A Longitudinal and Geospatial Analysis of COVID-19 Tweets During the Early Outbreak Period in the United States

Introduction: Local rates of COVID-19 cases and deaths may not accurately convey the variability in community-level concern about COVID-19 during the early outbreak period in the United States. Social media interaction may elucidate communication about COVID-19 in this critical period, during which communities may have formulated initial conceptions pertaining to the perceived gravity of the disease and potential behavioral strategies for prevention. Methods: : Scripts were written to obtain tweets related to COVID-19 from Twitter. Using manually-annotated tweets about symptom-related concerns from a prior study, a machine learning classifier was applied to obtain a subset of tweets about concerns relating to COVID-19. The longitudinal relationship between these social media posts and active COVID-19 cases was assessed using linear and exponential regression. Changes in the geospatial clustering of tweets was assessed for the top five most populous cities in the United States. Results: : Social media posts relating to COVID-19 concerns appeared more predictive of active COVID-19 cases as temporal distance increased. The distribution of tweets in New York City and Phoenix appeared concentrated in city centers, whereas tweets from other cities were more residential. Tweets from New York City became more highly concentrated, but the opposite trend was observed in tweets from Los Angeles. Conclusion: Clustering of social media posts about COVID-19 revealed discrepancies across major US cities. General concern about the COVID-19 pandemic may moderate the relationship between behavioral/environmental factors and COVID-19 transmission. The degree and modality of this moderating effect may differ across US areas.