This study adopted a longitudinal design with first-year students from a university in Hunan, selected through convenience sampling. The data collection protocol comprised 3 waves: baseline assessment in October 2019 (T0), first follow-up in December 2020 (T1), and second follow-up in December 2021 (T2). Initially, 3575 freshmen participated at T0 (female: 65.29% [n=2334], male: 34.71% [n=1241]). Subsequent assessments revealed an attrition rate of 17% at T1 (2967 retained) and 12% at T2 (3146 retained). The final analysis included 2193 participants (61.34% of the original cohort) who completed assessments at all 3 waves.

All procedures adhered to the ethical guidelines outlined in the Declaration of Helsinki and were approved by the Ethics Committee of the Second Xiangya Hospital of Central South University (approval number 81671341). Following the acquisition of informed consent, the questionnaires were administered in classrooms under the supervision of trained psychology postgraduate students. Data collection was conducted anonymously to ensure participants’ privacy and confidentiality. Participation was entirely voluntary, and participants were informed that they could withdraw from the study at any time without penalty. Counseling support was made available to participants.

We calculated the means, SDs, and correlations between main variables across 3 time points in SPSS (version 25.0; IBM). P values were adjusted using the false discovery rate correction with a significance threshold of P<.05 to control for multiple comparisons. According to the aims of this study, we conducted the following analyses in Mplus (version 8.3; Muthén and Muthén) [67]. First, IA trajectory classification was analyzed by LGMM. Second, the effects of FE and OBs on the initial level and development rate of IA trajectory were examined by the latent growth curve model (LGCM) [68]. Maximum Likelihood estimation with robust standard errors was employed to handle nonnormally distributed data and estimate model parameters, with missing data addressed via Full Information Maximum Likelihood in the LGMM and LGCM analyses [67]. The model fit indices and analysis process are detailed in the Supplemental Materials.

Furthermore, the Multivariate Logistic Regression Model was applied using the R3-step to examine the predictive effect of FE and OBs on trajectory classification after controlling for gender, residence, and parents’ education level [69]. Finally, structural equation modeling was used to explore the predictive effect of FE and OBs on the IA (T0, T1, and T2, respectively). A multicollinearity test showed that there was no evidence of multicollinearity among the independent variables. Sociodemographic characteristics were incorporated as control variables in the model. The mediation effects were validated with 5000 bias-corrected bootstrap samples (CI excluding 0 denoted significance).

The mean age of all participants at baseline was 18.00 years (SD 0.74), with 71.96% female. Attrition analyses indicated that missing data were random (Little’s MCAR test: P=.16) [70], and no substantial demographic differences were found between retained and dropped-out participants, except for gender and father’s education. Demographic characteristics are detailed in Table S1 in Multimedia Appendix 1. Across all 3 waves, IA was positively correlated with OBs and negatively correlated with FE, and other correlations are presented in Table S2 in Multimedia Appendix 1.

In this study, LGMMs with 1‐6 latent classes were separately estimated, and the summary of fit indices is presented in Table S3 in Multimedia Appendix 1. From the 1-class to the 4-class model, the information indices Akaike information criterion, Bayesian information criterion, and sample-size adjusted Bayesian information criterion gradually decreased, while the Bayesian information criterion value started to increase in the 5-class and 6-class models. Additionally, we found a significant bootstrap-based likelihood ratio test in 2-class to 4-class, while the Lo-Mendel-Rubin likelihood ratio test results suggested that the 2-class and 4-class solutions were suitable. Further observation confirmed that the entropy value (almost 0.80) was the highest when retaining the 4-class model. By comparing multiple indices as described above and comprehensively considering the parsimony, accuracy, and practical significance of the models, this study ultimately selected the 4-class classification model as the optimal one.

Figure 1 shows the estimated growth parameters and trajectories for each of the 4 classes, and Table S4 in Multimedia Appendix 1 presents the intercept and slope for each latent trajectory category, respectively. The first latent category, termed high-risk group (5.09% of the sample, n=112), maintained a high level of IA and exhibited an upward trend over time (M_intercept = 56.09, P<.001; M_slope = 7.97, P<.001); The second, labeled medium to high-risk group (29.85% of the sample, n=655), sustained a medium to high level of IA and showed a slow upward trend (M_intercept = 50.08, P<.001; M_slope =3.62, P<.001); The third, termed medium to low-risk group (35.95% of the sample, n=788), maintained a moderate to low level of IA and exhibited a slow downward trend (M_intercept =43.49, P<.001; M_slope =−0.68, P=.01); The fourth, termed low-risk group (29.11% of the sample, n=638), sustained a relatively low level of IA and displayed a downward trend (M_intercept =37.99, P<.001; M_slope =−6.09, P<.001).

Based on the correlation analysis results, this study integrated sociodemographic characteristics (such as gender, residence, and parents’ education level), FE factors, and OBs as covariates into the LGCM. A conditional model, illustrated in Figure 2, was developed to assess how these covariates predict both the initial levels and development rates of IA among college students. The conditional model exhibited a good fit, χ²₁₁=60.087, comparative fit index=0.976, Tucker-Lewis index=0.928, root mean square error of approximation=0.045, standardized root mean square residual=0.021.

OBs, along with FE characteristics and the mother’s education level, were significant predictors of the initial IA score. Specifically, higher levels of mother’s education (β_intercept=0.121, P<.05), OBs (RT and ICT) (β_intercept=0.129, P<.01; β_intercept=0.279, P<.001) and poor FE (cohesion, conflict) (β_intercept=−0.169 P<.001; β_intercept = −0.191, P<.001) were associated with higher initial IA score.

Furthermore, gender (β_slope=.123, P<.001), long-term residence (rural vs urban) (β_slope=−0.082, P<.05), mother’s education level (β_slope=−0.230, P<.001), ICT (β_slope=−0.165, P<.001) and family conflict (β_slope=0.073, P<.05) significantly predicted the growth rate of IA. This indicates that IA development changes more rapidly for females than males, and for individuals from rural rather than urban areas. Additionally, lower mother’s education level, ICT, and family conflict were associated with a more rapid change in the IA trajectory.

Tables 1-3 present the results of multivariate logistic regression analysis. Compared with the low-risk group, college students in the high-risk group (regression coefficient [B]=0.885, P<.01), medium to high-risk group (B=0.550, P<.001; odds ratio [OR] 1.734, P<.01) and medium to low-risk group (B=0.687, P<.001; OR 1.988, 95% CI 1.486-2.659 ; P<.001) were more likely to be female. College students in the medium to high-risk group were more likely to live in the rural areas than in the urban areas (B=–0.269, P<.05; OR 0.764, 95% CI 0.597-0.977; P<.05). The high-risk group reported poorer FE, characterized by lower cohesion (B=–0.186, P=.01; OR 0.831, 95% CI 0.721-0.957; P<.01), higher conflict (OR 0.866, 95% CI 0.745-1.006, P<.05), and reduced independence (B=−0.173, P<.05; OR 0.841, 95% CI 0.710-0.996; P<.05). The medium to high-risk group had higher family conflict (B=−0.117, P<.01; OR 0.890, 95% CI 0.826-0.959; P=.001), as well as elevated RT (B=0.020, P<.05; OR 1.020, 95% CI 1.003-1.037; P<.05) and ICT (B=0.027, P<.01; OR 1.028, 95% CI 1.010-1.045; P<.01). Students in the medium to low-risk group experienced higher family conflict (B=−0.094, P<.05; OR 0.911, 95% CI 0.841-0.986; P<.05) compared with the low-risk group.

As shown in Figure 3, the FE had a significant negative effect on both OBs (β=−0.113, P<.001) and IA at 3 times of survey (T0, T1, and T2) (β=−0.314; −.204; −.183, P<.001). While OBs positively predicted IA at 3 times of survey (β=0.251; .156; .126, P<.001). The bootstrap results with 5000 resamples revealed that FE exerted a significant indirect effect on IA in 3 waves (T0, T1, and T2), due to the mediating effect of OBs (effect_T0=−0.03, 95% CI −0.46 to −0.16; effect _T1 = −0.02, 95% CI −0.35 to −0.12; effect _T2 =−0.02, 95% CI −0.33 to −0.10). Thus, OBs partially mediated the relationship between FE and IA. The model fitting indexes and path coefficients are shown in Table S5 and Table S6 in Multimedia Appendix 1, respectively.

This study reveals a predominant declining trend of IA growth trajectory among college students, with 4 distinct categories emerging. Importantly, our study demonstrates that both FE and OBs significantly affect the initial level and development rates of IA developmental trajectory, further delineating the process of how environment and cognition influence IA development. Finally, consistent with our hypothesis, we show that OBs exhibit a significant mediating effect on the influence of FE on IA. The partial mediating model indicates that the FE not only influences the IA development directly but also modifies the individual cognitive style (ie, OB), which finally facilitates the IA progression indirectly.

Consistent with previous findings, first-year college students demonstrated relatively high baseline levels of IA [71], likely due to increased autonomy and greater access to the internet after graduating from high school. However, a gradual decline in IA severity was observed over time, which may reflect developmental transitions such as increased maturity, enhanced self-regulation, and adaptation to university life. This pattern aligns with prior research suggesting that the risk of IA typically rises during early adolescence but tends to decline in late adolescence or early adulthood as individuals gain better emotional and behavioral control. Additionally, the significant negative correlation between IA’s initial level and development rate suggests that individuals with higher baseline IA tend to show slower increases or decreases over time, while those with lower initial IA are more likely to escalate. This may reflect a ceiling effect or spontaneous remission in high-risk individuals and highlights the importance of monitoring low-level users who may develop IA later.

This study identified 4 distinct developmental trajectories of IA, which align with findings from Hong et al [72] and Zhou et al [73]. In contrast, some researchers have identified 3 subgroups among adolescents—low-decreasing, moderate-increasing, and high-increasing trajectories [12]—while others have reported only 2 IA trajectories (ie, low/decreasing and high/decreasing) [19]. Variations in the number and nature of identified trajectories across studies may be attributed to several factors, including differences in sample sociodemographic characteristics (eg, age, cultural contexts), personality and cognitive styles, as well as measurement instruments, assessment intervals, and the number of time points. These methodological and contextual differences may lead to the identification of either more nuanced or more general trajectory classifications. Understanding these developmental patterns is essential for designing targeted interventions and support systems that address the diverse needs of individuals at varying levels of IA risk.

In our sample, approximately 5.09% of students followed a high-risk trajectory of IA, characterized by an initially severe level of symptoms followed by a rapid escalation over time. This proportion closely aligns with the national prevalence of severe IA (approximately 6.3%) [74]. A similar high-risk trajectory has been observed in previous studies [12,75], lending further support to the existence of a small yet highly vulnerable subgroup. These individuals are likely to experience significant impairments in psychological well-being, academic performance, and social functioning. The steep and worsening trajectory underscores the urgent need for early identification and targeted interventions to prevent further deterioration and chronicity of IA.

In contrast, the medium to high-risk group (29.85%) exhibited a moderate level of IA symptoms at baseline that increased slightly over time, maintaining a state of “mild to moderate” addiction. Similar developmental patterns have been reported in prior studies [12,17,22,72,73]. This trajectory may reflect students who, under the influence of academic pressure, interpersonal difficulties, or emotional distress, gradually adopt more entrenched patterns of problematic internet use. Although not meeting clinical thresholds for IA, the upward trend indicates potential risk for future deterioration. Regular screening and early-stage psychosocial interventions—such as those embedded within school mental health services—are recommended to prevent escalation.

The medium to low-risk group (35.95%) demonstrated a moderate initial severity of IA that declined slightly over time, maintaining an overall “mild addiction” status. This trajectory suggests a degree of self-regulation and recovery. Comparable patterns have been reported in earlier work [22,72,76], suggesting that some students may possess strong internal coping capacities or benefit from external protective factors—such as peer support or successful academic adaptation—that contribute to behavioral improvement. While these students are not at immediate high risk, they may reside at the “risk margin,” warranting continued attention through psychoeducation, digital literacy programs, and developmental support to consolidate positive behavioral changes.

Finally, the low-risk group (29.11%) followed a decreasing trajectory, beginning with a low level of IA and trending toward remission. This pattern reflects a trajectory of psychological resilience and healthy development. Consistent with earlier findings [12,19,20,72], such individuals are typically classified as “stable-healthy.” The resilience demonstrated by this group may be attributed to various protective factors—such as effective emotion regulation, supportive family dynamics, or limited exposure to web-based risk environments—and warrants further exploration to inform universal prevention strategies.

We explored the influence of several sociodemographic characteristics, including gender, residence, and parents’ education level, on IA trajectory, although this is not the focus of this study. For gender, no significant differences in baseline IA levels were found between males and females. However, over the 2-year follow-up period, IA symptoms developed more rapidly among females compared with males, and females were more likely to fall into the high-risk IA trajectory group. This finding is consistent with previous studies [77-79], suggesting that females may be at greater risk for developing IA than males. Such differences may stem from gender-specific patterns of web-based activity and coping strategies. For instance, females tend to use the internet more frequently for maintaining social connections, making them more susceptible to emotionally driven triggers, whereas males are more likely to engage in gaming and competitive web-based activities [80]. Further, youths from rural areas exhibited a faster development of IA compared with those who grew up in urban areas and were more likely to fall into the medium-to-high-risk categories. The growing penetration rate of the internet, along with a large population of left-behind children in rural areas in China [81], may make the IA increasingly prominent in children and adolescents [82,83]. Moreover, we also showed that higher maternal education was associated with higher initial IA levels and slower progression of IA, suggesting that mothers with high education levels may provide earlier chances to bring their children into the internet world, but play a protective role against IA in the development of youths ultimately.

Consistent with the family system theory [84], our findings highlight FE as a critical determinant in the development of IA. Most notably, the observation of predictive effects of FE on IA initial level and development rate in this study highlights a potential mechanism in which FE may exert a persistent influence throughout the entire IA developmental process in youth. Thus, this study underscores the necessity of early and sustained attention to the FE as a cornerstone of intervention and prevention strategies. Subgroup analysis identifies protective factors such as cohesion and independence within the family, aligning with previous research [85] and underscoring the adaptive role of a positive FE in IA development [86]. Conversely, family conflict significantly elevates the risk of IA. Longitudinal studies further emphasize that individuals experiencing pronounced conflicts in parent-child relationships during early adolescence are more prone to developing IA symptoms in mid-adolescence [87]. A possible explanation is that adolescents may use excessive internet use as a coping mechanism for family strife, transforming into a maladaptive strategy over time [88].

Based on the relationship between obsessive-compulsive symptoms and IA [50], this study pioneers an exploration into the direct impact of OBs on the development of IA. We found that OBs significantly shape both the initial levels and development rates in IA, highlighting the critical need for early and ongoing attention to these cognitive patterns. Upon subgroup analysis, students with elevated control thinking were more likely to fall into “high-risk” or “medium-high risk” categories compared with their low-risk counterparts. The significance of these cognitive patterns underscores the notion that individuals bear responsibility for their ideas and their ensuing consequences. Excessive control thinking during web-based activities may impede individuals from discontinuing their behavior on the web, leading to addiction. Furthermore, college students expressing heightened levels of responsibility and threat assessment are prone to the “medium-to-high risk” group. This inflationary responsibility aligns with Salkovskis’ cognitive theory, positing that the evaluation of intrusive thoughts leads individuals to assume responsibility for the outcomes of such thoughts and perceived danger. Inflated responsibility, in turn, triggers automatic negative thoughts and discomfort [89,90]. The burden of excessive responsibility in the real world, coupled with the allure of the perceived simplicity and safety of the web-based realm, may propel individuals toward increased participation in the internet, thereby facilitating the emergence and progression of IA.

However, no significant relationship of IA development was observed with perfectionism in our data. Yang et al [91] identified a direct relationship between maladaptive perfectionism and IA among Chinese college students. A plausible explanation for this inconsistency could be the nuanced nature of perfectionism. Perfectionism encompasses various dimensions, such as self-oriented perfectionism, other-oriented perfectionism, and socially prescribed perfectionism, each exerting different influences on behaviors and attitudes [92]. In the specific context of IA development, it is conceivable that particular dimensions of perfectionism might have a more pronounced impact than others. For example, self-oriented perfectionism, characterized by setting high personal standards, may foster a conscientious approach to web-based activities without necessarily leading to addictive behaviors. Conversely, socially prescribed perfectionism, driven by perceived external expectations, might create pressure and stress, potentially fueling the development of IA. Further research could dissect specific aspects of perfectionism within the context of IA, examining their interplay with other cognitive elements for a nuanced comprehension of their contributions to the trajectories of IA.

In accordance with the social-cognitive model [93], the mediation analysis in this study confirms that FE, as a social background, can indirectly influence IA through the cognitive pathway of OBs. The findings reveal that OBs mediate the relationship between FE and IA at all 3 time points, indicating that FE persistently exerts a direct impact on college students’ IA, as well as an indirect effect by modifying cognitive patterns such as OBs.

This evidence suggests the importance of multisystemic efforts (ie, addressing social context and intrapersonal cognitive factors simultaneously) to prevent or reduce adolescents’ IA. Intervention strategies targeting IA among college students could include cognitive therapy to reduce OBs or family therapy to mitigate familial conflicts, thereby interrupting the developmental trajectory of IA.

This study offers 3 key contributions to the understanding of IA development. First, by employing LGMM, we identified 4 distinct IA trajectories (high-risk, medium to high-risk, medium to low-risk, and low-risk), revealing the heterogeneity in how IA unfolds among college students. Second, building on prior research linking OCD symptoms and IA, this study is among the first to systematically demonstrate the independent predictive role and mediating function of OBs in IA progression. Our findings highlight how cognitive vulnerabilities—such as inflated responsibility and excessive need for control—can exacerbate compulsive internet use and drive the longitudinal escalation of IA. Third, by integrating ecological systems theory with the social-cognitive model, we propose a dual-pathway mechanism: FE influences IA both directly—affecting the initial level and development rate—and indirectly, through its impact on OBs. This framework bridges the gap between macro-level contextual influences and micro-level cognitive processes, offering a more integrative etiological account of IA. Finally, the findings provide actionable insights for prevention and intervention. They underscore the necessity of tailoring strategies to specific risk profiles. For instance, individuals with highly OBs or those exposed to family conflict may benefit from cognitive-behavioral interventions targeting maladaptive cognitions, combined with family-based approaches that enhance cohesion and reduce conflict. Furthermore, the protective effect of maternal education suggests that empowering parents—especially mothers—through digital literacy and parenting programs could serve as a valuable preventive resource.

This study has several limitations. First, it was conducted in a single university in Hunan Province, which may limit the generalizability of the findings. Regional and cultural factors in China (such as rural-urban disparities in digital access and educational pressure) may influence students’ internet use patterns and IA risk [94,95]. For example, students from rural areas may have had less prior exposure to technology, while academic stress in Chinese universities may promote excessive internet use as a coping strategy. Future studies should include more diverse samples across different regions and cultural contexts. Second, although gender has been included as a covariate in the analysis, our findings showed a significant predictive effect of gender on the IA trajectory, suggesting the potential influence of gender imbalance (approximately 72% female) on the IA trajectory cannot be fully ruled out in this study. It calls for future studies with gender-balanced samples to clarify this issue more elegantly. Third, this study did not assess comorbid mental health issues such as depression and anxiety [17,96]. Future research could investigate how IA codevelops with other internalizing (eg, depressive symptoms) and externalizing (eg, substance use) disorders using multi-trajectory frameworks and explore bidirectional relationships between IA and comorbid symptoms to identify central nodes (ie, cognition distortions, sleep disturbances) as intervention targets. Fourth, due to the limited number of time points (3 waves), we were unable to estimate more complex developmental trajectories such as quadratic or nonlinear models. Future longitudinal studies with more frequent assessments are encouraged to address this limitation.

This study offers crucial insights into the development of adolescent IA, highlighting that IA is not a “one-size-fits-all” issue. It underscores the vital function of a supportive FE in mitigating IA risks among youth, while also revealing the detrimental impact of maladaptive cognitive patterns, like OBs, on the onset and development of IA. Our findings provide empirical evidence to help clinicians identify high-risk individuals by monitoring vulnerability factors, including elevated OBs, a lack of emotional warmth in the family, exposure to family conflict, and persistently high IA symptoms across follow-up assessments. Tailored interventions such as cognitive-behavioral therapy targeting maladaptive beliefs and family-based programs aimed at enhancing family cohesion may be effective for at-risk groups. Taken together, this research advances theoretical models of IA by integrating ecological and cognitive perspectives and offers practical implications for developing more individualized and effective prevention and intervention strategies for youth.