From the CDC to emerging infectious disease publics: The long-now of polarizing and complex health crises

TL;DR

This study investigates how CDC COVID-19 messaging and engagement on Twitter evolved over the first two years of the pandemic, focusing on how digital publics formed, polarized, and interacted with health authorities. It employs a longitudinal mixed-methods approach, combining BERTopic-based topic modeling (71 topics), clustering of 16 discourse publics, sentiment and credibility analysis, and temporal change-point detection on a dataset of 275,124 CDC-related tweets. The findings reveal pervasive echo chambers, with a notable exception in a marginalized racial health equity public (Cluster 1) that maintained engagement around equity, data justice, and accountability, even as overall discourse shifted. The authors propose design and policy interventions, including multi-agent AI assistants for diverse publics and long-term crisis communication planning, to support inclusive and adaptive health communication in extended public health crises.

Abstract

This study examines how public discourse around COVID-19 unfolded on Twitter through the lens of crisis communication and digital publics. Analyzing over 275,000 tweets involving the CDC, we identify 16 distinct discourse clusters shaped by framing, sentiment, credibility, and network dynamics. We find that CDC messaging became a flashpoint for affective and ideological polarization, with users aligning along competing frames of science vs. freedom, and public health vs. political overreach. Most clusters formed echo chambers, while a few enabled cross cutting dialogue. Publics emerged not only around ideology but also around topical and emotional stakes, reflecting shifting concerns across different stages of the pandemic. While marginalized communities raised consistent equity concerns, these narratives struggled to reshape broader discourse. Our findings highlight the importance of long-term, adaptive engagement with diverse publics and propose design interventions such as multi-agent AI assistants, to support more inclusive communication throughout extended public health crises.

Figures (5)

• Figure 1: This image shows public policy discourse framing analytical framework adapted from park2020stakeholder. We introduced new elements (in yellow) to the model based on our findings. We show how platform affordances, message and user properties, and platform governance are important factors to consider when communicating public health policy for emerging infectious diseases like COVID. Additionally, we discuss how emerging technologies (in gray), especially generative AI introduce both challenges and design opportunities for crisis organizations like the CDC. Finally, we discuss the effects of temporal crisis distance affects CDC messages over time, especially around flashpoints associated with major changes (e.g., new virus variants). The +/- signs represent how each of these components can make crisis communication more or less effective.
• Figure 2: BERTopic clustering results grouped into five major themes, including COVID mitigation, partisan discourse, and public health equity. Themes reflect how discussions clustered semantically and ideologically.
• Figure 3: Using the elbow method, we identified 16 clusters as being the best number based on distortion error.
• Figure 4: The figure on the left shows the co-occurrence of hashtags with the #CDC. The figure on the right shows the top hashtags in the discourse. We collapsed similar hashtags together. For example, ‘#coronavirus’, ‘#covid19’, ‘#covid-19’, ‘#sars’, ‘#coronaviruspandemic’, ‘#pandemic’, etc. were collapsed to simply to ‘#COVID-19’. ‘#CDC’ and ‘#cdc’ were collapsed to ‘#CDC’. The hashtags ‘#vaccine’, ‘#GetVaccinated’, ‘#vaccinated’, and ‘#VaccinesWork’ were collapsed to ‘#Vaccine’. Likewise, ‘#Trump’, ‘#TrumpVirus’, ‘#TrumpLiesAmericansDie’ were collapsed to ‘#Trump’. Lastly, ‘#Masks’, ‘#WearAMask’, ‘#MaskUp’, ‘#masks’, and ‘#mask’ were collapsed to ‘#Masks’.
• Figure 5: Time-series of race-related discourse across the full Twitter public (top) and Cluster 1 (bottom). Early decline in race discourse contrasts with episodic but sustained engagement by Cluster 1. Detected change-points (red lines) show how this marginalized public responded dynamically to external events, unlike broader publics that deprioritized racial equity over time.