Twitter (X) use predicts substantial changes in well-being, polarization, sense of belonging, and outrage

The study investigates whether and how everyday use of Twitter (now X) impacts users’ psychological states and social attitudes, focusing on subjective well-being, political affective polarization, moral outrage, and sense of belonging. The authors highlight that despite Twitter’s outsized societal influence due to its concentration of elites in entertainment, journalism, and politics, there have been few systematic attempts to assess its psychological impact using ecologically valid methods. Prior social media research has yielded mixed conclusions regarding effects on well-being, with some arguing effects are negligible or explained by confounds such as personality, and many studies relying on retrospective self-reports or public trace data that poorly represent typical users. The authors aim to overcome these limitations by using intensive experience sampling to distinguish within-person from between-person processes and by decomposing Twitter use into what users do (behaviors/affordances) and why they do it (functions/motivations), while also testing whether individual differences (e.g., personality, age, gender) moderate any associations.

The paper situates its inquiry within debates on social media’s effects on well-being and political outcomes. Prior work suggests small or contested links between social media use and well-being, with some arguing null or trivially small associations and concerns about confounding (e.g., personality) and measurement validity of self-reported use. Twitter-focused research often leverages public data (tweets/followings) to study polarization and outrage but such data are disproportionately produced by a small, more politically engaged minority, raising generalizability concerns. On polarization, experimental studies show mixed results: U.S. users who deactivated Facebook became less polarized, while replication in Balkan contexts found the opposite, implying contextual moderators. Outrage appears algorithmically amplified online; moral-emotional language increases diffusion (e.g., each moral-emotional word in political tweets boosts retweets by ~20%), but it is unclear whether such expressions reflect genuine emotions versus performative signaling. Social media can also bolster social capital and community belonging, though recent work has focused less on these potential benefits. A nuanced framework distinguishing structural features (affordances like scrolling, liking, messaging) from functional motivations (entertainment, escapism, social interaction, self-promotion, information seeking) is proposed to clarify heterogeneous effects and moderators, alongside possible roles for user characteristics such as Big Five traits.

Design: A one-week experience sampling method (ESM) study surveying participants five times daily between 9:00 and 22:00 at random times (≥2-hour intervals) to capture recent (past 30 minutes) Twitter use and current psychological states. Sampling and participants: Quota sampling on Prolific to approximate representativeness of U.S. Twitter users by age, gender, and race. Inclusion required using Twitter at least twice weekly. Of 3,058 pre-screened, 404 completed a baseline survey; 309 entered ESM; final analytic sample N=252 after excluding insufficient responses (<9) or failed attention checks. Demographics: Mean age 42.99 (SD=14.06), 51% male; 72% White, 14% Black, 7.5% Asian, 4% mixed race, 2.5% other. Political ideology: 67.5% liberal, 13.5% conservative, 16.3% moderate, 2.8% missing. Data collection: March–June 2021. Ethics: University of Toronto REB; informed consent obtained. Procedure and measures: Baseline survey included Big Five personality. ESM prompts delivered via SurveySignal/Qualtrics; response window 40 minutes. At each prompt, participants reported whether they used Twitter in the prior 30 minutes (yes/no). If yes, they reported: (a) behaviors (what they did): scrolling feed, liking, retweeting, tweeting, commenting, messaging, viewing trending topics, viewing others’ profiles (binary, multiple selections); and whether they interacted with someone holding different values/worldview (yes/no). (b) functions (why they used Twitter), assessed via a new Functions of Social Media Use scale (developed/piloted; five factors): entertainment, escapism, social interaction, self-promotion, information seeking (binary, multiple selections). Outcomes: momentary well-being (difference between mean positive emotions: positive, joyful, and mean negative emotions: negative, sad, afraid, angry) based on a modified SPANE; also single items for boredom, loneliness, anxiety. Outrage operationalized as sum of angry, disgusted, repulsed (α=0.80). Sense of belonging measured with two items adapted from sport belongingness scale (5-point Likert). Affective polarization measured as absolute difference between warmth toward Democrats vs Republicans (0–100 feeling thermometers). Positive control benchmark: social interaction in past 30 minutes (in-person vs no) and its quality. Compensation: $5 baseline + variable up to $25 for high completion and Twitter account verification. Response burden: mean 24.67 of 35 surveys (SD=7.15); median 26; 11 participants completed all 35 surveys; total 6,218 surveys. Statistical analysis: Bayesian multilevel models (brms in R 4.0.3) with random intercepts and random slopes. Outcomes z-scored for comparability. Predictors decomposed into between-person means and within-person deviations to distinguish levels of effect. Models controlled for lagged outcome (Y_{t−1}) and lagged predictor (X_{t−1}) to account for prior state and prior exposure, estimating associations between Twitter use in past 30 minutes and changes in outcomes. Four MCMC chains, 10,000 iterations each (5,000 warmup), total 20,000 post-warmup draws; default flat priors; report posterior means and two-sided 95% credible intervals (CrI). Practical significance gauged by comparison with the benchmark effect of in-person social interaction on well-being. Preregistration: initial protocol (osf.io/en92q) but analysis deviated substantially due to data-driven shifts and adoption of robust Bayesian approaches; hence treated as non-preregistered. Data/materials/code: osf.io/e8krz.

Description of use: Participants reported Twitter use in 26.7% of ESM surveys (SD=21%). Use was mostly passive: active behaviors (tweeting/retweeting) occurred in 18.2% of Twitter-use surveys (SD=29.4%). Most common behavior was scrolling (reported in 74% of Twitter-use surveys); least common was messaging (7%). Functions: entertainment 66%, information seeking 49%, social interaction 23%, escapism 18%, self-promotion 2%. Encountering differing views occurred 14% of the time (SD=24%). Well-being and emotions: Within-person, Twitter use predicted a −0.10 SD change in well-being (b_within = −0.10, 95% CrI [−0.15, −0.04]); the benchmark in-person social interaction predicted +0.15 SD well-being (b_within = 0.15, 95% CrI [0.02, 0.28]). Twitter use related to increased boredom: within-person +0.22 SD (95% CrI [0.14, 0.30]) and between-person +0.50 SD (95% CrI [0.18, 0.83]). Loneliness increased at the between-person level only: +0.42 SD (95% CrI [0.05, 0.78]). Anxiety showed no reliable association (within b=0.02, 95% CrI [−0.03, 0.08]; between b=0.16, 95% CrI [−0.18, 0.50]). Sense of belonging: Within-person, Twitter use predicted +0.11 SD in sense of belonging (95% CrI [0.04, 0.17]). Polarization: Within-person, Twitter use predicted a small +0.03 SD increase (95% CrI [0.00, 0.06]). Encountering differing views was not reliably associated with polarization (CrIs included zero at both levels). Outrage: Within-person, Twitter use predicted +0.19 SD in outrage (95% CrI [0.10, 0.28]), exceeding the benchmark magnitude. Encountering differing views was not reliably related to outrage (CrIs included zero). What (behaviors): Scrolling predicted lower well-being within-person (−0.08 SD, 95% CrI [−0.15, −0.01]). Replying to others’ tweets (+0.23 SD, 95% CrI [0.06, 0.40]), visiting trending topics (+0.15 SD, 95% CrI [0.05, 0.25]), and viewing others’ profiles (+0.15 SD, 95% CrI [0.01, 0.28]) predicted higher sense of belonging within-person. Retweeting predicted higher polarization at the between-person level (+1.15 SD, 95% CrI [0.21, 2.15]). No specific behaviors reliably predicted changes in outrage. Why (functions): Escapism predicted decreased well-being within-person (−0.25 SD, 95% CrI [−0.39, −0.12]) and between-person (−1.43 SD, 95% CrI [−2.78, −0.25]). Entertainment predicted higher polarization within-person (+0.04 SD, 95% CrI [0.01, 0.08]). Social interaction function predicted higher sense of belonging between-person (+0.70 SD, 95% CrI [0.11, 1.31]). Outrage increased with escapism at the between-person level (positive association; CrI reported as >0) and with information seeking within-person (+0.14 SD, 95% CrI [0.03, 0.25]). Who (moderators): Despite substantial heterogeneity in individual slopes, no evidence that Big Five traits, age, or gender moderated effects; model fit comparisons supported null interactions. Overall: Many effects were primarily within-person (well-being, belonging, polarization, outrage), with boredom and loneliness showing stronger between-person associations. Effect sizes ranged from about one-fifth to larger than a benchmark positive effect of in-person social interaction on well-being.

The findings indicate that everyday Twitter use is associated with immediate changes in users’ psychological states: reduced momentary well-being and increased sense of belonging, outrage, and (slightly) affective polarization within the ensuing 30 minutes. By separating within- from between-person processes, the study clarifies that several associations (well-being, belonging, outrage, polarization) reflect short-term intra-individual dynamics rather than stable differences between heavier and lighter users. The analysis of what people do and why they do it helps explain these effects: passive consumption (e.g., scrolling) and using Twitter to cope via escapism relate to diminished well-being, consistent with theories of negative social comparison and the costs of avoidant coping. Information seeking is linked to heightened outrage, consistent with greater exposure to moral-violation content online and genuine experienced moral emotions rather than purely performative expressions. Socially oriented behaviors (replying, viewing profiles, checking trending topics) and the social interaction motive correspond to stronger sense of belonging, suggesting that interactive, community-focused uses can yield psychosocial benefits. The polarization findings are nuanced: the immediate within-person effect is small, and behavioral markers like frequent retweeting track greater between-person polarization, aligning with the notion that a small, highly engaged minority generates most political content. Together, these results demonstrate practically meaningful, heterogeneous effects that depend on usage patterns and momentary context rather than on fixed user traits.

Using intensive experience sampling and multilevel Bayesian modeling, the study shows that Twitter use is associated with both costs and benefits: short-term decreases in well-being alongside increases in sense of belonging, outrage, and modest increases in affective polarization. Disaggregating use by behaviors and functions reveals that passive consumption and escapism relate to poorer well-being, information seeking relates to greater outrage, while socially interactive uses foster a greater sense of belonging. Individual differences in personality, age, and gender did not explain effect heterogeneity. These contributions underscore the value of parsing within- versus between-person dynamics and of classifying social media use by affordances and motivations to understand heterogeneous outcomes. Future work should test causal mechanisms experimentally, examine longer-term effects and broader platforms, and identify contextual or systemic moderators (e.g., network structures, algorithmic exposure) to inform healthier engagement with social media.

Causal inference is limited: despite lagged controls, unmeasured time-varying confounders and reverse causality cannot be ruled out. The ESM window captures short-term dynamics over hours and a week; effects may differ over months or years. The sample consists of active Twitter users, so those most negatively affected may have already left the platform, potentially underestimating harms. Although more representative than typical convenience samples, the sample is not strictly representative of all Twitter users or the U.S. population. Power to detect interaction moderation may be limited, and some null findings (e.g., encountering differing views) may reflect low power. Cross-platform generalizability was not tested, limiting comparative conclusions across social media ecosystems.