The dynamics of Twitter users' gun narratives across major mass shooting events

The paper examines whether and how beliefs about gun policy, as expressed on Twitter, changed across several major mass shooting events from 2016–2018. Motivated by the persistent polarization of U.S. gun policy debates and the limited understanding of how ordinary citizens construct and evolve their narratives online, the authors analyze how users’ tweets reflect causal attributions (what is the problem, who is to blame, what are the causes and solutions), policy stances, rhetoric, and emotions. Using large-scale longitudinal data from ideology-identifiable users (liberal-leaning vs. conservative-leaning), the study investigates narrative shifts surrounding the 2016 Orlando nightclub shooting, the 2017 Las Vegas shooting, and the 2018 Parkland shooting. The key research questions probe differences between ideological camps in causal narrative construction, stance and rhetoric shifts around events, and the association between emotions and causal narratives.

Background covers several strands: (1) Gun debate: rights vs. control—Americans’ high gun ownership, longstanding polarized beliefs grounded in cultural values, and the intensification of debates around mass shootings. (2) Collective action and framing—social media’s roles as mobilizing and opportunity structures and in framing processes; most prior work focuses on elite framing rather than ordinary users’ discourse and frame variation. (3) Social media and public deliberation—platforms enable everyday sociopolitical talk and can go beyond opinion polling to examine deliberation; however, little work has unpacked users’ causal attributions on contentious issues like guns. (4) Narrative analysis of policy discussion—adapting the Narrative Policy Framework (NPF) to decompose stories into components (villain, victim, hero, plot, moral) to recover causal attributions in public discourse. (5) Rhetoric with emotion—emotion is used in policy persuasion; prior work emphasizes elites, while this study explores emotions within public narratives and their interaction with causal attributions in the context of mass shootings.

Data source and design: Twitter timelines were collected for over 155,000 ideology-identifiable users extended from a prior corpus of exclusive followers of 2016 U.S. presidential candidates (Trump-supporters labeled Con; Clinton-supporters labeled Lib) validated via supportive/disapproving campaign hashtags. The expanded collection spans over 30 months and includes more than 66 million non-retweet tweets from users whose complete timelines were available (approximately 77,954 Con and 77,355 Lib). Retweets were excluded. To avoid dominance by highly active accounts, the top 10% most active users (“super users”) were excluded, focusing analyses on the bottom 90% who produced 30% of tweets, representing more typical users. Known IRA-linked accounts and likely bots were removed or assessed; Botometer estimates indicated negligible bot contribution and no overrepresentation of elites by follower/retweet metrics. Event windows and controls: For each of three mass shootings (Orlando 2016; Las Vegas 2017; Parkland 2018), tweets from a 4-week Event period centered on the incident were compiled. For causal attribution against exogenous factors, a matched 4-week Control period in a preceding/succeeding year without a major shooting near that date was selected for each event. Relevance filtering: Gun-related tweets were identified using and iteratively extending keyword/hashtag lists from Benton et al. (2016), with separate seed sets for pro-control, anti-control, and ambiguous terms, expanded via co-occurrence within the dataset and filtered to remove advertising or off-topic tokens. Relevant tweets (#Relevant) were those matching curated gun-related terms. Linguistic features: Psycholinguistic categories from LIWC were extracted—cause (causal connectives and cause/effect terms) and negative affect categories: sadness (sad), anxiety (anx), and anger. Manual verification on sampled tweets showed high precision (most above 90%, lowest 73.3% for anger). Statistical analysis of change: Occurrences of features within time windows were modeled as Poisson counts. Before-after rate ratios were computed as post-event rate over the 2-week pre-event base rate. Differences were compared against control periods and between groups using probability-based exact methods to obtain p-values and confidence intervals. Temporal granularity included immediate 48 hours post-event, days 2–7, and week 2. Narrative coding: To recover narrative structures, a mixed-methods approach adapted NPF with a coding taxonomy: (1) Stance (support gun control, oppose, mixed, none), (2) Moral Solution (presence of call to action), (3) Rhetoric Scheme (seven categories: Argumentation with Plot [five plot types], Argumentative Assertion, Evidence-based Argumentation, Sarcasm, Rhetorical Question/Imperative, Personal/Community Relatedness, Scripted Message/Petition), and (4) Attributional Elements (Hero, Victim, Villain, Blame Target, Nature of Problem). Tweets for coding were stratified by group, event, and stage, focusing on those containing cause words to enrich for causal attributions. Coding proceeded in three iterative phases (exploratory, pilot with reliability testing, final). Inter-rater reliability (Cohen’s kappa) for the finalized scheme ranged from 0.61 to 0.79 across components, indicating substantial agreement. In total, 480 tweets were coded (raw frequencies: Relevancy 414; Moral Solution 117; Rhetoric Scheme 358; Stance 370; Victim 64; Villain 104; Hero 44; Problem Nature 121; Blame Target 177). Analyses examined distributions by group, stage, and event. Data access: The dataset is publicly available at https://github.com/picsolab/gun-narrative-tweets.

Linguistic shifts: After mass shooting events, gun-relevant tweeting and causal/emotional language rose sharply. Across events, event period before-after rate ratios were: #Relevant 12.99 (p<0.001), cause 16.42 (p<0.001), sad 15.72 (p<0.001), anxiety 16.13 (p<0.001), anger 16.21 (p<0.001). Control periods showed no corresponding increases (ratios ~0.60–0.70), and ratio differences indicated ≥10x increases attributable to events. Group-wise, both camps exhibited similar proportional increases, with Lib slightly higher for cause and sad, and Con slightly higher for anxiety. Across events, cause-word surges grew larger over time (e.g., Event-1: Con 8.99, Lib 10.21; Event-2: Con 16.58, Lib 15.88; Event-3: Con 36.38, Lib 43.17; all p<0.001), and elevated levels persisted for up to two weeks. Attributional elements: After events, discussions of Problem Nature and Blame Target intensified, notably in Con. Problem Nature: Con increased from 9.4% pre-event to 39.0% within 48h (p=0.003); Lib had been higher pre-event (37.8%) but also rose post-event. Blame Target: Con rose 25.0%→55.9% (p=0.008); Lib 29.7%→47.4% (ns). Combined, blame mentions rose 27.5%→45.8% (p=0.005). Character focus shifted post-event toward Villain and Victim in both camps, with later increases in Hero discussion (Con 3.4%→10.3%; Lib 5.3%→20.0% from immediate 48h to week 2). Policy stance shifts: 89.6% of coded gun-relevant tweets conveyed a stance (Lib 89.4%, Con 89.9%). Majorities were opposing control (Con) vs supporting control (Lib) consistently across events. Yet immediately post-event, adherence to majority stances declined: Con opposing control dropped 96.9% pre-event to 69.1% post-event; Lib supporting control declined 86.5%→84.1% (Con’s drop ~4.3× Lib). Minority stances rose: Con support for control 0.0%→17.0%; Lib opposition 2.7%→5.3%. Mixed stances emerged in Con (2.3%) during immediate 48h and week 1 (none in Lib). Yearly aggregates across camps showed pro-control tweets increasing to a majority: 48.3% (2016), 44.3% (2017), 55.7% (2018), while anti-control remained ~38–42%; non-stance decreased to 5.7% in 2018. In Con, stance-bearing tweets post-event were more likely to include Problem Nature (9.7%→33.5%, p=0.009) and Blame Target (25.8%→48.4%, p=0.029), suggesting increased causal reasoning accompanying stance. Moral Solution (calls to action): Overall 28.3% of tweets invoked a solution/call; more common in Lib (30.4%) than Con (26.1%). Timing differed: Con’s calls surged within 48h then declined; Lib’s calls rose after 48h and continued increasing through week 2. Stance within calls: Lib calls overwhelmingly supported control (96.8%); a small 9.5% opposing appeared only in week 2. Con calls shifted post-event from 100% anti-control pre-event to split within 48h (52.4% anti-control), then partially rebounded (70.0% week 1; 68.8% week 2). Some mixed-stance calls (3.0%) appeared post-event. Rhetoric schemes: 86.5% of tweets employed identifiable rhetoric. Dominant schemes were Argumentative Assertion (28.0%) and Sarcasm (20.8%). Immediately post-event, rhetoric diversity narrowed toward these two schemes. Con: pre-event Sarcasm was most used (34.4%) but dropped to 20.3% within 48h, while Argumentative Assertion rose 15.6%→33.9% (becoming dominant). Lib: pre-event distribution was more even; within 48h, both Argumentative Assertion and Sarcasm increased to 28.1% each. Yearly trends showed growing Argumentative Assertion (22.4%→27.5%→34.3% from 2016 to 2018) and declining Sarcasm (22.4%→21.4%→18.6%). Emotions and attribution: Negative emotions spiked post-event and associated differently with attributional elements across camps. Con: anxiety and sadness were dominant—anxiety was more likely when Villain (8.5% vs 0.7%, p=0.008), Victim (11.1% vs 1.2%, p=0.009), or Blame Target (7.0% vs 0.0%, p=0.005) were present; sadness often linked to perceived powerlessness under gun laws. Lib: anger dominated—more likely when Villain (37.8% vs 10.5%, p<0.001) or Victim (32.1% vs 14.0%, p=0.026) were present, reflecting grievance and a moral imperative to assert control (e.g., via regulation). Interpretively, Con’s anxiety tied to unpredictability of “villains” and sadness to low perceived control; Lib’s anger tied to perceived injustice and desire to increase control through policy.

The study shows that mass shooting events can momentarily and, in some respects, persistently shift public discourse on Twitter toward less polarized expressions. Immediately after events, adherence to majority stances within camps declined, mixed and minority stances surfaced (especially in Con), and more tweets coupled stances with causal reasoning (Problem Nature, Blame Target), potentially facilitating cross-camp understanding. Over years, rhetoric trended away from antagonistic sarcasm toward argumentative assertions, suggesting discourse that is more assertive but less derisive, aligning with broader polling that indicates growing support for certain gun regulations. Coupling stance, attributional elements, and emotions revealed camp-specific emotional-cognitive patterns tied to perceived control (Con: anxiety/sadness; Lib: anger), offering richer interpretability and potential predictive signals of stance shifts beyond ideological identity cues. Practically, these patterns highlight time windows (especially within 48 hours post-event and the following week) as opportunity structures for policymakers and communicators to bridge frames, introduce alternatives, and support deliberation. Design implications include developing civic technologies that surface diverse narratives and framing across ideological boundaries, mitigating echo-chamber dynamics, and improving interpretability of linguistic signals for decision-makers and media.

Using a mixed-methods approach blending large-scale psycholinguistic analysis (LIWC) with narrative coding adapted from the Narrative Policy Framework, the paper demonstrates measurable shifts in both language and narrative structures of Twitter users discussing gun policy around major mass shootings. Contributions include: (1) empirical evidence that events penetrate entrenched ideological discourse, leading to transient reductions in polarization (stance flips, mixed stances, increased causal reasoning); (2) a scalable methodology to retrieve and compare narrative components at scale; and (3) implications for reconciling conflicts, leveraging rhetoric schemes that favor constructive argumentation, enhancing interpretability of linguistic signals, and informing the design of inclusive civic media that elevate diverse frames. Future directions include integrating richer network-based sampling with stratified approaches, incorporating demographic information via linked surveys, and developing context-sensitive linguistic tools to improve precision in emotion and attribution detection across heterogeneous events.

Several limitations qualify the findings: (1) Sampling and activity bias—analyses excluded the top 10% most active users to capture more typical discourse; while increasing diversity, this may underrepresent highly influential narratives and diffusion dynamics. (2) Network and diffusion effects were not explicitly modeled; users may adopt elite or organizational frames, affecting originality and diversity of narratives. (3) Limited demographic resolution—lack of reliable user demographics precluded equity-focused analyses (e.g., differing community impacts). (4) Event heterogeneity vs. temporal trends—the study focuses on the deadliest events (2016–2018), making it difficult to fully disentangle event nature from long-term trends. (5) Lexicon-based measurement—LIWC categories can include context-specific terms (e.g., event descriptions like “kill*”) with varying connotations; although some terms were manually filtered, context sensitivity remains limited. (6) Human coding scale—the narrative analysis relied on 480 stratified tweets; while reliability was substantial, unobserved narrative patterns may exist beyond this sample.