Content analysis

• Cross-sectional observational study
• Used Snscrape library to get tweets from February 1st to December 11th, 2020 that contain the keywords “covid” or “coronavirus” and “vaccine”
• Used natural language processing to analyze tweet content – focused on 300 most commonly mentioned terms

Sentiment analysis

• Used Valence Aware Dictionary and Sentiment Reasoner (VADER) for determining sentiments behind tweets. Vader provides a normalized, weighted composite score, taking into account emojis, punctuations, and syntax in addition to pure text. A score that is larger than or equal to 0.05 is considered positive sentiment, otherwise it is of negative sentimental rating.
• Used Mann-Kendall trend test and TextBlob library to rate tweets on a scale of objective (0) to subjective (1)
• Used NRCLex library to match specific words in tweets to emotions

Topic modeling

• Used Correlation Explanation (CorEx) to identify 20 topic clusters and determine which of the topics were most informative
• Picked 15 topics for the model with highest correlation to the tweets

User demographics

• Obtained user information including account setup date, followers, follows, posts, likes, descriptions, verified status
• Used m3inference library to make predictions about whether the accounts were from organizations or individuals
• Used Mann-Whitney U and Chi-squared tests for significance