CSH researcher Hannah Metzler will give a virtual talk (about research with Bernard Rimé, Max Pellert, Thomas Niederkrotenthaler, Anna Di Natale and David Garcia) at the 7th International Conference on Computational Social Science IC2S2 2021.
The COVID-19 pandemic has exposed the world’s population to sudden and long-lasting challenges, including many changes to everyday life and events that elicited strong emotional reactions. While investigations of responses to tragic one-off events exist, studies on the evolution of collective emotions during a pandemic are missing.
We analysed digital traces of emotional expressions in a sample of 8.3 billion tweets from 18 countries during the first five weeks after the start of the outbreak (day with 30 cases). We included only tweets from users located in these countries, and excluded tweets from users with very low and high follower numbers to exclude bots and mass media. We used emotion classification methods for tweets, including unsupervised dictionaries in 6 languages (LIWC; 4) and a supervised model based on the RoBERTa model finetuned for English tweets with an additional linear layer that allowed training on multiple emotion categories. Next, we calculated the daily proportion of emotion tweets for anxiety, sadness, anger and positive emotions. To assess common changes in emotions across countries while accounting for individual language- and country-variations, we applied generalized linear mixed-effects logistic regression models.
We found a strong upsurge of anxiety-related terms at the beginning of the pandemic consistent across all countries, with peak increases of 52 % on average across countries. At the start of the outbreak (week 1), this increase clearly correlated with COVID-19 cases across countries, and thus likely reflects people’s strong initial uncertainty about how to cope with the new threat. About two weeks later, as casualties rose and the stringency of social distancing measures increased, sadness rose and anger decreased in most countries (with peaks of +23% and –9 % on average across countries), both reaching their most extreme levels about three weeks after the outbreak. As sadness increased at the time of social distancing measures, and in countries with few COVID-19 deaths, it may have been a response to the (anticipated) loss of contact and daily routines during lockdowns. Decreased levels of anger may indicate relative acceptance of protective measures, less discussions on controversial or political topics than before the pandemic, as well as reductions of violence and crime during lock-downs.
Positive emotions remained relatively stable, suggesting a certain degree of resiliencein this early phase of the pandemic. A memory pattern in all emotion timelines suggests that the changes were in part collective emotional phenomena, during which people reactivated each other’s emotional experiences via social interaction. This interaction likely contributed to the long time spans for which emotions remained significantly below or above the baseline in many countries, which are some of the most enduring changes in emotional expressions observed in long periods of social media data.
This kind of time-sensitive analyses of large-scale samples of emotional expression have the potential to inform the planning of mental health support and to help tailor risk communication.