X平台上气候变化运动中AI推断的表达幸福感与集体行动话语

# AI-inferred expressed well-being and collective-action discourse
Source: [https://arxiv.org/html/2609.22096](https://arxiv.org/html/2609.22096)
Climate campaigns are often evaluated through attention and mobilization, but less is known about the well\-being language that accompanies them\. Whether campaign periods alter positive affect and hope, and whether happiness aligns with action language, remains unresolved\. We analysed 364,118 public Twitter/X posts from Earth Day, Earth Hour, Global Climate Action Day and World Environment Day in 19 occurrence\-years, using 30\-day pre\-event, event and post\-event windows\. A versioned weighted lexical model estimated happiness, future\-oriented hope, collective capability, distress and action language\. Event\-period happiness prevalence was 9\.02 percentage points higher than the pre\-event baseline \(95% CI 1\.32–16\.73; nominalp=0\.0218p=0\.0218; BHq=0\.1516q=0\.1516\), whereas paired occurrence contrasts showed a 10\.75\-point decline in action language, indicating a happiness–action divergence\. The happiness estimate remained positive across composition and text\-deduplication checks, but was less precise under a 19\-cluster wild bootstrap \(p=0\.0565p=0\.0565\)\. Happier source posts had lower odds of an observed matched retweet cascade \(OR 0\.457, 95% CI 0\.233–0\.896\);

###### keywords

climate communication, expressed well\-being, happiness, hope, collective efficacy, Twitter, retweet cascades, computational social science

## 1Introduction

Climate change is a biophysical threat, a public\-health problem and a collective communication problem\. Recent assessments describe effects on physical health, displacement, livelihoods and mental health, while emphasizing that vulnerability is socially patterned and that responses depend on institutions and collective capacity[romanello2023countdown](https://arxiv.org/html/2609.22096#bib.bib1);[lawrance2022mental](https://arxiv.org/html/2609.22096#bib.bib2);[ostrom2009framework](https://arxiv.org/html/2609.22096#bib.bib3)\. Public communication is part of this response system\. It transmits information about temperature, emissions and policy, but it also signals whether a risk is shared, whether a desirable future remains imaginable and whether people can act together\. These cues matter because climate concern is widespread while sustained attention and action are uneven\.

Emotion research treats emotions as appraisal processes that organize attention, meaning and action readiness\. Several emotions can occur in the same episode or utterance[scherer2005emotions](https://arxiv.org/html/2609.22096#bib.bib4);[lazarus1991adaptation](https://arxiv.org/html/2609.22096#bib.bib5);[gross1998regulation](https://arxiv.org/html/2609.22096#bib.bib6)\. Positive emotions can broaden thought and support social resources, while positive affect can coexist with danger, dissatisfaction or inaction[fredrickson2001positive](https://arxiv.org/html/2609.22096#bib.bib7)\. Climate communication research has moved beyond a simple fear\-versus\-hope opposition\. Concern, grief, anger, moral outrage, solidarity, hope and efficacy can coexist, and people with different identities or experiences can read the same message differently[chapman2017reassessing](https://arxiv.org/html/2609.22096#bib.bib8);[markowitz2014psychology](https://arxiv.org/html/2609.22096#bib.bib9);[brosch2021affect](https://arxiv.org/html/2609.22096#bib.bib10)\.

This distinction matters for work on well\-being\. Climate anxiety is associated with persistent worry, impairment and uncertainty about institutional responses; climate\-related worry also encompasses experiences beyond a mental\-health disorder[clayton2020anxiety](https://arxiv.org/html/2609.22096#bib.bib11);[pihkala2020anxiety](https://arxiv.org/html/2609.22096#bib.bib12);[hickman2021anxiety](https://arxiv.org/html/2609.22096#bib.bib13)\. Population studies also show that climate concern can coexist with hope, meaning and engagement\. Among young people, coping strategies that combine hope with problem\-focused engagement are related to environmental action and well\-being, whereas guilt, denial or helplessness can undermine both[ojala2012coping](https://arxiv.org/html/2609.22096#bib.bib14);[ojala2012hope](https://arxiv.org/html/2609.22096#bib.bib15);[ojala2015hope](https://arxiv.org/html/2609.22096#bib.bib16)\. Hope is multidimensional\. Snyder’s hope theory separates agency, the will to pursue a goal, from pathways, the routes perceived to reach it[snyder2002hope](https://arxiv.org/html/2609.22096#bib.bib17)\. In climate contexts, constructive hope and passive optimism can lead to different forms of language and action[cohenchen2019hope](https://arxiv.org/html/2609.22096#bib.bib18);[marlon2019mobilization](https://arxiv.org/html/2609.22096#bib.bib19);[mortreux2025hope](https://arxiv.org/html/2609.22096#bib.bib20)\.

Evidence about climate\-message framing remains mixed\. Images and text that communicate attainable solutions can increase hope and policy support, but progress\-focused messages can also weaken motivation when they make the problem appear solved or when they fail to name responsibility and agency[feldman2018hope](https://arxiv.org/html/2609.22096#bib.bib21);[hornsey2016hope](https://arxiv.org/html/2609.22096#bib.bib22);[morris2020endings](https://arxiv.org/html/2609.22096#bib.bib23)\. Recent preregistered and large\-scale message studies likewise report heterogeneous effects across emotional frames[lammers2026communicating](https://arxiv.org/html/2609.22096#bib.bib24);[voelkel2026megastudy](https://arxiv.org/html/2609.22096#bib.bib25)\. The implication is methodological as well as theoretical: a study should measure positive affect, future\-oriented hope and action language separately, and should test whether their temporal patterns converge or diverge\. A single sentiment score can obscure precisely the distinction that climate\-communication theory seeks to explain\.

Collective efficacy provides a bridge between emotion and action\. Social identity models of collective action predict that people act when they perceive a shared grievance, identify with a group and believe that coordinated effort can produce change[vanzomeren2008integrative](https://arxiv.org/html/2609.22096#bib.bib26);[vanZomeren2012perspective](https://arxiv.org/html/2609.22096#bib.bib27)\. The social identity model of pro\-environmental action further proposes that group norms, efficacy beliefs, environmental appraisal and perceived costs jointly shape action[fritsche2018simpea](https://arxiv.org/html/2609.22096#bib.bib28)\. Empirical studies link collective efficacy to pro\-environmental intentions through both self\-efficacy and group identification[bamberg2015collective](https://arxiv.org/html/2609.22096#bib.bib29);[jugert2016efficacy](https://arxiv.org/html/2609.22096#bib.bib30);[reese2019social](https://arxiv.org/html/2609.22096#bib.bib31)\. The theory of planned behaviour and value\-belief\-norm approaches add behavioural intentions, perceived control and moral obligation to this account[ajzen1991planned](https://arxiv.org/html/2609.22096#bib.bib32);[stern2000environmental](https://arxiv.org/html/2609.22096#bib.bib33);[fielding2008planned](https://arxiv.org/html/2609.22096#bib.bib34)\. In text, collective efficacy concerns shared ability, coordination, institutional leverage and a community’s capacity to produce a climate outcome\.

Well\-being and action may follow different trajectories\. Positive language can support participation, mark solidarity, express a rhetorical position or accompany low\-cost symbolic involvement\. Anger or distress can motivate action when paired with efficacy and a target, but can also produce withdrawal when paired with futility[gifford2011dragons](https://arxiv.org/html/2609.22096#bib.bib35);[cohenchen2019hope](https://arxiv.org/html/2609.22096#bib.bib18);[fritsche2018simpea](https://arxiv.org/html/2609.22096#bib.bib28)\. Surveys and experiments have measured these mechanisms in controlled settings\. They are needed for psychological validity, but they usually expose participants to researcher\-selected messages and sample smaller populations\.

Social media provide a second source of evidence\. Twitter/X posts are public, time\-stamped and connected to retweet and reply metadata, so researchers can study the timing and visibility of campaign discourse\. Computational social science has shown that digital traces can reveal aggregate temporal patterns that are difficult to observe with conventional surveys[lazer2009computational](https://arxiv.org/html/2609.22096#bib.bib36);[gentzkow2019text](https://arxiv.org/html/2609.22096#bib.bib37);[grimmer2013promise](https://arxiv.org/html/2609.22096#bib.bib38)\. Hedonometer\-style dictionaries estimate word pleasantness and have been used to track expressed happiness in songs, blogs and Twitter streams[dodds2010happiness](https://arxiv.org/html/2609.22096#bib.bib39);[dodds2011twitter](https://arxiv.org/html/2609.22096#bib.bib40)\. LIWC supplies psychologically motivated categories such as positive emotion, negative emotion, social processes and agency[tausczik2010liwc](https://arxiv.org/html/2609.22096#bib.bib41);[pennebaker2007liwc](https://arxiv.org/html/2609.22096#bib.bib42)\. These tools produce reproducible language measures\.

Social\-media aggregates have several limitations\. Word meanings depend on context, and automated lexicons can reproduce cultural and demographic biases[caliskan2017semantics](https://arxiv.org/html/2609.22096#bib.bib43);[hutto2014vader](https://arxiv.org/html/2609.22096#bib.bib44)\. User activity is unequal, platform audiences are selected, and a burst of posts may be driven by a small group of highly active accounts\. Seasonal rhythms and news cycles can create apparent mood changes unrelated to a campaign[golder2011diurnal](https://arxiv.org/html/2609.22096#bib.bib45)\. Keyword rules can overrepresent organized communities and miss indirect or image\-based communication\. Big\-data studies also fail when a digital proxy is treated as the construct itself, as the Google Flu case illustrates[lazer2014paradigm](https://arxiv.org/html/2609.22096#bib.bib46)\. We address these limits by freezing the corpus before outcome analysis, fixing the campaign rules, measuring several constructs, weighting and deduplicating authors in sensitivity analyses and reporting uncertainty with 19 occurrence clusters\.

The network analysis has a separate limitation\. Retweets create observable source–child links, but a retweet usually reproduces the source text, so source and child receive the same lexical score\. A retweet edge records exposure or cascade selection; recipient emotional uptake remains unobserved\. Experiments have shown that social\-network exposure can alter emotional expression under particular platform conditions[kramer2014contagion](https://arxiv.org/html/2609.22096#bib.bib47), while observational studies link emotion and moral language to diffusion[brady2017emotion](https://arxiv.org/html/2609.22096#bib.bib48)\. Exposure is also shaped by algorithmic ranking, homophily and ideological selection[bakshy2015exposure](https://arxiv.org/html/2609.22096#bib.bib49);[bail2018exposure](https://arxiv.org/html/2609.22096#bib.bib50);[vosoughi2018false](https://arxiv.org/html/2609.22096#bib.bib51)\. We therefore model whether a matched cascade is observed and, conditional on observation, its size and timing\.

We focus on recurring climate campaigns\. Earth Day, Earth Hour, Global Climate Action Day and World Environment Day provide repeated temporal anchors with different institutional histories, participation styles and levels of public visibility\. The repeated\-event design allows within\-occurrence contrasts while preserving differences across campaign families and years\.

The temporal design also sets the limits of the comparison\. A pre\-event window provides a local reference for the same occurrence, the event window covers the period when campaign cues are most salient, and the post\-event window tests persistence after the focal date\. The design remains vulnerable to coincident news, seasonal routines and changes in account composition\. We treat the pre/event/post contrast as an observational comparison\. A sign that recurs across occurrence\-years suggests a repeated discourse pattern; a sign found in one campaign is event\-specific\.

The design distinguishes three levels of inference\. At the post level, a score describes one message\. At the occurrence\-day level, a prevalence describes the composition of messages on a day\. At the occurrence\-year level, a paired contrast describes change relative to that occurrence’s local baseline\. These measures describe messages and their temporal composition, leaving individual psychological states unobserved\. A campaign can change who posts, which organizations are visible or which genres are shared while readers’ well\-being remains unchanged\. We report the unit of inference with each estimate to avoid reading an ecological association as an individual effect\.

These considerations motivate three connected research questions\.RQ1\.Do recurring climate campaigns coincide with changes in expressed happiness, hope, distress and collective\-capability language across the pre\-event, event and post\-event periods?RQ2\.If happiness language changes, does it move in parallel with explicit climate\-action language, or do campaign periods show a measurable happiness–action decoupling?RQ3\.Finally, how is happiness\-labelled climate discourse represented in observed retweet cascades across the same periods?

We estimate changes in expressed well\-being language in a versioned corpus and measure how those signals move with action language\. External lexical controls and a reproducibility audit assess measurement stability\. Paired occurrence contrasts, small\-cluster resampling and linkage checks assess inferential sensitivity\. The main theoretical possibility is a happiness–action difference: campaign discourse may contain more positive or future\-oriented language without a parallel rise in explicit action language\.

## 2Methods

### 2\.1Data Collection

The analysis uses a versioned local archive of public Twitter posts collected by the project team\. The source archive contains approximately 170 million posts retrieved from November 1, 2010 to November 30, 2022 with the broad climate\-query terms “climate change”, “global warming”, “climate crisis”, “climate issue”, “carbon neutrality” and “low\-carbon”\. From this source archive, the formal analytic table contains 364,118 posts from four pre\-specified climate campaign families: Earth Day, Earth Hour, Global Climate Action Day and World Environment Day\. Campaign membership is assigned by an outcome\-blind keyword/hashtag scanner defined before the well\-being outcomes were inspected\. The scanner relies exclusively on exact and normalized campaign names and their documented hashtags\. Posts are retained with a stable post identifier, author identifier, timestamp, text, language, campaign family, occurrence\-year and period label\. Retweet, reply and quote identifiers are retained for the separate network audit\. Supporting Information, Table S1, provides the frozen data design and analysis units\.

Each occurrence has a 30\-day pre\-event window, an event window defined by the campaign’s operational date range, and a 30\-day post\-event window\. The event boundaries are fixed in the frozen manifest and are not optimized to maximize a result\. The primary temporal unit is an occurrence\-day, producing 1,008 daily observations; paired change analyses use the 19 occurrence\-years as the independent occurrence units\. This repeated\-event design allows the same campaign family to be observed across years while preserving differences in scale and institutional context\. Occurrence\-year is the unit of inference\.

The large input files were scanned in a columnar, chunked workflow\. Only the fields needed for campaign assignment, text scoring, temporal aggregation and linkage were read into memory\. Exact duplicate identifiers were removed before aggregation, and text\-template deduplication is used only as a pre\-specified robustness check\. The frozen manifest records the input file hashes, row counts, column selections, campaign rules and scoring version, so all main tables can be regenerated without loading the complete archive into memory\.

### 2\.2Constructs and text\-level measurement

The primary construct is*expressed happiness*: whether a post contains a weighted lexical signal associated with pleasant affect, positive appraisal or social enjoyment\. Secondary constructs are hope\-waypower, collective capability, distress and action language\. Hope\-waypower represents future\-oriented improvement coupled with a route, effort or feasible change\. Collective capability represents language about shared ability, coordination, institutions or a community’s capacity to produce a climate outcome\. Distress captures threat, worry, loss and helplessness language\. Action captures explicit behavioural, political or policy verbs\. These are text\-level language measures; they provide no diagnosis, trait measure or direct reading of subjective well\-being\.

For postiiand constructkk, the frozen scorer computes a weighted lexical scoresi​k=∑w∈ini​w​vw​ks\_\{ik\}=\\sum\_\{w\\in i\}n\_\{iw\}v\_\{wk\}, whereni​wn\_\{iw\}is the token count andvw​kv\_\{wk\}is the versioned token weight\. The primary outcome is a binary signalIi​k=1​\(si​k\>0\)I\_\{ik\}=1\(s\_\{ik\}\>0\), and the continuous score is retained for sensitivity analyses\. Token contributions and dictionary versions are stored with each score, allowing a post\-level result to be audited without re\-estimating the weights\. Scores are aggregated within occurrence\-day as prevalence,Yd​k=Nd−1​∑iIi​kY\_\{dk\}=N\_\{d\}^\{\-1\}\\sum\_\{i\}I\_\{ik\}, whereNdN\_\{d\}is the number of retained posts on daydd\. Supporting Information, Table S2, gives the operational definitions and interpretation boundaries for all five constructs\.

The lexical model was constructed from psychologically motivated seed dictionaries and corpus\-derived weights, with action terms kept separate from the outcome construct\. The happiness score excludes action terms, and the action score excludes happiness terms\. LIWC\-2007 positive\-emotion prevalence and labMT pleasantness are external lexical controls\. Their correlations with the primary score provide convergent language evidence; shared text and vocabulary limitations preclude independent criterion validation\.

To assess scoring reproducibility, two independent model coders apply the frozen codebook to a random set of 451 English posts while ignoring the existing scores\. Agreement is summarized with nominal Krippendorff’s alpha and Cohen’s kappa\. This audit measures the consistency of codebook application and offers no human validation of the text scores as measures of a user’s psychological state\.

### 2\.3Primary temporal model and estimands

The primary estimand is the event\-period change in happiness prevalence relative to the pre\-event period\. We fit a linear\-probability occurrence\-day model:

Hd=\\displaystyle H\_\{d\}=\{\}β0\+β1​\(Eventd×Campaignd\)\+β2​Postd\\displaystyle\\beta\_\{0\}\+\\beta\_\{1\}\(\\mathrm\{Event\}\_\{d\}\\times\\mathrm\{Campaign\}\_\{d\}\)\+\\beta\_\{2\}\\mathrm\{Post\}\_\{d\}\(1\)\+γ​DayOfWeekd\+δ​log⁡\(1\+Nd\)\+η​Yeard\+εd,\\displaystyle\+\\gamma\\mathrm\{DayOfWeek\}\_\{d\}\+\\delta\\log\(1\+N\_\{d\}\)\+\\eta\\mathrm\{Year\}\_\{d\}\+\\varepsilon\_\{d\},
whereHdH\_\{d\}is daily happiness prevalence,NdN\_\{d\}is the daily post count, and campaign\-family indicators allow the period contrast to vary across families\. Standard errors are clustered by occurrence\-year\. We report percentage\-point estimates, 95% confidence intervals, nominalppvalues and Benjamini–Hochberg adjusted values for the pre\-specified model family[benjamini1995false](https://arxiv.org/html/2609.22096#bib.bib52)\. A logistic model for post\-level presence and a beta\-binomial sensitivity for daily counts are included in Supporting Information\.

The event\-study specification adds relative day, event\-step and pre\-event trend terms\. It characterizes temporal shape without invoking a sharp causal discontinuity\. The occurrence\-level paired analysis computes pre\-to\-event and pre\-to\-post changes for each of the 19 occurrence\-years\. Because the number of clusters is small, we use conventional cluster\-robust standard errors, a Rademacher wild\-cluster bootstrap and exact sign\-flip tests\. The sign\-flip tests permute the sign of each occurrence contrast and preserve the repeated\-event structure\. The sensitivity family excludes retweets, restricts to English non\-retweets, weights authors equally within day, trims authors above the frozen 99th\-percentile activity cutoff, and removes exact or normalized duplicate text\.

### 2\.4Direct happiness–action decoupling

For occurrenceooand contrastc∈\{event−pre,post−pre\}c\\in\\\{\\mathrm\{event\\\!\\\!\-\\\!\\\!pre\},\\mathrm\{post\\\!\\\!\-\\\!\\\!pre\}\\\}, the direct decoupling index is

Do,c=\(Ho,c−Ho,pre\)−\(Ao,c−Ao,pre\),D\_\{o,c\}=\\left\(H\_\{o,c\}\-H\_\{o,\\mathrm\{pre\}\}\\right\)\-\\left\(A\_\{o,c\}\-A\_\{o,\\mathrm\{pre\}\}\\right\),\(2\)
whereAAis action\-language prevalence\. A positive value means that happiness language increased relative to action language\. We report the distribution ofDDacross occurrence\-years, the number of positive contrasts, exact sign\-flipppvalues and Wilcoxon signed\-rank sensitivities with BH adjustment\. Supporting Information, Table S3, reports the occurrence\-level paired contrasts\. This descriptive divergence estimand addresses neither mediation nor causation, and it leaves symbolic participation and changes in campaign\-talk composition indistinguishable\.

### 2\.5Observed retweet cascades

The network analysis uses retained retweet and quote identifiers to match child posts to in\-corpus source posts\. The linkage audit first reports unresolved identifiers and identifiers affected by scientific\-notation precision loss\. Only edges with a resolvable source and a common campaign occurrence are used for the formal analysis\. The primary network estimand is the association between source\-post happiness and the observation of a matched cascade\. A hurdle model separates \(i\) the probability that any matched retweet is observed and \(ii\) the conditional count of matched retweets, with occurrence\-clustered uncertainty\. A source\-level model of log time to first matched retweet serves as a timing sensitivity, with right censoring left uncorrected\.

The data\-generating process sets the interpretation: a retweet copies the source text, so source and child happiness scores are identical on matched edges\. The network model traces exposure and cascade selection\. An emotional\-contagion claim would require an independent recipient outcome, an exposure denominator, reliable source–child linkage and a design that addresses homophily, ranking and confounding[kramer2014contagion](https://arxiv.org/html/2609.22096#bib.bib47);[brady2017emotion](https://arxiv.org/html/2609.22096#bib.bib48);[bakshy2015exposure](https://arxiv.org/html/2609.22096#bib.bib49);[vosoughi2018false](https://arxiv.org/html/2609.22096#bib.bib51)\.

### 2\.6Missingness, multiplicity and quality controls

Missing text, malformed timestamps and rows without a stable identifier are removed before scoring and are counted in the frozen data\-quality table\. Missing text and construct scores remain unimputed\. Daily prevalence is calculated from the posts observed in that day; therefore, a low\-volume day contributes a noisier estimate and adjacent dates supply no replacement values\. Volume enters the primary model as a covariate, and author\-equal and text\-deduplicated variants test whether high\-volume accounts or repeated templates drive the result\. Campaign rules, event boundaries and the 99th\-percentile activity cutoff are stored before the final outcome table is generated\.

The confirmatory family contains the event\-period happiness estimate, the pre\-specified event\-study step and the two direct decoupling contrasts\. Secondary constructs, campaign\-specific estimates, network hurdles and M3 temporal\-precedence scans are labelled exploratory or Supporting Information analyses\. Benjamini–Hochberg correction is applied within each declared family\. We report effect sizes and intervals even when adjusted or small\-clusterppvalues fail conventional significance thresholds\. This prevents lexical and account\-group comparisons from being treated as independent discoveries\.

### 2\.7Exploratory account\-group analysis and reproducibility

M3 text\-only predictions[wang2019demographic](https://arxiv.org/html/2609.22096#bib.bib53)are retained for an exploratory Supporting Information analysis of aggregate account groups\. Age and gender labels come from the model; the archive contains no self\-reports or verified demographics\. Sparse categories are collapsed or excluded under pre\-specified rules\. Granger\-style tests use first\-differenced daily happiness series and a three\-day lag to test temporal precedence, with BH correction within the testing family[granger1969investigating](https://arxiv.org/html/2609.22096#bib.bib54)\. Supporting Information, Fig\. S1, reports the complete account\-group temporal\-precedence scan\. These tests provide no causal evidence and sit outside the primary happiness claim\.

All reported numbers are linked to machine\-readable CSV/JSON tables, a frozen manifest and a dependency\-free audit\. Refresh scripts resolve the analysis root from the file location or the environment variableANALYSIS\_ROOT; the formal input is supplied throughFORMAL\_PARQUET\. Raw post text and user\-level exports fall outside the redistributable release\. The code release contains the scoring, aggregation, model and figure scripts, while reproduction of the full text\-level analysis requires an authorized local archive\. The reproducibility code, derived intermediate data and publication figures are archived at Figshare:[https://doi\.org/10\.6084/m9\.figshare\.33207624](https://doi.org/10.6084/m9.figshare.33207624)\.

## 3Results

### 3\.1RQ1: campaign\-period changes in expressed well\-being

Figure[1](https://arxiv.org/html/2609.22096#S3.F1)shows the campaign\-period pattern across the four families\. In the primary linear\-probability model, the event\-period coefficient is\+9\.02\+9\.02percentage points \(95% CI1\.321\.32–16\.7316\.73; nominalp=0\.0218p=0\.0218\)\. The corresponding BH value for the model family isq=0\.1516q=0\.1516, and we report the finding as an uncertain positive adjusted association\. The post\-period coefficient is\+0\.69\+0\.69percentage points \(95% CI−6\.82\-6\.82–8\.198\.19;p=0\.858p=0\.858\)\. The segmented event\-study finds no pooled linear pre\-event trend \(−0\.019\-0\.019percentage points per relative day;p=0\.848p=0\.848\) and an event step of\+5\.41\+5\.41percentage points \(p=0\.056p=0\.056\)\. Supporting Information, Table S4, provides the complete temporal estimates and small\-cluster checks\.

At the paired occurrence\-year level, happiness prevalence increases by 4\.94 percentage points from pre\-event to event and by 3\.39 percentage points from pre\-event to post\-event\. The construct\-change summary in Figure[2](https://arxiv.org/html/2609.22096#S3.F2)shows that hope\-waypower and collective\-capability language also move upward descriptively, while distress remains close to zero in the aggregate contrasts\. Campaign\-specific estimates are heterogeneous, precluding any single campaign from serving as a universal proxy for climate communication\.

Figure 1:Expressed\-happiness prevalence across campaign families and temporal periods\. Points and intervals summarize the frozen occurrence\-period cells; the figure reports aggregate language prevalence\.Figure 2:Construct changes across occurrence\-years\. The heatmap displays paired period contrasts for happiness, hope\-waypower, collective capability, distress and action language\. The secondary constructs are lexical indicators and are interpreted more cautiously than happiness\.
### 3\.2RQ2: happiness–action decoupling

For each occurrence\-year, we calculated the change in happiness prevalence and the change in action prevalence relative to the pre\-event period\. The direct decoupling index isD=Δ​H−Δ​AD=\\Delta H\-\\Delta A\. During the event contrast, mean happiness increases by 4\.94 percentage points while action language decreases by 10\.75 percentage points, giving a mean divergence of 15\.69 percentage points\. The divergence is positive in 16 of 19 occurrence\-years\. For the post contrast, the corresponding changes are \+3\.39 and−14\.03\-14\.03percentage points, with a mean divergence of 17\.42 percentage points and positive divergence in 17 of 19 occurrence\-years\. Exact occurrence\-level sign\-flip tests are below 0\.011 for the six pre\-specified happiness, action and decoupling contrasts; Supporting Information, Table S3, reports these contrasts in full\.

Within the frozen campaign windows, happiness\-labelled language and explicit action language show divergent trajectories\. The pattern is compatible with several mechanisms, including symbolic participation, collective coping and changes in the composition of campaign talk\. The design leaves the contribution of these mechanisms unidentified, and causal claims about positive language and action lie beyond its scope\.

### 3\.3Measurement checks and robustness

Happiness mean correlates with the labMT happiness mean \(r=0\.412r=0\.412, bootstrap 95% CI0\.3780\.378–0\.4420\.442\) and with LIWC positive\-emotion proportion \(r=0\.551r=0\.551, bootstrap 95% CI0\.4910\.491–0\.5970\.597\)\. These convergent lexical checks share a language basis, which limits their use for criterion validation\. Supporting Information, Table S5 and Fig\. S2, provide the full external lexical comparisons\.

The event\-period estimate remains positive across every pre\-specified composition and text\-template sensitivity, ranging from \+8\.43 to \+9\.04 percentage points\. Figure[3](https://arxiv.org/html/2609.22096#S3.F3)summarizes these checks, and Supporting Information, Table S6, gives the full numerical results\. A Rademacher wild\-cluster bootstrap over 19 occurrence\-years givesp=0\.0565p=0\.0565for the common event\-period coefficient\. The wild\-bootstrap result is less precise than the conventional clustered result\.

Two independent model coders annotated the same 451 English posts using the frozen instructions\. Nominal Krippendorff’s alpha is 0\.811 for happiness, 0\.519 for distress, 0\.321 for hope, 0\.374 for collective capability and 0\.344 for action\. Supporting Information, Table S7, gives the full annotation audit\. The audit documents reproducibility of the coding instructions, especially for happiness; human criterion validity remains unestablished\. The lower agreement for the secondary constructs motivates their treatment as lexical indicators with limited psychological interpretation\.

Figure 3:Robustness and placebo checks for the event\-period happiness association\. The direction is stable across composition and text\-template variants, while the small\-cluster wild\-bootstrap result is appropriately more conservative\.
### 3\.4RQ3: observed retweet\-cascade exposure

Many source identifiers in the formal edge table remain unresolved, including identifiers affected by scientific\-notation precision loss; matched cascade counts are therefore lower bounds\. Supporting Information, Table S8, provides the complete linkage\-coverage audit\. A hurdle model separates whether any matched cascade is observed from the number of matched retweets conditional on a positive count\. Happier source posts have lower odds of any observed matched cascade \(OR 0\.457, 95% CI 0\.233–0\.896; clusteredp=0\.0225p=0\.0225\)\. Conditional cascade size is directionally similar but uncertain after occurrence clustering \(IRR 0\.302, 95% CI 0\.047–1\.928;p=0\.206p=0\.206\)\. The timing sensitivity yieldsβ=0\.089\\beta=0\.089for log time to the first matched retweet \(95% CI−0\.190\-0\.190–0\.3680\.368;p=0\.532p=0\.532\)\.

The same text is copied by a retweet, and source and child happiness scores are identical for every matched edge\. The matched data contain no independent recipient emotional outcome\. Figure[4](https://arxiv.org/html/2609.22096#S3.F4)therefore reports observed exposure and cascade selection, with emotional contagion and causal diffusion outside the estimand\.

Figure 4:Observed retweet\-edge representation of happiness\-labelled climate discourse\. The figure reports matched exposure and cascade structure; duplicated retweet text leaves changes in recipient emotion unobserved\.

## 4Discussion

The main finding is a difference between two forms of climate discourse\. Happiness\-labelled language rises around campaign periods in the primary panel, while explicit action language decreases in paired occurrence contrasts\. A campaign can coincide with more positive\-affect or social\-possibility language without a corresponding increase in explicit action language\. These data describe an observed discourse pattern; causal effects of campaigns on happiness or action remain outside the analysis\.

The result is consistent with theories that distinguish constructive hope from passive optimism\. Hope can signal that a desirable future is imaginable, but its relationship with action depends on whether the text also names agency, pathways and collective capacity[ojala2012hope](https://arxiv.org/html/2609.22096#bib.bib15);[marlon2019mobilization](https://arxiv.org/html/2609.22096#bib.bib19)\. Our lexical indicators cannot recover the full psychological appraisal or lived context behind a post\. They can identify when these semantic components co\-occur or diverge in a large public corpus\. A happiness word records wording in a post; the user’s happiness remains unobserved\.

The network analysis answers a narrower question\. Happiness\-labelled source posts are less likely to have an observed matched cascade in the hurdle model, but the conditional cascade\-size and timing estimates are uncertain\. Source and child texts are identical, and the graph contains no independent recipient outcome after exposure\. The analysis therefore concerns selection into observed cascades and the visibility of source language\. Causal diffusion claims would require independent recipient outcomes, an exposure denominator, reliable source identifiers and a design that addresses confounding and censoring\.

Five limitations qualify these results\. First, the design is observational and contains no randomized campaign exposure\. The event\-period coefficient is an adjusted temporal association, and the wild\-cluster sensitivity is borderline with only 19 occurrence\-years\. Second, campaign families were selected through observed keywords and hashtags, which confines the corpus to campaign\-labelled Twitter/X discourse\. Third, the happiness, hope, collective\-capability, distress and action measures are lexical/model\-inferred indicators\. LIWC and labMT provide convergent language checks, and the two\-model audit tests reproducibility, but neither establishes human criterion validity\. Fourth, the retweet network is incomplete because source identifiers are unresolved for many rows and the outcome counts only matched edges\. Fifth, M3 age and gender outputs are model\-inferred aggregate groups with possible systematic error; the associated temporal\-precedence scans remain exploratory Supporting Information results\.

The analysis provides a reproducible way to study positive well\-being language in climate communication while keeping social\-media text separate from clinical measurement\. Its contribution is methodological and descriptive: repeated campaign windows, paired happiness–action contrasts, lexical controls, author and text\-template sensitivities, small\-cluster inference and a network\-linkage audit\. Future work should add independently validated human annotations, broader non\-campaign comparison periods, multilingual measurement and longitudinal designs that distinguish affective expression from individual well\-being and behavioral change\.

#### Data and code availability

Reproducibility code, the dependency specification, final statistical audit and machine\-readable derived outputs are archived at Figshare:[https://doi\.org/10\.6084/m9\.figshare\.33207624](https://doi.org/10.6084/m9.figshare.33207624)\. Raw Twitter/X text and the full user\-level archive are excluded from redistribution\.

## References

- \(1\)Romanello, M\.*et al\.*The 2023 report of the lancet countdown on health and climate change: The imperative for a health\-centred response in a world facing irreversible harms\.*The Lancet*402, 2346–2394 \(2023\)\.
- \(2\)Lawrance, E\. L\., Thompson, R\., Fontana, G\. & Jennings, N\.The impact of climate change on mental health and emotional wellbeing: Current evidence and implications for policy and practice\.*The Lancet Planetary Health*6, e726–e738 \(2022\)\.
- \(3\)Ostrom, E\.A general framework for analyzing sustainability of social\-ecological systems\.*Science*325, 419–422 \(2009\)\.
- \(4\)Scherer, K\. R\.What are emotions? and how can they be measured?*Social Science Information*44, 695–729 \(2005\)\.
- \(5\)Lazarus, R\. S\.*Emotion and Adaptation*\(Oxford University Press, 1991\)\.
- \(6\)Gross, J\. J\.The emerging field of emotion regulation: An integrative review\.*Review of General Psychology*2, 271–299 \(1998\)\.
- \(7\)Fredrickson, B\. L\.The role of positive emotions in positive psychology: The broaden\-and\-build theory of positive emotions\.*American Psychologist*56, 218–226 \(2001\)\.
- \(8\)Chapman, D\. A\., Lickel, B\. & Markowitz, E\. M\.Reassessing emotion in climate change communication\.*Nature Climate Change*7, 850–852 \(2017\)\.
- \(9\)Markowitz, E\. M\. & Shariff, A\. F\.Climate change and moral judgement\.*Nature Climate Change*4, 243–247 \(2014\)\.
- \(10\)Brosch, T\.Affect and emotions as drivers of climate change perception and action: A review\.*Current Opinion in Behavioral Sciences*42, 15–21 \(2021\)\.
- \(11\)Clayton, S\. & Karazsia, B\. T\.Development and validation of a measure of climate change anxiety\.*Journal of Environmental Psychology*69, 101434 \(2020\)\.
- \(12\)Pihkala, P\.Anxiety and the ecological crisis: An analysis of eco\-anxiety and climate anxiety\.*Sustainability*12, 7836 \(2020\)\.
- \(13\)Hickman, C\.*et al\.*Climate anxiety in children and young people and their beliefs about government responses to climate change: A global survey\.*The Lancet Planetary Health*5, e863–e873 \(2021\)\.
- \(14\)Ojala, M\.How do children cope with global climate change? coping strategies, engagement, and well\-being\.*Journal of Environmental Psychology*32, 225–233 \(2012\)\.
- \(15\)Ojala, M\.Hope and climate change: The importance of hope for environmental engagement among young people\.*Environmental Education Research*18, 625–642 \(2012\)\.
- \(16\)Ojala, M\.Hope in the face of climate change: Associations with environmental engagement and well\-being\.*The Journal of Environmental Education*46, 133–148 \(2015\)\.
- \(17\)Snyder, C\. R\.Hope theory: Rainbows in the mind\.*Psychological Inquiry*13, 249–275 \(2002\)\.
- \(18\)Cohen\-Chen, S\., van Zomeren, M\., Saguy, T\., Halperin, E\. & Bos, G\.Is hope good for motivating collective action in the context of climate change? differentiating hope’s emotion\- and problem\-focused coping functions\.*Global Environmental Change*58, 101915 \(2019\)\.
- \(19\)Marlon, J\. R\.*et al\.*How hope and doubt affect climate change mobilization\.*Frontiers in Communication*4, 20 \(2019\)\.
- \(20\)Mortreux, C\., O’Neill, S\.*et al\.*Hope as an enabler of climate change adaptation\.*Communications Psychology*3, 147 \(2025\)\.
- \(21\)Feldman, L\. & Hart, P\. S\.Is there any hope? how climate change news imagery and text influence audience emotions and support for mitigation policies\.*Risk Analysis*38, 585–602 \(2018\)\.
- \(22\)Hornsey, M\. J\., Harris, E\. A\. & Fielding, K\. S\.A cautionary note about messages of hope: Focusing on progress in reducing carbon emissions weakens mitigation motivation\.*Global Environmental Change*39, 26–34 \(2016\)\.
- \(23\)Morris, B\. S\., Chrysochou, P\., van Echelt, T\. & Molder, H\.Optimistic versus pessimistic endings in climate change appeals\.*Humanities and Social Sciences Communications*7, 82 \(2020\)\.
- \(24\)Lammers, J\. & Formanānski, K\.Communicating the need for climate action\.*Nature Climate Change*\(2026\)\.
- \(25\)Voelkel, J\. G\.*et al\.*A registered report megastudy on the persuasiveness of the most\-cited climate messages\.*Nature Climate Change*\(2026\)\.
- \(26\)van Zomeren, M\., Postmes, T\. & Spears, R\.Toward an integrative social identity model of collective action: A quantitative research synthesis of three socio\-psychological perspectives\.*Psychological Bulletin*134, 504–535 \(2008\)\.
- \(27\)van Zomeren, M\.An integrative perspective on social and collective action\.*European Review of Social Psychology*23, 1–35 \(2012\)\.
- \(28\)Fritsche, I\., Barth, M\., Jugert, P\., Masson, T\. & Reese, G\.A social identity model of pro\-environmental action \(simpea\)\.*Psychological Review*125, 245–269 \(2018\)\.
- \(29\)Bamberg, S\., Rees, J\. & Seebauer, S\.Collective climate action: Determinants of participation intention in community\-based pro\-environmental initiatives\.*Journal of Environmental Psychology*43, 155–165 \(2015\)\.
- \(30\)Jugert, P\., Greenaway, K\. H\., Barth, M\., B”ackstr”om, M\. & Sch”uler, J\.Collective efficacy increases pro\-environmental intentions through increasing self\-efficacy\.*Journal of Environmental Psychology*48, 12–23 \(2016\)\.
- \(31\)Reese, G\.*et al\.*Social identity and pro\-environmental action: A meta\-analytic review\.*Journal of Environmental Psychology*65, 101328 \(2019\)\.
- \(32\)Ajzen, I\.The theory of planned behavior\.*Organizational Behavior and Human Decision Processes*50, 179–211 \(1991\)\.
- \(33\)Stern, P\. C\.New environmental theories: Toward a coherent theory of environmentally significant behavior\.*Journal of Social Issues*56, 407–424 \(2000\)\.
- \(34\)Fielding, K\. S\., McDonald, R\. & Louis, W\. R\.Theory of planned behaviour, identity and intentions to engage in environmental activism\.*Journal of Environmental Psychology*28, 318–326 \(2008\)\.
- \(35\)Gifford, R\.The dragons of inaction: Psychological barriers that limit climate change mitigation and adaptation\.*American Psychologist*66, 290–302 \(2011\)\.
- \(36\)Lazer, D\.*et al\.*Computational social science\.*Science*323, 721–723 \(2009\)\.
- \(37\)Gentzkow, M\., Kelly, B\. & Taddy, M\.Text as data\.*Journal of Economic Literature*57, 535–574 \(2019\)\.
- \(38\)Grimmer, J\. & Stewart, B\. M\.Text as data: The promise and pitfalls of automatic content analysis methods for political texts\.*Political Analysis*21, 267–297 \(2013\)\.
- \(39\)Dodds, P\. S\. & Danforth, C\. M\.Measuring the happiness of large\-scale written expression: Songs, blogs, and presidents\.*Journal of Happiness Studies*11, 441–456 \(2010\)\.
- \(40\)Dodds, P\. S\., Harris, K\. D\., Kloumann, I\. M\., Bliss, C\. A\. & Danforth, C\. M\.Temporal patterns of happiness and information in a global\-scale social network: Hedonometrics and twitter\.*PLoS ONE*6, e26752 \(2011\)\.
- \(41\)Tausczik, Y\. R\. & Pennebaker, J\. W\.The psychological meaning of words: Liwc and computerized text analysis methods\.*Journal of Language and Social Psychology*29, 24–54 \(2010\)\.
- \(42\)Pennebaker, J\. W\., Booth, R\. J\. & Francis, M\. E\.*Linguistic Inquiry and Word Count: LIWC2007*\(LIWC\.net, 2007\)\.
- \(43\)Caliskan, A\., Bryson, J\. J\. & Narayanan, A\.Semantics derived automatically from language corpora contain human\-like biases\.*Science*356, 183–186 \(2017\)\.
- \(44\)Hutto, C\. J\. & Gilbert, E\.Vader: A parsimonious rule\-based model for sentiment analysis of social media text\.*Proceedings of the International AAAI Conference on Web and Social Media*8, 216–225 \(2014\)\.
- \(45\)Golder, S\. A\. & Macy, M\. W\.Diurnal and seasonal mood vary with work, sleep, and daylength across diverse cultures\.*Science*333, 1878–1881 \(2011\)\.
- \(46\)Lazer, D\.*et al\.*The parable of google flu: Traps in big data analysis\.*Science*343, 1203–1205 \(2014\)\.
- \(47\)Kramer, A\. D\. I\., Guillory, J\. E\. & Hancock, J\. T\.Experimental evidence of massive\-scale emotional contagion through social networks\.*Proceedings of the National Academy of Sciences*111, 8788–8790 \(2014\)\.
- \(48\)Brady, W\. J\., Wills, J\. A\., Jost, J\. T\., Tucker, J\. A\. & Van Bavel, J\. J\.Emotion shapes the diffusion of moralized content in social networks\.*Proceedings of the National Academy of Sciences*114, 7313–7318 \(2017\)\.
- \(49\)Bakshy, E\., Messing, S\. & Adamic, L\. A\.Exposure to ideologically diverse news and opinion on facebook\.*Science*348, 1130–1132 \(2015\)\.
- \(50\)Bail, C\. A\.*et al\.*Exposure to opposing views on social media can increase political polarization\.*Proceedings of the National Academy of Sciences*115, 9216–9221 \(2018\)\.
- \(51\)Vosoughi, S\., Roy, D\. & Aral, S\.The spread of true and false news online\.*Science*359, 1146–1151 \(2018\)\.
- \(52\)Benjamini, Y\. & Hochberg, Y\.Controlling the false discovery rate: A practical and powerful approach to multiple testing\.*Journal of the Royal Statistical Society: Series B*57, 289–300 \(1995\)\.
- \(53\)Wang, Z\.*et al\.*Demographic inference and representative population estimates from multilingual social media data\.*Proceedings of The World Wide Web Conference*2056–2067 \(2019\)\.
- \(54\)Granger, C\. W\. J\.Investigating causal relations by econometric models and cross\-spectral methods\.*Econometrica*37, 424–438 \(1969\)\.