Understanding the buffering effect of social media use on anxiety during the COVID-19 pandemic lockdown

The COVID-19 pandemic created an unprecedented global health crisis and an existential risk, marked by uncertainty, fear of contamination and death, and widespread lockdowns. These conditions amplified information overload and anxiety while social media usage surged to historic levels. The study asks whether and how social media use buffers (reduces) anxiety by fostering resilience during the pandemic. Drawing on the Looming Vulnerability Model (LVM), the authors posit that anxiety arises from overestimation of the proximity and growth of threats. They hypothesize that social media may function as a placeholder for collective resilience processes, modulated by cognitive (knowledge, appraisal) and emotional components, thereby shaping a positive perception toward stressors and mitigating fear. The work situates social media within stress-buffering mechanisms and virtual collective consciousness to understand how online interactions could help individuals cope with a dynamically evolving, uncertain threat like COVID-19.

Prior research shows quarantine measures can trigger psychological distress, including panic, anxiety, depression, PTSD symptoms, confusion, and anger (Qiu et al., 2020; Brooks et al., 2020). During the pandemic, social media and internet use increased markedly (Effenberger et al., 2020; Fischer, 2020), with mixed roles in disseminating information, expressing fear, and building support. The LVM suggests anxiety stems from perceived imminence and dynamism of threats (Riskind & Rector, 2018). Cognitive psychology highlights gaps between perceived and actual knowledge (Radecki & Jaccard, 1995), likely widened by social media in crises. Social support literature (Cohen & Wills, 1985; Wills & Filer, 2001) proposes stress-buffering via self-esteem, informational, belonging, and instrumental support. Work on Virtual Collective Consciousness (VCC) demonstrates how social platforms can create consensus and mobilize collective responses (Marzouki et al., 2012; Marzouki & Oullier, 2012, 2015; Alperstein, 2019). Uncertainty reduction theory (Berger, 1986) and crisis communication research (Burke et al., 2010) indicate information seeking reduces uncertainty. Emerging evidence suggests media exposure increases perceived (but not necessarily actual) knowledge about COVID-19 (Granderath et al., 2020).

Design and sampling: Cross-sectional online survey conducted during global lockdown using snowball sampling. Participants were anonymous and completed an online questionnaire distributed via social media. Data collection spanned March–May 2020, defining nine weekly cross-sections (week 4 dropped due to n=4). Sample: Total N=1408 responses from 3/20/2020 to 5/21/2020. Weekly n: W1=477, W2=62, W3=141, W4=4 (dropped), W5=88, W6=369, W7=72, W8=117, W9=78. Verbatim keywords provided by 1279 participants (9.16% missing verbatim); language distribution: Arabic 70.21%, English 22.36%, French 7.43%. Measures and SEM: The structural equation model (SEM) tested the impact of perceived knowledge on fear (anxiety). A latent variable Perceived knowledge was indicated by three observed exogenous variables: Social media use for social knowledge about the pandemic, Self-perceived knowledge (subjective knowledge), and Threat perception (death saliency and estimated probability of death). The endogenous variable was Anxiety (fear). SEM estimated weekly models (except weeks 4 and 7 due to limited responses). Data were complete (forced responses). Estimation used maximum likelihood in R (lavaan) with 50 bootstrap draws; variables treated as continuous. Model fit indices reported per week. Text corpus analysis: Participants listed up to five keywords reflecting their condition. The corpus (N=862 participants with usable text; bi-grams with min frequency ≥4) underwent: - Unsupervised clustering: Constructed context-by-lexical matrix, applied TF-IDF normalization and unit-length scaling, then bisecting K-means clustering (seed selection via iterative K-means bisection) to derive thematic clusters. For each partition: built lexical-by-cluster contingency tables, chi-square testing, and correspondence analysis to characterize clusters. - Co-occurrence analysis: Generated bi-gram co-occurrence matrix (N=165 frequent bi-grams) to visualize thematic clustering. - Discriminant analysis: Per week, performed discriminant function analysis to test whether fear, social media use, self-knowledge, and threat predicted membership in thematic clusters assigned to each participant. - Sentiment and psycholinguistic analysis: Applied LIWC to quantify categories (valence, social ties, risk, cognitive processing, space, time, death) across weeks to detect shifts in emotional vs cognitive language. - Entropy analysis: Computed normalized social entropy (0–1) over weeks using cluster membership probabilities for words (Balch, 2000; Shannon & Weaver, 1975) to assess lexical diversity and consensus dynamics.

SEM results: Across weeks, models showed good fit: CFI > 0.98 in most weeks (range ~0.980–0.993), RMSEA < 0.08 (range ~0.023–0.051), non-significant chi-square with df=2 and p-values > 0.44, indicating adequate model fit. The latent Perceived knowledge negatively predicted Fear each week (standardized path coefficients approximately −0.40 to −0.84), indicating higher perceived knowledge associates with lower anxiety. Indicator weights varied by week, revealing a trade-off between Self-perceived knowledge and Threat perception, conditioned by Social media use’s contribution to the latent factor. Example indicator loadings: Week 1—Social media use .58, Self-knowledge .23, Threat perception .59; Knowledge→Fear −.84. Week 6—SM use .64, Self-knowledge .17, Threat .41; Knowledge→Fear −.79. Week 9—SM use .53, Self-knowledge −.20, Threat .65; Knowledge→Fear −.81. Text clustering: Four thematic clusters emerged: Global knowledge, Anxiety, Symptoms, and Coping. Over time, the relative presence of 'Anxiety' and 'Coping' clusters showed opposing trajectories, with a notable shift around week 5; 'Coping' increased as 'Anxiety' decreased. Discriminant analysis: In weeks 6–8, fear, social media usage, self-knowledge, and threat significantly discriminated among the four thematic clusters, indicating distinct combinations of these variables mapped onto different emergent themes (see Supplementary Table S1). Sentiment/psycholinguistics: Words related to risk were most frequent initially, followed by social ties. A clear turning point occurred at week 5, marked by a peak in cognitive processing words and a decrease in negatively valenced words. Positive emotion words increased over time, coinciding with reduced risk-related terms, consistent with a shift from emotional to cognitive appraisal. Entropy: Normalized entropy averaged ~0.85 across the first 5 weeks, rising to ~0.94 across the last 4 weeks, trending toward an asymptote. Early lower entropy reflected more coherent, consensus expressions centered on negative emotion; later higher entropy reflected increased lexical diversity aligned with cognitive reappraisal and expanded coping discourse. Overall: Findings triangulate that social media use plays a dynamic, granular role in fostering a buffering effect: it conditions the balance between self-knowledge and perceived threat within perceived knowledge, reduces fear, and supports a shift from emotional reactions to cognitively oriented coping as lockdown progressed.

The results support the hypothesis that social media use can buffer anxiety during a large-scale, dynamic threat by facilitating collective resilience processes. Perceived knowledge, partly shaped by social media, consistently reduced fear across weeks. The trade-off between self-perceived knowledge and threat perception, conditioned by social media use, indicates that online interactions help recalibrate appraisals of risk. The emergence of a Virtual Collective Consciousness (VCC) is suggested by early consensus around negatively valenced narratives (lower entropy) and subsequent transition to cognitively framed coping (higher entropy), mirroring the stress-buffering model where informational, esteem, belonging, and instrumental supports mitigate threat appraisals and enable coping strategies. Consistent with uncertainty reduction theory, heightened information seeking via social media reduced uncertainty and reorganized collective interpretations of the pandemic. The observed turning point around weeks 4–5 marks a shift from an emotionally dominated narrative to cognitive appraisal and coping, aligning with the decline in negative emotion words and rise in cognitive processing and positive valence. These dynamics extend prior findings that media exposure increases perceived (but not necessarily actual) knowledge by demonstrating how social media specifically contributes to resilience and anxiety reduction during lockdown.

This study provides an empirically grounded framework explaining how social media use buffers anxiety during COVID-19 lockdowns by shaping perceived knowledge and enabling collective resilience. Triangulating SEM, text clustering, sentiment analysis, and entropy metrics shows that social media conditions the balance between self-knowledge and perceived threat, reduces fear, and supports a shift from emotional responses to cognitive coping around mid-study. Practical recommendations for future lockdowns include: (1) reduce uncertainty by promoting trustworthy information sources; (2) tailor online content to emotional cycles to accelerate collective resilience; (3) reinforce social norms that promote protective behaviors; (4) structure and guide online socialization (e.g., scheduled social activities) to enhance buffering effects. Future research should examine mediating and interacting effects of social media with other variables (e.g., platform differences such as Facebook, YouTube, Twitter), adopt comparative and cross-cultural analyses, and incorporate socio-demographic moderators to refine models of online collective coping and resilience.

Key limitations include the non-probabilistic (snowball) sample, precluding precise incidence estimates and limiting generalizability. The study focused on structural relationships derived from the Looming Vulnerability Model and did not collect socio-demographic variables (e.g., education, gender, digital literacy, socioeconomic status) that could serve as important covariates in SEM analyses. Week-level sample fluctuations led to excluding week 4 (n=4) and not modeling week 7 in SEM due to limited responses.