TL;DR
This study uses topic modeling on Twitter data from six countries to understand public perceptions of COVID-19 nonpharmaceutical interventions (NPIs). It reveals that while less restrictive NPIs generally receive broad support, public attitudes towards more restrictive measures vary significantly by country, influenced by factors like implementation timeliness and communication strategies.
What Problem Does This Paper Address?
Governments worldwide implemented nonpharmaceutical interventions (NPIs) like mask-wearing and social distancing to curb the spread of COVID-19. For these measures to be effective, public adherence is crucial. However, understanding public perceptions and attitudes towards these NPIs, and the factors influencing them, is challenging. Traditional methods like surveys can be slow, especially during rapidly evolving public health crises. This paper addresses the need for a rapid and scalable method to gauge public sentiment on NPIs, leveraging the vast and immediate data available on social media platforms like Twitter.
Methods
The researchers employed a hybrid computational and qualitative approach. They collected over 777,000 English-language tweets related to COVID-19 NPIs from six countries (Australia, Canada, New Zealand, Ireland, the United Kingdom, and the United States) between January and April 2020.
Key techniques included:
- Data Collection and Preprocessing: Tweets were gathered using specific hashtags and then cleaned by removing duplicates, converting to lowercase, expanding contractions, and removing non-alphabetic characters, URLs, and mentions. Standard natural language processing (NLP) techniques like tokenization, lemmatization, and stop-word removal were applied.
- Topic Modeling: The study utilized MetaLDA, a topic modeling algorithm empirically shown to perform well on sparse and noisy tweet data, to identify latent semantic structures and group tweets into coherent topics related to NPIs. They explored multiple topic model configurations (different values of 'k') and selected the best models based on statistical (NPMI coherence) and qualitative evaluations.
- Comparative Analysis: A coding schema was developed to identify and label NPI-related topics. This involved a qualitative review of tweets within each topic. Heatmaps and chord diagrams were used to visualize the proportion of tweets discussing specific NPIs and the relationships between discussions of different NPI categories across countries.
Datasets
The primary dataset consisted of 777,869 English-language tweets related to COVID-19 NPIs collected from:
- Australia
- Canada
- New Zealand
- Ireland
- United Kingdom
- United States
This dataset was collected between January 1 and April 30, 2020. The paper does not introduce a new benchmark dataset but rather utilizes publicly available Twitter data.
Key Insights and Ideas
- Twitter as a Real-time Barometer: The study powerfully demonstrates that Twitter can provide rapid, timely feedback on public sentiment regarding public health interventions, complementing slower traditional methods.
- Varying Public Appetite for Restrictiveness: While less restrictive NPIs (like personal protection and social distancing) generally saw widespread public support and encouragement for adherence, more restrictive measures (like lockdowns and travel bans) elicited more varied and country-specific reactions.
- The Criticality of Timeliness: Delayed or staggered implementation of restrictive NPIs often led to public frustration and demands for stricter measures, suggesting that swift and decisive action can foster greater public acceptance.
- Communication is Key: The style and clarity of government communication campaigns significantly impacted public perception. Empathy, clarity, and a sense of collective action (as seen in New Zealand's "Unite against COVID-19" campaign) were more positively received than purely instructive messages.
- Information Asymmetry Fuels Distrust: When public understanding of NPIs, particularly more intrusive ones, conflicted with government advice, it led to skepticism and a reliance on alternative information sources, impacting confidence in national regimes.
Key Results
- Public Attention to NPIs: New Zealand showed the highest proportion of NPI-related tweets, while the US showed the lowest.
- Correlation with Case Numbers: The number of daily tweets was statistically significantly correlated with the number of daily confirmed cases only in Australia (r=0.837) and New Zealand (r=0.747). This correlation was stronger when considering the entire study period compared to just the date of the first confirmed case.
- Topic Breakdown: Topic modeling identified 131 topics related to 22 specific NPIs, categorized into seven broader NPI types (Personal Protection, Social Distancing, Testing and Tracing, Gathering Restrictions, Lockdown, Travel Restrictions, and Workplace Closures).
- Country-Specific Trends:
- Personal Protection: Face masks were perceived as important in Australia despite government advice to the contrary.
- Social Distancing: Public spaces were a concern across most countries, with dissatisfaction noted regarding capacity limits on public transport.
- Testing and Tracing: Negative perceptions of testing efforts were prevalent in Canada and the US due to issues like inconsistent criteria and backlogs. Privacy concerns were raised regarding contact tracing apps in Australia and New Zealand.
- Gathering Restrictions: The US saw a disproportionately high discussion of public gatherings (30.42% of NPI tweets), often linked to election implications.
- Lockdowns: While generally supported, the timing of lockdowns was crucial. New Zealand's swift implementation led to overwhelming support, whereas delayed lockdowns elsewhere caused frustration.
- Travel Restrictions: New Zealand and Ireland expressed outrage that travel bans were not enacted sooner.
- Workplace Closures: Concerns about children's health were a major driver for school closure discussions. Working from home was generally viewed positively, with surprise at the ease of transition.
- Factors Influencing Adherence: Four key characteristics were identified as influencing public adherence: timeliness of implementation, NPI campaign strategies, inconsistent information, and enforcement strategies.
Limitations and Future Work
- Demographic Representation: The authors acknowledge that Twitter users may not perfectly represent the general population of each country.
- Language Bias: The study was limited to English-language tweets, excluding non-anglophone perspectives.
- Future Work: The paper suggests that ongoing social media analysis can provide valuable insights for governments and health authorities to improve public health responses to current and future pandemics. Integrating this feedback into communication strategies and NPI program management is crucial. Further research could explore the nuances of language, sentiment, and the impact of misinformation on public adherence.
Summary updated