Twitter Public Health Sentiment
Basic Overview
Status: Completed
Timeline: 1 Month April 2022
Technology: R
Note: Source code for the twitter data mining is not provided out of an abundance of caution for my own developer account, but I am happy to talk through my process. I used Twitter's API accessed via the rtweet package; nothing particularly complicated nor technically worth bragging about.
Download Source CodeWhy Twitter for Public Health?
80%+ of Americans Get News Primarily Online
A large majority of Americans get news at least sometimes from digital devices, according to a Pew Research Center survey conducted Aug. 31-Sept. 7, 2020.
Social media has a disproportionate effect on the disbursal of false information
The Center for Countering Digital Hate, a nonprofit whose work focuses on misinformation and hate disseminated online, conducted the study to examine the origins of anti-vaccine sentiment that has gained momentum on social networking platforms during the coronavirus pandemic. Results pinpointed a group of 12 individuals, collectively referred to as "the disinformation dozen" in the CCDH's conclusory report, who are at the forefront of false information campaigns targeting COVID-19 vaccines on Facebook, Instagram and Twitter.
Method
- Identify a variety of news organizations representing various viewpoints in American politics
- Find “high influence” followers of these news organizations on twitter (top 50 number of follower users per org)
- Draw 1000 tweets from these followers' timelines
- Isolate tweets with public health terms
- Perform sentiment analysis on these public health related tweets
- Test for any differences in results by news org they follow using one-way ANOVA and grouping
Summary of Results
| News Org | Number of Tweets | Average Sentiment | Median Sentiment |
|---|---|---|---|
| ABC News | 1258 | -0.11 | 0 |
| BBC News (World) | 1224 | -0.02 | 0 |
| Bloomberg | 582 | -0.17 | 0 |
| CBS News | 1953 | -0.2 | 0 |
| CDC | 2660 | -0.01 | 0 |
| CNBC | 1234 | 0.15 | 0 |
| CNN | 423 | 0.13 | 0 |
| Forbes | 1460 | 0.05 | 0 |
| Fox News | 1094 | -0.22 | 0 |
| NBC News | 733 | 0.22 | 1 |
| Reuters | 1835 | 0.07 | 0 |
| The Associated Press | 2566 | -0.12 | 0 |
| The Economist | 1315 | 0.11 | 0 |
| The Guardian | 2077 | -0.06 | 0 |
| The New York Times | 1255 | -0.03 | 0 |
| The Wall Street Journal | 923 | 0.16 | 0 |
| The Washington Post | 1075 | -0.08 | 0 |
| TIME | 415 | -0.05 | 0 |
| World Health Organization (WHO) | 1450 | 0.11 | 0 |
.png)
Tukey Summary Groups
The above reults may be hard to interpret and lack confidence intervals, so to determine significant difference I performed a one way ANOVA between each news org results with Tukey's honest significant difference wit ha 95% confidence interval. Each news org was the nassigned to one of several possible groups, represented by a letter; if a news org shares a letter with another news org, then there is nothing in the data to suggest a significant difference between the groups. However, if the news orgs share no letters, then there is evidence that the difference in sentiment is significant.
| News Org | Groups |
|---|---|
| ABC News | abce |
| BBC News (World) | abcdef |
| Bloomberg | abce |
| CBS News | a |
| CDC | bcdef |
| CNBC | df |
| CNN | abcdef |
| Forbes | bdef |
| Fox News | ac |
| NBC News | d |
| Reuters | bdef |
| The Associated Press | ace |
| The Economist | bdf |
| The Guardian | abcef |
| The New York Times | abcdef |
| The Wall Street Journal | bdf |
| The Washington Post | abcdef |
| TIME | abcdef |
| World Health Organization (WHO) | bdf |
Conclusions
There are some significant difference between groups in my findings that suggest which news sources a person turns to may affect their perception of various public health efforts.
However, this represents a tiny cross section of data, only ~38,000 tweets that were determined to be relevant from a pool of ~750,000. With a longer term project and a larger pool of data, there may be more clarity and confidence in the results. The Twitter API time limit severely limited my ability to mine for a significant number of tweets quickly.
May also need to work on a paired down list of public health terms, and do more data exploration for relevance (especially outliers!)