This study employed a cross-sectional survey method to assess the conceptual model. Data were collected by a reputable market research company, which first used screening questions to confirm that participants had experience using social media, and system control measures were applied to prevent duplicate responses. To minimize potential interference arising from regional differences in social media usage habits and network infrastructure (Cheng et al., 2021), Taiwan was chosen as the research context, given its global leadership in information and communication technology (ICT) applications. Taiwan performed well in the World Digital Ranking, placing ninth among 67 economies in 2024 (IMD World Competitiveness Center, 2024). Prior to the study, sufficient information was provided to ensure that all participants gave informed consent.

The sample size was calculated based on statistical power analysis to ensure sufficient participants during the formal data collection phase. A priori analysis was conducted using the G-Power tool (Faul et al., 2009) to confirm the adequacy of the sample size. Following Cohen’s (1988) recommendations, the final sample of 532 participants achieved a statistical power of 99%, significantly exceeding the standard requirement.

As shown in Table 1, the survey collected 532 valid samples, with an average participant age of 37.19 years (SD = 9.06). Of all participants, 37.22% were male, and 62.78% were female. The majority of participants had a university degree (74.81%), with the remaining participants having a high school or vocational education (14.10%) and master’s or above (11.09%). According to the criteria suggested by Pontes et al. (2018), 88 participants (16.54%) met the threshold for SMA, which aligns with the 8%–19% prevalence range reported in previous meta-analyses (Cheng et al., 2021). This indicates that the study’s data are consistent with broader population distributions, thereby enhancing the external validity of the results.

The survey utilized several established scales from previous literature, all of which were measured using a 5-point Likert scale ranging from 1 (strongly disagree) to 5 (strongly agree). The questionnaire was professionally translated into Chinese by bilingual experts, followed by a semantic validation process with ten native Chinese speakers and a back-translation procedure to ensure fidelity to the original texts. The scales are described as follows.

GCV was measured using the GCV scale developed by Haws et al. (2014), which aims to capture the extent to which individuals express environmental protection values through their purchasing and consumption behaviors. The scale consists of six items on a single dimension, with sample items including, “My purchasing habits are affected by my concern for the environment”.

GIS was measured using the social media information sharing (SMIS) scale developed by Bock et al. (2005), which consists of five items on a single dimension and primarily assesses users’ attitudes toward information sharing on social media. In this study, the focus was specifically on green-related information. A sample item includes, “Sharing environmentally friendly green information with others on social media is valuable to me”.

Self-orientation was adapted from Whiting and Williams (2013), which measures the self-related motivations for using social media. The scale consists of four items on a single dimension, with sample items such as, “On social media, I enjoy gaining attention and praise through posts, status updates, or showcasing achievements”, and “I often seek likes, comments, and shares to enhance my emotional satisfaction”.

Other-orientation was adapted from Apuke and Omar (2021), primarily measuring the other-oriented use of social media. The scale consists of five items on a single dimension, with sample items such as, “I engage, share, and respond on social media because I want to remind others of important issues”.

SMA was assessed using the Bergen Social Media Addiction Scale (BSMAS), which consists of six items evaluating participants’ levels of SMA over the past 12 months. The six items reflect typical elements of addiction, including salience, mood modification, tolerance, withdrawal, conflict, and relapse (Griffiths, 2005; Schou Andreassen et al., 2016). The scale has been validated across different language versions. In social media research, addictive behavior is a clinically diagnosable term, and the prevalence of SMA has been validated using various threshold criteria, though these thresholds may change over time (Cheng et al., 2021). For this study, a more stringent threshold of 4 was adopted to test Hypothesis 5, as recommended by Pontes et al. (2018).

This study employed Partial Least Squares Structural Equation Modeling (PLS-SEM) software for statistical analysis, as it is capable of estimating models with smaller sample sizes and handling complex relationships between multiple variables. For common method variance (CMV) testing, the Harman single-factor test explained 36.142% of the variance, which is well below the 50% threshold. Multicollinearity analysis showed that all VIF values ranged between 1.56 and 2.10, well below the recommended threshold of 3.3 proposed by Kock and Lynn (2012), indicating that CMV was not a major concern in this study. In the subsequent analysis, the measurement model was first evaluated, followed by the validation of the structural equation model (Sarstedt et al., 2017) to test the impact pathways among the variables for Hypotheses 1 to 4. Finally, Partial Least Squares Multi-Group Analysis (PLS-MGA) was employed to examine the differences in the pathways between the groups of social media addicts and non-addicts.

The reliability and validity of the five reflective constructs included in the research model were evaluated. To ensure good convergent validity, all factor loadings were required to exceed 0.6 (Hair et al., 2017). Based on this criterion, the constructs were unidimensional, and only items with loadings above 0.6 were retained in the measurement model (Shi and Maydeu-Olivares, 2020; Hair et al., 2010). As shown in Table 2, without removing any items, all Cronbach’s alpha values ranged from 0.835 to 0.848, and all composite reliability (CR) values exceeded the recommended threshold of 0.7, indicating sufficient internal consistency reliability (Hair et al., 2010). The average variance extracted (AVE) was used to assess convergent validity, with AVE values ranging from 0.554 to 0.688, all above the recommended threshold of 0.5 proposed by Fornell and Larcker (1981), indicating that the constructs explained over half of the variance in their respective items. Additionally, as shown in Table 3, all item loadings exceeded 0.7. These results confirm the discriminant validity of the constructs.

The overall model fit is acceptable. SRMR is 0.055 for the saturated model and 0.082 for the estimated model-below or close to the 0.08 guideline. The discrepancy based on Unweighted Least Squares (d_ULS) (0.639; 1.428) and the discrepancy based on Geodesic distance (d_G) (0.202; 0.241) are small, indicating the covariance structure is well reproduced. The chi-square values (632.725; 719.294) are large, which is common with large samples or complex models. NFI is 0.872 and 0.854, suggesting a moderate comparative fit (Hair et al., 2010). In sum, the model shows a satisfactory and methodologically acceptable fit. The f-square results indicate medium effects of OO on both endogenous constructs (GCV = 0.251; GIS = 0.272). By contrast, self-oriented motivation (SO) shows minor effects (GCV = 0.007; GIS = 0.011). R-squared/adjusted R-squared for GCV are 0.315/0.312, and for GIS are 0.344/0.341, exhibiting moderate explanatory strength.

The HTMT and CI (confidence intervals) results support satisfactory discriminant validity as shown in Table 4. Pairwise HTMT values range from 0.449 to 0.728, all below the commonly used 0.85 threshold, indicating that the constructs are well differentiated (Henseler et al., 2015). The bootstrap 95% confidence intervals span 0.354–0.804, with upper bounds not approaching 1.00. Overall, the constructs do not exhibit excessive overlap, providing empirical support for the discriminant validity of the measurement model.

The common Fornell-Larcker criterion for discriminant validity, as seen in Table 5, shows that the square root of each construct’s AVE was greater than the correlation with any other construct (Fornell and Larcker, 1981).

As shown in Table 6, the upper part of the table presents the path analysis for the impact of self- and other-orientation on GIS. The effect of other-orientation on GIS was positive and significant ($\beta$ = 0.519, p 0.001), while the effect of self-orientation on GIS was also positive and significant ($\beta$ = 0.105, p 0.05).

The lower part of the table presents the path analysis for the impact of both orientations on GCV. The effect of other-orientation on GCV was positive and significant ($\beta$ = 0.509, p 0.001), whereas the effect of self-orientation on GCV did not reach statistical significance ($\beta$ = 0.083, p $>$ 0.05).

Overall, the results indicate that both GIS and GCV are associated with other-orientation, while self-orientation is associated only with GIS.

To verify the generalizability of the concept and compare whether SMA negatively impacts GIS and values, this study conducted multi-group analysis (MGA). To ensure the validity of between-group comparisons prior to MGA, we conducted the MICOM procedure in SmartPLS with 5000 permutations to assess measurement invariance between the social media addicts and non-addicts. As shown in Table 7, Step 2 (compositional invariance) held for all constructs (GCV = 0.393; GIS = 0.518; OO = 0.288; SO = 0.233; all p $>$ 0.05), indicating that the composites are formed equivalently across groups. Step 3a (equality of means) was not supported (all four constructs p = 0.000), evidencing significant mean differences. Step 3b (equality of variances) was supported for GCV (p = 0.963), GIS (p = 0.180), and OO (p = 0.265), but not for SO (p 0.000). Taken together, the results establish partial measurement invariance, which is sufficient to proceed with PLS-MGA comparisons of structural paths (Henseler et al., 2016).

The MGA pathway test results are shown in Table 8; no significant differences were found between the groups based on gender, age, income, or SMA (p $>$ 0.05).

On the other hand, this study further employed independent sample t-tests to examine differences in the four main variables between social media addicts and non-addicts. The results, shown in Table 9, indicate significant differences between social media addicts (M = 4.04, SD = 0.55) and non-addicts (M = 3.24, SD = 0.73) in self-orientation (t = –11.52***, p 0.001). Significant differences were also found in other-orientation between social media addicts (M = 4.13, SD = 0.49) and non-addicts (M = 3.61, SD = 0.54) (t = –8.34***, p 0.001), in GIS (SMA: M = 4.25, SD = 0.46; non-SMA: M = 3.86, SD = 0.51; (t = –6.49***, p 0.001), and in GCV (SMA: M = 4.18, SD = 0.51; non-SMA: M = 3.82, SD = 0.51; (t = –5.99***, p 0.001).

In summary, social media addicts scored significantly higher than non-addicts on all four main variables. However, in terms of the effect pathways, the SMA group did not exhibit stronger effects compared to the non-addiction group.

This study has several limitations. First, the cross-sectional design limits causal inference, and reverse causality cannot be ruled out. Second, potential interaction effects between self- and other-orientation were not examined and may be associated with prosocial engagement in more complex ways. Third, the study focused on GIS as one form of visible online prosocial behavior; future work should explore whether similar motivational patterns hold for other digital prosocial acts, such as online volunteering or mental health advocacy. Fourth, the sample was drawn from a single region. Although Taiwan offers a compelling context given its global leadership in ICT applications, future research should replicate the analyses across diverse cultural settings to evaluate the robustness and generalizability of the findings. Finally, SMA is a multifaceted construct, and its nuanced links to motivation and behavior merit further investigation.

This study has several limitations. First, the cross-sectional design limits causal inference, and reverse causality cannot be ruled out. Second, potential interaction effects between self- and other-orientation were not examined and may be associated with prosocial engagement in more complex ways. Third, the study focused on GIS as one form of visible online prosocial behavior; future work should explore whether similar motivational patterns hold for other digital prosocial acts, such as online volunteering or mental health advocacy. Fourth, the sample was drawn from a single region. Although Taiwan offers a compelling context given its global leadership in ICT applications, future research should replicate the analyses across diverse cultural settings to evaluate the robustness and generalizability of the findings. Finally, SMA is a multifaceted construct, and its nuanced links to motivation and behavior merit further investigation.