Compensatory internet use theory focuses on explaining why individuals become addicted to the internet (Kardefelt-Winther, 2014a). The theory suggests that individuals may compensate in the online world for unmet needs in real life. When people’s negative emotions cannot be effectively vented in real life, individuals may use the internet as an escape or coping strategy to reduce the impact of negative emotions, which may lead to internet addiction in the long run (Kardefelt-Winther, 2014a). According to the compensatory internet use theory, the degree of an individual’s short video addiction is influenced by negative emotional states, particularly stress perception (Liu et al., 2021). Individuals could experience various forms of stress in daily life, including academic pressure, employment challenges, work-related stress, and familial obligations. To cope with these pressures, individuals tend to relieve themselves by using short video applications, which leads to excessive usage of short videos or short video addiction.

The compensatory internet use theory also holds the importance of self-compensatory motivations, and highlights the variations in different contexts. For instance, the compensatory purposes of video games include achievement, social interaction, and immersion (Kardefelt-Winther, 2014b), while the compensatory purposes of social network service include information search and entertainment (Luchman et al., 2014), as well as maladaptive mood regulation (LaRose et al., 2003; Caplan, 2005). Short video applications such as TikTok, Kwai, and Bilibili allow users to upload short videos of 10 of seconds or minutes from their daily lives, such as cooking, singing, dancing, traveling, and health, for other users to watch (Zhang et al., 2019). These UGC contents are interesting and have addictive hedonic value (Tian et al., 2023; Cui et al., 2022; Zhang et al., 2019). These applications also provide functions such as “like,” “comment,” or “forward” to meet users’ social needs (Da-yong and Zhan, 2022). Therefore, this study classifies the compensatory motives of users who choose to use short video applications as relaxing entertainment and social interaction, and takes them as mediating factors to analyze the mechanisms between stress and short video addiction.

The compensatory internet use theory reveals the mediating role of compensatory motivations (e.g., social interaction and relaxing entertainment) in the relationship between negative emotions (i.e., stress) and internet addiction behaviors (Kardefelt-Winther, 2014a, b), but overlooks the possibility of other factors related to affective responses. Prior studies have found that motives positively predict affective responses (Karagiannidis et al., 2015), suggesting that once users perceive the favorable value of using mobile service, they respond with certain positive experiences. For instance, Pang (2021) demonstrated that individuals’ positive attitude and gratification toward social media usage was impacted by hedonic and utilitarian values. Simultaneously, affective responses have a significant positive impact on short video addiction, such as curiosity, and affinity (Dong et al., 2024). Therefore, we contend that it is necessary to integrate other affective factors into the original CIU model and construct an extended CIU model to improve the explanatory power of the short video addiction behavior.

Attitude is an important factor that effectively predicts online addictive behavior (Can and Kaya, 2016; Jeong and Kim, 2011; Tsai and Lin, 2001). Previous studies indicated that attitude should be conceptualized as a tripartite structure, primarily consisting of affective (e.g., emotion or feeling), cognitive (e.g., beliefs, judgments, or thoughts), and behavioral information domains (Edwards, 1990; Bizer and Krosnick, 2001). Nevertheless, scholars’ definitions of attitude show their recognition of the affective component. For example, a handful of extant studies suggested that attitude refers to the positive and negative views held by an individual toward a specific object, such as an action, event, situation, issue, or people (Howarth, 2006; Marcinkowski and Reid, 2019). Ozel et al. (2013, p. 13) also indicated that attitude reflects “a general liking or disliking, or more specific affective reactions toward the object.” Similarly, Petty and Briñol (2015, p. 2) have posited that “attitudes are not only based on thoughts and beliefs but also feelings and emotions.” In addition, existing studies have also regarded attitude as a factor of affective response when constructing theoretical models (Henter, 2014; Pang, 2021; Zhang, 2013). Therefore, attitude is one of the affective factors considered in this study.

Generally, immersion has been recognized as one of the most significant conceptions for understanding users’ affective experience in the field of media effect (Cheng and Tsai, 2020; Lin et al., 2020). Immersion refers to “a state of deep mental involvement in which their cognitive processes (with or without sensory stimulation) cause a shift in their attentional state such that one may experience dissociation from the awareness of the physical world” (Agrawal et al., 2020, p. 407). Jennett et al. (2008) analyzed the differences between immersion and relevant concepts such as flow, cognitive absorption, and presence in detail. They argued that the key to the construct of immersion lay in emphasizing the loss of awareness of time and the real world, involvement, and focused attention. Building upon this foundation, this study believes that when individuals watch short videos they like or prefer, they have a good viewing experience and a sense of being “lost to the world” (Seah and Cairns, 2008). At present, most studies on immersion and addictive behaviors focus on video games (Lee et al., 2021; Seah and Cairns, 2008), virtual reality services (Saneinia et al., 2022), and gambling (Rémond and Romo, 2019), the role of immersion in short video addiction also needs to be empirically tested. Thus, this study proposes attitude and immersion as the affective factors, we aim to empirically investigate how stress perception affects addictive outcomes through motivational and affective paths in the context of short video applications.

Following rigorous translation and back-translation, all English measurement items involved in this research were translated into Chinese by the authors. The questionnaire used in this survey mainly consists of two parts and two additional screening question. First, participants need to answer one screening questions, which asks whether they have the habit of watching short videos in the past 6 months. Participants who choose “Yes” are invited to fill in the remaining survey questions. In contrast, those who choose “No” are denied to participate. The first part measures the model variables, and the second part is descriptive statistical analysis, including variables such as gender, age, monthly income, education level, and average daily usage time (Table 1). Additionally, an attention-check question was inserted in the middle of a questionnaire that helps to screen out invalid questionnaires. In total, 30 questions of the questionnaire were encompassed in this paper.

To further enhance the validity of the questionnaire and ensure that each item could represent a specific construct, we invited four experienced researchers and four master students in the field to modify it and obtained their approval. Subsequently, we conducted a pre-survey with 82 participants to evaluate whether the reliability and validity of the six scales met the thresholds proposed by previous studies. We undertook additional modifications to the questionnaire scale in response to the confusion and feedback provided by participants following their completion of the survey. The results of the preliminary survey are satisfactory, and then we can proceed to distributing the formal version’s questionnaire. In the first part of the questionnaire, the measurement items were evaluated using a 5-point Likert scale, ranging from “1 = strongly disagree” to “5 = strongly agree.” Specifically, attitude was measured in three items using Ajzen’s (2002) and Eagly and Chaiken’s (1993) scale. Following Khan (2017), relaxing entertainment was measured by using two items. The measurement of immersion was adopted from Hamari et al. (2016). Social interaction was measured in three items using Zadeh et al. (2023) scale. The measurement of perceived stress was based on Cohen et al.’s (1995) scale. Moreover, short video addiction was adopted from Chen et al. (2003). All measurement items are included in the Table 2.

The formal survey was conducted from June 3rd to July 5th, 2021, utilizing the Sojump¹ online platform, which boasts a user base of over 10 million. This study obtained approval from the institutional review board (IRB) of the author’s affiliated institution. A total of 356 questionnaires were distributed, and after excluding invalid responses such as those from non-short video users and those with identical consecutive answers exceeding 10 instances, a total of 319 valid questionnaires were collected, resulting in a response rate of 89.60%. Table 1 presents the sample characteristics. Among all 319 participants, 46.4% (n = 148) were males and 53.6% (n = 171) were females. 68.9% of them had a bachelor’s degree or higher and 46.4% of the participants engaged in daily consumption of short videos for over 1 h.

SPSS 24.0 and SmartPLS 4.0 were used to analyze the data. This study used SPSS 24.0 to conduct a descriptive analysis of respondent demographic characteristics analysis. Then, a two-step approach was used for PLS-SEM analysis. In the initial step, the reliability, content validity, and discriminant validity of the measurement model were evaluated. In the second step, path analysis and multigroup analysis was applied to test the proposed research hypotheses in this study.

The validity and reliability of the measurement were assessed through confirmatory factor analysis using SmartPLS 4.0. As shown in Table 3, the Cronbach’s alpha values exceeded a marginal value of >0.7, and Composite Reliability (CR) for each construct was higher than 0.7, which is a widely accepted threshold in research, indicating good internal consistency reliability (Hair et al., 2017). The Average Variance Extracted (AVE) values for all constructs ranged from 0.685 to 0.880, surpassing the threshold of >0.5, and all Outer Loadings (OL) were greater than the suggested value of 0.7, leading to the conclusion that convergent validity is not a concern (Bagozzi and Yi, 1988). The findings, as presented in Table 4, confirmed the discriminant validity by showing that the square root of the AVE (highlighted in bold and displayed on the diagonal figure) surpassed its correlation coefficients with other factors. Hence, the present study has favorable discriminant validity (Fornell and Larcker, 1981).

In the analysis of the reflective measurement model, we employed the Heterotrait Monotrait Ratio (HTMT) criterion proposed by Henseler et al. (2015) to evaluate discriminant validity among six constructs (see Table 5). The scores of HTMT for all constructs are below the threshold value of 0.85, confirming the discriminant validity of the model.

The structural analysis represents the second step of PLS-SEM. To determine the significance of each path coefficient, bootstrapping with 5,000 samples was used. The result of the structural model is depicted in Figure 2, displaying path coefficients, significance levels of paths, VIF, f², R² values, and Q² values. The model explains 24.7% of the variance for relaxing entertainment, 23.9% of the variance for social interaction, 38.5% of the variance for immersion, 45.6% of the variance for attitude, and 57.9% of the variance for short video addiction. This finding showed that the explanatory power of our model is acceptable (Chin, 1998). Furthermore, relaxing entertainment has 0.197, social interaction has 0.186, immersion has 0.254, attitude has 0.317, and short video addiction 0.384 as Q² values, all Q² values exceed zero and indicated that the predictive relevance of the structural model was satisfied (Geisser, 1975). In addition, we examined the model fit using the standardized root mean square residual (SRMR) measure. Our model had an SRMR value of 0.072, which is below the threshold of 0.08 (Henseler et al., 2016), and can be considered an acceptable value for a PLS-SEM-based model.

As shown in Figure 2 and Table 6, the findings from the SEM analysis revealed several key relationships. Perceived stress is positively related to the short video addiction (β = 0.424, p < 0.001), supporting H1. Perceived stress is positively associated with relaxing entertainment and social interaction (β = 0.497, p < 0.001; β = 0.489, p < 0.001). Hence, H2 and H3 were confirmed. Relaxing entertainment displayed significant and positive associations with attitude and immersion (β = 0.335, p < 0.001; β = 0.593, p < 0.001). H4 and H6 were supported. Social interaction exhibited noteworthy positive correlations with attitude (β = 0.486, p < 0.001), but failed to be related to immersion (β = 0.073, p > 0.05). Therefore, H5 was supported, but H7 was not supported. The effect of attitude on short video addiction is significantly positive (β = 0.287, p < 0.001). H8 was supported. The results reveal a significant and positive correlation between immersion and short video addiction (β = 0.226, p < 0.001). Hence, H9 was confirmed.

Following the preceding findings, it is evident that further analysis is warranted to investigate the mediating effects in the relationship between perceived pressure and short video addiction. Mediation analyses using bootstrapping were performed in SmartPLS 4.0 to examine two specific pathways as outlined in Table 7. Given that the positive correlation between social engagement and immersion remains unverified, this study focuses on three specific paths. As indicated by the path analysis results presented in Table 7, perceived pressure influences short video addiction through a serial mediating pathway comprising relaxation entertainment, social interaction, immersion, and attitude.

We use the categorical variable “age group” as a moderator. More specifically, multigroup analysis has sample size requirements: (1) the sample size for each group must be almost equal; (2) no group can use less than 80% of the recommended sample size (Matthews, 2017). Our basic idea is to combine the descriptive statistics of 26–30, 31–35, and 36 and more into one group category “25 and more,” which further divide the sample into three groups: 14–20 (n = 70), 21–25 (n = 178), and 25 and more (n = 71). However, we found that the sample size in each group was not equal, which would lead to weakened statistical power and the underestimation of the moderating effect (Hair et al., 2017; Aguinis et al., 2017). According to the recommendations of Aguinis et al. (2017) and Matthews (2017), we chose almost equal sample sizes. The main method was to use the random sampling method in the Excel software from the 21–25 (n = 178) group and select 71 participants.

Following the procedure suggested by Cheah et al. (2020), we first used the measurement invariance of composite models (MICOM) to assess measurement equivalence, which can be effectively explained if the research data standards are acceptable. More precisely, this study aimed to ascertain whether construct measurements are understood similarly across the three age groups. MICOM consists of three steps, the first step is the configurational invariance assessment, the second is the establishment of compositional invariance assessment and the third is the assessment of equal means and variances. The results as shown in Table 8, the MGA’s group-specific differences in the PLS-SEM findings are feasible to compare and interpret (Henseler et al., 2016) and are acceptable to perform path analysis among three groups.

From Table 9, our results show a significant difference between Group 14–20, Group 21–25, and Group >25. First, the comparisons between Group 14–20 and Group >25 (|diff| = − 0.574, p < 0.01) and Group 21–25 and Group >25 (|diff| = − 0.540, p < 0.001) show a significant relationship between perceived stress and short video addiction. Second, Group >25 significantly differs from Group 14–20 (|diff| = − 0.288, p < 0.01) and Group 21–25 and Group >25 (|diff| = − 0.334, p < 0.01) for the relationship between perceived stress and relaxing entertainment. Third, the result shows a significant difference between Group 14–20 and Group >25, and Group 21–25 and Group >25 for the relationship between attitude and short video addiction (|diff| = − 0.288, p < 0.01, |diff| = − 0.334, p < 0.01). Last, the result of the relationship between immersion and short video addiction showed that the Group 14–20 sample differs significantly from the Group >25 sample (|diff| = − 0.334, p < 0.05).

The empirical findings of the present study have two theoretical implications. First, this study introduces an E-CIU framework to elucidate users’ addiction toward short video applications. Previous research has predominantly focused on explaining addictive behavior using a motivational approach, neglecting a comprehensive investigation into affective path. In our framework, we integrate both psychological (i.e., stress), compensatory motivations (i.e., social interaction and relaxing entertainment), and affective responses (i.e., immersion and attitude) into an extended model to examine the impact paths of stress on addictive behavior. Consequently, this research establishes an innovative theoretical framework for understanding the underlying mechanisms in the relationship between stress and short video addiction. Secondly, we conducted a detailed analysis of the intricate interplay between these influencing factors among different groups. The results of this study contribute to the short video addiction literature by highlighting that stress has a great influence on short video addiction among different age groups.

This research also has two key practical implications. Firstly, the study highlights the role of stress perception in contributing to short video addiction. The findings reveal that stress directly influences users’ propensity for addiction to short videos. Stress affects users’ affective experiences, which in turn indirectly influence their immersion in short video content. Given that short videos have become a significant outlet for stress relief in modern life, it is crucial for designers of these applications to implement features that help users monitor and manage their viewing time. Providing prompts or tools to encourage users to take breaks could address their need for intermittent disconnection.

Secondly, targeted interventions should be developed for different user groups. For adolescents, strategies could include limiting their mobile and short video viewing time to prevent excessive use. For college students, educational programs that raise awareness about the risks and consequences of short video addiction could enhance self-awareness and self-control. For working professionals, promoting effective time management and encouraging alternative stress-relief activities, such as reading, exercising, or socializing, could help mitigate perceived stress and reduce reliance on short videos.

This study possesses certain limitations. Firstly, the quantitative methods employed in online surveys may not fully capture the comprehensive nature of users’ short video addiction reasons in China. Future studies could consider using interviews and grounded theory to obtain a more systematic and profound understanding of the underlying reasons for users’ addiction to short videos. Secondly, our sample predominantly comprised young individuals (aged 14–30), whereas China’s short video users represent a diverse demographic with varying addictive inclinations across different age groups. In the future, it would be advantageous to recruit a broader and more diverse range of participants from various age groups (i.e., Gen X; Gen Y; Gen Z). Finally, the present study adopted a cross-sectional survey design, which presents challenges in establishing causal relationships between variables. Future research efforts could use long-term longitudinal studies to investigate the impact of these variables over time.