Twitter Data Revealed Distinct COVID-19 Risk Perception Patterns Across the Pandemic
Social media left a dense behavioral trace of how Americans thought about COVID-19, and those traces did not move in lockstep with case counts or death curves. In a U.S. analysis of more than 502 million geotagged tweets from over 4.9 million users, researchers separated perceived susceptibility, perceived severity, and negative emotion into three distinct dimensions and tracked each from January 2020 through December 2021. The result was a public health signal that shifted over time, but never in a single direction.
Perceived severity stayed elevated across the full study period. Susceptibility and negative emotion followed a sharper arc: both peaked on March 11, 2020, the day COVID-19 was declared a global pandemic by the World Health Organization, then fell and remained comparatively stable until the Omicron period pushed them upward again. The pattern suggests that public attention to danger was most reactive at moments of symbolic or informational shock, while broader concern about how severe the disease was likely to be remained persistently high.
Three Risk Dimensions Moved on Separate Timelines
The study’s core contribution is conceptual as much as statistical. Risk perception is often treated as a single construct, but this analysis pulled it apart into three biologically and psychologically different responses. Susceptibility captures the belief that infection could happen to oneself or one’s community. Severity reflects expectations about consequences once infected. Negative emotion reflects the affective burden of fear, worry, and distress.
Those dimensions did not behave identically. Severity maintained a sustained plateau, implying that once the public absorbed COVID-19 as a potentially serious illness, that appraisal did not fully recede even when other concerns eased. Susceptibility and negative emotion were more volatile, rising abruptly at the start of the pandemic and again during Omicron, then easing between those waves. That separation matters because behavior is rarely driven by a single perception. A person may believe disease is serious without feeling personally exposed, or feel anxious without translating that emotion into preventive action.
A sudden peak in attention did not last.
The social media signal also failed to closely mirror epidemic trends in cases and deaths. Tweet frequency about risk perceptions did not track the outbreak curves in a simple one-to-one fashion, which is clinically important for communicators who assume that rising disease burden automatically produces rising public concern. Real-world behavior depends on more than incidence. Media coverage, political framing, local experience, and fatigue all shape whether risk is mentally available enough to drive protective choices such as vaccination, masking, testing, and isolation.
Socioeconomic Vulnerability Was Linked to Lower Attention to Severity and Susceptibility
County-level social determinants of health added another layer. Users from socioeconomically vulnerable counties showed lower attention to COVID-19 severity and susceptibility than users from wealthier counties. The finding points to a communication gap that likely reflects more than simple information access. Structural disadvantage can blunt the visibility of risk through crowded living conditions, essential work that limits online engagement, reduced trust in institutions, and competing daily priorities that crowd out preventive messaging.
The analysis does not suggest that vulnerable communities were less affected by the virus. It shows something more concerning: the digital conversation around threat was weaker where the structural burden was likely greater. In public health terms, that mismatch can widen inequities. If the populations at highest risk are also the least exposed to risk-oriented messaging, the uptake of mitigation measures may lag exactly where they are most needed.
Why the Separation Between Severity and Emotion Matters
The divergence between severity and negative emotion is especially useful for designing interventions. Fear can spike fast, then fade; severity beliefs may persist longer and may be easier to align with long-term prevention messages. Public health campaigns often lean on urgency, but sustained behavior change usually requires a more durable appraisal of harm, personal vulnerability, and effective action. The study suggests that communicators should not assume a single message will move all three dimensions in parallel.
This distinction also helps explain why some periods generated intense online anxiety without a proportional shift in preventive behavior, while other periods sustained a high sense of seriousness with less emotional intensity. The brain does not process threat in a uniform way. Cognitive appraisal and affective response are separable systems, and they can be activated by different cues. A novel variant, a dramatic announcement, or a visible policy change may stir concern quickly, but repeated exposure can normalize the threat even when disease remains clinically serious.
Geotagged Tweets Offer Reach, but Not a Full Clinical Census
The dataset was enormous, yet it still reflects the behavior of people who use Twitter and enable geotagging. That introduces a built-in selection effect. Social media users are not the population as a whole, and online expression is not identical to private belief. Some groups post more often, some less, and some communities may be underrepresented despite bearing substantial disease burden.
Still, the scale of the data gives the analysis real value. More than 4.9 million users and a time window spanning the first two years of the pandemic make it possible to see changes that would be difficult to capture in small surveys or episodic polling. Public health surveillance rarely gets access to such a large, time-sensitive behavioral signal. Even with its limits, the platform offers a window into how collective attention rises, plateaus, and decays.
Tailored Messaging May Be More Effective Than Broad Alerts
The practical implication is straightforward: risk communication during a pandemic cannot rely on a single generic message. Messages aimed at the public need to track the specific dimension that is lagging. If susceptibility is low, campaigns should emphasize personal exposure and transmission pathways. If severity is fading from view, communication should focus on clinical consequences, especially for vulnerable groups. If negative emotion is dominating, messaging should avoid amplifying panic while still reinforcing concrete preventive steps.
A tailored approach is even more important in counties facing socioeconomic vulnerability. Those communities may not receive or prioritize the same signals as wealthier areas, and broad national messaging can miss local realities. Public health agencies that monitor social media in real time may be able to identify where attention has dropped, where concern remains high, and where a different tone is needed to support behavior uptake.
The study leaves an important limitation in place: online attention is not the same as disease incidence, and it is not a direct measure of preventive behavior. Even so, the work shows that COVID-19 risk perception in the United States was not one sentiment but three, moving at different speeds and across different social contexts. The remaining challenge is to turn that richer map of public concern into communication that reaches the communities most likely to be missed.