Investigators analyzed 207,647 first-person posts announcing COVID-19 vaccination on Twitter, covering December 14, 2020, through January 1, 2022. The study team used a mixed deductive-inductive coding approach anchored in the disclosure decision model (DDM) to classify content and motivations. From that process they derived 11 discrete disclosure themes. Reported theme prevalences included community identity (31,847 posts, 15.34%), encouraging others (29,619, 14.26%), and gratitude (29,511, 14.21%). Intercoder reliability for the thematic coding ranged from Krippendorff α = 0.72 to 0.93, as the authors noted. The dataset and coding framework were then extended across the full corpus using machine classifiers trained to label each theme at scale.
The authors divided the rollout into four phases and examined how theme frequencies changed over time. Gratitude dominated during phase 1 of the US rollout, according to the reported counts. Community identity grew more prominent during phases 2 and 4. Encouraging others rose in prominence in phase 3. These temporal shifts suggest that disclosure content evolved alongside the public health context and the changing makeup of vaccine-eligible populations. The source provided these phase-level trends but did not report additional phase boundary details beyond the study timeframe and phase labeling in the abstract.
The team evaluated associations between disclosure themes and public engagement, defined as the sum of likes, reposts, quotes, and replies. They used negative binomial regression models that controlled for follower count, presence of media, word count, sentiment, and other post-level covariates. Emotional and identity-focused disclosures were tied to higher engagement: gratitude was associated with a 25% increase in engagement (incidence rate ratio [IRR] = 1.25; 95% CI 1.23-1.28), and relief showed a similar effect (IRR = 1.26; 95% CI 1.19-1.33). Community identity disclosures correlated with a 19% higher engagement rate (IRR = 1.19; 95% CI 1.17-1.21), investigators reported.
By contrast, posts centering on side effects, political framing, or social benefits tended to receive less engagement. Specifically, side effects were linked to decreased engagement (IRR = 0.85; 95% CI 0.83-0.87), political disclosures to lower engagement (IRR = 0.90; 95% CI 0.87-0.94), and social benefits disclosures to modestly reduced engagement (IRR = 0.92; 95% CI 0.90-0.95). The authors also observed that each additional theme present in a post correlated with a 7% increase in engagement (IRR = 1.07; 95% CI 1.06-1.08), suggesting multi-theme messages attracted more interaction.
Network analysis mapped how disclosure themes co-occurred within posts across each rollout phase. The analysis sought to identify clusters of motives and rhetorical frames that tended to appear together in single messages. The abstract reports that co-occurrence patterns shifted by phase—community and identity themes coalesced more in later phases, while gratitude dominated early messaging—yet it does not provide the specific network metrics or visualizations in the summary. The study team trained separate classifiers for each of the 11 themes to enable this phasewise network labeling at scale.
Automated identification of disclosure posts relied on a RoBERTa binary classifier trained on 6,348 human-coded posts; the classifier achieved an F1 score of 0.85 for identifying disclosure posts, the investigators reported. Eleven additional RoBERTa classifiers were trained to label the specific disclosure themes across the full dataset. The thematic codebook was developed from the DDM’s motivational goals and refined through deductive and inductive coding. Krippendorff α values for coded themes ranged from 0.72 to 0.93, indicating acceptable to high intercoder agreement. Beyond these performance metrics, the abstract does not report hyperparameter settings, training-validation splits, or error analyses in this summary.
The authors concluded that disclosure content on Twitter was thematically patterned and responsive to changing public health contexts, and they argued identity-based and emotionally resonant disclosures—particularly gratitude and community identity—drove higher engagement. The abstract does not report causal inferences about behavior change or vaccine uptake resulting from disclosures; investigators reported associations with engagement only. The source also did not provide detailed classifier error analysis or the exact phase demarcations in the abstract. Open questions remain about whether higher engagement translates into measurable changes in vaccine attitudes or uptake, how these patterns behave on other platforms, and how demographic factors shape disclosure effects. These points represent documented limitations and areas for further investigation rather than extrapolated conclusions.