Psychological Network Analysis for Risk and Protective Factors of Problematic Social Media Use

Abstract

Identifying when and which adolescents are at increased risk of developing problematic social media use (PSMU) is critical for effective prevention and early intervention. Previous research has examined risk and protective factors using theory-driven (confirmatory-explanatory) approaches, such as regression models. However, few studies have simultaneously considered personal, peer, and parent characteristics to assess their relative contributions, and none have explored how these factors are structurally interrelated using data-driven (inductive–exploratory) approaches. To address these gaps, this study combines logistic regression and psychological network analysis to examine which personal, parent, and peer factors are most relevant in identifying at-risk/problematic social media use among adolescents. Using three waves of data analyzed cross-sectionally from N = 2441 secondary school students, adolescents were classified as normative (0–1 symptoms) or at-risk/problematic (2+ symptoms) users based on the Social Media Disorder Scale. Logistic regression showed that fear of missing out, impulsivity, depressive symptoms, intensity of meeting with friends, and reactive parental rules uniquely predicted at-risk/problematic use. Psychological network analysis identified self-esteem, attention problems, impulsivity, depressive symptoms, and life satisfaction as central, highly interconnected nodes. These findings show that theory- and data-driven approaches illuminate different aspects of PSMU risk, and that network analysis can generate novel hypotheses about underlying processes.

1. Introduction

Online social media, including social networking sites and instant messengers, have become a central part of everyday life. Today, young people grow up in a hybrid world where offline and online environments, interactions, and experiences are deeply intertwined. Indeed, recent research among North American and European youth found that almost four in ten adolescents use popular social media platforms almost constantly throughout the day [1,2,3]. Although social media can help adolescents navigate critical developmental tasks—such as constructing a self- and social identity and maintaining relationships with peers—there is a growing recognition that some youth are unable to regulate their social media use (SMU) and, as a result, experience difficulties in important life domains (e.g., [4,5,6]).

Problematic social media use (PSMU)—also referred to in the literature as social media addiction, social media disorder, and compulsive social media use—is generally defined as maladaptive use of social media characterized by addiction-like symptoms and/or reduced self-regulation despite unfavorable consequences [6,7]. Studies have linked PSMU to lower mental well-being and life satisfaction (e.g., [6,8,9]), attention difficulties (e.g., [10,11]), lower sleep quality (e.g., [4,12,13]), and poorer academic performance (e.g., [9,14]). Furthermore, recent evidence highlights the importance of considering a broader spectrum of problematic use patterns beyond those meeting clinical thresholds. A representative study among 6,626 Dutch adolescents [15] found that while only 3.6% met criteria for PSMU (six to nine symptoms), a much larger group (34.8%) reported at-risk SMU (two to five symptoms). Critically, this study demonstrated that not only problematic users but also at-risk users were significantly more likely than normative users to experience adverse consequences from their SMU.

Considering these negative impacts, it is essential to identify when and which youth are at increased risk of developing PSMU, as well as the key factors that should be targeted for prevention and early intervention. However, significant gaps remain in our understanding of how multiple risk and protective factors across different domains of influence interact to shape PSMU. Most existing studies rely on theory-driven approaches that focus on a single domain or a small set of variables, which may obscure complex dynamics within and across domains and constrain our understanding of PSMU risk processes. To address this limitation, this study combines theory-driven and data-driven approaches to examine a broad set of intrapersonal and social-contextual factors. The theory-driven component tests the unique contributions of established predictors, while the data-driven component uncovers broader interrelations and emergent patterns. Although our analyses are correlational (not causal), this combined approach offers richer insights into PSMU risk processes than either method applied independently.

This paper is structured as follows. Section 2 reviews prior research on PSMU, introduces the two analytical approaches applied in the current study, and summarizes the specific goals of the study. Section 3 outlines the data, measures, and analytical procedures. Section 4 presents descriptive statistics for our data and empirical results for the two approaches we contrast and combine here. We discuss our findings at length in Section 5 and conclude in Section 6.

2. Related Work

2.1. Predictors of Problematic Social Media Use

Until now, most empirical work on factors involved in PSMU has concentrated on personal characteristics. For instance, in line with the body of evidence linking attention-deficit/hyperactivity disorder (ADHD) to addictions more generally (e.g., [16]), studies have demonstrated that ADHD symptoms (e.g., attention problems, impulsivity, hyperactivity) predict higher levels of PSMU [10,11,17]. Adolescents with ADHD symptoms may be more easily distracted by the constant stream of information and notifications inherent in social media platforms and have more difficulty inhibiting their immediate impulses, resulting in more PSMU. In addition, the association with psychological distress and mental well-being factors has been extensively analyzed. Studies have shown that individuals with depression, low life satisfaction, and low (physical) self-esteem are more likely to engage in PSMU (e.g., [6,17,18,19]). These relationships are frequently described as cyclic or escalating in nature: Adolescents may initially turn to social media as a coping mechanism to manage emotional distress or psychosocial vulnerabilities. The immediate rewards this use provides, such as mood enhancement and a sense of self-efficacy, can encourage continued and increased SMU. Over time, however, this increased engagement may reinforce the preexisting issues that prompted the behavior in the first place. As adolescents increasingly rely on social media to cope, this feedback loop can intensify, potentially resulting in the development of PSMU [5,6]. Others, however, have argued that the relationships between mental well-being factors and PSMU are highly complex, shaped not only by culturally relevant factors such as gender differences [20] but also by intervening and mediating variables with poorly understood underlying mechanisms [19,21].

In addition to psychological functioning and well-being factors, several personality traits have been linked to PSMU. For instance, individuals with elevated narcissistic traits are thought to be more prone to developing PSMU because specific attributes of social media—such as carefully curated self-presentation, explicit validation through “likes” and followers, and access to a potentially large audience—may fulfill narcissistic individuals’ self-enhancement and self-validation needs [18,22]. Moreover, a growing body of research has highlighted the role of fear of missing out (FoMO) in the etiology of PSMU. FoMO, defined as “a pervasive apprehension that others might be having rewarding experiences from which one is absent” [23], may drive PSMU through a strong desire to remain continuously connected with what others are doing [23,24]. FoMO has also been shown to mediate the relationship between deficits in basic psychological needs (i.e., competence, autonomy, and relatedness, as outlined in Self-Determination Theory [25]) and PSMU [23,24], thereby helping to explain observed associations between depression or low life satisfaction and PSMU [19,21].

Surprisingly, the role of the social environment in PSMU has received far less attention, despite theoretical perspectives (e.g., ecological systems theory [26]; transactional model [27]) and empirical work emphasizing the crucial role of the family and peer context in developmental processes and outcomes during adolescence, including problematic behaviors [28]. Moreover, social media use is inherently interpersonal and central to forming and maintaining relationships. Studies have suggested that adolescents who suffer from loneliness and poor (perceived) social competencies tend to prefer online social interactions over offline encounters with peers because they find the former less threatening and more conducive to positive self-presentation. The perceived benefits and control experienced through social media use may subsequently foster maladaptive beliefs (e.g., that one can only be successful online) that can ultimately escalate into PSMU [15,29].

In the family context, only a few studies have explored how parental mediation influences PSMU, and the findings for problematic Internet use more broadly remain mixed, partly due to measurement differences. Nevertheless, available evidence suggests that restrictive parental rule-setting (i.e., clear advance rules about adolescents’ access to social media) may be more effective in protecting against PSMU than reactive parental rule-setting (i.e., ad hoc restrictions [30,31]). A recent study further showed that restricting Internet access through strict parental rule-setting may no longer have the desired protective effect once adolescents have become highly engaged or problematic users [32]. Lakić and colleagues likewise reported minimal effects of parental mediation on gaming addiction [33]. Finally, research suggests that parent–adolescent communication about online behavior marked by mutual respect—where adolescents feel comfortable, understood, and taken seriously by their parents—may help prevent PSMU [31].

In sum, prior research has identified a range of intrapersonal and social-contextual factors potentially involved in the development of PSMU. However, to our knowledge, no study to date has assessed the relative contribution of personal, peer, and parental factors in a single model.

2.2. Theory-Driven vs. Data-Driven Approaches

In quantitative social sciences, efforts to understand social phenomena such as PSMU have traditionally been guided by confirmatory–explanatory approaches that test theory-derived hypotheses about causal mechanisms (e.g., risk and protective factors) by collecting empirical data and evaluating how well the data fit hypothesized models [34]. Because they are constrained by theory, research designs and analytic tools within this tradition (e.g., randomized controlled trials, regression models) prioritize parsimony, focusing on a limited set of factors presumed to influence the outcome of interest [35,36]. Such theory-driven, parsimonious models can help explain individual effects, refine theory, and guide interventions. However, they often fail to provide a comprehensive understanding of complex phenomena.

With the emergence of big data analysis approaches, social scientists have grown increasingly interested in more inductive–exploratory approaches to data analysis. These alternative approaches rely on computer-aided analysis and methods from computer science (e.g., expert systems, machine learning algorithms, data mining) to iteratively estimate complex models of social phenomena that best represent the observed data [34]. Data-driven techniques enable the exploration of different objectives and questions; for example, they can be used to analyze the structure of multivariate data in the absence of theory on how variables are related and identify patterns of risk and resilience that theory-driven approaches to model building may have missed. These structures may then generate new hypotheses that can be tested empirically [35,36].

From a theoretical perspective, the various risk and protective factors involved in the development of PSMU do not operate in isolation but, rather, within a complex system of multiple pathways and dynamic interactions [26]. That is, components of the system likely not only influence PSMU itself but also exert positive, negative, unidirectional, or reciprocal effects on each other. However, traditional analytical approaches to studying PSMU often focus on discrete aspects of this system. Even when multiple factors are examined together, covariances among these factors are either not the focus of attention or are attributed to a common underlying construct (e.g., PSMU) rather than being interpreted as direct interactions between the factors themselves. For example, multiple regression models describe the linear relationship between a key outcome variable and several predictors by estimating the unique variance explained by each predictor while controlling for the effects of the others. Similarly, at the symptom level, principal component analysis and sum scores treat symptoms (e.g., preoccupation, withdrawal, problems) as passive indicators of an underlying common cause, PSMU. Although such methods yield valuable insights, they do not capture the complex interrelations among risk and protective factors or the potential for changes in one factor to influence other components within the system.

2.3. Psychological Network Analysis: A Data-Driven Approach to Understanding PSMU

An alternative analytical approach that has gained increasing traction in recent years is psychological network analysis, a promising method for understanding the structure of psychopathology and related phenomena, and one that can better accommodate the complex, interconnected nature of PSMU. This approach conceptualizes psychological phenomena as complex systems of causally interacting components (e.g., [37,38]). From this perspective, PSMU consists of networks of relationships among cognitions, traits, emotional states, behaviors, and contextual factors; it emerges from the dynamic interactions among these components rather than being the root cause of all of them.

Psychological network analysis offers insight into the structural organization of psychological phenomena and the roles of specific factors within the network using a data-driven approach. The network structures can be graphically depicted as nodes representing observed variables and edges indicating pathways on which nodes affect each other [37,39]. This approach also identifies “central” factors defined by strongly connected nodes. Due to their strong relations with other nodes and/or their high interconnectedness, such factors are likely to have a greater influence over the entire network [40].

Understanding how networks of potential PSMU predictors are organized helps to gain insight into how changes in one factor could activate pathways of connected factors (e.g., lower self-esteem may reduce life satisfaction, which, in turn, may elevate depressive symptoms), ultimately contributing to shifts in the system as a whole. As such, psychological network analysis may be particularly useful to explore potential mechanisms through which PSMU develops that can inform theory and clinical practice. Although network edges are conceptualized as causal pathways between factors, psychological network analysis remains a fundamentally exploratory technique; the connections it estimates should, therefore, be interpreted as hypothesis-generating for potential causal pathways that require further statistical testing [40].

Despite its promise, psychological network analysis has only rarely been applied to PSMU. Existing studies have focused either on the organization of PSMU symptoms (e.g., [13,41]) or on associations between PSMU and personal factors (e.g., [42,43]). To our knowledge, no study has yet explored patterns of relationships among potential personal, peer, and parent predictors of PSMU using this approach. However, psychological network analysis has been successfully applied in related healthcare domains, such as predicting treatment response [44], underscoring its potential value for PSMU research.

2.4. The Current Study

Our study adds to the existing literature on PSMU risk and protective factors in at least three ways. First, we distinguish between normative social media users and those who show signs of at-risk or problematic social media use (at-risk/problematic SMU), based on the number of symptoms reported. This approach enables us to examine a broader population of adolescents who experience adverse consequences from their social media use, providing valuable insights for identifying early indicators of PSMU.

Second, because many interrelated factors may be involved in the development of PSMU, we simultaneously investigate a set of personal, peer, and parent characteristics that previously have been linked to PSMU or, more generally, to Internet-use disorders. Specifically, within the personal domain, we investigate the role of ADHD symptoms, depressive symptoms, life satisfaction, self-esteem, physical self-esteem, narcissism, and FoMO. Within the peer domain, we explore perceived social competence and the intensity or frequency of meeting with friends face-to-face. Finally, within the parent domain, we investigate restrictive and reactive parental rules regarding SMU and the quality of communication about Internet-related matters. Studying these three domains of influence together yields a more comprehensive profile of youth at risk for developing PSMU and may improve predictive accuracy.

Third, given that traditional statistical approaches to studying psychopathology are less suited for capturing the interrelatedness of risk and protective factors that drive PSMU, the current study utilizes two complementary approaches to explore which factors are particularly relevant in identifying at-risk/problematic SMU among adolescents. We use logistic regression analysis to evaluate the relative contribution of the personal, peer, and parent factors in predicting the likelihood of being an at-risk/problematic social media user in three consecutive years. In addition, we use psychological network analysis to explore the network structure of potential risk and protective factors and identify central factors in the network. These two complementary approaches provide novel insights into key factors and processes that distinguish at-risk/problematic social media users from normative users.

Based on our review of the literature, we hypothesize that higher levels of ADHD symptoms, depressive symptoms, FoMO, and narcissism, along with lower life satisfaction, self-esteem, and physical self-esteem, will be associated with greater odds of being classified as an at-risk/problematic user. Within the peer domain, we expect that lower perceived social competence and lower frequency of face-to-face meetings with friends will also increase risk. Regarding the parent domain, we anticipate that more reactive parental rules and a lower quality of communication will increase risk, whereas more restrictive parental rules are expected to be protective against at-risk/problematic SMU. However, the relative contribution of these factors when examined simultaneously across domains remains an open empirical question, as does the identification of central factors through psychological network analysis, which is fundamentally exploratory and hypothesis-generating.

3. Methodology

3.1. Participants and Procedure

Data for this study came from the Digital Youth Project project, a longitudinal study on online behaviors and well-being among Dutch secondary school students [9]. For the present study, data from the second, third, and fourth waves were used, which took place in February and March of 2016, 2017, and 2018, respectively. Data from the first assessment (2015) were excluded, as some of the variables of interest were not measured in this wave. The waves included in the current study are further referred to as T1, T2, and T3. In total, n = 3799 respondents participated in at least one of the three assessments (n = 644 contributed data at all three waves, n = 1622 at two waves, and n = 1533 at one wave). Non-response and variation in sample sizes between waves were mainly due to entire schools or classes (e.g., exam classes) dropping out of the project and new schools or classes joining the project.

At T1, 1928 adolescents participated in the study (56.7% boys). Respondents were between 10 and 16 years of age (M = 13.31, SD = 0.91), and 37.1% were first year (7th grade) students. Most adolescents (74.3%) had a Dutch ethnic background. Students were enrolled in pre-vocational (68.9%), higher general (27.1%), and pre-university (4.0%) education. At T2, 2708 adolescents between 11 and 17 years of age (M = 13.94, SD = 1.20) completed the survey. Sample characteristics at T2 were similar to those at T1 (53.9% boys; 74.6% Dutch ethnic background; 62.6% pre-vocational, 27.4% higher general, 10.0% pre-university levels). At T3, 2073 adolescents between 11 and 18 years of age (M = 14.37, SD = 1.49) participated. Due to practical reasons, two pre-vocational level schools and several pre-vocational level classes from mixed schools dropped out at this wave, resulting in a sample with a higher educational level than T1 and T2 (52.1% boys; 78.7% Dutch ethnic background; 45.2% pre-vocational, 39.9% higher general, 14.9% pre-university levels).

Data were collected through online self-report questionnaires administered in the classroom setting using Qualtrics survey software. Trained research assistants were present to supervise data collection, answer questions, and ensure respondents’ privacy. Prior to the survey assessment, parents were provided information about the study and the opportunity to decline their child’s participation. Adolescents with passive informed parental consent were informed about the purpose of the study, that participation was voluntary and anonymous, and that they could withdraw their participation at any time.

3.2. Measures

3.2.1. At-Risk/Problematic SMU

Adolescents’ PSMU was assessed with the Social Media Disorder Scale (SMDS) [45]. This scale consists of nine dichotomous items (0 = no, 1 = yes) that cover symptoms of addiction to social media, including preoccupation, withdrawal, tolerance, persistence, displacement, conflict, deception, escape, and problems. For example, to measure withdrawal, respondents were asked, “During the past 12 months, have you often felt bad when you could not use social media?” The SMDS corresponds to the nine diagnostic criteria for internet gaming disorder outlined in the appendix of the Diagnostic and Statistical Manual of Mental Disorders, 5th edition [46] and has been found to have solid structural, convergent, and criterion validity and good reliability among adolescents [15,45].

A sum score was computed to reflect the number of present criteria. Based on a recent representative Dutch study [15], respondents who reported zero or one symptom were classified as normative users, and respondents who reported two or more symptoms were classified as at-risk/problematic users. Given the dichotomous nature of the SMDS, internal consistency was calculated using the tetrachoric correlation matrix. This yielded an ordinal $\alpha$ that varied between 0.83 and 0.86 across waves.

3.2.2. Predictors of At-Risk/Problematic SMU

To examine which factors are particularly relevant in identifying at-risk or problematic SMU among adolescents, a set of personal, peer, and parent characteristics was assessed. Table 1 summarizes the instruments used to measure these constructs. Unless stated otherwise, mean scores were calculated for each predictor variable, with higher scores indicating a higher level of the measured construct.

3.3. Analytical Approach

3.3.1. Imputation of Missing Data

Due to the aforementioned variation in sample sizes across waves, the amount of missing data on the study variables was substantial. To retain respondents with partially missing data in the analysis, multiple imputation (MI) via chained equations was used to impute missing values. MI has been shown to be superior to listwise deletion as it improves statistical power and reduces potential bias related to missing data [47].

Prior to imputation, steps were taken to ensure that only intended values were imputed. First, as the Digital Youth Project focuses on secondary school students, only cases who were in secondary school during all three waves were eligible for imputation. Respondents who did not participate in earlier waves because they were still in primary school during that time and respondents who did not participate in later waves because they had already left secondary school were excluded, as imputation of missing data would not be justified in their case. This resulted in a final sample of N = 2441 respondents.

Table 1. Instruments used to assess personal, peer, and parent predictors of at-risk/problematic social media use (SMU).

Concept	Instrument	# of Items, Example Item	Cronbach’s $\mathsf{\alpha }$ (Range)	Likert Scale
Personal variables
ADHD symptoms	ADHD Questionnaire [48]	Subscale Attention Problems: 9 items, e.g., “I am often distracted or have unimportant thoughts”	0.87–0.88	1 = never, 5 = very often
		Subscale Impulsivity: 6 items, e.g., “I go from one task to another without completing the first one”	0.80–0.83
		Subscale Hyperactivity: 6 items, e.g., “I have trouble sitting still when this is expected from me”	0.83–0.86
Depressive symptoms	Depressive Mood Inventory [49]	6 items, e.g., “How often did you have the following feelings in the last 12 months … felt unhappy, sad, or depressed?”	0.81–0.86	1 = never, 5 = very often
Life satisfaction	Student’s Life Satisfaction Scale [50]	7 items, e.g., “I have a good life”	0.83–0.85	1 = strongly disagree, 6 = strongly agree
Self-esteem	Rosenberg Self-Esteem scale [51]	5 items, e.g., “I feel that I have a number of good qualities”	0.81–0.83	1 = completely untrue, 5 = completely true
Physical self-esteem	Self-Perception Profile for Adolescents [52]	Subscale Physical Appearance: 5 items, e.g., “I am happy with how I look”	0.84–0.85	1 = completely untrue, 5 = completely true
Narcissism	Childhood Narcissism Scale [53]	10 items, e.g., “Kids like me deserve something extra”	0.84–0.85	1 = completely untrue, 4 = completely true scores summed
Fear of missing out (FoMO)	Fear of Missing Out scale [23]	5 items, e.g., “I get worried when I find out that my friends are having fun without me”	0.82–0.84	1 = completely untrue, 5 = completely true
Peer variables
Perceived social competence	Self-Perception Profile for Adolescents [52]	Subscale Close Friendship: 5 items, e.g., “I can keep a friendship for a long time”	0.65–0.70	1 = completely untrue, 5 = completely true
Intensity of meeting with friends	[54]	4 items, e.g., “How often are you at your friend’s house?”	0.85–0.87	1 = never, 6 = very often
Parent variables
Restrictive parental rules regarding SMU	[31]	5 items, e.g., “How often are you allowed to use the Internet/games or your smartphone/tablet … while your homework isn’t finished yet?”	0.76–0.84	1 = never, 5 = very often reverse coded
Reactive parental rules regarding SMU	[31]	4 items, e.g., “How often do your parents tell you that you have to turn off the computer/tablet or smartphone?”	0.83–0.87	1 = (almost) never, 5 = more than five times a day
Quality of communication	[55]	3 items, e.g., “I feel comfortable talking about Internet or game use with my parents	0.91	1 = completely untrue, 5 = completely true

Table 1. Instruments used to assess personal, peer, and parent predictors of at-risk/problematic social media use (SMU).

Concept	Instrument	# of Items, Example Item	Cronbach’s $\mathsf{\alpha }$ (Range)	Likert Scale
Personal variables
ADHD symptoms	ADHD Questionnaire [48]	Subscale Attention Problems: 9 items, e.g., “I am often distracted or have unimportant thoughts”	0.87–0.88	1 = never, 5 = very often
		Subscale Impulsivity: 6 items, e.g., “I go from one task to another without completing the first one”	0.80–0.83
		Subscale Hyperactivity: 6 items, e.g., “I have trouble sitting still when this is expected from me”	0.83–0.86
Depressive symptoms	Depressive Mood Inventory [49]	6 items, e.g., “How often did you have the following feelings in the last 12 months … felt unhappy, sad, or depressed?”	0.81–0.86	1 = never, 5 = very often
Life satisfaction	Student’s Life Satisfaction Scale [50]	7 items, e.g., “I have a good life”	0.83–0.85	1 = strongly disagree, 6 = strongly agree
Self-esteem	Rosenberg Self-Esteem scale [51]	5 items, e.g., “I feel that I have a number of good qualities”	0.81–0.83	1 = completely untrue, 5 = completely true
Physical self-esteem	Self-Perception Profile for Adolescents [52]	Subscale Physical Appearance: 5 items, e.g., “I am happy with how I look”	0.84–0.85	1 = completely untrue, 5 = completely true
Narcissism	Childhood Narcissism Scale [53]	10 items, e.g., “Kids like me deserve something extra”	0.84–0.85	1 = completely untrue, 4 = completely true scores summed
Fear of missing out (FoMO)	Fear of Missing Out scale [23]	5 items, e.g., “I get worried when I find out that my friends are having fun without me”	0.82–0.84	1 = completely untrue, 5 = completely true
Peer variables
Perceived social competence	Self-Perception Profile for Adolescents [52]	Subscale Close Friendship: 5 items, e.g., “I can keep a friendship for a long time”	0.65–0.70	1 = completely untrue, 5 = completely true
Intensity of meeting with friends	[54]	4 items, e.g., “How often are you at your friend’s house?”	0.85–0.87	1 = never, 6 = very often
Parent variables
Restrictive parental rules regarding SMU	[31]	5 items, e.g., “How often are you allowed to use the Internet/games or your smartphone/tablet … while your homework isn’t finished yet?”	0.76–0.84	1 = never, 5 = very often reverse coded
Reactive parental rules regarding SMU	[31]	4 items, e.g., “How often do your parents tell you that you have to turn off the computer/tablet or smartphone?”	0.83–0.87	1 = (almost) never, 5 = more than five times a day
Quality of communication	[55]	3 items, e.g., “I feel comfortable talking about Internet or game use with my parents	0.91	1 = completely untrue, 5 = completely true

Second, missing values in demographic variables (e.g., gender, age, educational level) were replaced or calculated using observed information from other waves. The percentage of missing data on study variables ranged from 25.7% (PSMU, T2) to 57.2% (quality of communication, T3). Using R’s mice 3.9.0 package [56], five multiple imputed datasets were created using fully conditional specification (i.e., iterative, based on a predefined imputation model). The imputation models included gender, age, educational level, and the study variables as predictors. Variables were only included in the imputation model if they had at least 50% usable cases and a minimum correlation of 0.2 with the variable to be imputed, to ensure relevant predictors. For continuous variables, missing values were imputed based on predictive mean matching, which replaces missing values with observed values from similar cases based on regression predictions. For dichotomous variables, logistic regression imputation was used to impute binary outcomes.

3.3.2. Data Analysis

Although data were collected longitudinally from the same cohort across three waves, the current study treated each wave as an independent cross-sectional snapshot. Because our primary aim was to identify which factors are particularly relevant in distinguishing at-risk/problematic users—either by uniquely contributing to at-risk/problematic SMU or by occupying a central position in the broader network of risk and protective factors—a cross-sectional strategy was appropriate. Additionally, analyzing the waves separately enabled us to assess the consistency of findings across time points, thereby strengthening the robustness of our results.

First, descriptive statistics were calculated for the total sample and normative and at-risk/problematic social media users separately. Bivariate correlations between study variables were obtained for the total sample. Second, logistic regression analyses were performed in R to identify personal, peer, and parent factors associated with higher odds of at-risk/problematic SMU at each wave. All predictor variables were entered simultaneously into the regression model with at-risk/problematic SMU as dependent dichotomous variable (0 = normative SMU, 1 = at-risk/problematic SMU). Third, psychological networks were estimated using Gaussian graphical models [57]. In these models, nodes represent personal, peer, and parent variables and edges represent their partial correlations (i.e., the relationship between two variables after controlling for the effect of all other variables in the network). Given the large number of parameters typically estimated in psychological networks, it is recommended to regularize partial correlations using the graphical LASSO algorithm (glasso [58]). This algorithm reduces small or unstable edges to zero to obtain a parsimonious network that only reflects the most robust interactions between nodes.

Separate networks were estimated for normative and at-risk/problematic users at each wave. Since it is not possible to perform psychological network analyses with Gaussian graphical models on multiple imputed datasets, we estimated networks on a single (i.e., the first) imputed dataset. Networks were estimated in R using the qgraph package [59], which implements the glasso regularization in combination with the extended Bayesian information criterion (EBIC) to select the most optimal estimation [58]. The Fruchterman–Reingold algorithm was used to visualize networks so that closely spaced nodes represent strongly connected variables and thicker edges represent stronger associations. To interpret the estimated networks, stability and centrality indices were calculated using R’s bootnet package. Stability indices provide insight into (a) the accuracy of edge weights by calculating their bootstrapped confidence intervals (CIs), and (b) the stability of centrality indices by applying a case-dropping subset bootstrap and calculating a correlation-stability (CS) coefficient [40]. Centrality indices reflect the importance of nodes in the network. For each node, strength (the sum of the absolute edge weights connected to a node), betweenness (the number of times a node is on the shortest path between two other nodes, indicating a potential bridging factor), and closeness (how closely located, on average, a node is to other nodes) were calculated. We refer the reader to [39,40] for more details on network estimation, regularization, stability, and centrality.

4. Results

4.1. Descriptive Statistics

At T1, 34.5% of respondents were classified as at-risk/problematic social media users (of which 1.7% were problematic users with six or more symptoms); at T2, 33.7% (1.7% problematic users); and at T3, 36.6% (0.9% problematic users). Correlations between the study variables are available in Supplementary Material, Table S1.

Table 2. Descriptive statistics for study variables, specified for the total sample and normative (Norm.) and at-risk/problematic (Risk) social media users.

	T1			T2			T3
	Total	Norm. SMU	Risk SMU	Total	Norm. SMU	Risk SMU	Total	Norm. SMU	Risk SMU
	(N = 2441)	(n = 1598)	(n = 843)	(N = 2441)	(n = 1619)	(n = 822)	(N = 2441)	(n = 1548)	(n = 893)
Variable	M (SD)	M (SD)	M (SD)	M (SD)	M (SD)	M (SD)	M (SD)	M (SD)
Age	13.26 (0.95)	13.26 (1.10)	13.27 (1.18)	14.37 (0.96)	14.40 (1.10)	14.31 (1.24)	15.46 (0.97)	15.50 (1.25)	15.40 (1.26)
Attention prob.	2.24 (0.74)	2.10 (0.85)	2.51 (0.82)	2.39 (0.74)	2.28 (0.77)	2.61 (0.77)	2.44 (0.72)	2.36 (0.80)	2.58 (1.02)
Impulsivity	1.93 (0.71)	1.80 (0.77)	2.18 (0.86)	2.00 (0.70)	1.90 (0.69)	2.21 (0.77)	2.01 (0.64)	1.94 (0.72)	2.14 (0.83)
Hyperactivity	2.23 (0.88)	2.09 (0.90)	2.51 (1.12)	2.22 (0.84)	2.12 (0.85)	2.43 (0.89)	2.24 (0.79)	2.15 (1.09)	2.38 (1.05)
Depres. symp.	2.21 (0.75)	2.07 (0.81)	2.48 (0.80)	2.31 (0.77)	2.18 (0.80)	2.55 (0.86)	2.39 (0.77)	2.28 (1.20)	2.59 (1.01)
Life satisfaction	4.73 (0.84)	4.85 (0.82)	4.51 (1.00)	4.56 (0.85)	4.67 (0.95)	4.34 (1.04)	4.44 (0.82)	4.54 (1.12)	4.26 (1.18)
Self-esteem	3.88 (0.71)	3.97 (0.87)	3.71 (0.76)	3.78 (0.72)	3.87 (0.77)	3.59 (0.82)	3.73 (0.67)	3.81 (1.27)	3.59 (1.17)
Phys. self-esteem	3.61 (0.81)	3.71 (0.86)	3.43 (1.06)	3.47 (0.79)	3.54 (0.87)	3.31 (0.88)	3.43 (0.75)	3.50 (0.86)	3.32 (0.89)
Narcissism	2.31 (0.53)	2.29 (0.57)	2.34 (0.65)	2.21 (0.54)	2.19 (0.61)	2.25 (0.57)	2.20 (0.49)	2.18 (0.59)	2.22 (0.74)
FoMO	1.73 (0.68)	1.60 (0.71)	1.99 (0.88)	1.78 (0.69)	1.68 (0.71)	1.98 (0.78)	1.85 (0.67)	1.76 (0.78)	2.00 (1.04)
Perc. soc. comp.	4.33 (0.66)	4.37 (0.74)	4.26 (0.87)	4.31 (0.67)	4.34 (0.79)	4.25 (0.72)	4.27 (0.63)	4.32 (0.92)	4.19 (1.03)
Int. meet. friends	3.45 (1.04)	3.39 (1.13)	3.57 (1.15)	3.40 (1.08)	3.35 (1.11)	3.50 (1.22)	3.44 (1.05)	3.44 (1.24)	3.44 (1.26)
Restr. par. rules	3.25 (0.98)	3.32 (1.03)	3.10 (1.00)	2.94 (1.03)	2.97 (1.10)	2.89 (1.15)	2.63 (0.98)	2.65 (1.03)	2.60 (1.27)
React. par. rules	1.72 (0.72)	1.65 (0.69)	1.84 (0.82)	1.64 (0.72)	1.56 (0.73)	1.79 (0.96)	1.54 (0.63)	1.49 (0.89)	1.62 (0.85)
Qual. comm.	3.19 (1.06)	3.25 (1.13)	3.07 (1.31)	3.42 (1.03)	3.51 (1.16)	3.23 (1.20)	3.51 (0.94)	3.57 (1.45)	3.39 (1.52)

Bold denotes significance (p < 0.05).

presents descriptive statistics for the total sample and for normative and at-risk/problematic social media users. Independent samples t-tests (Benjamini–Hochberg corrected, Q = 0.05) revealed significant differences (shown with bold on the table) between normative users and at-risk/problematic users on most personal, peer, and parent variables. Across waves, at-risk/problematic users reported significantly more ADHD symptoms, depressive symptoms, FoMO, intensity of meeting with friends, and reactive parental rules regarding SMU than normative users. At-risk/problematic users also consistently reported lower life satisfaction, self-esteem, physical self-esteem, perceived social competence, and quality of communication with parents.

4.2. Logistic Regression Analyses

Table 3. Logistic regression analyses examining personal, peer, and parent predictors of at-risk/problematic social media use (N = 2441).

Variable	T1		T2		T3
	OR	95% CI	OR	95% CI	OR	95% CI
Attention problems	1.22	[0.97, 1.53]	1.15	[0.94, 1.40]	1.02	[0.82, 1.27]
Impulsivity	1.26 *	[1.01, 1.58]	1.27 *	[1.03, 1.56]	1.22	[0.94, 1.59]
Hyperactivity	1.07	[0.91, 1.27]	1.00	[0.86, 1.17]	1.02	[0.84, 1.23]
Depressive symptoms	1.17	[0.97, 1.41]	1.32 **	[1.10, 1.58]	1.29 **	[1.07, 1.55]
Life satisfaction	0.91	[0.74, 1.12]	0.89	[0.75, 1.05]	0.92	[0.70, 1.22]
Self-esteem	0.87	[0.67, 1.13]	0.76 *	[0.60, 0.97]	0.81	[0.57, 1.15]
Physical self-esteem	0.85	[0.70, 1.03]	0.94	[0.79, 1.11]	0.96	[0.80, 1.14]
Narcissism	1.21	[0.90, 1.63]	1.36 *	[1.03, 1.78]	1.30	[0.96, 1.77]
Fear of missing out (FoMO)	1.54 ***	[1.23, 1.92]	1.24 *	[1.05, 1.46]	1.23 *	[1.03, 1.47]
Perceived social competence	1.00	[0.84, 1.19]	1.05	[0.88, 1.27]	0.87	[0.71, 1.05]
Intensity of meeting friends	1.20 ***	[1.08, 1.33]	1.20 **	[1.07, 1.35]	1.08	[0.97, 1.20]
Restrictive parental rules	0.84 **	[0.75, 0.94]	0.93	[0.84, 1.03]	0.95	[0.84, 1.08]
Reactive parental rules	1.32 ***	[1.15, 1.51]	1.39 ***	[1.20, 1.61]	1.26 **	[1.08, 1.47]
Quality of communication	0.94	[0.84, 1.05]	0.80 ***	[0.71, 0.90]	0.94	[0.82, 1.08]

OR = odds ratio; CI = confidence interval. Reference category: normative users. Nagelkerke R2 = 0.19 (T1), 0.17 (T2), 0.10 (T3). *** p < 0.001; ** p < 0.01; * p < 0.05.

displays results of the logistic regression analyses examining the contribution of personal, peer, and parent factors in predicting adolescents’ at-risk/problematic SMU. Across all waves, higher levels of FoMO and reactive parental rules regarding SMU were associated with higher odds of being an at-risk/problematic social media user. Odds ratios ranged between 1.23–1.54 for FoMO and 1.26–1.39 for reactive parental rules. Other associations were less consistent across waves. Regarding personal factors, higher levels of impulsivity (OR_T1 = 1.26; OR_T2 = 1.27), depressive symptoms (OR_T2 = 1.32; OR_T3 = 1.29), and narcissism (OR_T2 = 1.36) were associated with significantly higher odds of at-risk/problematic SMU. Conversely, higher self-esteem levels (OR_T2 = 0.76) were associated with lower odds of at-risk/problematic SMU. Among the peer variables, increased intensity of meeting with friends was associated with higher odds of being an at-risk/problematic social media user (OR_T1 = 1.20; OR_T2 = 1.20). Finally, regarding parent variables, more restrictive parental rules regarding SMU (OR_T1 = 0.84) and higher communication quality (OR_T2 = 0.80) were associated with lower odds of at-risk/problematic SMU. The set of personal, peer, and parent variables accounted for 10–19% of the variance in at-risk/problematic SMU.

4.3. Psychological Network Analyses

4.3.1. Network Structure

Figure 1 presents the estimated psychological networks at each time point for normative (left) and at-risk/problematic social media users (right). Supplementary Material Figure S2A–F show their bootstrapped edge weight means and 95% CIs. Overall, the networks had much of their structure in common. Two clusters of tightly connected personal variables were consistently observed: one cluster representing attention problems, impulsivity, and hyperactivity (i.e., ADHD symptoms), and another consisting of narcissism, physical self-esteem, self-esteem, life satisfaction, depressive symptoms, and FoMO. Additionally, stable intra-domain relations between perceived social competence and intensity of meeting with friends and between reactive and restrictive parental rules were present in all networks. As expected, the strongest edges were observed within the personal clusters: attention problems and impulsivity; impulsivity and hyperactivity; self-esteem and physical self-esteem; self-esteem and life satisfaction; depressive symptoms and FoMO. Bridging these clusters, stable connections were found between attention problems and depressive symptoms and between impulsivity and narcissism. Inter-domain edges appeared less consistent and stable (based on bootstrapped CIs); however, stable links between self-esteem and perceived social competence, and between life satisfaction and quality of communication, were present for both groups at T3.

4.3.2. Centrality

Before identifying the most influential nodes in the networks, the centrality indices’ stability—and, thus, reliability—was evaluated. Strength was highly stable across networks with CS coefficients > 0.67 (in most networks exceeding 0.75), indicating that dropping at least 67% of the samples would retain, with 95% probability, a strength correlation of 0.7 (default) or higher with the full samples. Consistent with previous studies (e.g., [60,61,62], betweenness was unstable under subsetting (CS coefficients between 0.00–0.36). Closeness stability was sufficient to high for the normative SMU networks (CS coefficients between 0.45–0.75) but could not be determined for T1 and T2 at-risk/problematic SMU networks (see Supplementary Material Figure S3A–F for all stability analyses). Considering these results and earlier debates on the interpretation of betweenness and closeness in psychological networks (e.g., [61]), we focus our analysis of central nodes solely on strength.

Figure 2 presents standardized strength values for normative SMU (red) and at-risk/problematic SMU (blue) networks at each time point. Self-esteem, attention problems, impulsivity, and depressive symptoms consistently appeared as the most central nodes in the normative SMU networks, with the highest sum of absolute edge weights connected to these nodes. In the at-risk/problematic SMU networks, self-esteem, impulsivity, life satisfaction (particularly at T1 and T2), and depressive symptoms (particularly at T1 and T3) were the most central factors.

5. Discussion

5.1. General Findings

Problematic social media use has been linked to a range of adverse outcomes. It is, therefore, critical to identify youths at elevated risk for developing PSMU, as well as the factors that can be targeted in prevention and intervention efforts. To our knowledge, this study is the first to combine complementary analytical approaches to examine key factors distinguishing at-risk/problematic social media users from normative users: a theory-driven logistic regression analysis to assess the relative contribution of various personal, peer, and parent factors in predicting risk status, and a data-driven psychological network analysis to explore the structure of relationships among these potential risk and protective factors. Our findings show that the two approaches emphasize different factors involved in the emergence of at-risk/problematic SMU among adolescents, demonstrating how psychological network analysis can yield unique insights into complex social phenomena beyond those offered by traditional regression modeling techniques. In what follows, we discuss the main findings and their implications.

Logistic regression analyses revealed that FoMO, impulsivity, depressive symptoms, intensity of meeting with friends, and reactive parental rules explained unique variance in at-risk/problematic SMU across multiple waves. Our study is among the first to assess a broad set of 14 personal, peer, and parent factors simultaneously within a comprehensive model. Although previous research has linked most of these factors individually to PSMU, our multi-domain approach shows that only five demonstrate unique predictive value when examined together. This finding suggests that many previously identified risk and protective factors may share variance with other predictors and overlap in their explanatory power, underscoring the importance of integrated modeling for identifying the most robust predictors of PSMU.

When controlling for other variables, higher levels of these factors increased the likelihood of adolescents being classified as at-risk/problematic users. Among them, FoMO and reactive parental rules emerged as the most robust predictors. These findings align with prior studies and theoretical perspectives suggesting that adolescents with psychological vulnerabilities are prone to developing PSMU, either as a maladaptive coping strategy for managing negative emotions and/or due to a reduced capacity to resist social media impulses [6,15]. Furthermore, the finding that reactive parental rule-setting was associated with increased odds of at-risk/problematic SMU corresponds with previous research describing differential effects of restrictive versus reactive parental mediation on adolescents’ SMU [30,31]. It has been suggested that ad hoc restrictive intervention reflects more inconsistent parenting, and that the resulting unpredictability of reactive measures may exacerbate symptoms of PSMU (e.g., tolerance, withdrawal, conflict [30,55]). It is also possible that reactive parental rule-setting is a response to at-risk/problematic SMU, rather than a causal factor, with parents imposing restrictions when they observe PSMU tendencies in their child’s behavior (e.g., problems, displacement). As the cross-sectional design of this study prohibits conclusions about the direction of effects, future research should investigate whether reactive parental rule-setting has (short-term effects) on PSMU or vice versa.

Conversely, our findings challenge the social compensation hypothesis, which posits that adolescents engage in PSMU to compensate for limited face-to-face interactions or poor social skills [29,63]. In our sample, perceived social competence did not explain unique variance in at-risk/problematic SMU, and meeting with friends more frequently increased, rather than decreased, the odds of being an at-risk/problematic user. This pattern aligns with a recent study of American college students, which found that greater social activity was positively related to symptoms of Snapchat addiction (but not Facebook or Instagram addiction) and that communicating with friends was the primary motivation for using various social media platforms, suggesting that SMU intensifies rather than replaces offline interactions [64]. Notably, that study assessed PSMU on a continuum, and participants reported low average addiction scores for each social media platform. In our data, most adolescents classified as at-risk/problematic users also reported relatively few PSMU symptoms. It is, therefore, possible that the social compensation hypothesis applies primarily to adolescents with more severe PSMU, but not to those in the at-risk range. Future research should investigate whether the relationship between peer factors and at-risk/problematic social media use varies across severity levels. For instance, within the at-risk/problematic group, adolescents with a high intensity of “meeting with friends” may show a relatively lower symptom level.

Psychological network analyses revealed a relatively stable overall structure of personal, peer, and parent factors over time and across both normative and at-risk/problematic SMU groups. In all networks, two clusters of tightly connected personal variables emerged: one with ADHD symptoms and another involving mental well-being and personality factors. The two clusters were linked via bridge pathways between attention problems and depressive symptoms, and between impulsivity and narcissism. These links are consistent with recognized comorbidities between ADHD and depression [65], as well as between narcissism and impulsivity [66]. Based on strength centrality indices, self-esteem, attention problems, impulsivity, depressive symptoms, and life satisfaction were identified as the most central factors in the networks.

The finding that factors from the two personal clusters were connected and most influential in the networks aligns with the Interaction of Person–Affect–Cognition–Execution (I-PACE) model of specific Internet-use disorders [67]. According to this model, PSMU develops as a result of interactions between predisposing characteristics (such as personality and psychopathology) and response mechanisms (including cognitive and attention biases, coping styles, and reduced inhibitory control), with the latter serving as mediators and moderators between predisposing variables and PSMU. Through conditioning processes, these associations are expected to intensify over time. The I-PACE model may also explain why network structures did not markedly differ between normative and at-risk/problematic users: Rather than forming entirely distinct networks, elevated levels of predisposing and response variables in at-risk/problematic users, along with significant associations between these factors, may trigger network activation and initiate causal risk mechanisms in this group. Future studies with larger sample sizes and longitudinal designs are needed to further explore these potential mechanisms. Our study did not find evidence for conditioning processes (i.e., stronger associations or interconnectedness among factors) in the networks of at-risk/problematic users compared to normative users (although differences in edge weights’ strength were not tested); however, this may be due to the grouping of at-risk and problematic social media users. Given that fewer than 2% of adolescents met criteria for problematic SMU, any differences in interconnectedness were likely obscured.

5.2. Logistic Regression vs. Psychological Network Analysis

This study is the first to combine theory-driven and data-driven approaches using the same comprehensive set of personal, peer, and parent predictors, providing an opportunity to examine how different analytical techniques inform our understanding of PSMU risk processes. Although logistic regression and psychological network analyses serve different purposes and are not directly comparable, they reveal intriguing distinctions that merit further consideration (see

Table 4. Key personal, peer, and parent predictors involved in at-risk/problematic social media use according to two analytical methods. [+/−] symbols represent significant [positive/negative] predictors of at-risk/problematic social media use (based on logistic regression analyses) and/or more influential nodes in the structure of problematic social media use predictors (based on higher [positive/negative] strength centrality).

Variable	Logistic Regression			Psychological Network
	T1	T2	T3	T1	T2	T3
Attention problems				+	+	+
Impulsivity	+	+		+	+	+
Hyperactivity
Depressive symptoms		+	+	+	+	+
Life satisfaction			−	−	−	−
Self-esteem		−		−	−	−
Physical self-esteem
Narcissism	+					+
Fear of missing out (FoMO)	+	+	+
Perceived social competence
Intensity of meeting friends	+	+
Restrictive parental rules	−
Reactive parental rules	+	+	+
Quality of communication		−

). First, FoMO consistently emerged as a risk factor for at-risk/problematic SMU in the logistic regression analyses, yet it appeared more peripheral in the network structures. Indeed, most significant bivariate associations between FoMO and other PSMU correlates (see Supplementary Material Table S1) did not manifest as direct edges in the networks but were connected through depressive symptoms (and subsequently life satisfaction).

This finding underscores a key contribution of psychological network analysis beyond logistic regression: By offering a structured method to evaluate relationships among variables, it facilitates the identification of potential mediating processes and conditional probabilities. Compared to other factors that have been theoretically linked to PSMU due to their relevance in other behavioral addictions (e.g., low self-esteem, depressive symptoms, impulsivity), FoMO has been introduced as an SMU-specific construct that helps explain why certain individuals are particularly drawn to social media [24]. Some studies have suggested that, in addition to acting as an independent predictor, FoMO may mediate relationships between certain personality traits or predisposing psychopathology and PSMU [19,24]. FoMO has also been found to be negatively associated with life satisfaction [23]. Our psychological network analysis findings further support the idea that FoMO’s association with PSMU may operate in a broader context of other psychological mechanisms and predisposing psychopathology, potentially as a response category variable in the I-PACE model.

Second, reactive parental rule-setting emerged as a consistent risk factor for at-risk/problematic SMU in logistic regression analyses and showed positive associations with other personal, peer, and parent factors in bivariate correlations. However, it was not connected to other risk and protective factors in the network models (i.e., no stable edges apart from its link with restrictive parental rules). This may indicate that reactive parental rule-setting does not share PSMU variance with other factors. Indeed, in graphical models, unconnected nodes are interpreted as conditionally independent given all other nodes in the network. As noted earlier, reactive rule-setting may reflect a parental response to their children’s PSMU tendencies, rather than a risk factor for the development of at-risk/problematic SMU. Thus, although reactive parental rule-setting may be part of the broader risk profile, it might not be involved in the causal pathways leading to PSMU. Nevertheless, we cannot rule out the possibility that reactive parental rule-setting is indirectly linked to other parts of the network through unmeasured third variables. For instance, general parenting practices, including warmth, support, supervision, consistent rule-setting, and autonomy granting, are essential for meeting basic psychological needs and fostering an optimal pedagogical environment for healthy development [25]. Parenting styles characterized by low levels on these core dimensions have been associated with a wide range of adverse outcomes, including PSMU (e.g., [30]). Therefore, the broader parenting context may connect SMU-specific practices like reactive rule-setting to PSMU risk and protective factors in other domains.

Third, self-esteem, life satisfaction, and attention problems emerged as some of the most central factors in the network analyses; however, unlike depressive symptoms and impulsivity, they did not significantly predict at-risk/problematic SMU in the logistic regression analyses. This discrepancy highlights a key limitation of traditional regression approaches when examining complex, interconnected phenomena. The absence of significant associations may stem from the substantial variance shared among these factors (depressive symptoms and self-esteem, r = 0.40; depressive symptoms and life satisfaction, r = 0.50; and impulsivity and attention problems, r = 0.60), which limits the ability of regression models to isolate the effect of each variable on at-risk/problematic SMU. This methodological limitation of logistic regression could have important implications for prevention and intervention decisions. Research has pointed out that high centrality nodes are vital in the etiology and maintenance of psychopathology networks, and that prioritizing these nodes in prevention and intervention efforts might be more effective and efficient in reducing overall network activation than targeting peripheral nodes [39,60]. Consequently, our results suggest that interventions targeting factors such as self-esteem, life satisfaction, and attention problems—which would not have been prioritized based on regression results alone—may initiate broader changes across the network of PSMU-related factors. For example, addressing attention problems might reduce (activation of) depressive symptoms, which could in turn reduce (activation of) FoMO and boost life satisfaction and related protective factors (e.g., self-esteem, physical self-esteem), leading to further reductions in depressive symptoms and FoMO. This sequence may ultimately deactivate large parts of the PSMU risk pathways. Although our network analyses are exploratory and our cross-sectional design precludes causal inference, future studies should further investigate and validate these potential causal pathways.

Together, our findings suggest that theory-driven and data-driven approaches to studying PSMU and other social phenomena can offer complementary insights into potential PSMU risk processes unlikely to emerge from either method alone, with important implications for theory and practice. From a scientific perspective, psychological network analysis provides a valuable tool for investigating adolescent PSMU as the emergent phenomenon from a complex system of causally interacting cognitions, traits, emotional states, behaviors, and contextual factors. It allows insight into the structural organization of this system and the relative importance of each variable in it that more conventional statistical approaches within the social sciences cannot provide.

Unlike regression models, psychological network models do not require a priori assumptions about unidirectional causal relations. Additionally, although both approaches in the current study used the same set of personal, peer, and parent variables hypothesized to be involved in the emergence of PSMU, psychological network analysis is not constrained by the number of parameters included in the model. Hence, it is possible to explore interrelations between large sets of potential risk and protective factors in a data-driven manner and identify patterns of risk and resilience that traditional approaches to model building may have missed [35,36]. Considering that the logistic regression models in our study explained only 10–19% of the variance in at-risk/problematic SMU, psychological network analysis holds promise in identifying unexplored predictors and mechanisms involved in PSMU.

Network insights can be used to generate novel hypotheses about causal processes, which can then be tested using confirmatory methods. Consequently, inductive–exploratory and confirmatory–explanatory approaches can complement one another to advance our understanding of when and which youth are at a higher risk of developing PSMU. From a clinical perspective, identifying central factors in networks of risk and protective factors can assist in the early identification of at-risk users and guide targeted interventions that substantially disrupt the system as a whole. As previously noted, targeting central nodes may be more effective and efficient in reducing overall risk than targeting peripheral nodes [60,68].

Our results indicate that factors identified as significant predictors in multiple regression models may not necessarily be central nodes—a finding that could ultimately alter prevention and intervention priorities. For example, whereas FoMO appeared as the most consistent predictor of at-risk/problematic SMU in the logistic regression analysis, targeting depressive symptoms—a central and bridging factor—may affect multiple causal pathways involved in PSMU. This is an important novel insight with implications for prevention and intervention design, suggesting that targeting variables identified through regression models may not optimize systemic change, while comprehensive approaches focusing on network-central factors may yield broader effects across interconnected risk and protective factors. In the network analysis, FoMO seems to be a proximal factor influencing at-risk/problematic use through psychosocial well-being factors (depression and life satisfaction), yet it may still be important that intervention programs aimed at preventing PSMU also focus on reducing feelings of FoMO. This could be achieved by educating young people on how specific social media habits may reinforce feelings of FoMO and by raising their awareness of this phenomenon. Probably, feelings of FoMO are more amenable to intervention than depressive symptoms or overall life satisfaction.

It should be noted that network edges were undirected and, thus, do not indicate whether factors were central because they had a greater influence on other nodes in the network or because they were particularly susceptible to the influence of other nodes. Furthermore, the degree to which risk and protective factors are modifiable by means of intervention may vary. Thus, it is crucial that researchers not only identify central factors but also verify their causal role and assess the effects of behavioral interventions on these factors through longitudinal and experimental studies.

6. Conclusions

Social media platforms offer important developmental opportunities for adolescents, including social connection, identity exploration, and access to information and support networks. However, a subset of youth is particularly vulnerable to developing maladaptive use patterns, marked by compulsive engagement, poor self-regulation, and interference with daily functioning. Considering the growing body of literature documenting adverse outcomes associated with PSMU, it is critical to pinpoint when and which youth are at increased risk and identify key factors to target in prevention and intervention. This study addresses critical gaps in our understanding of how multiple risk and protective factors across different domains interact to shape PSMU. While previous research has typically examined risk factors in isolation or within single domains, the current study provides the first comprehensive examination of personal, peer, and parent factors associated with PSMU through both theory-driven and data-driven analytical approaches. Our findings illustrate how combining these domains and approaches yields novel insights into potential mechanisms involved in PSMU.

Logistic regression analyses on personal, peer, and parent factors revealed that FoMO, impulsivity, depressive symptoms, intensity of meeting with friends, and reactive parental rules explained unique variance in at-risk/problematic SMU, suggesting that many of the previously identified risk and protective factors may overlap in their explanatory power. Psychological network analysis provided additional insights by illuminating structural relationships among risk and protective factors, identifying self-esteem, attention problems, impulsivity, depressive symptoms, and life satisfaction as central factors most strongly connected in the entire network. These findings together point to potential causal pathways involved in the development of PSMU that warrant further investigation. Understanding the dynamic risk and resilience patterns that comprise PSMU can yield a more comprehensive profile of at-risk youth, enhance predictive accuracy, and inform the development of more effective prevention and intervention strategies.

Several limitations of our study should be noted. First, our cross-sectional design does not permit conclusions about the direction of relationships. While we utilized multiple waves of data, our study aimed not to assess longitudinal associations but to identify, employing two different data-analytical approaches, the most relevant factors in distinguishing at-risk/problematic social media users from normative users. Cross-sectional designs were better suited for this purpose. Nevertheless, our design allows the generation of hypotheses regarding causal processes involved in PSMU, which can be evaluated in future longitudinal and experimental studies. Splitting the symptoms we investigated into finer-grained subgroups could be an informative direction for future research, although our current sample size limits such stratification.

Second, although we evaluated a broad range of personal, peer, and parent factors in relation to adolescents’ at-risk/problematic SMU, the predictor set was not exhaustive. For instance, the logistic regression models explained only 10-19% of the variance in at-risk/problematic SMU. Future research should investigate other factors and mechanisms that may be involved in the emergence and maintenance of PSMU. In this regard, data-driven approaches such as psychological network analysis could help identify risk and resilience patterns that have not yet been explored.

Third, the prevalence of problematic SMU in our sample was low (<2%). To retain power, we grouped at-risk and problematic social media users together. Although previous research among a representative Dutch sample found that at-risk users were more likely than normative users to experience adverse consequences due to their SMU [15], the inclusion of less extreme cases may have resulted in weaker associations and/or different network structures for the at-risk/problematic group compared to when problematic users would have been singled out.

Fourth, due to classes and schools entering and exiting the study over time, there was a considerable amount of missing data on the study variables. To enhance statistical power and mitigate potential bias, we imputed missing values using multiple imputation; however, the psychological networks were estimated based on only one of the imputed sets. Different network structures might have emerged if alternative sets had been used. That being said, we evaluated the stability of edge weights and centrality indices using bootstrap techniques and only interpreted results meeting both accuracy and reliability criteria.

Finally, our results are based on a sample of adolescents from the Netherlands. The extent to which our results can be generalized to adolescents in other cultural or regional contexts requires further investigation.