relative probabilities — as ‘relative probability ratios’ rather than odds ratios. The data supporting the findings of this study are freely available in the Open Science Framework (OSF) at Fotiou (2026).

The sample for analysis consisted of 3,202 adolescent students (50.8% female), with a mean age of 15.7 years (range 15.3–16.2). Table 1 presents data on Internet use and the prevalence of adverse family characteristics, for the total sample and by gender. Nearly 70% of participants reported using the Internet every day in the week preceding the survey. Average use was 5.8 days (standard deviation [SD] = 2.1). The average number of days of Internet use ranged from less than one day per week for gambling and for selling/buying or searching for products, to more than five days per week for communicating on social media.

The numbers of days per week across the various Internet activities were positively correlated with each other and statistically significant at p ≤ .001, except for the use of social media with gaming (p = .030) and with gambling (p = .47). Overall, significantly more females (72.0%) than males (65.4%) reported daily use of the Internet, whereas males reported higher rates of gaming, gambling, and selling/buying or searching for products. A significantly higher percentage of females (19.0%) also scored high on the CIUS, compared with males (11.0%).

The best-fitting LCA model including gender, as indicated by the lowest values of the BIC and corrected AIC, comprised four classes (see Table 2). Figure 1 presents the conditional item response probabilities within each latent class. Adolescents in the first class — labelled “heavy-with-gaming/gambling” users — had the highest probabilities of reporting frequent use of the Internet for every activity and made up 14.3% of the sample. The second class — “heavy-without-gaming/gambling” users (44.1% of the sample) — consisted of adolescents with relatively high probabilities of reporting all types of Internet use except gaming and gambling. Adolescents in the third class — “moderate” users (21.4%) — reported moderate to frequent use of several types of Internet activity, including gaming but neither gambling nor streaming/downloading. The fourth class — “light” Internet users (20.3%) — included adolescents with high probabilities of reporting moderate or frequent use of social media and moderate use of reading, surfing, and searching for information, but low probabilities of using the Internet for other activities.

Boys had a higher probability than girls of belonging to the “heavy-with-gaming/gambling” class (.295 for boys versus .033 for girls) and the “moderate” Internet user class (.410 for boys versus < .001 for girls) (not shown in Table). Girls had a higher probability than boys of belonging to the “heavy-without-gaming/gambling” (.698 versus .158) and “light” (.269 versus .137) classes.

Logistic regression analyses for CIUS status by latent class (not shown in Table) showed that membership of the “heavy-with-gaming/gambling” Internet user class (Class 1) was associated with a sevenfold higher probability of compulsive Internet use [gender-adjusted odds ratio (aOR) and corresponding 99% confidence interval (CI): aOR = 6.6, 99% CI = 3.7–11.7], compared with “light” Internet users. Membership of the “heavy-without-gaming/gambling” (Class 2) and “moderate” Internet user classes (Class 3) was associated with probabilities of compulsive Internet use increased by approximately two- and threefold, respectively (aOR = 2.0, 99% CI = 1.3–3.3; and aOR = 2.7, 99% CI = 1.4–5.0), compared with “light” Internet users.

Table 3 presents the results of multinomial logistic regression analyses of latent class membership in relation to family characteristics. Values of goodness-of-fit statistics (pseudo R²) were .472 (Cox and Snell) and .510 (Nagelkerke). Adolescents in households with low parental monitoring were twice as likely to be “heavy” Internet users (with or without gambling/gaming) [Relative Probability Ratio (RPR)_Class1 = 2.0, 99% CI = 1.3–3.2 and RPR_Class2 = 1.8, 99% CI = 1.2–2.6)] and 60% more likely to be “moderate” Internet users (RPR_Class3 = 1.6, 99% CI = 1.0–2.3) than “light” users. Students with low satisfaction in their relationship with their father were less likely to be “heavy-without-gaming/gambling” or “moderate” users (RPR_Class2 = 0.5, 99% CI = 0.2–0.9 and RPR_Class3 = 0.5, 99% CI = 0.3–0.9, respectively) than “heavy-with-gaming/gambling” users.

Among the strengths of the present study is its use of mixture modelling to classify adolescents into latent groups of Internet users with differentiated profiles relating to the frequency of their online activity and the role of family factors. However, the study is not without limitations. First, the entropy level of .71 in the present study may arguably be considered to indicate some uncertainty in class separation. Entropy values closer to 1 indicate clearer demarcation of classes (Celeux & Soromenho, 1996). Although no generally agreed cut-offs exist, a rule of thumb is that values above .80 may be regarded as indicating acceptable entropy. In this respect, the entropy level reached in the present study suggests that at least some of the classes can be reasonably well differentiated.

A key limitation is the cross-sectional design of the research, which allows for the detection of associations but not the determination of ætiology or temporality between heavy Internet use and family functioning. Therefore, we were unable to establish whether a poor relationship with the father preceded or followed adolescents’ heavy use of the Internet. Importantly, our research design could not capture the complex and, in many respects, bidirectional effects of heavy Internet use among adolescents and the Internet-specific attitudes and behaviours of their parents (e.g., Koning et al., 2018; Lin et al., 2020).

In addition, as per standard procedure (Guttormsson et al., 2015), we used self-report questionnaires for data collection. This means that the data for both the dependent and explanatory variables were drawn from the same respondents, thereby increasing the risk of systematic measurement error that could account for a considerable proportion of the observed variance (known as common method variance; Podsakoff et al., 2012). However, it should be noted that the items measuring the key constructs in our study — for example, the relationship with mother or father — were not perceptually linked to Internet use, were presented in formats different from those measuring Internet use and appeared in a separate part of the questionnaire. All of these features are thought to help mitigate bias.

An additional limitation of the present study concerns the nature of the measures used. As with all analyses based on large-scale surveys such as ESPAD, constructs could not be assessed in depth due to constraints of space and scope. Measures of internet use frequency relied on simple self-reports of days of use per activity, and although the CIUS is psychometrically robust, the appropriate cut-off point for dichotomisation among adolescent internet users remains under debate (see, e.g., Laconi et al., 2014). The > 28 threshold was applied for reasons of comparability and interpretability. Similarly, family-related variables were measured through single items, which provide only a limited perspective on the broader family context. These measures should therefore be considered simplistic proxies. Nevertheless, retaining individual family items rather than aggregating them into a composite variable preserved detail and allowed identification of which specific family aspects were most relevant to compulsive internet use.

Another limitation is the lack of detail in the study regarding the devices that adolescents use to access the Internet, ranging from PCs to mobile devices such as smartphones, as these may affect not only the type of activity undertaken (Jeong et al., 2016) but also, very likely, the level of parental monitoring. Our findings should also be considered in the context of patterns of family management, parenting, and Internet use that are specific to Greece. Heavy Internet use is growing rapidly with the expansion of wireless access to fast and inexpensive Internet and with the proliferation of devices through which adolescents can go online. However, Greece remains among the European countries with the lowest levels of access to, and daily use of, the Internet (Eurostat, 2024). Replication of these findings with large, cross-national samples is important in order to enhance generalisability.

The present study extends the limited body of research that has sought to identify latent subgroups of adolescents who share common Internet use attributes and to examine how membership of these groups relates to family characteristics such as family support, quality of parental relationships, and parental monitoring. We identified two groups of adolescents characterised by higher-risk Internet use. These groups may warrant interventions that are flexible and tailored to their particular characteristics: boys who engage in frequent gaming/gambling together with online communication, and girls who engage in frequent online communication. Our results also suggest that poor quality of the father–adolescent relationship is associated with heavy Internet use which includes online gaming and gambling. Family-focused prevention programmes and interventions may therefore be needed to enhance family functioning and strengthen parent–child relationships, with particular emphasis on the dynamic interactions between adolescents and their fathers. Fathers, too, need support to become more actively involved in their children’s online behaviour.