A Meta-Analytical Review of the Determinants of Social Media Discontinuance Intentions

ABSTRACT

The use of social media has grown tremendously, but a considerable number of individuals have stopped using it. This meta-analysis aims to examine the factors that contribute to discontinuing social media use by reviewing 88 studies with a cumulative sample size of 42,159, including 33 effect sizes. Our study reveals that various stressors, including messaging overload (CO), social overload (SO), information collection overload (IO), system feature overload (SFO), privacy concerns (PC), and negative emotions such as technostress, fatigue (SNF), guilt, and dissatisfaction, are significantly correlated with social media discontinuance (DUIN). It is worth noting that only gratifications were negatively associated with both discontinuance and fatigue, but not with all other inhibitors. Furthermore, self-disclosure (S-disc), social comparison (SC), and fear of missing out (FoMo), as well as addiction, were significantly associated only with fatigue. Theoretical and practical implications are discussed, emphasizing the importance of social media operators balancing content supply with actual user demand to prevent overload, negative emotions, and discontinuance.

The worldwide Internet penetration rate has surpassed the critical mass phase (Miniwatts Marketing Group, 2021). Consequently, the focus of current interest has extended beyond access or adoption. Despite the proliferation of social media apps and users (Auxier & Anderson, 2021), many individuals have disengaged from social media for myriad reasons (Ellison, 2021). For example, Facebook usage has declined from 79% in 2017 to 57% in 2021 among 12–34 years old Americans according to Edison Research (2021). Similar trends have been observed across other social media platforms. Although the dynamics of technology adoption or diffusion have been widely studied in both primary and meta-analyses, they (e.g., the technology acceptance model (Davis et al., 1989)) say little of the post-adoption stage. Not only has discontinuance been understudied, but the limited number of studies have also been based on diverse configurations and dissimilar theories.

Studies on social media use have been abundant. Despite the voluminous account of the positive effects of social media use (e.g., Boulianne, 2015; Gil de Zúñiga et al., 2012), the serious consequences brought by the excessive use of social media or other forms of problematic use, particularly for users’ psychological well-being (Faelens et al., 2021), have been well documented. Some users are even advised to take corrective actions against social media use for their own good. In addition, some people who may be able to use social media properly but who have unpleasant experiences associated with social media have also discontinued their social media use. Therefore, for some people, the discontinuance of social media has been a necessary coping measure to avoid the dark side of social media for the good of their well-being (Farooq et al., 2023).

Due to the nascent nature of the field, this line of research is fraught with conceptual and operational problems; a plethora of questions remain unanswered. First, for some correlates in previous studies, there were mixed and even contradictory results, but other correlates dominated the variance in the discontinuance intention. What are the more important correlates of the outcome? Second, what are the effects and magnitude of the predictions? Lastly, are the predictions or the magnitude of the effect size consistent across methodological artifacts? These crucial questions cannot be answered by a single study and require a meta-analysis methodology to quantitatively determine the real effects and effect sizes based on previous research findings on this topic (Hunter & Schmidt, 2015).

Literature review and hypotheses

Understanding discontinuance and its determinants

Despite the extensive research on the topic, the conceptualization of discontinuance in social media remains ambiguous and lacks consensus among scholars (Farooq et al., 2023; Soliman & Rinta-Kahila, 2020). In this study, we have clearly defined discontinuance as a voluntary decision to either temporarily or permanently quit all previously used social media applications (e.g., Cao et al., 2021), excluding the decision to switch to alternative social media applications or other forms of nonuse, as defined by prior studies (Rogers, 2003; Soliman & Rinta-Kahila, 2020; Turel, 2014).

The exploration of discontinuance in the existing literature is multifaceted, drawing on diverse theoretical perspectives and attributing it to a variety of factors. As posited by the Diffusion of Innovation Theory, discontinuance is a deliberate decision made during the confirmation phase of the innovation-decision process (Rogers, 2003, pp. 164–165). This process extends beyond adoption, as depicted in the user transformation model proposed by Maier et al. (2015), which outlines a progression where exposure leads to adoption and then continued use continued use, which in turn proceeds to discontinued use. It’s important to note that the factors influencing discontinuance are distinct from those driving adoption, which primarily revolve around beliefs about the benefits, costs, and other attributes of a new technology (Davis et al., 1989; Rogers, 2003, pp. 164–165). These factors are not fully encapsulated by Rogers’ theory alone (Bhattacherjee, 2001), highlighting the need for a more comprehensive understanding of discontinuance.

Some scholars argue that discontinuance occurs when the characteristics of applications do not meet users’ expectations and needs, leading to dissatisfaction, which may result in switching or discontinuance (Bhattacherjee et al., 2012). The Uses and Gratifications Theory (UGT) (Blumler & Katz, 1974; see; Liu & Xiao, 2014) and the Technology Acceptance Model (TAM) (Davis et al., 1989; Feng et al., 2020) have been used to explain this reasoning (see Chaouali, 2016). However, questions remain. If the continuance or discontinuance of social media requires a voluntary decision, perceived usefulness and perceived ease of use, which are core variables in the TAM, do not significantly influence this process because people evaluate their previous adoption decision under the influence of cognitive and emotional factors, including individual, relational, and environmental aspects (Zheng & Ling, 2021). Furthermore, some scholars (e.g., Farooq et al., 2023) have contended that satisfaction is distinct from its opposite, dissatisfaction. Gratifications and satisfaction lead to usage when specific needs or motivations are fulfilled (Ku et al., 2013), while only dissatisfaction causes discontinuance.

Technostress-related frameworks, such as the Stress–Strain–Outcome (S–S–O) model (Koeske & Koeske, 1993), the Stimulus–Organism–Response (S–O–R) paradigm (Mehrabian & Russell, 1974), and the Transactional Theory of Stress and Coping (TTSC) (Lazarus & Folkman, 1984), offer a more comprehensive lens to examine discontinuance. These frameworks allow for a dual-factor perspective, distinguishing between the triggers and inhibitors of discontinuance, a perspective adopted by several researchers (Cao et al., 2021; Farooq et al., 2023; Turel, 2014, 2016). In an extension of this dual-factor perspective, Farooq et al. (2023) further divided these triggers and inhibitors into individual, relational, and platform-related factors. This division mirrors the categories proposed by Zheng and Ling (2021), which are individual, interpersonal/relational, and environmental factors.

Drawing upon these insightful categorizations, we have structured our analysis accordingly. Effect sizes have been grouped into individual and relational factors, while platform-related factors have been considered as potential moderators that may influence all effect sizes. Moreover, acknowledging the limitations associated with the theoretical frameworks discussed, we propose that the direct drivers of discontinuance are primarily individual negative emotions, predominantly those induced by stress-related factors.

Drivers of discontinuance

The drivers of discontinuance refer to factors that prompt users to discontinue the use of social media applications. The previously tested direct drivers have included technostress, dissatisfaction, guilt feelings, regret, and, most importantly, fatigue. Among these drivers, guilt and regret mainly result from addiction or disconfirmation (Liao et al., 2011). Both fatigue and dissatisfaction are primarily accounted for by techno-stressors (Maier et al., 2015). Consequently, negative emotional arousal mediates the effects of various stressful statuses of use (overload, addiction, etc.) on discontinuance intention.

Relational factors such as cyberbullying, social comparison, and peer pressure to discontinue are key drivers of social media discontinuance (Cao et al., 2019). However, these factors do not directly cause discontinuance. Instead, they foster negative emotional states such as distress and fatigue, which subsequently lead to the decision to discontinue social media use. Although the literature review does not delve into relational factors in detail, their impact will be examined in the ensuing meta-analysis.

While the relational factors are the primary drivers of negative emotional states, it is ultimately these negative emotional states that lead to discontinuance. In the following section, we review the major mediator, i.e., fatigue, followed by its predictors, i.e., techno-stressors.

Social media fatigue

The concept of social media fatigue refers to the negative feeling of exhaustion and burnout that arises from social media use (Ravindran et al., 2014). Fatigue has been found to be the principal factor directly leading to the discontinuance of social media. When individuals become fatigued with social media, they are less likely to continue using it. Consequently, the following hypothesis is proposed:

H1:

The greater level of social media fatigue, the greater the likelihood of discontinuing social media use.

Additionally, we propose the following research question:

RQ1:

What are the direction and magnitude of the relationships between negative emotional responses (most notably social media fatigue) and discontinuance?

Social media fatigue mainly results from the cognitive effort needed to continuously pay attention to social media (Ravindran et al., 2014). Prior studies have identified a range of stress-inducing causes that may contribute to social media fatigue (Kang et al., 2020).

Stressors

Prior studies have identified numerous stressors of fatigue, including privacy invasion (Ayyagari et al., 2011; Dhir et al., 2019), self-disclosure and parental supervision strategies (Dhir et al., 2019), and fear of missing out (FoMo) (Bright & Logan, 2018; Tugtekin et al., 2020), among others (Maier et al., 2015). However, the majority of studies have focused on overload as a primary stressor.

Milgram’s (1970) overload theory, originally formulated to explain urban-rural differences, posits that urban dwellers limit their social contacts to manage stimulus overload from dense populations. Despite predating social media, this theory provides key insights into the interplay between stimulus overload and social media use, serving as a foundational framework for such studies.

Previous research has focused on various types of overload and their effects on stress or social media fatigue. Social overload occurs when individuals perceive that they are providing too much social support to friends through social media (Maier et al., 2015), or when they experience a sense of inadequacy of time and attention to maintain relationships with the growing number of contacts on an online social network (Zhang et al., 2016). Many studies have confirmed the effect of social overload on fatigue (Maier et al., 2015; Zhang et al., 2020).

Information overload has been defined as the state in which an individual cannot accommodate and process excessive information received (Rogers, 2003, p. 315). Some studies have found that only information overload affects fatigue, frustration, and dissatisfaction among social media users (Niu et al., 2020). However, it has been found by others (Cao & Sun, 2018; Chuang & Liao, 2021) to contribute to fatigue together with other types of overload.

System feature overload occurs when the features afforded by social media technology are so complex and complicated that they outweigh what users can manage (Karr-Wisniewski & Lu, 2010; Lee et al., 2016). Empirical evidence supports its effect on fatigue (Fu et al., 2020; Lee et al., 2016). Previous studies have examined the effects of one or two types of overload, with some scholars (Cao & Sun, 2018; Lin et al., 2021) identifying three types of overload, i.e., information, communication, and social overload. Fewer studies have examined both social and information overload (Chaouali, 2016), while others have additionally examined system feature overload (Fu et al., 2020; Karr-Wisniewski & Lu, 2010; Lee et al., 2016; Zhang et al., 2016).

The term communication overload has been interpreted in various ways across the literature (Cho et al., 2011; Lee et al., 2016; Reinecke et al., 2016; Zhang et al., 2016). Given these diverse interpretations, we propose the term communication and information overload as a more encompassing descriptor. Furthermore, we suggest refining the terminology used in previous studies. Specifically, we propose renaming the variables used in prior studies that refer to communication overload as messaging overload and those referring to information overload as information collection overload, provided these terms more accurately reflect their intended meanings (Fu et al., 2020; Ji et al., 2014; Karr-Wisniewski & Lu, 2010; Lee et al., 2016; Schmitt et al., 2021).

According to prior studies, fatigue plays the role of a mediator (Guo et al., 2020; Niu et al., 2020; Zhang et al., 2020) between stressors and discontinuance. Stressors or indicators of technostress predict fatigue and may also directly predict discontinuance intention. Based on previous findings, we propose the following hypotheses:

H2:

Experiencing communication and information overload in the areas of a) messaging, b) social interaction, c) information collection, and d) system features, as a result of social media use, will lead to heightened levels of fatigue.

H3:

Experiencing communication and information overload in the areas of a) messaging, b) social interaction, c) information collection, and d) system features, as a result of social media use, will increase the propensity for discontinuing social media use.

We would also like to explore other correlations among variables based on the integrated theoretical framework. Therefore, we raise the following research question:

RQ2:

What are the magnitude and direction of the effects between social media experiences (including both drivers and inhibitors), negative emotions, and discontinuance?

Inhibitors of discontinuance

Individual factors

Inhibitors of social media discontinuance refer to factors that prevent users from ceasing the use of social media applications (Farooq et al., 2023). These inhibitors are largely based on studies investigating factors that contribute to technology adoption (Venkatesh et al., 2012). It is hypothesized that these factors, which contribute to continued use, might prevent users from discontinuing the use of a technology.

One of the most commonly studied inhibitors of discontinuance is satisfaction, although it has been overshadowed by its opposite, i.e., dissatisfaction, in prior studies. Satisfaction refers to the positive feelings that users experience when using a technology by comparing its perceived performances with their expectations (Kim, 2012). Studies have shown that satisfaction can prevent discontinuance (Chaouali, 2016; Turel, 2014, 2016).

Perceived self-efficacy, which refers to the user’s belief in their ability to use a technology effectively, is an important inhibitor of discontinuance that has been analyzed in prior studies. Studies have shown that users who have a high level of perceived self-efficacy are less likely to discontinue technology use (Adhikari & Panda, 2020; Vaghefi et al., 2020). Likewise, Ji et al. (2014) revealed that information searching efficiency was negatively associated with information overload. The two factors are related to the construct of digital literacy, which has not been explicitly tested in this line of research (Chetty et al., 2018).

Emotional attachment is also a vital inhibitor of discontinuance. Emotional attachment refers to the emotional bond that users form with a technology (Mamun et al., 2022). Users who are emotionally attached to a social media application are less likely to discontinue its use (Cao et al., 2021).

An additional factor that may hinder discontinuance is flow experience. Flow refers to the state of complete immersion and focus that users experience when using a social media application (Zha et al., 2018). Users who experience flow when using a technology are less likely to discontinue its use (Lin et al., 2021; for other inhibitors, see Cramer et al. (2016); Guan et al. (2022); Liu et al. (2018); Maier et al. (2015); and Turel (2014, 2016)).

Relational factors

While there are relatively few relational inhibitors, controlled motivation and social support are notable among them. For instance, Lo (2019) discovered that proper social support can help alleviate the exhaustion that users experience from excessive social media use and enhance their satisfaction with social media. This increased satisfaction, in turn, discourages users from discontinuing their use of social media. Similarly, Luqman et al. (2018) observed that controlled motivation has a negative impact on social norms, attitudes, and perceived behavioral control related to discontinuing social media use.

Proposed theoretical framework

Theoretical frameworks addressing discontinuance have been a source of confusion and conflicts. In order to address this issue, we have summarized the existing frameworks and integrated them into a more coherent framework, as presented in Figure 1. The proposed framework centers on discontinuance intention, along with the direct drivers of discontinuance, which include negative emotions or experiences, stressors or stress stimuli, and the perceptions of others. It also considers discontinuance inhibitors, which are also known as or coping resources according to Lazarus and Folkman (1984, pp. 157–164). Unpleasant experiences, stressors, and the perceptions of others are beliefs or cognitions concerning previous social media use. These factors predict negative emotional arousals such as technostress, fatigue, and regret, which in turn predict discontinuance (behavioral) intention. Moreover, we suggest that most inhibitors should moderate the mediation process.

Figure 1. The summarized theoretical framework.

Note. 1) The arrows in dashed lines represent potential moderation effects. Lines that originate from a large box containing multiple constructs indicate that these constructs share the same relationships with other variables. 2) The term “overload” refers to communication and information overload, which encompasses messaging overload, social overload, information collection overload, and system feature overload. “Perceptions of others” includes social norms, fear of missing out, social comparison, social surveillance, and self-disclosure. 3) “social norms” refers to norms for the continuance of an application. However, it could also encompass norms for the discontinuance of an application, which would then act as a driver of discontinuance (Turel, 2016).

While the integrated framework we propose is clear and parsimonious, we leave the task of testing its validity, explicability, and predictability to future scholars. This is due to the limited number of studies providing effect sizes for these predictions and the constraints on space, which preclude us from performing a viable model-based meta-analysis at present. However, this framework has guided our selection of key correlates of social media discontinuance for the current meta-analysis.

Effects of moderators

As noted above, both the significance and magnitude of the predictions vary from study to study, which suggests the influence of moderators.

Consequently, the present study tested eight major type of moderators: (a) participant characteristics¹ (including age,² gender (the proportion of female), income, and education level), (b) study design (experiments vs. surveys), (c) sample type (students vs. general adults), (d) geographical location (Confucian vs. non-Confucian regions³), (e) social media platforms⁴ (social networking sites vs. others), (f) social media medium-types (text vs. multimedia)⁵ (Anonymous reviewer, 2023), (g) publication type (conference papers vs. journal articles), and (h) measurement type (measurement scale vs. the single item). A research question is posed to address the effects of these moderators.

RQ3:

How do the above-mentioned moderators influence the relationships between social media discontinuance and fatigue and their correlates?

Method

Selection criteria

Because social media fatigue and discontinuance intentions are the focus of our study, we tried various combinations of the following keywords in five databases (i.e., Web of Science, PsycINFO, ScienceDirect, Scopus, and Google Scholar): “social media fatigue,” “social media burnout,” “social media exhaustion,” “social network* fatigue,” “social network* exhaustion,” “social network* burnout,” “social media discontinu*,” “social network* discontinu*,” “SNS fatigue,” “SNS burnout,” “SNS exhaustion,” “social media techn* stress,” “social media techn* strain,” “SNS techn* stress,” “SNS techn* strain,” “social network techn* stress,” and “social network techn* strain.”

The search process was conducted between February 2021 and October 2021. We implemented multiple rounds of search following the selection criteria and checklist described in the PRISMA statement (Moher et al., 2009). See Figure 2 in the online supplemental materials for more information.

The initial search yielded 1,703 studies, and we further searched through the reference lists of all obtained studies, contacted relevant authors for full texts that were unavailable in the databases, and acquired 49 additional eligible articles. Multiple filtering steps were then applied to the pool of studies based on the selection criteria. After applying these steps, a total of 88 eligible articles were identified, which provided us with a cumulative sample size of 42,159 (M = 468, SD = 319).

Inclusion criteria of variables

The majority of primary studies tested discontinuance intention rather than actual behavior as the ultimate dependent variable. Consequently, the ultimate outcome variable is intention rather than actual discontinuance behavior. Moreover, we selected variables from the integrated framework reviewed above on account of their theoretical significance, data availability, and quality.

We then inspected the underlying meanings of the variables in the included studies and allocated the variables to the three major classes: discontinuance intention, drivers of discontinuance, and inhibitors of discontinuance. Drivers included negative emotions or experiences (such as fatigue, dissatisfaction, and guilt), stressors (social, messaging, information collection, and system feature overload and privacy concerns), and the perceptions of others (social norms, fear of missing out, social comparison, social surveillance, and self-disclosure). Inhibitors include emotional attachment or involvement, perceived usefulness, flow experience, self-efficacy, and social support. This process was done on a close fit basis using a theory coding scheme⁶ (Michie & Prestwich, 2010).

Coding

The second and third authors independently coded the studies based on the codebook, found in the online supplemental materials, which was made by and extensively discussed among all the authors. After obtaining the coding results, the two coders randomly selected 30% of their counterpart’s data to code again to check the intercoder reliability of the major variables, estimated via the “irr” package of R 4.0. The results of the intercoder reliability estimation using Krippendorff’s $\mathit{\alpha }$ ranged between .89 and 1.00 (Feng, 2014, 2015; Feng & Zhao, 2016). See the online supplemental materials for all testing results. Partial discrepancies were resolved through discussion.

Analysis

Unit of analysis

The unit of analysis is Pearson’s correlation (r). Some effect sizes for which only standardized regression betas were reported were converted to r following Peterson and Brown (2005).

Analytical procedures

Effect sizes were weighted by sample size and corrected for reliability and range restriction using the method proposed by Hunter and Schmidt (2015). Heterogeneity was assessed using the barebones model, which estimates residual standard deviations (SD_res and SD_ρ). Further assessment of heterogeneity was conducted using the conventional Q-test (Higgins & Thompson, 2002). To accommodate heterogeneity across studies, a random-effects (RE) model was utilized (Schmid et al., 1991). To conserve space, we have relocated the detailed description and results of our publication bias estimation to the online appendix.

Results

The present study addressed both RQ1 and RQ2 by estimating the average effect sizes. The procedures used for estimation are illustrated in Figure 3 stored in the online supplemental materials. The forest plots in the online supplemental materials present the estimated effect sizes with discontinuance and fatigue as outcome variables. According to Cohen (2013), Pearson’s r values of .10, .30, and .50 indicate small, medium, and strong effects, respectively (also cf. Rosenthal, 1991).

Average effect sizes

Drivers

Individual correlates

The meta-analysis, which excluded the moderators, revealed significant associations between discontinuance and several key variables (for details, see Table 1). Strong correlations were identified for fatigue, dissatisfaction, and messaging overload. Conversely, medium effect sizes were found for information collection overload, social overload, privacy concerns, technostress, and guilt.

Table 1. Estimation results of the null model.

Construct Y	Construct X	k	N	$\bar{r}$	$\mathit{S}{D}_r$	$\mathit{S}{D}_{\mathit{res}}$	$\bar{\rho }$	$\mathit{S}{D}_{r_c}$	$\mathit{S}{D}_{\mathit{\rho }}$	95% CI
DUIN	SNF	42	16 609	.50	.18	.17	.57	.20	.20	[0.51, 0.64]
DUIN	Norms	3	1 140	.24	.20	.19	.28	.22	.21	[−0.27, 0.82]
DUIN	Flow	8	3 000	−.10	.36	.35	−.14	.42	.41	[−0.49, 0.21]
DUIN	Satis	8	4 451	−.53	.11	.10	−.60	.13	.12	[−0.71,−0.50]
DUIN	PU	8	3 575	−.15	.35	.34	−.18	.40	.40	[−0.52, 0.15]
DUIN	Gender	5	2 149	.07	.08	.06	.07	.08	.07	[−0.04, 0.17]
DUIN	Age	5	2 149	.03	.04	.00	.03	.04	.00	[−0.03, 0.08]
DUIN	Dissat	9	3 172	.59	.13	.12	.66	.16	.15	[0.53, 0.78]
DUIN	Guilt	10	4 375	.35	.18	.17	.39	.19	.19	[0.25, 0.53]
DUIN	SO	29	11 361	.27	.18	.18	.31	.21	.21	[0.23, 0.39]
DUIN	SFO	15	5 148	.26	.30	.29	.31	.35	.34	[0.12, 0.51]
DUIN	IO	27	11 018	.38	.12	.11	.44	.14	.13	[0.39, 0.50]
DUIN	Support	3	2 335	−.15	.12	.11	−.17	.14	.13	[−0.52, 0.18]
DUIN	Efficacy	9	5 477	.08	.20	.19	.09	.23	.22	[−0.08, 0.27]
DUIN	Stress	6	2 261	.38	.17	.17	.45	.21	.20	[0.23, 0.67]
DUIN	PC	12	4 779	.38	.18	.17	.46	.23	.23	[0.31, 0.61]
DUIN	S_disc	2	402	.15	.14	.12	.16	.15	.13	[−1.17, 1.49]
DUIN	FoMO	3	906	.03	.28	.27	.04	.33	.32	[−0.79, 0.86]
DUIN	Addiction	7	3 069	.09	.20	.19	.10	.22	.21	[−0.10, 0.30]
DUIN	CO	4	1 596	.46	.13	.13	.55	.19	.19	[0.24, 0.86]
DUIN	Monitor	3	1 263	−.09	.29	.29	−.09	.34	.33	[−0.92, 0.74]
DUIN	SOCC	2	1 404	.36	.31	.31	.44	.38	.38	[−3.01, 3.90]
DUIN	Involvement	3	1 024	−.05	.33	.33	−.05	.38	.38	[−1.00, 0.90]
SNF	Norms	2	653	.42	.08	.07	.47	.10	.09	[−0.44, 1.38]
SNF	Flow	8	2 846	.03	.26	.26	.03	.31	.30	[−0.23, 0.29]
SNF	Satis	8	4 446	−.32	.24	.23	−.34	.28	.28	[−0.58,−0.11]
SNF	PU	13	5 643	−.02	.24	.24	−.02	.28	.27	[−0.19, 0.15]
SNF	Gender	3	1 052	.01	.03	.00	.01	.03	.00	[−0.06, 0.09]
SNF	Age	5	3 694	.02	.05	.03	.02	.05	.03	[−0.04, 0.08]
SNF	Dissat	7	2 541	.50	.17	.16	.56	.18	.17	[0.39, 0.72]
SNF	Guilt	4	1 714	.47	.13	.12	.52	.16	.16	[0.26, 0.78]
SNF	SO	36	14 442	.37	.20	.19	.42	.23	.22	[0.34, 0.50]
SNF	SFO	23	7 774	.39	.29	.29	.47	.33	.33	[0.33, 0.62]
SNF	IO	41	15 524	.45	.16	.16	.53	.20	.20	[0.47, 0.60]
SNF	Support	5	3 249	.09	.19	.18	.10	.21	.21	[−0.16, 0.36]
SNF	Efficacy	9	5 514	−.01	.30	.30	−.00	.35	.35	[−0.27, 0.27]
SNF	Stress	11	6 650	.45	.20	.20	.54	.22	.22	[0.39, 0.69]
SNF	PC	21	10 255	.41	.17	.17	.48	.20	.19	[0.39, 0.57]
SNF	S_disc	6	3 658	.38	.18	.18	.45	.22	.22	[0.21, 0.68]
SNF	FoMO	9	6 362	.20	.16	.15	.24	.19	.19	[0.09, 0.38]
SNF	Addiction	11	6 114	.43	.21	.21	.51	.27	.26	[0.33, 0.69]
SNF	CO	15	6 058	.43	.27	.27	.49	.34	.33	[0.30, 0.68]
SNF	Monitor	5	2 008	.22	.24	.23	.26	.28	.27	[−0.09, 0.60]
SNF	SOCC	8	4 227	.45	.10	.09	.56	.12	.12	[0.45, 0.66]
SNF	Involvement	4	1 715	.16	.22	.22	.18	.25	.24	[−0.22, 0.57]

1) The rows highlighted in bold represent effect sizes that are statistically significant.

2) k = number of studies contributing to meta-analysis; N = total sample size; = mean observed correlation estimated using the Hunter and Schmidt method; = observed standard deviation of $\mathit{r}$; $\mathit{S}{D}_{\mathit{res}}$ = residual standard deviation of $\mathit{r}$; $\bar{\rho }$ = mean true-score correlation accounting for measurement error; $\mathit{S}{D}_{r_c}$ = observed standard deviation of corrected correlations ($r_c$); $\mathit{S}{D}_{\mathit{\rho }}$ = residual standard deviation of $\mathit{\rho }$; CI = confidence interval around $\bar{\rho }$; CR = credibility interval around $\bar{\rho }$. Correlations corrected individually.

3) It is important to note that effect sizes with fewer than five studies should be interpreted with caution due to concerns over statistical power, as highlighted by Jackson and Turner (2017).

4) IO, SO, CO, and SFO represents the overload on information collection, social, messaging, and system feature, respectively. In addition, PU, Dissat, Satis, Support, Efficacy, S_disc, FoMO, Monitor, DUIN, PC, SNF, and SOCC stands for perceived usefulness, dissatisfaction, social support, self-efficacy, self-disclosure, fear of missing out, self-monitor, discontinuance intention, privacy concerns, social media fatigue, social comparison, and use intensity, respectively.

Strong effects were observed for fatigue in relation to variables such as dissatisfaction, information collection overload, guilt, social comparison, addiction, and technostress. Medium effect sizes were also identified between fatigue and other variables, including messaging overload, privacy concerns, and social overload.

Relational correlates

The analysis showed an absence of significant correlations between discontinuance and various social perceptions. However, medium effects were noted for system feature overload and a specific type of social perception, i.e., self-disclosure.

Pertaining to relational correlates of fatigue, a small effect size was found in relation to fear of missing out, and no significant correlations were observed between fatigue and other types of social perceptions.

Inhibitors

The correlation between discontinuance and gratifications (ρ = −.60, k = 8, n = 4,451, 95% CI [−.71, −.50]) was a significant and strong effect, while there were medium effects between fatigue and gratifications (ρ = −.34, k = 8, n = 4,446, 95% CI [−.58, −.11]). Although the correlations between discontinuance and other inhibitors (self-efficacy, emotional attachment, flow experience, and perceived usefulness) including the relational inhibitor (proper social support) were negative as expected, they were not significant.

It is important to note that effect sizes with fewer than five studies, such as the correlations between discontinuance and message overload, and between fatigue and guilt, should be interpreted with caution due to concerns over statistical power, as highlighted by Jackson and Turner (2017).

Moderator analysis

The moderators comprise of categorical and continuous variables. Subgroup analysis was used to evaluate the impact of the former, whereas meta-regression was utilized to analyze the effect of the latter.

Subgroup analysis

The results of subgroup analysis are presented in Table 2 of the online supplemental materials. Each analysis tested a single moderator. However, two moderators—study design and measurement type—have highly skewed distributions between categories (89% vs. 11% and 98% vs. 2%, respectively), rendering the subgroup analysis inapplicable for them.

Origin of a study

Corresponding to RQ3, the moderating effects of the origin of a study (or culture) on all the effect sizes were significant for the correlations between fatigue and perceived usefulness (F (1,9.549) = 9.543, M_difference = −.320, 95% CI [−.547, −.092]), between discontinuance and satisfaction (F (1,5.752) = 15.521, M_difference = .171, 95% CI [.065, .277]), and between discontinuance and privacy concerns (F (1,9.580) = 5.758, M_difference = −.177, 95% CI [−.341, −.013]). Specifically, studies conducted in non-Confucian cultures had stronger effects between fatigue and perceived usefulness, even in the positive direction, discontinuance and gratifications, and discontinuance and privacy concerns.

Sample composition

The moderating effects of the sample composition on the estimated correlations between discontinuance and guilt (F (1,4.937) = 6.526, M_difference = −.232, 95% CI [−.441, −.023]) and between fatigue and systematic feature overload (F (1, 8.825) = 7.182, M_difference = .302, 95% CI [.068, .536]) were significant. The effects between discontinuance and guilt from general adult samples were higher than those from student samples, but studies recruiting student samples had stronger effects between fatigue and system feature overload.

Social media platforms

The moderating effects of social media platforms on the estimated correlations between discontinuance and satisfaction (F (1,5.752) = 15.521, M_difference = .171, 95% CI [.065, .277]) and between fatigue and self-efficacy (F (1,6.131) = 6.455, M_difference = .296, 95% CI [.021, .571]) were significant. That is, studies whose participants used SNSs had a stronger effect between discontinuance and satisfaction. In contrast, those whose participants used non-SNSs had a stronger effect between fatigue and self-efficacy than those using SNSs.

Social media medium-types

The moderating effects of social media medium types on the estimated correlations between discontinuance and messaging overload (F (1, 7.305) = 13.689, M_difference = −.366, 95% CI [−.580, −.152]) and between fatigue and self-efficacy (F (1,6.131) = 6.455, M_difference = −.296, 95% CI [−.571, −.020]) were significant. Studies that examined multimedia-based SNSs had stronger effects between discontinuance and messaging overload and between fatigue and self-efficacy than those examining text-based SNSs.

Publication Type

Lastly, the moderator of publication type had significant effects on the correlations between discontinuance and flow experience (F (1,5.599) = 14.350, M_difference = −.534, 95% CI [−.878, −.189]), between discontinuance and perceived usefulness (F (1,5.013) = 10.040, M_difference = −.433, 95% CI [−.768, −.099]), and between fatigue and system feature overload (F (1,19.046) = 5.810, M_difference = .210, 95% CI [.029, .391]). Specifically, compared to journal articles, conference papers had stronger effects between discontinuance and flow, between discontinuance and perceived usefulness, and between fatigue and system feature overload.

Meta-regression

Among the demographic variables, only gender had a significant effect on the effect sizes of the relationship between dissatisfaction and fatigue, $\mathit{\beta }=.023$ (95% CI [.004, .042]$,\mathit{\hspace{0.17em}}\mathit{t}\left(5\right)=3.086,\mathit{\hspace{0.17em}}\mathit{p}=.027)$ and between dissatisfaction and discontinuance, $\mathit{\beta }=.016,\left(95\%\mathrm{CI}\left[.003,\mathit{\hspace{0.17em}}.030\right],\mathit{\hspace{0.17em}}\mathit{t}\mathit{\hspace{0.17em}}\left(7\right)=2.846,\mathit{\hspace{0.17em}}\mathit{p}=.025\right)$. Women are more likely than men to experience fatigue and to discontinue the use of social media due to dissatisfaction.

Discussion

Average effects sizes

Our meta-analysis, which tested a series of individual effect sizes, revealed that only a handful of stressors including four types of overload and privacy concerns, and negative emotions including technostress, fatigue, guilt, and dissatisfaction significantly correlated with discontinuance intentions, thereby acting as drivers of discontinuance. Conversely, most inhibitors of discontinuance, excluding gratifications, along with perceptions of others and demographic variables, did not significantly correlate with discontinuance.

Neither discontinuance nor fatigue was associated with most inhibitors previously examined. While factors such as flow experience, social support, and perceived usefulness might promote social media use and extend its usage, they appeared to have little impact on inhibiting discontinuance. This could suggest that these inhibitors, particularly self-efficacy, may moderate the relationships between drivers and discontinuance (Lin et al., 2021; Zhang et al., 2020). Discontinuance seems primarily driven by negative emotional arousal, and the motivators of continuance, gratifications, are negatively associated with fatigue and the intention to discontinue social media use. These findings underscore the importance and dominance of stress-related theoretical frameworks and extend the applicability of UGT to the area of discontinuance.

Furthermore, most perceptions of others (self-disclosure, social comparison, and fear of missing out) were significantly associated only with fatigue, whereas two perceptions of others (social influence and social surveillance) did not significantly associate with either fatigue or discontinuance. This discrepancy suggests that relational factors should be further divided into intrinsically driven and extrinsically driven orientations. The former, causing strain or fatigue, indirectly affects discontinuance through fatigue, while the latter may precipitate social media use rather than discontinuance and strain (Marino et al., 2020). This classification is a novel contribution to the literature and warrants empirical examination in future research.

However, the dynamics of the associations between perceptions of others and fatigue may not be straightforward. More attention to others may prompt more social media usage, leading to communication and information overload, and subsequently fatigue. This chain of effects requires validation in future primary and model-based meta-analyses with sufficient empirical data, as the current results based on correlations cannot provide conclusive evidence. Finally, addiction was associated with fatigue rather than discontinuance, suggesting that the effect of addiction on discontinuance may be fully mediated through fatigue or other negative emotions (cf. Vaghefi & Qahri-Saremi, 2017; Vaghefi et al., 2020). Therefore, excessive use does not necessarily lead to discontinuance unless it induces negative feelings.

Effects of moderators

The findings of the meta-analysis were in general robust across moderators, but some effect sizes did vary by study-level moderators. For instance, the effects between fatigue and perceived usefulness (PU), between discontinuance and gratifications, and between discontinuance and privacy concerns were stronger in non-Confucian cultures than in Confucian cultures (cf. Ku et al., 2013). Although the inhibitors were not associated with fatigue and discontinuance, two of them, i.e., PU and gratifications, seem to stand out only in non-Confucian cultures, which suggests that the TAM and UGT are applicable to social media discontinuance only in these contexts. Compared to collectivistic cultures, individualistic cultures cherish personal freedom, hedonism, and rights (Joshanloo & Jarden, 2016). Therefore, any perceived threats to these values could provoke a stronger aversion to continued social media use.

Although student and nonstudent samples did not differ systematically in most effect sizes, the effect between discontinuance and guilt from nonstudent samples was higher than that from student samples. This finding may initially appear counterintuitive, given that guilt is often more prominent among students who share communal relationships (Baumeister et al., 1994). Students may feel guilty about excessive social media use due to its potential negative impact on academic performance and social relationships (Umberson & Montez, 2010). However, despite these feelings of guilt, students are unlikely to discontinue social media use because of its integral role in peer communication (Ellison et al., 2007; Valenzuela et al., 2009). Consequently, the variations in feelings of guilt and discontinuance intention among students are relatively low. This results in a smaller product of deviations between individual students and the average student for both variables, as per the formulae for covariance and correlation (Cohen et al., 2015). Thus, the observed higher effect size in non-student samples may be a reflection of these dynamics.

The correlation between system feature overload (SFO) and fatigue was stronger among student samples than among general adult samples. This result is probably because the ability to handle complex system features among students from different disciplines is very diverse (cf. Henderson et al., 2017). Therefore, the relatively high variation in the perception of SFO among students results in a higher correlation between SFO and fatigue.

The effect between fatigue and self-efficacy was higher for people using nonsocial networking sites (non-SNSs) than for those using social networking sites (SNSs). However, the negative correlation between discontinuance and gratifications was stronger for studies employing people using SNSs. The required ability to be proficient in using non-SNSs is more demanding than that in using SNSs, so the perceived self-efficacy in using non-SNSs would vary more than that in using SNSs. Consequently, the correlation between self-efficacy and fatigue is higher for people using non-SNSs. In addition, the needs of using SNSs are so diverse (Pai & Arnott, 2013; Park et al., 2009) that the gratifications resulting from SNS use and the resulting discontinuance intention are also more varied than those resulting from the use of non-SNSs.

Similarly, multimedia-based SNSs presented stronger effects between discontinuance and messaging overload and between fatigue and self-efficacy than those examining text-based SNSs. The higher levels of media richness and a more immersive experience offered by videos lead to higher engagement with content and relatively higher levels of sense of messaging overload [e.g., Danmaku or bullet chatting (Liu et al., 2016)]. In addition, video-based social media platforms are more likely to attract users who are seeking entertainment or escapism, rather than social connection and other needs. If users on video-based social media platforms are less motivated by social connection and more motivated by entertainment, they may be less likely to tolerate messaging overload and more likely to discontinue use when it becomes overwhelming (Al-Menayes, 2015).

As for the moderation effect on the correlation between fatigue and self-efficacy, text-based social media platforms, compared to video-based social media platforms, may provide more opportunities for information-seeking and social support, which can enhance self-efficacy and reduce feelings of fatigue (Fardouly et al., 2015).

For the three effects above, conference papers had larger effect sizes than journal articles. Moreover, conference papers seem to support the negative effects of flow experience and perceived usefulness, which were as nonsignificant as other inhibitors in journal articles. The findings regarding the stronger effects of conference papers are consistent with prior studies in this area (Feng et al., 2019). In general, conference papers undergo a less rigorous peer review process than journal articles. Hence, some of the stronger effects may result from some artifacts that are neglected by both the authors and reviewers.

Practical implications

Overload will cause users’ negative emotional responses and discontinuance based on the findings of this meta-analysis. Although the antecedents of overload are too diverse to be exhaustive (Ayyagari et al., 2011), what social media operators can do to make a difference is clear: strike a balance between the supply of content, both the number and complexity of features, and the actual demand level of specific users. Operators often use algorithms to provide tailored content, and they should also smartly use algorithms to retain users. Regarding the optimal level of demand, this might be another invaluable question. Operators may need to frequently conduct field surveys and even experiments to answer such an important question. Operators can try to disable some advanced features by default and limit the maximum number of friends, the number of subscriptions, and the number of postings received and posted for ordinary users. If a user exceeds a certain level in stress testing or risk profile assessment, (s)he will then be allowed to receive a higher level of content features. This practice will ensure a healthy and sustainable ecosystem, ultimately benefiting all stakeholders within the society. All these abovementioned findings and suggestions constitute the practical implications of the study.

Limitations and future research directions

This study inevitably has limitations. The influence of two moderators, i.e., the study design and measurement type, on effect sizes could not be empirically tested due to the seriously unequal distributions between the categories. Therefore, the role of the two moderators remains unknown. However, with the accumulation of sufficient primary studies, this limitation should be mitigated in the future.

This meta-analysis only included studies published in English. This limitation may have affected the generalizability of the results to non-English-speaking populations. Furthermore, the proposed theoretical framework in this paper was not examined with model-based meta-analysis due to space and other limitations. Model-based meta-analysis could have allowed for a more in-depth theory-based exploration of the relationships between variables, which may reveal additional insights that the conventional meta-analysis does not capture. Future studies could address the limitations by including non-English publications and more effect sizes with complete correlation-matrices involved in the theoretical framework.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/15205436.2023.2263872.

Additional information

Funding

This work was supported by the Hong Kong Baptist University [Start-up grant]; National Social Science Fund of China [18BXW082].

Notes on contributors

Guangchao Charles Feng

Guangchao Charles Feng, PhD is an associate professor at School of Communication, Hong Kong Baptist University. His main research areas are research methods, new media studies, computational advertising, message effects, and political communication.

Xianglin Su

Xianglin Su and Yiru He are both master’s students under the supervision of Dr. Feng and were primarily involved in the data collection and visualization aspects of this research.

Notes

¹ If the primary studies report demographics as effect sizes, they will be tested as such. However, if the primary studies report demographics as sample characteristics rather than effect sizes, they will be tested as moderators.

² To address the inconsistency in the number of points on the scales, the means reported were adjusted by subtracting the medians of the corresponding scales. For instance, if a study employed a 7-point scale and reported an average income level of 3, the adjusted mean would be −1 (calculated as 3 minus 4).

³ Since most studies were conducted in China, Korea, Japan, and Taiwan (52%), we classified these regions as Confucian culture. Zhang et al. (2005), whereas the rest of the regions were termed non-Confucian cultures.

⁴ Since most studies studied social networking sites (SNS) (70%), we classified social media platforms as social network sites vs. other forms of social media. For a definition of social media and social network sites, see Boyd and Ellison (2007), Feng et al. (2019), and Liu et al. (2018)

⁵ Chu and Kim (2011).

⁶ Some variables that are conceptually similar were unified into the same name. For instance, guilt, shame, and regret are combined according to Kugler and Jones (1992).