Temporal relationship between physical activity, social anxiety, and internet addiction among adolescents: a cross-lagged analysis Authors: Xianbang Huang, Ling Li, Song Wang Published: 2026 ABSTRACT Physical activity (PA), social anxiety (SA), and internet addiction (IA) are closely related to adolescents’ psychological and behavioral development, yet their longitudinal directional association Responses to Editor and Reviewers. This study examined the temporal predictive relationships among PA, SA, and IA in adolescents using a cross-lagged panel model. A two-wave longitudinal study with a 16-week interval was conducted among 877 adolescents from three middle schools in Jiangxi Province. PA, SA, and IA were assessed using the Physical Activity Rating Scale (PARS-3), Social Anxiety Scale for Adolescents (SAS-A), and Internet Addiction Test (IAT). Cross-lagged panel analyses were applied to test temporal predictive associations and the role of SA. PA was negatively associated with IA at both time points. PA at T1 positively predicted PA at T2 (β = 0.78, p < 0.001) and negatively predicted IA at T2 (β = − 0.12, p < 0.001). SA at T1 negatively predicted PA (β = − 0.03, p < 0.001) and positively predicted both SA (β = 0.50, p < 0.001) and IA at T2 (β = 0.10, p < 0.01). IA at T1 negatively predicted PA (β = − 0.05, p < 0.001) and positively predicted IA at T2 (β = 0.13, p < 0.001), but not SA. PA, SA, and IA form a longitudinal developmental system in adolescence. PA functions as a protective factor, whereas elevated SA and IA may undermine subsequent PA engagement. Promoting sustained PA may support adolescents’ mental health and reduce problematic internet use. Introduction In the context of the global digital expansion, the internet has become deeply integrated into adolescents’ learning, social interaction, and leisure activities. Despite its benefits for information access and communication, internet addiction (IA) has emerged as a significant psychological and behavioral health concern. Cross-national evidence indicates substantial levels of IA across countries and regions, with particularly high prevalence among adolescents and student populations. Notably, a recent meta-analysis covering 38 countries reported a pooled prevalence of problematic internet use of 41.84% among students, highlighting the widespread nature of IA-related problems1. Similarly, a meta-analysis targeting middle school and university students estimated a pooled prevalence of IA of 15.2%. Although variations in regions, measurement instruments, and sample characteristics contributed to heterogeneity in the findings, the results consistently indicate that IA is gradually becoming a normative phenomenon within adolescent and student populations2. More alarmingly, a growing body of empirical and review studies has consistently demonstrated that IA is closely associated with a range of adverse psychological and functional outcomes. These consequences include a significantly elevated risk of anxiety and depression3–7, poorer sleep quality8, increased aggression and impulsivity9, diminished creativity and constrained optimal motor and sensory development10, and even heightened suicide risk11,12. Neuroimaging research further suggests that individuals with IA exhibit abnormalities in brain function and functional connectivity, which may exert profound effects on executive control and emotion regulation13. Within this context, if IA persists during adolescence, it may gradually consolidate into a maladaptive behavioral trajectory by undermining real-world social engagement and reinforcing avoidant coping strategies, with potential harms that are neither short-term nor transient. The World Health Organization’s Mental Health Gap Action Programme (mhGAP), released in 2023, identifies physical activity(PA) as a non-pharmacological intervention that can effectively promote mental health14. As a form of positive health behavior, PA has been shown to effectively alleviate internet addiction. Previous research examining the relationship between PA and IA has predominantly relied on randomized controlled intervention trials and cross-sectional analyses. For instance, intervention studies conducted by Chinese scholars have demonstrated the effectiveness of Tai Chi exercise programs15,16, while researchers in Europe have reported beneficial effects of combined virtual reality and aerobic exercise interventions17. In addition, cross-sectional studies conducted globally across diverse populations—including adolescents and university students—have consistently shown that PA significantly and negatively predicts IA18–22. More recently, a meta-analysis focusing on PA interventions among adolescents revealed that PA interventions significantly reduce symptoms of IA, providing robust evidence for the therapeutic effects of PA on IA in this population23 However, the long-term mechanisms through which physical activity exerts its effects within natural developmental contexts remain insufficiently elucidated. Based on the cognitive–behavioral model, IA is not directly caused by the internet medium itself but rather emerges from the interaction among individuals’ pre-existing psychological problems, maladaptive cognitive processing styles, and reinforced patterns of internet use, through which a self-maintaining negative cycle is formed24. Previous studies have indicated that IA tends to occur among individuals with existing psychological problems, such as depression, anxiety, social functioning impairments, and deficits in impulse control. These pre-existing psychological problems have been collectively conceptualized in subsequent research as individuals’ antecedent psychological vulnerabilities25. Social anxiety(SA), as a form of “antecedent psychological vulnerability,” may lead individuals to use the internet as an avoidant strategy for coping with social stress, thereby increasing the risk of IA26,27. Within traditional causal frameworks, IA has primarily been conceptualized as an outcome of SA and deficits in health-related behaviors, whereas its feedback effects on psychological and behavioral functioning have received limited attention. From a stress–coping perspective, exposure to social stressors may prompt avoidant coping strategies that temporarily alleviate distress but ultimately maintain or exacerbate stress by failing to address underlying problems28. Empirical evidence indicates that individuals with IA are more likely to adopt emotion-focused and avoidant coping strategies29, which are positively associated with IA30 and may impair real-world social functioning, thereby intensifying SA. Moreover, prolonged immersion in online activities may displace time and psychological resources for health-promoting behaviors such as PA, further undermining social competence and psychological adaptability and contributing to a self-reinforcing cycle. Accordingly, interventions targeting IA should emphasize addressing core issues related to SA31. Concurrently, participation in PA can reduce suppressed emotional expression and systematically alleviate SA through multiple pathways and mechanisms32,33 However, existing studies have largely relied on cross-sectional designs, which limits understanding of the temporal ordering and directional relationships among PA, SA, and IA. Such approaches inevitably entail substantial limitations, given that adolescence represents a critical developmental period characterized by rapid changes in social functioning, emotion regulation capacities, and health behavior patterns. Behavioral and emotional patterns formed during adolescence tend to exhibit strong continuity and may exert long-term influences on mental health and lifestyle in adulthood34. Therefore, clarifying the longitudinal directional relationships among PA, SA, and IA during adolescence is of significant practical importance for identifying key risk pathways and optimal windows for intervention. PA may be closely associated with SA and IA. However, most existing intervention studies have relied on short-term experimental designs and have primarily focused on immediate pre–post intervention effects, making it difficult to elucidate the dynamic mechanisms through which PA, emotional variables, and IA interact over time within naturalistic contexts. Moreover, prior research has largely adopted a unidirectional perspective, emphasizing the effects of PA on IA, while paying limited attention to potential reverse pathways—specifically, whether IA may predict subsequent levels of PA. Empirical evidence examining cross-temporal predictive relationships among these therefore remains notably scarce. To address these gaps, this study employed a two-wave cross-lagged panel model to examine the bidirectional longitudinal relationships among PA, SA, and IA. This analytical approach clarifies their temporal ordering and potential reciprocal effects. Moreover, if SA or IA were found to negatively predict subsequent PA, such findings would suggest that PA-focused interventions alone may be insufficient for at-risk adolescents. Thus, the present study provides empirical evidence to inform the development of more targeted and comprehensive intervention strategies. Materials and methods Participants A cluster random sampling method was used to recruit students from three middle schools in Jiangxi Province. Intact classes within each school were regarded as clusters and randomly selected to ensure the sample was representative of the target population. Students with physical motor impairments, emotional regulation disorders, sleep disorders, or clinically diagnosed psychological disorders were excluded prior to data collection. A total of 1,130 students initially completed questionnaires at T1. Before formal analysis, 55 participants were excluded due to invalid responding: 21 had missing key data, 18 showed invariant or careless response patterns, and 16 failed to submit complete questionnaires. This resulted in 1075 valid responses at T1. The 877 participants at T2 were the same individuals who provided valid data at T1. Of the 1075 valid T1 participants, 198 were lost during follow-up, yielding a sample attrition rate of 18.42%. Attrition was mainly attributable to: 71 cases with mismatched identification information that could not be linked to T1 data, 35 withdrawals from the study, 22 cases of careless or invariant responding, 25 cases of absenteeism at T2 assessment, 23 incomplete questionnaire submissions, and 22 cases due to temporary leave or transfer to another school. Participant demographic characteristics are reported in Table 1.Table 1Demographic characteristics of the sample.VariablesCategoriesParticipants (%)GenderBoy395 (45.03)Girl482 (54.96)One-childYes688 (78.44)No189 (21.55)GradeGrade 1397 (45.26)Grade 2306 (34.89)Grade 3174 (19.84)AgeTwelve114 (12.99)Thirteen258 (29.42)Fourteen259 (29.53)Fifteen188 (21.44)Sixteen58 (6.61) Table 1Demographic characteristics of the sample.VariablesCategoriesParticipants (%)GenderBoy395 (45.03)Girl482 (54.96)One-childYes688 (78.44)No189 (21.55)GradeGrade 1397 (45.26)Grade 2306 (34.89)Grade 3174 (19.84)AgeTwelve114 (12.99)Thirteen258 (29.42)Fourteen259 (29.53)Fifteen188 (21.44)Sixteen58 (6.61) Demographic characteristics of the sample. Procedures From September 2025 to January 2026, participants were assessed across two waves with a 16-week interval using the Physical Activity Rating Scale (PARS-3), Social Anxiety Scale for Adolescents (SAS-A), and Internet Addiction Test (IAT). Trained researchers explained the study objectives and procedures prior to data collection. A 16-week interval (approximately one full academic semester) was adopted in this study. This time frame is sufficient to capture stable longitudinal changes in adolescents’ PA, SA, and IA. Furthermore, using a full academic semester avoids the dramatic shifts in daily routines and lifestyles caused by summer or winter vacations, as well as the potential disturbances from intensive examinations, thereby ensuring the reliability and stability of the assessment results. Participation was approved by class teachers, and informed consent was obtained from students and their parents. Surveys were completed anonymously, and questionnaires with excessively short completion times (< 200 s), uniform response patterns, or high response homogeneity were excluded to ensure data quality. The study protocol was approved by the Ethics Committee of Xi’an Physical Education University (XAIPE202584) and conducted in accordance with the Declaration of Helsinki. Measurements Physical activity Adolescents’ levels of physical activity were assessed using the Physical Activity Rating Scale–3 (PARS-3)35. This scale was revised by Liang Deqing and colleagues based on the original Physical Activity Rating Scale and is designed to evaluate individuals’ levels of physical exercise in daily life. The PARS-3 is concise and easy to administer, is applicable to both adolescent and adult populations, and has been widely used in research in the fields of education, psychology, and public health to assess the intensity and frequency of physical activity and thereby determine overall levels of participation. In the present study, the Cronbach’s α coefficients of the PARS-3 were 0.811 at T1 and 0.828 at T2. Social anxiety Social anxiety was measured using the Social Anxiety Scale for Adolescents (SAS-A), originally developed by La Greca and Lopez36 to assess adolescents’ experiences of anxiety, tension, self-focused attention, and avoidance in social situations. The SAS-A is one of the most widely used instruments internationally for the assessment of adolescent social anxiety. The scale consists of 18 items rated on a five-point Likert scale, ranging from 1 (“not at all true”) to 5 (“completely true”), with higher scores indicating higher levels of social anxiety. In the present study, the Cronbach’s α coefficients were 0.910 at T1 and 0.968 at T2. Internet addiction Internet addiction was assessed using the Internet Addiction Test (IAT), developed by Young37. The IAT is one of the most widely used instruments internationally for evaluating whether individuals’ internet use has reached an addictive level and has been extensively applied in studies of internet-related behaviors among adolescents and adults. The original IAT consists of 20 items rated on a five-point Likert scale. Total scores are obtained by summing all item scores, yielding a range from 20 to 100, with higher scores indicating more severe tendencies toward internet addiction. The response options for each item range from 1 (“rarely”) to 5 (“very frequently”). In the present study, the Cronbach’s α coefficients of the IAT were 0.965 at T1 and 0.969 at T2. Data analysis IBM SPSS Statistics 29.0 was used to conduct descriptive statistics and correlation analyses for adolescents’ physical activity, social anxiety, and internet addiction. Harman’s single-factor test was performed to examine potential common method bias. Amos 26.0 was employed to conduct a cross-lagged panel model to analyze the longitudinal temporal relationships among PA, SA, and IA. In the model, autoregressive paths from T1 to T2 were estimated for each variable to control for stability effects, and concurrent correlations among the three variables at T1 and at T2 were freely estimated. Maximum likelihood estimation was used for model estimation. Four commonly used fit indices were adopted to evaluate model fit: the chi-square ratio (χ2/df), goodness-of-fit index (GFI), comparative fit index (CFI), and root mean square error of approximation (RMSEA). Result Common method bias test Harman’s single-factor test was conducted to examine the presence of common method bias in the data38. To assess whether systematic bias attributable to common method variance existed, two separate tests were performed, corresponding to the T1 and T2 measurement waves. Specifically, exploratory factor analyses with a single-factor solution were conducted for the T1 and T2 datasets, respectively. The results indicated that, at both time points, the variance explained by the first factor was below 40%, which falls within the acceptable threshold suggested in the literature. These findings suggest that common method bias was not a serious concern in the present study. Independent-samples t tests and ANOVA Independent-samples t tests were conducted at both T1 and T2 using gender as the grouping variable (see Table 2). At both time points, boys exhibited significantly higher levels of PA than girls (both p < 0.001), significantly lower levels of SA (both p < 0.001), and significantly lower levels of IA (T1: p < 0.05; T2: p < 0.001).Group comparisons based on only-child status were also performed (see Table 3). The results indicated that, at both T1 and T2, there were no significant between-group differences in PA, SA, or IA (all p > 0.05).Furthermore, analyses of grade-level differences were conducted using one-way analysis of variance (ANOVA) (see Table 4). At T1, significant grade-level differences were observed in PA (p < 0.05), with levels increasing as grade level advanced. SA also differed significantly across grades (p < 0.05), showing an upward trend with increasing grade level. No significant grade-level differences were found for the remaining variables.Table 2Gender differences in the study variables.M ± SDTPBoyGirlT1 PARS-327.82 ± 22.6020.68 ± 12.354.79 < 0.001T1 SAS34.44 ± 11.1840.11 ± 12.55− 7 < 0.001T1 IAT35.8 ± 1639.60 ± 19.49− 4.30.03T2 PARS-323.59 ± 22.3616.48 ± 19.924.98 < 0.001T2 SAS33.35 ± 14.4539.33 ± 16.40− 5.66 < 0.001T2 IAT33.75 ± 16.3939.10 ± 18.62− 4.46 < 0.001PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction TestTable 3Only-child status differences in the study variables.M ± SDTPYesNoT1 PARS-322.94 ± 21.8724.16 ± 22.3− 0.670.5T1 SAS38.39 ± 13.5037.32 ± 11.921.060.29T1 IAT36.38 ± 17.1537.60 ± 18.29− 0.820.41T2 PARS-318.57 ± 20.1019.99 ± 21.67− 0.810.42T2 SAS37.10 ± 16.3736.52 ± 15.680.420.67T2 IAT35.84 ± 16.9236.92 ± 18.10− 0.740.46PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test.Table 4Grade-level differences in the study variables.(M ± SD)FPGrade 1Grade 2Grade 3T1 PARS-321.87 ± 21.0825.32 ± 23.5426.12 ± 22.013.170.04T1 SAS36.37 ± 11.4838.10 ± 12.8639.36 ± 12.794.020.01T1 IAT36.75 ± 18.0137.14 ± 17.9138.98 ± 18.430.940.37T2 PARS-317.81 ± 20.3121.33 ± 23.0621.15 ± 20.252.870.06T2 SAS36.02 ± 15.1237.17 ± 16.6837.09 ± 15.950.550.57T2 IAT36.44 ± 17.2036.91 ± 18.6936.89 ± 17.900.070.93PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test. Table 2Gender differences in the study variables.M ± SDTPBoyGirlT1 PARS-327.82 ± 22.6020.68 ± 12.354.79 < 0.001T1 SAS34.44 ± 11.1840.11 ± 12.55− 7 < 0.001T1 IAT35.8 ± 1639.60 ± 19.49− 4.30.03T2 PARS-323.59 ± 22.3616.48 ± 19.924.98 < 0.001T2 SAS33.35 ± 14.4539.33 ± 16.40− 5.66 < 0.001T2 IAT33.75 ± 16.3939.10 ± 18.62− 4.46 < 0.001PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test Gender differences in the study variables. PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test Table 3Only-child status differences in the study variables.M ± SDTPYesNoT1 PARS-322.94 ± 21.8724.16 ± 22.3− 0.670.5T1 SAS38.39 ± 13.5037.32 ± 11.921.060.29T1 IAT36.38 ± 17.1537.60 ± 18.29− 0.820.41T2 PARS-318.57 ± 20.1019.99 ± 21.67− 0.810.42T2 SAS37.10 ± 16.3736.52 ± 15.680.420.67T2 IAT35.84 ± 16.9236.92 ± 18.10− 0.740.46PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test. Only-child status differences in the study variables. PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test. Table 4Grade-level differences in the study variables.(M ± SD)FPGrade 1Grade 2Grade 3T1 PARS-321.87 ± 21.0825.32 ± 23.5426.12 ± 22.013.170.04T1 SAS36.37 ± 11.4838.10 ± 12.8639.36 ± 12.794.020.01T1 IAT36.75 ± 18.0137.14 ± 17.9138.98 ± 18.430.940.37T2 PARS-317.81 ± 20.3121.33 ± 23.0621.15 ± 20.252.870.06T2 SAS36.02 ± 15.1237.17 ± 16.6837.09 ± 15.950.550.57T2 IAT36.44 ± 17.2036.91 ± 18.6936.89 ± 17.900.070.93PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test. Grade-level differences in the study variables. PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test. Descriptive statistics and correlation analysis Descriptive statistics and correlation analyses were conducted for adolescents’ PA, SA, and IA, and the detailed results are presented in Table 5. The correlation analysis indicated that PA at T1 was significantly negatively correlated with SA and IA at T1 (p < 0.001). SA at T1 was significantly positively correlated with IA at T1 (p < 0.001). PA at T1 was significantly positively correlated with PA at T2 (p < 0.001), and was also significantly negatively correlated with SA and IA at T2 (p < 0.001). In addition, SA at T2 was significantly positively correlated with IA at T2 (p < 0.001).Table 5Correlation analysis.VariablesM ± SDT1 PARS-3T1 SAST1 IATT2 PARS-3T2 SAST1 IATT1 PARS-323.90 ± 22.201T1 SAS37.55 ± 12.27− 0.42***1T1 IAT37.33 ± 18.05− 0.42***0.52***1T2 PARS-319.68 ± 21.340.83***− 0.42***− 0.38***1T2 SAS36.64 ± 15.83− 0.49***0.61***0.35***− 0.48***1T2 IAT36.69 ± 17.84− 0.42***0.47***0.38***− 0.41***0.58***1(***p < 0.001,** p < 0.01, *p < 0.05) PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test. Table 5Correlation analysis.VariablesM ± SDT1 PARS-3T1 SAST1 IATT2 PARS-3T2 SAST1 IATT1 PARS-323.90 ± 22.201T1 SAS37.55 ± 12.27− 0.42***1T1 IAT37.33 ± 18.05− 0.42***0.52***1T2 PARS-319.68 ± 21.340.83***− 0.42***− 0.38***1T2 SAS36.64 ± 15.83− 0.49***0.61***0.35***− 0.48***1T2 IAT36.69 ± 17.84− 0.42***0.47***0.38***− 0.41***0.58***1(***p < 0.001,** p < 0.01, *p < 0.05) PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test. Correlation analysis. (***p < 0.001,** p < 0.01, *p < 0.05) PARS-3 = Physical Activity Rating Scale-3; SAS = Social Anxiety Scale for Adolescents; IAT = Internet Addiction Test. Cross-lagged analysis Cross-lagged panel analysis was employed in this study. This method can control the autoregressive effects of baseline variables, clarify the temporal order and directional predictive relationships among variables39, which is suitable for exploring the longitudinal directional among PA, SA and IA. (reviewer 2 comment 8). The cross-lagged panel analysis was performed using AMOS 27.0 with the maximum likelihood (ML) estimation method. The model fit indices were calculated and presented in Table 6, and the detailed fit indexes were as follows: χ2/df = 0.609, GFI = 0.998, CFI = 1.000, RMSEA = 0.001, indicating that the proposed cross-lagged panel model demonstrated an adequate overall fit to the data. The cross-lagged results showed that PA at T1 significantly and positively predicted PA at T2 (β = 0.90, p < 0.001), and significantly and negatively predicted SA at T2 (β = − 0.32, p < 0.01) and IA at T2 (β = − 0.12, p < 0.001). SA at T1 significantly and positively predicted SA at T2 (β = 0.50, p < 0.001), and SA at T1 negatively predicted PA at T2 (β = − 0.03, p < 0.001) as well as IA at T2 (β = 0.10, p < 0.01). In addition, IA at T1 significantly and negatively predicted PA at T2 (β = − 0.05, p < 0.001), whereas its positive predictive effect on SA at T2 was not significant (p > 0.05). These results are illustrated in Fig. 1.Table 6Model-fit indices for the cross-lagged model.χ2/dfGFICFIRMSEAModel0.6090.9981.0000.001Fig. 1Cross-lagged panel model of physical activities, social anxiety, and internet addiction across two time points. Note All values represent standardized regression coefficients. Solid lines indicate significant paths (***p < 0.001,** p < 0.01, *p < 0.05), and the dashed line indicates a non-significant path. T1 = Time 1, T2 = Time 2. Table 6Model-fit indices for the cross-lagged model.χ2/dfGFICFIRMSEAModel0.6090.9981.0000.001 Model-fit indices for the cross-lagged model. Cross-lagged panel model of physical activities, social anxiety, and internet addiction across two time points. Note All values represent standardized regression coefficients. Solid lines indicate significant paths (***p < 0.001,** p < 0.01, *p < 0.05), and the dashed line indicates a non-significant path. T1 = Time 1, T2 = Time 2. Discussion The present study examined the longitudinal relationships among adolescents’ PA, SA, and IA to advance understanding of their dynamic interplay in natural developmental contexts. Using a two-wave cross-lagged panel design, 877 middle school students in Jiangxi Province were followed over 16 weeks to assess bidirectional predictive relationships among these variables. The cross-lagged model demonstrated good fit and confirmed significant longitudinal associations. Demographic factors such as gender and grade showed differential effects, aligning with prior cross-sectional findings. Overall, this study extends the literature by providing novel longitudinal evidence on the dynamic interrelations among PA, SA, and IA. Boys exhibited higher PA (p < 0.001) and lower SA (p < 0.001) and IA (p < 0.001) than girls, consistent with prior research40. In the present study, girls reported higher levels of SA severity than boys, which is also in line with prior empirical evidence41. Higher SA in girls may reflect increased internalizing symptoms during adolescence, influenced by pubertal hormonal changes and gender-specific psychosocial stressors such as body dissatisfaction and restrictive gender roles42,43. Although traditional studies often report higher IA in boys due to gaming, the present study found higher IA levels among girls. This discrepancy may stem from prior measures emphasizing gaming, whereas girls tend to engage more with short-video platforms and social media, which involve strong emotional engagement44–48. Gender differences in cognitive control, family monitoring, and the regulatory environment—such as China’s “anti-addiction systems” for gaming but not for social media—may further explain why boys show lower IA despite higher gaming activity, while girls’ predominant use of social media contributes to elevated IA levels49. At T1, adolescents’ PA and SA increased significantly with grade level (p < 0.05), likely reflecting greater academic pressure, increased use of PA for stress relief, and developmental factors such as heightened self-consciousness and sensitivity to peer evaluation50,51. By T2, these grade-related differences were no longer significant, possibly because pervasive examination-related stress and increased academic workload attenuated variability across grades. These findings suggest that grade-level differences in PA and SA are context-dependent and highlight the importance of considering dynamic fluctuations in psychological and behavioral variables across academic periods when designing interventions. Contrary to expectations, only-child status did not significantly affect PA, SA, or IA (p > 0.05). This aligns with contemporary social and parenting changes, as recent evidence indicates that, after controlling for socioeconomic and family factors, only children do not differ from non-only children on most mental health indicators and may even show favorable outcomes on certain emotional and stress measures52. These findings suggest that traditional assumptions of greater psychological vulnerability among only children may no longer apply. Greater equality in family resources and strengthened social support systems likely reduce the influence of only-child status, while family relationship quality—such as parent–child interactions, family functioning, and parental monitoring—appears more closely linked to adolescents’ addictive behaviors53. From a longitudinal perspective, this study systematically examined the bidirectional cross-temporal relationships among adolescents’ PA, SA, and IA using a two-wave cross-lagged panel model. The results indicated that T1 PA negatively predicted T2 SA (β = − 0.32, p < 0.001) and T2 IA (β = − 0.12, p < 0.001). T1 SA positively predicted T2 SA (β = 0.50, p < 0.001) and T2 IA (β = 0.10, p < 0.01). Conversely, both T1 SA and T1 IA negatively predicted T2 PA (β = − 0.03, p < 0.001; β = − 0.05, p < 0.001). These findings suggest that during adolescence, PA, emotional experiences, and problematic internet use do not follow a unidirectional causal pattern but instead constitute a dynamically interactive and mutually influential developmental system. Based on these results, the following sections further discuss the longitudinal mechanisms underlying the relationships among the variables by focusing on specific cross-lagged pathways. First, T1 PA significantly and positively predicted T2 PA (β = 0.78, p < 0.001). The relatively large autoregressive coefficient indicates substantial temporal stability in adolescents’ PA levels over the 16-week interval, which is consistent with prior research suggesting continuity in PA trajectories as a health behavior54. Moreover, given that the measurement interval covered an entire academic semester—during which school-based physical education curricula and opportunities for extracurricular PA were relatively stable—this autoregressive effect likely reflects the realistic persistence of PA behaviors within the school context. Notably, such behavioral stability may imply that adolescents with low levels of PA are unlikely to naturally increase their participation in exercise over the short term. This finding suggests that relying solely on the passage of time or natural environmental changes may be insufficient to promote improvement among low-activity individuals. Therefore, future research should further explore targeted intervention strategies for adolescents with persistently low levels of PA in order to disrupt stable low-activity trajectories. The present study also found that T1 PA negatively predicted T2 SA (β = − 0.32, p < 0.001) and significantly negatively predicted T2 IA (β = − 0.12, p < 0.001). These findings are consistent with previous research55,56. From the perspective of the cognitive–behavioral model, the development and maintenance of SA are often influenced by a range of antecedent vulnerability factors, including excessive expectations of negative social evaluation, deficits in emotion regulation, and a tendency toward avoidance behaviors24. Regular PA has been shown to enhance emotion regulation capacity and stress-coping strategies, enabling adolescents to better manage tension and anxiety in social situations57. In addition, group-based PA typically involve frequent interaction and cooperation, which may naturally reduce behavioral avoidance and gradually weaken the negative cycles associated with social withdrawal58. Within the cognitive–behavioral framework, these factors jointly attenuate the vulnerability basis of SA, thereby accounting for the observed negative cross-lagged effect of PA on subsequent SA. Although prior studies have widely documented a significant association between PA and IA, most have relied on cross-sectional designs, making it difficult to disentangle temporal stability from cross-time effects. In this context, relatively few studies have employed cross-lagged panel models to systematically examine the predictive role of PA on subsequent IA. The present study found that T1 PA negatively predicted T2 IA significantly, although the effect size was small and warrants cautious interpretation According to the cognitive–behavioral model, the development and persistence of IA are often driven by maladaptive cognitions and the repeated reinforcement of avoidant coping strategies59. Within this framework, PA may serve as a potential antecedent protective factor by attenuating these vulnerability bases prior to the onset of problematic internet use. Empirical evidence suggests that regular PA is associated with enhanced perceived control over real-life situations and increased self-efficacy60, which may help reduce irrational beliefs related to seeking compensatory gratification through online activities. Furthermore, by improving psychological resilience and alleviating negative emotional states such as anxiety and depression55, PA may also reduce adolescents’ tendency to rely on the internet as a means of emotional escape. From a cognitive–behavioral perspective, these cognitive, emotional, and behavioral pathways may jointly contribute to the observed longitudinal association. Nevertheless, given the small effect size, due to multiple testing, and possible methodological constraints, this result may provide only preliminary evidence of a temporal association rather than a strong predictive effect—these factors are speculative and may not be the sole or definitive causes. Further studies with larger samples and replicated designs are needed to confirm its robustness and explore potential factors underlying the small effect size. The present study found that T1 SA significantly and positively predicted T2 SA, indicating a high degree of temporal stability in social anxiety across the 16-week interval. This finding is consistent with previous research suggesting that social anxiety in adolescence exhibits relatively stable trait-like characteristics61. According to the cognitive–behavioral model, social anxiety is maintained through self-perpetuating mechanisms such as negative self-cognitions, heightened self-focused attention, and avoidant coping strategies62,63, making it difficult for anxiety symptoms to remit naturally in the absence of external intervention. Moreover, within relatively stable school environments, persistent pressures related to social evaluation may further reinforce the continuity of social anxiety symptoms. From the perspective of stress–coping theory, when individuals are continuously exposed to psychological stressors and perceive their coping resources as insufficient, they are more likely to adopt emotion-focused and avoidant coping strategies to temporarily alleviate distress28. In addition, T1 SA significantly and positively predicted T2 IA, but the effect was extremely small (β = 0.04), suggesting limited practical relevance despite statistical significance. Although individuals with elevated SA may be more likely to perceive online environments as safer alternatives to face-to-face interactions24,64, due to features such as anonymity, controllability, and reduced risk of negative evaluation, this mechanism may appear to contribute minimally to the development of IA in the current study, though this interpretation remains tentative. Considering potential non-response bias and multiple comparisons, this association should be interpreted conservatively and not overstated, as these factors may or may not be the primary drivers of the observed result. Future research is needed to verify its stability using more representative samples and clarify potential influencing factors. T1 SA and T1 IA both significantly and negatively predicted T2 PA (β = − 0.03, p < 0.001; β = − 0.05, p < 0.001, respectively), yet these effects were again small in magnitude and require cautious interpretation. Within the stress–coping framework, SA may be associated with a tendency to adopt avoidant coping strategies, which could, to some extent, reduce individuals’ motivation to engage in PA. For adolescents with higher levels of SA, PA contexts—often involving social interaction and perceived evaluation—may be experienced as relatively stressful, potentially leading to avoidance responses, particularly among girls65. According to stress–coping theory28, individuals who rely more on avoidant coping may be inclined to reduce exposure to perceived high-threat situations while shifting toward lower-threat alternatives66. In this context, PA participation may be modestly reduced over time. Similarly, frequent internet use may provide a readily accessible and low-threshold coping alternative. For individuals inclined toward avoidant coping, online activities may serve as a convenient way to manage stress or negative affect, which could be weakly associated with reduced engagement in PA over time59,67. From this perspective, the observed associations may reflect a potential displacement tendency, although the strength of this relationship appears limited. Importantly, the small effect sizes observed in these cross-lagged paths may be attributable to both substantive and methodological factors. On the one hand, adolescents’ PA levels are often constrained by relatively stable school-based physical education curricula and structured extracurricular programs, which may limit short-term variability. On the other hand, prior methodological research suggests that, after controlling for autoregressive effects, cross-lagged associations in longitudinal models are typically small in magnitude, even when statistically significant68. Accordingly, these findings may provide only preliminary evidence of temporal associations rather than robust predictive relationships. Caution is warranted in interpreting their practical or clinical significance, as the factors underlying the small effect sizes have not been definitively identified, and future longitudinal studies with longer intervals are needed to examine generalizability and clarify these potential factors. In the present study, PA demonstrated relatively high temporal stability, a finding that may be closely related to the academically oriented educational system embedded within the Chinese cultural context. In East Asian societies, adolescents’ daily routines are typically characterized by strong structural constraints and external regulation, with PA participation relying largely on scheduled school-based physical education classes or fixed extracurricular training programs rather than being autonomously adjusted according to individuals’ psychological or physical states69. Within this context, PA levels tend to remain stable over short time intervals, which may, in turn, attenuate the predictive effects of psychological variables on changes in PA behavior. Moreover, within China’s collectivistic cultural framework, individuals are generally more attentive to social evaluation and group norms, which may amplify the behavioral impact of SA. Compared with cultural contexts that emphasize individual expression, adolescents in East Asian cultures are more likely to experience heightened social pressure in public and evaluative situations70. Consequently, the inherent features of PA contexts—such as peer comparison, performance evaluation, and public exposure—may culturally reinforce the inhibitory effect of SA on PA participation. Under conditions of high academic pressure and relatively restrained emotional expression, the online environment may be more readily perceived by adolescents as a low-cost regulatory space with reduced evaluation risk. This cultural backdrop may facilitate the transformation of SA or broader psychological stress into increased online engagement, thereby longitudinally elevating the risk of IA.Taken together, these findings provide empirical support for understanding cross-cultural differences in the relationships among PA, psychological factors, and IA. Future research should employ multi-wave longitudinal designs across diverse cultural contexts to examine whether these associations differ between cultures that emphasize individual autonomy and those that prioritize collective norms. Limitations Several limitations should be noted. First, the sample was drawn from three secondary schools in Jiangxi Province and included a high proportion of only children (78.4%), which may limit generalizability to adolescents from other regions or demographic backgrounds. Second, although a two-wave design with a 16-week interval was adopted, this period may be insufficient to capture long-term developmental trajectories, and relevant confounders were not fully controlled. Third, all variables were assessed via self-report, rendering the results susceptible to response biases. Third, the PARS-3 and IAT did not distinguish specific PA types or IA subtypes, limiting the assessment of behavioral heterogeneity. Fourth, the study was conducted within a collectivistic and academically structured cultural context in China, which may constrain the predictive effects of psychological variables and limit cross-cultural generalizability. Furthermore, some cross-lagged effects were small, and the absence of an experimental design precludes causal inference; thus, the findings should be interpreted as predictive associations rather than definitive causal effects. Conclusion This study identifies dynamic prospective associations among adolescents’ physical activity (PA), social anxiety (SA), and internet addiction (IA) using two-wave longitudinal data and a cross-lagged panel design, extending prior cross-sectional and unidirectional findings. The findings reveal that PA, SA, and IA operate within a complex relational system, with predictive links that align with cognitive–behavioral and stress–coping frameworks—specifically highlighting avoidant coping as a potential mechanism linking these variables. Methodologically, the findings confirm that although cross-lagged effects are modest in magnitude, they yield meaningful theoretical insights and reinforce the utility of cross-lagged models in examining adolescent developmental processes. Collectively, these findings advance our understanding of the prospective PA–IA relationship and provide a more nuanced perspective on adolescent health behavior. Furthermore, within an academically structured cultural context, the temporal stability of PA and the constrained predictive effects of psychological factors underscore the need for cross-cultural perspectives in adolescent behavioral research. Future studies should employ longer follow-up periods, incorporate contextual and cultural variables, and adopt experimental or intervention designs to further clarify the long-term role of PA in promoting adolescent mental health and preventing problematic internet use.