Dynamic Topic Alignment and Sentiment between Official Health Communication and General Public Discourse during COVID-19: A Comprehensive Infoveillance Framework

This study introduces a comprehensive infoveillance framework to analyze the dynamic alignment and sentiment between CDC communications and public discourse on social media during the COVID-19 pandemic, revealing that increased topic alignment correlates with heightened negative emotional responses rather than public agreement.

The Gist

Imagine the COVID-19 pandemic as a massive, chaotic concert where two different groups are trying to lead the music: the CDC (the official conductors) and the general public (the audience).

This paper is like a high-tech sound engineer's report. It tries to answer a simple but tricky question: Are the conductors and the audience actually playing the same song, or are they just shouting over each other?

Here is how the researchers broke it down, using simple analogies:

1. The Two Playlists (Topic Modeling)

First, the researchers collected two huge playlists of tweets:

• Playlist A: 17,524 tweets from official CDC accounts (the conductors).
• Playlist B: 67,895 tweets from regular people (the audience).

They used a smart computer program (called a "Biterm Topic Model") to listen to these playlists and group the words into themes.

• The CDC Playlist sounded like a medical textbook: "hospitalization rates," "vaccination protocols," "non-pharmaceutical interventions," and "clinical aspects."
• The Public Playlist sounded like a mix of a town hall, a bar, and a living room: "economic impact," "politics," "sports," "family protection," and "personal struggles."

The Finding: While both groups were talking about the virus, they were often speaking different languages. The CDC was focused on technical metrics, while the public was focused on how the virus affected their daily lives, jobs, and feelings.

2. The "Alignment Score" (Topic Consistency)

The researchers invented a new way to measure how well the two playlists matched up. Imagine a thermometer that doesn't measure heat, but measures "topic alignment."

• How it works: They didn't just look at whether the words were similar; they looked at how much the public was talking about a specific topic on any given day.
• The Result: Over time, the "alignment thermometer" went up. This means the public started talking about the same topics the CDC was talking about (like school responses or family protection).
• The Catch: Just because the public was talking about the same topics as the CDC does not mean they agreed with them. It just means they were paying attention to the same things. It's like two people staring at the same storm cloud; one is worried about the rain, and the other is worried about the wind. They are looking at the same thing, but they aren't necessarily on the same page.

3. The Mood Ring (Sentiment Analysis)

To understand how the public felt about these topics, the researchers used two different "Mood Rings":

• Mood Ring #1 (Expected Sentiment): This measures the average tone. If you picked a random tweet, would it be happy, sad, or angry?

  • The Result: Even when the public was talking about the exact same topics as the CDC, the average tone remained consistently negative. The more the public aligned with the CDC's topics, the more negative the tone became. It suggests that when people focus on the CDC's warnings, they feel more worried, not more reassured.

• Mood Ring #2 (Net Sentiment): This measures the intensity or volume of the emotion. It's not about what they feel, but how loudly they feel it.

  • The Result: When the public aligned with CDC topics, the emotional intensity spiked immediately. It was like a sudden roar of emotion. The public reacted fast and hard to the topics the CDC brought up.

4. The Time Machine (ARIMAX Models)

Finally, the researchers used a time-traveling math model to see how long it took for the public's mood to change after the CDC spoke.

• The "Average Mood" (Expected Sentiment): This changed slowly. When the CDC talked about a topic, the public's average mood shifted negatively immediately, but it also kept shifting negatively for days afterward, and even followed a weekly pattern (like a recurring headache).
• The "Emotional Intensity" (Net Sentiment): This was a lightning bolt. The intensity of the public's emotion spiked immediately on the same day the CDC spoke, but didn't have the same delayed effect.

The Big Takeaway

The main lesson from this study is a bit of a reality check for health officials:

"Alignment" does not equal "Agreement."

When the public starts talking about the same things the CDC is talking about, it doesn't mean they are nodding in agreement or feeling better. In fact, the study found that higher alignment was actually linked to more negative feelings and higher anxiety.

The public was paying close attention to the CDC's messages, but they were doing so with fear and concern, not with calm acceptance. The study suggests that when health agencies and the public are "on the same topic," it often means the public is deeply worried about the issues the agency is highlighting.

In short: The public and the CDC were dancing to the same beat, but the public was dancing to a sad song.

Technical Summary

Technical Summary: Dynamic Topic Alignment and Sentiment between Official Health Communication and General Public Discourse during COVID-19

Problem Statement
During the COVID-19 pandemic, social media served as a critical channel for both official health communication (e.g., Centers for Disease Control and Prevention, CDC) and general public discourse. While existing research has examined public sentiment or isolated topics, there is a lack of systemic investigation into the dynamic temporal alignment between focal messages from official health agencies and the evolving topics of the broader public. Previous studies were often limited to single-day snapshots, failed to model the evolution of topic alignment and emotions over time, or did not distinguish between public attention and public agreement. This study addresses the gap in understanding how public opinions, expressed as content topics and sentiments, correlate with and diverge from official health communications over the course of a health emergency.

Methodology
The authors developed an end-to-end infoveillance framework integrating Natural Language Processing (NLP), topic modeling, and multivariate time series analysis. The study utilized two large-scale Twitter/X datasets:

1. CDC Communications: 17,524 tweets from 7 official CDC-affiliated accounts (Dec 17, 2019 – Jan 15, 2022).
2. General Public Discourse: 67,895 tweets collected via BrandWatch (Feb 28, 2020 – June 1, 2023).

The methodological pipeline consisted of four key stages:

1. Topic Modeling (Biterm Topic Model - BTM):

   • BTM was applied separately to both corpora to handle short-text data sparsity, modeling word co-occurrence across the entire corpus rather than within individual documents.
   • A shared vocabulary was constructed to allow comparability.
   • The optimal number of topics was determined using Rényi entropy (stabilizing at 10 for public discourse and 9 for CDC) and validated through expert qualitative review.
   • This yielded two distinct sets of topics: 9 CDC topics (e.g., epidemic risk, local responses, family protection) and 10 public discourse topics (e.g., economic impact, politics, sports).

2. Topic Consistency Scoring System:

   • A novel metric was developed to quantify dynamic alignment. It integrates cosine similarity (measuring semantic similarity between topic-word probability distributions) with daily topic prominence (the proportion of a specific topic in the public discourse on a given day).
   • The system calculates a weighted alignment score ($C_{j,d}$) for each CDC topic $j$ on day $d$, reflecting how closely the public discourse aligns with that specific official topic, weighted by how prominent that topic is in the public sphere.

3. Sentiment Analysis:

   • Two complementary sentiment measures were incorporated to capture different dimensions of public emotion:
     • Expected Sentiment: The average emotional tone (normalized), representing the tone of a randomly selected post regardless of volume.
     • Net Sentiment: The aggregate emotional intensity (unnormalized sum of positive minus negative), capturing the magnitude and concentration of emotional surges.

4. Temporal Dynamics (ARIMAX Modeling):

   • Autoregressive Integrated Moving Average with Exogenous Variables (ARIMAX) models were employed to examine the temporal relationship between topic consistency (exogenous variable) and public sentiment (dependent variable).
   • Models were run for both expected and net sentiment, incorporating lag terms up to 7 days to capture immediate, delayed, and weekly cyclical patterns.

Key Results

• Topic Alignment Trends: Topic alignment between CDC communications and public discourse increased over time for most topics, particularly for "local and state responses" and "family protection." However, alignment varied significantly across topics; technical topics like "non-pharmaceutical interventions" showed lower alignment, likely due to differences in language and framing between official and public discourse.
• Sentiment Dynamics: Despite increasing topic alignment, expected public sentiment remained consistently negative throughout the study period. This indicates that higher alignment reflects public attention and concern rather than agreement or positive reception of official messaging.
• Temporal Effects (ARIMAX Findings):
  • Expected Sentiment: Higher topic consistency was significantly associated with immediate decreases in average sentiment (more negative tone) and showed significant lagged effects at 2, 3, and 7 days, suggesting a weekly cyclical pattern in public emotional response.
  • Net Sentiment: Higher topic consistency was associated with a significant immediate increase in emotional intensity (magnitude), but without the same delayed or weekly lagged effects observed in expected sentiment. This suggests that when the public focuses on the same topics as the CDC, the emotional reaction is intense and immediate, often driven by negative spikes (e.g., during early pandemic phases or reports of deaths).

Key Contributions

1. Novel Infoveillance Framework: The study presents a scalable, end-to-end framework that moves beyond static analysis to model the dynamic, temporal relationship between official health communication and public discourse.
2. Topic Consistency Scoring System: The authors introduce a weighted similarity metric that accounts for both semantic alignment and the daily prominence of topics, offering a more nuanced measure of alignment than simple cosine similarity.
3. Dual Sentiment Metrics: By distinguishing between expected sentiment (average tone) and net sentiment (emotional intensity), the study provides a more comprehensive view of public reaction, revealing that high alignment can coexist with negative tone and high emotional volatility.
4. Empirical Insight on Alignment: The findings challenge the assumption that topic alignment equates to successful communication or public agreement. Instead, the study demonstrates that alignment often reflects heightened public concern and attention to the same issues, frequently accompanied by negative emotional responses.

Significance and Claims
The paper claims that this framework provides a generalizable approach for investigating and evaluating public engagement with official health communication. The primary significance lies in the ability to distinguish between public attention (alignment) and public agreement (sentiment). The authors assert that their results suggest health agencies should interpret topic alignment not as a sign of message acceptance, but as an indicator of public concern that may require specific emotional context management. The framework is designed to be modular, allowing for the integration of different topic modeling approaches or sentiment metrics, and is intended to support more effective communication strategies during future public health emergencies. The study explicitly notes limitations regarding the exclusive use of Twitter/X data and the inability of current NLP methods to fully capture sarcasm or mixed sentiments, framing these as areas for future expansion rather than current capabilities.