Assessing Differential Sentiment across Counties through Crisis Progression
Code for Twitter Analysis of Coronavirus-related Tweets
To understand how local level responses to the COVID-19 pandemic can be better understood and whether they are associated with markers of health disparities.
We collected a dataset of coronavirus-related tweets mined from Twitter from January 23rd to March 25th 2020, as well as county-level data of social deprivation statistics and population-adjusted hotspots of the virus at the end of March, We cleaned and geocoded the Twitter data and performed word-level analysis with TF-IDF and Word2Vec and topic-level analysis with LDA to extract salient words and topics and inform how tweet context changes over time, across counties with different levels of social deprivation, and across counties considered hotspots for the crisis compared to the general population.
Both the word analysis and topic analysis confirmed that tweets over the progression of the crisis will vary in sentiment and topics, and these variations will differ across counties due to SDI and case numbers. Through the months, we see tweet topics shifting away from China, growing dissatisfaction with governmental response and negative sentiment, and low SDI areas considerably more concerned with national level politics, while high SDI levels more concerned with local and personal difficulties. Hotspots are more concerned with the Care Act and Trump, Twitter through COVID-19 and also display personal issues surrounding job security, as well as other negative emotions as compared to the population.
Integrating computational natural language processing analysis of Twitter data nationwide with an understanding of county-level disparities and virus hotspots, we can glean an important interpretation of nationwide sentiment through the duration of the crisis.