Our daily lives are characterized by order, but epidemics dramatically disrupt this order, posing significant challenges to public health and mental well-being. Philip Strong's model of epidemic psychology, developed during the AIDS epidemic, suggests that the response to a disease outbreak involves three interwoven psycho-social epidemics: fear, moralization, and action. The fear epidemic centers around the anxiety of infection and suspicion of carriers. Moralization involves judgments about the epidemic and responses to it, potentially leading to cooperation or stigmatization. Finally, the action epidemic describes the measures individuals take. Strong's model, based on historical epidemics including the Black Death, highlights the crucial role of language in propagating these epidemics. While previous research has examined social media during outbreaks, this study is the first large-scale empirical test of Strong's model, utilizing Twitter data to analyze public responses to the COVID-19 pandemic in the US during 2020. Despite limitations inherent in using Twitter data, such as representativeness and self-presentation biases, the scale and granularity of social media data offer unique insights into the collective psychological response to a global event.

Existing research on social media and epidemics primarily focuses on information diffusion, misinformation, and behavioral markers during outbreaks like Zika, Ebola, and H1N1. Studies have analyzed social media content and behavioral patterns, tracked information spread, and explored misinformation. Research on psychological responses has mainly relied on surveys. However, no large-scale studies have directly tested Strong's model of epidemic psychology. This work fills this gap by utilizing a massive dataset of tweets to examine the evolution of public discourse concerning COVID-19 in the United States during 2020, focusing on how the language used reflects Strong's proposed psycho-social epidemics.

This study analyzed 122 million English-language tweets from 6.27 million unique US-based Twitter users posted between February 1st and December 31st, 2020. The researchers operationalized Strong's model using three steps. First, they hand-coded Strong's original paper to identify keywords associated with fear, moralization, and action. Second, they mapped these keywords to language categories from four existing lexicons: Linguistic Inquiry and Word Count (LIWC), EmoLex, a pro-social behavior lexicon, and moral foundations lexicon. Third, they analyzed the temporal dynamics of these language categories by calculating the fraction of users mentioning each category daily. To account for variations in daily tweet volume, the researchers standardized the fractions using z-scores. The study also validated its findings using alternative methods, including a keyword-search approach, a deep learning model for identifying social interaction types, and a deep learning model for medical entity extraction. Data on user mobility was also sourced from Foursquare to verify findings. The analysis also involved a thematic analysis of the most retweeted tweets to identify overarching themes within each phase. Change-point detection was used to identify phases, and logistic regression was employed to construct a parsimonious composite measure for real-time applications. Smoothing was applied to the time series of indicators to clarify trends.

The analysis revealed three distinct phases in the evolution of language use on Twitter related to COVID-19 in the U.S. during 2020. **Refusal Phase:** This initial phase, characterized by anxiety and fear, showed relatively little change in the use of other language categories reflecting a lack of change in behavior despite the growing number of deaths worldwide. Death was frequently mentioned, peaking around 11 February. This initial period exhibited limited variations in language related to the other dimensions (moralization and action). **Anger Phase:** Beginning after the announcement of the first US infection, this phase exhibited negative emotions and discussions of health concerns, particularly risk perception, bodily failures, and conflict, as expressions related to motion, social activities, and leisure increased. The declaration of a national emergency and social distancing guidelines intensified these sentiments. This phase was marked by conflict, particularly directed toward foreigners, political opponents, and individuals not adhering to suggested guidelines. Increased mentions of “home” and self-isolation reflected tensions between individual responsibility and community behavior. Science and religion were debated vigorously. **Acceptance Phase:** This phase started when physical distancing measures were legally imposed. Mentions of words associated with concepts such as power and authority likely reflected discussions about new policies and government mandates. As deaths increased, conflict subsided, and sadness became predominant. Prosocial behavior and caring for others increased. Mentions of work-related activities reflected job losses and the shift to working from home. Throughout the year, the acceptance phase dominated Twitter conversations, while the anger phase re-emerged cyclically with each new wave of infections. Two specific peaks in anger were observed: when the US death toll reached 100,000, and when President Trump tested positive for COVID-19. Thematic analysis corroborated these findings. Validation using alternative datasets and methodologies reinforced the validity of these findings. Correlation analysis showed that the phases were significantly associated with behavioral markers of conflict, health concerns, mobility patterns, and prosocial behaviors.

The findings strongly support Strong's model of epidemic psychology, demonstrating the presence and cyclical nature of the three psycho-social epidemics. The study extends Strong's model by identifying three distinct temporal phases (refusal, anger, acceptance), showing that the epidemics are not strictly sequential but rather cyclical, triggered by increases in the rate of virus transmission. The study also differentiates between initial and later stages of the epidemic, showing larger changes in the initial cycle of these epidemics. This approach adds a temporal dimension to Strong's model, making it more robust and predictive. The findings have significant implications for understanding public responses to epidemics and inform the development of more effective public health interventions. The study also demonstrates that the method can capture short-term variations in public sentiment which could be further investigated.

This study provides the first large-scale empirical validation of Strong's model of epidemic psychology using Twitter data. The identification of three phases—refusal, anger, and acceptance—and their cyclical nature extends Strong's model, offering a more nuanced understanding of public responses to epidemics. This real-time operationalization allows for the integration of epidemic psychology into existing epidemiological and mobility models, facilitating more comprehensive and predictive modeling of disease outbreaks. Future research could compare findings across different epidemics and cultures, employ finer-grained language analysis, and investigate the interplay between online and offline behavior.

The study's reliance on Twitter data introduces limitations, including its non-representativeness of the overall US population and the potential influence of bots. The geographical focus on the entire US might obscure regional variations, and the analysis's linguistic focus limits the capture of nonverbal aspects of online communication. Future studies should account for these limitations by using diverse datasets and incorporating multi-modal data to paint a richer picture of epidemic psychology.