We conducted a two-wave study (an interval of 1 week, during which the control measures remained the same) using an online questionnaire during the epidemic. To match participants’ data of the two measuring points of the study, participants were instructed to enter several pieces of personal information consisting of (1) real name, (2) phone number, (3) social media accounts (QQ and WeChat), and (4) Alipay accounts (channels for payment of partition fees). Informed consent was obtained from all participants, and all respondents were informed that all their data would be used for scientific purposes only and will be kept confidential. They could terminate participation at any time during the study. The initial sample included 1,189 participants, with 859 participants completing the surveys in both waves. Participants who only completed the survey at Time 1 did not significantly differ from the longitudinal participants in terms of demographic and study variables. At Time 1, 71% of the 859 participants (45% male) were in the low-risk area,¹ and at Time 2, 78% of participants (46% male) were in the low-risk area. The study received ethical approval from the host university to launch the study.

Confirmatory factor analysis (CFA) was conducted in Mplus 8.3 (Muthén and Muthén, 2017) to test the measurement models for the present study. All the results in two waves revealed a good model fit for all scales (χ²/df < 5.5, CFIs>0.9, RMSEAs<0.08, SRMRs<0.08), indicating that all scales had sound measurement properties.

We computed a cross-lagged structural model using Mplus 8.3. The default estimation method of maximum likelihood was used. To study the adequacy of the estimated model, we used χ²/df, the RMSEA, the CFI, the TLI, and the SRMR. We employed Mclust, a popular R package for model-based clustering, classification, and density estimation based on finite Gaussian mixture modeling (Scrucca et al., 2016), to develop profiles of emotion regulation strategies based on the data collected at T1 and T2. This method does not require the researcher to make assumptions about the shape and size of the clusters that are difficult to justify. In the model-based clustering approach, each component of a finite mixture density is usually associated with a group or cluster, and in the model, it is assumed that the mixture components share the same orientation matrix. The number of mixing components and the covariance parameterization are selected using the Bayesian Information Criterion (Scrucca et al., 2016).

The data collected in this study were controlled for common method bias through anonymous collection and inverse item scoring. We used Harman’s single-factor test to check for common method bias. and the result showed that the first factor only explained 22.16% at T1 and 26.6% at T2 of the variance, below 40%, which means that common method bias did not threaten the interpretation of our results.

Descriptive statistics of the variables are shown in Table 1. We employed paired sample T-test analysis of variance and found that positive emotion and life satisfaction were improved in T2, and negative emotion and loneliness were decreased in T2. Additionally, there were also changes in emotion regulation strategies except cognitive reappraisal, enhancing positive emotion, and soothing. Specifically, the use of experiential avoidance and retweeting willingness were decreased in T2, and the use of rest strategies was increased in T2. These results suggest that participants’ mental health improved over time, and emotion regulation was a dynamically changing process.

Results of correlative analysis (see Supplementary Table S1) showed that all strategies except experiential avoidance (AAQ) positively related to positive emotion and life satisfaction, and negatively related to negative emotion and loneliness. Conversely, AAQ exhibited a negative correlation with positive emotion and life satisfaction and a positive correlation with negative emotion and loneliness. In addition, the correlation between different strategies was also significant.

In cross-lag analysis, we controlled for the effects of gender, age, and risk level at two measure points. The cross-lagged model fitted the data well, with χ²/df = 4.79, p < 0.001, TLI = 0.911, CFI = 0.972, RMSEA = 0.065 (95%CI = [0.059, 0.072]), SRMR = 0.069. Standardized path coefficients are presented in Table 2 (Auto-regressive paths are significant for all variables, ps < 0.001. For brevity, auto-regressive paths, control variables paths, and insignificant coefficients are not presented in the table).

The use of cognitive reappraisal at T1 was positively related to positive emotion at T2. The use of experiential avoidance was reciprocally positively related to negative emotions over time, perspective-taking was reciprocally positively related to life satisfaction over time, and the saving behavior of humorous memes was reciprocally positively related to positive emotions over time. The experiential avoidance at T1 was also positively related to loneliness at T2, and negatively related to positive emotion and life satisfaction at T2. The saving behavior of humorous memes at T1 was also negatively related to loneliness at T2. These results may indicate that the use of experiential avoidance was a risk factor for mental health while the rest three strategies played protective roles in mental health.

Positive emotions at T1 were positively related to the usage of experiential avoidance, perspective-taking, soothing, the willingness to retweet, and the saving behavior of humorous memes at T2. Life satisfaction at T1 was positively related to the use of enhancing positive emotions, social modeling, and the willingness to retweet at T2. The loneliness at T1 was negatively related to the use of all the interpersonal emotion regulation strategies except perspective-taking. These results suggested that certain psychological states promote the different use of emotion regulation strategies.

Control variables only affected a few emotion regulation strategies scores and clustering analysis is more powerful with larger sample sizes, so we elected to regress out these effects rather than conduct the analysis separately by age, gender, and risk level as other researchers do (Xie et al., 2022). The results indicated the presence of nine groups displaying different profiles of emotion regulation strategies (BIC = −30418.75; ICL = −30559.66). To enhance profile interpretability, we computed the standardized deviation from the means for each strategy within each profile (Figure 1). The bar chart of the average scores can be found in Supplementary Figure S1.

In profile 1 (22.5%, n = 201), all emotion regulation strategies except experiential avoidance deviated negatively (implying they were used below the mean level), so we labeled this profile “low user.” In profile 2, humorous meme saving deviated negatively, and the remaining strategies approached the mean (11.4%, n = 102), so we labeled it as “low hyper.” Profile 3 (13%, n = 116), 6 (6.1%, n = 55), and 8 (18.4%, n = 165) had similar patterns but to different degrees, in which all strategies except experiential avoidance deviated positively, so we labeled them as “multi-user.” In profile 4 (2.3%, n = 21), intra-personal strategies deviated positively, and the remaining strategies except social modeling, which approached the mean, deviated negatively, so we labeled it “independent.” In profile 5 (17.4%, n = 156), none of the ER strategies deviated noticeably from the mean; therefore, this profile was named “average user.” In profile 7 (5.9%, n = 53), the hyper-personal strategies deviated positively from the mean and the remaining approached the mean, so we labeled it as “high hyper.” In profile 9 (2.9%, n = 26), the interpersonal strategies deviated negatively and the remaining approached the mean, so we labeled it as “low inter.”

We then compared emotion and mental health scores based on the cluster analysis results, and there were significant main effects of the group at T1 on the emotion and mental health scores at two measure points (ps < 0.001). Post-hoc comparisons revealed that at T1 positive emotions were highest in the “multi-user-high” group, negative emotions were lowest in the “multi-use-high” groups, and the highest in the “low inter” group. Life satisfaction was the highest in the “multi-user-high” group, and loneliness was the lowest in the “multi-user-high” group. There were similar patterns at T2. We also computed the standardized deviation from the means for each mental health indicator within each profile at T1 (Figure 2). The bar chart of the average scores can be found in Supplementary Figure S2. It is not hard to notice that the “multi-user” groups displayed better psychological states, with positive emotions and life satisfaction deviated positively and negative emotions and loneliness deviated negatively, and the “low user” group and the “low hyper” groups displayed the opposite outcomes from the “multi-user” groups. The “independent” group was characterized by negatively deviated positive mental health indicators (positive emotions and life satisfaction) and positively deviated loneliness. These findings indicated that only individuals who use multi-level strategies can maintain mental health when facing critical public events.

We further compared the differences in the use of emotion regulation strategies between the “multi-user” groups. Although the three “multi-user” groups had similar patterns, the mental health in the “multi-user-high” group was significantly better than the others. Compared to the “multi-user-low” group and the “multi-user-medium” group, the use of experiential avoidance was lower (ps < 0.001), and the use of soothing and humorous meme saving was higher (ps < 0.001) in the “multi-user high” group. These results were in line with the results in cross-lag analysis, which indicated that the use of experiential avoidance was a risk factor for mental health while humorous-meme-saving played a protective role in mental health.

Cross-lagged analysis on the intrapersonal level strategies indicated that cognitive reappraisal at T1 was positively related to positive emotion at T2, implying that cognitive reappraisal is an effective affect-improving strategy, as found in previous studies (Aldao et al., 2010). Experiential avoidance was reciprocally positively related to negative emotions, confirming that experiential avoidance is a risk factor for mental health (Akbari et al., 2022). This reciprocal relationship raises the question of why people continue to use this strategy after experiencing negative feelings following its use. It may be explained by individuals’ motives for regulating emotions and affective forecasting errors, the ability to predict future emotions (Pilin, 2020). Avoidance of negative emotions was both consistent with the hedonic motive and utility motive (Tamir et al., 2007), but individuals may have a false expectation of its effect, which leads them to use experiential avoidance again, resulting in increased negative emotions (Akbari et al., 2022).

Regarding interpersonal-level strategies, results showed that perspective-taking was reciprocally positively related to life satisfaction. Research has demonstrated that individuals high in IER tendency are more socially connected (Williams et al., 2018), so perspective-taking may contribute to the development of supportive relationships, which benefit life satisfaction. Furthermore, perspective-taking refers to using others to remind oneself not to worry and that others have it worse, which involves talking to others and the modification of maladaptive appraisals and beliefs, may result in increasing their life satisfaction.

The study also found that humorous-meme-saving behavior was reciprocally positively related to positive emotions, suggesting that it is a protective factor for mental health. This finding is consistent with research on humor strategy, which indicates that humor strategy can effectively upregulate positive emotions and has a good intervention effect in depressed people and elderly people (Braniecka et al., 2022). Meanwhile, humor, also serves as a frequent and highly valued element of online communication, thus flourishing on social media (Laineste and Voolaid, 2017). This means although regulating emotions with digital technologies is a ‘double-edged sword’ (Erfani and Abedin, 2018; Smith et al., 2022), it remains a valuable channel for emotion regulation when it is related to positive content, as we found in our study.

Cluster analysis suggested that there were 9 profiles with different usage preferences, which vary in terms of mental health indicators. Compared the mental indicators of individuals who primarily use intrapersonal strategies versus those who use a combination of intrapersonal and hyper-personal strategies, we found that the former exhibited high levels of loneliness and low levels of positive emotion and life satisfaction while the latter featured high negative emotions, high life satisfaction and average loneliness. This may indicate that intrapersonal strategies were effectively down-regulating negative emotions (Lepore et al., 2004) while associated with high loneliness, and hyper-personal strategies had a counter effect on loneliness since they helped to strengthen social connection (Hamilton et al., 2023) while having the risk of negative contagion (Prikhidko et al., 2020). It should be noted that people who balanced use of strategies at all three levels tended to be associated with greater positive emotions and life satisfaction while with lower negative emotions and loneliness, with the “multi-user-high” group significantly better than the others. This finding aligns with previous studies, indicating that frequent use of multiple strategies represents higher flexibility of emotion regulation, which is more adaptive for well-being (Grommisch et al., 2020).

Based on the result mentioned above, this may be attributed to the fact that different levels of emotion strategies serve as different resources and compensate for each other, allowing individuals to better cope with critical social events (Troy et al., 2013). For example, interpersonal emotion regulation is closely associated with the formation of intimate relationships that serve as relational resources to help people better cope with stressors (Schoebi and Randall, 2015), at the same time digital technologies provide more strategy choices that can be deployed with less effort (Smith et al., 2022). That is, the balanced use of strategies at all three levels expands solutions of emotion regulation, and improves the overall effectiveness of emotion regulation.

This study offers new insights into the relationship between emotion regulation and mental health. We examined the use of emotion regulation strategies from a multilevel perspective in the context of a real-life critical public event and re-examined the effectiveness of these strategies. These findings enrich the existing research on emotion regulation in mental health. Moreover, these findings provide empirical evidence of how individuals can maintain mental health when confronted with critical public events. The current study has identified the profile of emotion regulation strategies that can help individuals cope effectively with major social events, encouraging people to actively use strategies at all three levels.

We must acknowledge several limitations in this research. Firstly, the participants were only from China, and the dynamic zero-COVID policy in China stopped the spread of the virus efficiently, so more participants lived in low-risk areas which may potentially influence the outcomes of the study. Secondly, it is important to acknowledge the limitations related to the cross-sectional design of the study. This design hinders the ability to definitively establish cause-effect relationships, and the complexity of the model made it challenging to enforce strict invariance. Therefore, readers should interpret the findings with caution, and it would be ideal for future research to launch longitude studies with more measurement points and observe the changes. Thirdly, only self-report measures were used in this research, and the strategies for hyper-personal emotion regulation are not yet standardized in emotion research. Future studies should incorporate more objective measures. Finally, this study focused on a relatively longer time episode, but emotion regulation is a dynamic process, future research should opt for new methods such as ESM approaches to reflect the temporal aspects of emotion regulation, or employ new technologies such as virtual reality to explore emotion regulation in a higher ecological validity by simulating real-life settings (Colombo et al., 2019).